diff --git a/v2.1/french_24l/cond_prefill_ane.mlmodelc/analytics/coremldata.bin b/v2.1/french_24l/cond_prefill_ane.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..2bd8d34039d5e18ded6e8e750b172a919c5f8a6e --- /dev/null +++ b/v2.1/french_24l/cond_prefill_ane.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1770a57b4ca0390b6b0adc62437e3bfbeb24a1e48b21046f91dfeda87f634de7 +size 243 diff --git a/v2.1/french_24l/cond_prefill_ane.mlmodelc/coremldata.bin b/v2.1/french_24l/cond_prefill_ane.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..9323079cbe3568291dee608083cb5f8de5f9c407 --- /dev/null +++ b/v2.1/french_24l/cond_prefill_ane.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c6c73ec0cdb34f5d5b26a725128e9440ca9eed583727acae93b45fd7123c0f20 +size 4496 diff --git a/v2.1/french_24l/cond_prefill_ane.mlmodelc/model.mil b/v2.1/french_24l/cond_prefill_ane.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..f102b6e147b28d002e028bb659c273f9e9c7f4c9 --- /dev/null +++ b/v2.1/french_24l/cond_prefill_ane.mlmodelc/model.mil @@ -0,0 +1,3198 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.9.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor conditioning, tensor k_cache0, tensor k_cache1, tensor k_cache10, tensor k_cache11, tensor k_cache12, tensor k_cache13, tensor k_cache14, tensor k_cache15, tensor k_cache16, tensor k_cache17, tensor k_cache18, tensor k_cache19, tensor k_cache2, tensor k_cache20, tensor k_cache21, tensor k_cache22, tensor k_cache23, tensor k_cache3, tensor k_cache4, tensor k_cache5, tensor k_cache6, tensor k_cache7, tensor k_cache8, tensor k_cache9, tensor position0, tensor position1, tensor position10, tensor position11, tensor position12, tensor position13, tensor position14, tensor position15, tensor position16, tensor position17, tensor position18, tensor position19, tensor position2, tensor position20, tensor position21, tensor position22, tensor position23, tensor position3, tensor position4, tensor position5, tensor position6, tensor position7, tensor position8, tensor position9, tensor v_cache0, tensor v_cache1, tensor v_cache10, tensor v_cache11, tensor v_cache12, tensor v_cache13, tensor v_cache14, tensor v_cache15, tensor v_cache16, tensor v_cache17, tensor v_cache18, tensor v_cache19, tensor v_cache2, tensor v_cache20, tensor v_cache21, tensor v_cache22, tensor v_cache23, tensor v_cache3, tensor v_cache4, tensor v_cache5, tensor v_cache6, tensor v_cache7, tensor v_cache8, tensor v_cache9, tensor valid_len) { + tensor token_indices_to_fp16 = const()[name = tensor("token_indices_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor position0_to_fp16_dtype_0 = const()[name = tensor("position0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position0_to_fp16 = cast(dtype = position0_to_fp16_dtype_0, x = position0)[name = tensor("cast_470")]; + tensor ts_cast_fp16 = add(x = token_indices_to_fp16, y = position0_to_fp16)[name = tensor("ts_cast_fp16")]; + tensor var_231 = const()[name = tensor("op_231"), val = tensor([256, 1, 1])]; + tensor var_232_cast_fp16 = reshape(shape = var_231, x = ts_cast_fp16)[name = tensor("op_232_cast_fp16")]; + tensor var_227_to_fp16 = const()[name = tensor("op_227_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(640)))]; + tensor angles_cast_fp16 = mul(x = var_227_to_fp16, y = var_232_cast_fp16)[name = tensor("angles_cast_fp16")]; + tensor rotr_cast_fp16 = cos(x = angles_cast_fp16)[name = tensor("rotr_cast_fp16")]; + tensor roti_cast_fp16 = sin(x = angles_cast_fp16)[name = tensor("roti_cast_fp16")]; + tensor _inversed_237_y_0_to_fp16 = const()[name = tensor("_inversed_237_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_237_cast_fp16 = mul(x = ts_cast_fp16, y = _inversed_237_y_0_to_fp16)[name = tensor("_inversed_237_cast_fp16")]; + tensor var_238_cast_fp16 = floor(x = _inversed_237_cast_fp16)[name = tensor("op_238_cast_fp16")]; + tensor var_239_to_fp16 = const()[name = tensor("op_239_to_fp16"), val = tensor(0x1p+9)]; + tensor var_240_cast_fp16 = mul(x = var_238_cast_fp16, y = var_239_to_fp16)[name = tensor("op_240_cast_fp16")]; + tensor write_idx_1_cast_fp16 = sub(x = ts_cast_fp16, y = var_240_cast_fp16)[name = tensor("write_idx_1_cast_fp16")]; + tensor valid_len_to_fp16_dtype_0 = const()[name = tensor("valid_len_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor valid_len_to_fp16 = cast(dtype = valid_len_to_fp16_dtype_0, x = valid_len)[name = tensor("cast_469")]; + tensor var_246_cast_fp16 = less(x = token_indices_to_fp16, y = valid_len_to_fp16)[name = tensor("op_246_cast_fp16")]; + tensor token_valid_to_fp16_dtype_0 = const()[name = tensor("token_valid_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_246_cast_fp16_to_fp16 = cast(dtype = token_valid_to_fp16_dtype_0, x = var_246_cast_fp16)[name = tensor("cast_468")]; + tensor var_252_cast_fp16 = mul(x = write_idx_1_cast_fp16, y = var_246_cast_fp16_to_fp16)[name = tensor("op_252_cast_fp16")]; + tensor var_253_to_fp16 = const()[name = tensor("op_253_to_fp16"), val = tensor(0x1p+0)]; + tensor var_255_cast_fp16 = sub(x = var_253_to_fp16, y = var_246_cast_fp16_to_fp16)[name = tensor("op_255_cast_fp16")]; + tensor var_256_to_fp16 = const()[name = tensor("op_256_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_257_cast_fp16 = mul(x = var_255_cast_fp16, y = var_256_to_fp16)[name = tensor("op_257_cast_fp16")]; + tensor write_idx_cast_fp16 = add(x = var_252_cast_fp16, y = var_257_cast_fp16)[name = tensor("write_idx_cast_fp16")]; + tensor var_262 = const()[name = tensor("op_262"), val = tensor([256, 1])]; + tensor var_263_cast_fp16 = reshape(shape = var_262, x = write_idx_cast_fp16)[name = tensor("op_263_cast_fp16")]; + tensor var_267_to_fp16 = const()[name = tensor("op_267_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(768)))]; + tensor var_268_cast_fp16 = equal(x = var_263_cast_fp16, y = var_267_to_fp16)[name = tensor("op_268_cast_fp16")]; + tensor var_278_axes_0 = const()[name = tensor("op_278_axes_0"), val = tensor([0])]; + tensor var_278_keep_dims_0 = const()[name = tensor("op_278_keep_dims_0"), val = tensor(false)]; + tensor assign_to_fp16_dtype_0 = const()[name = tensor("assign_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_268_cast_fp16_to_fp16 = cast(dtype = assign_to_fp16_dtype_0, x = var_268_cast_fp16)[name = tensor("cast_467")]; + tensor var_278_cast_fp16 = reduce_sum(axes = var_278_axes_0, keep_dims = var_278_keep_dims_0, x = var_268_cast_fp16_to_fp16)[name = tensor("op_278_cast_fp16")]; + tensor const_0_to_fp16 = const()[name = tensor("const_0_to_fp16"), val = tensor(-inf)]; + tensor var_280_to_fp16 = const()[name = tensor("op_280_to_fp16"), val = tensor(0x1p+0)]; + tensor clip_0_cast_fp16 = clip(alpha = const_0_to_fp16, beta = var_280_to_fp16, x = var_278_cast_fp16)[name = tensor("clip_0_cast_fp16")]; + tensor var_284 = const()[name = tensor("op_284"), val = tensor([1, 1])]; + tensor var_285_cast_fp16 = reshape(shape = var_284, x = position0_to_fp16)[name = tensor("op_285_cast_fp16")]; + tensor var_289_to_fp16 = const()[name = tensor("op_289_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1856)))]; + tensor q_positions_cast_fp16 = add(x = var_285_cast_fp16, y = var_289_to_fp16)[name = tensor("q_positions_cast_fp16")]; + tensor var_297_cast_fp16 = add(x = position0_to_fp16, y = valid_len_to_fp16)[name = tensor("op_297_cast_fp16")]; + tensor var_300 = const()[name = tensor("op_300"), val = tensor([1, 1])]; + tensor valid_end_cast_fp16 = reshape(shape = var_300, x = var_297_cast_fp16)[name = tensor("valid_end_cast_fp16")]; + tensor var_302_cast_fp16 = less(x = var_267_to_fp16, y = valid_end_cast_fp16)[name = tensor("op_302_cast_fp16")]; + tensor var_303_cast_fp16 = less_equal(x = var_267_to_fp16, y = q_positions_cast_fp16)[name = tensor("op_303_cast_fp16")]; + tensor var_304 = logical_and(x = var_302_cast_fp16, y = var_303_cast_fp16)[name = tensor("op_304")]; + tensor mask_to_fp16_dtype_0 = const()[name = tensor("mask_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_311_to_fp16 = const()[name = tensor("op_311_to_fp16"), val = tensor(0x1p+0)]; + tensor var_304_to_fp16 = cast(dtype = mask_to_fp16_dtype_0, x = var_304)[name = tensor("cast_466")]; + tensor var_312_cast_fp16 = sub(x = var_304_to_fp16, y = var_311_to_fp16)[name = tensor("op_312_cast_fp16")]; + tensor var_313_to_fp16 = const()[name = tensor("op_313_to_fp16"), val = tensor(0x1.388p+13)]; + tensor mask_bias_cast_fp16 = mul(x = var_312_cast_fp16, y = var_313_to_fp16)[name = tensor("mask_bias_cast_fp16")]; + tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; + tensor conditioning_to_fp16_dtype_0 = const()[name = tensor("conditioning_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor norm0_1_weight_to_fp16 = const()[name = tensor("norm0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2432)))]; + tensor norm0_1_bias_to_fp16 = const()[name = tensor("norm0_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4544)))]; + tensor var_316_to_fp16 = const()[name = tensor("op_316_to_fp16"), val = tensor(0x1.5p-17)]; + tensor conditioning_to_fp16 = cast(dtype = conditioning_to_fp16_dtype_0, x = conditioning)[name = tensor("cast_465")]; + tensor input_3_cast_fp16 = layer_norm(axes = input_3_axes_0, beta = norm0_1_bias_to_fp16, epsilon = var_316_to_fp16, gamma = norm0_1_weight_to_fp16, x = conditioning_to_fp16)[name = tensor("input_3_cast_fp16")]; + tensor attn0_in_proj_weight_to_fp16 = const()[name = tensor("attn0_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6656)))]; + tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6298176)))]; + tensor linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn0_in_proj_weight_to_fp16, x = input_3_cast_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor var_330 = const()[name = tensor("op_330"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_1_cast_fp16 = reshape(shape = var_330, x = linear_0_cast_fp16)[name = tensor("qkv_1_cast_fp16")]; + tensor x_1_begin_0 = const()[name = tensor("x_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_1_end_0 = const()[name = tensor("x_1_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_1_end_mask_0 = const()[name = tensor("x_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_1_squeeze_mask_0 = const()[name = tensor("x_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_1_cast_fp16 = slice_by_index(begin = x_1_begin_0, end = x_1_end_0, end_mask = x_1_end_mask_0, squeeze_mask = x_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("x_1_cast_fp16")]; + tensor x_3_begin_0 = const()[name = tensor("x_3_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_3_end_0 = const()[name = tensor("x_3_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_3_end_mask_0 = const()[name = tensor("x_3_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_3_squeeze_mask_0 = const()[name = tensor("x_3_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_3_cast_fp16 = slice_by_index(begin = x_3_begin_0, end = x_3_end_0, end_mask = x_3_end_mask_0, squeeze_mask = x_3_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("x_3_cast_fp16")]; + tensor v_1_begin_0 = const()[name = tensor("v_1_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_1_end_0 = const()[name = tensor("v_1_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_1_end_mask_0 = const()[name = tensor("v_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_1_squeeze_mask_0 = const()[name = tensor("v_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_1_cast_fp16 = slice_by_index(begin = v_1_begin_0, end = v_1_end_0, end_mask = v_1_end_mask_0, squeeze_mask = v_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("v_1_cast_fp16")]; + tensor var_375 = const()[name = tensor("op_375"), val = tensor([256, 16, 32, 2])]; + tensor pairs_1_cast_fp16 = reshape(shape = var_375, x = x_1_cast_fp16)[name = tensor("pairs_1_cast_fp16")]; + tensor xr_1_begin_0 = const()[name = tensor("xr_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_1_end_0 = const()[name = tensor("xr_1_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_1_end_mask_0 = const()[name = tensor("xr_1_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_1_squeeze_mask_0 = const()[name = tensor("xr_1_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_1_cast_fp16 = slice_by_index(begin = xr_1_begin_0, end = xr_1_end_0, end_mask = xr_1_end_mask_0, squeeze_mask = xr_1_squeeze_mask_0, x = pairs_1_cast_fp16)[name = tensor("xr_1_cast_fp16")]; + tensor xi_1_begin_0 = const()[name = tensor("xi_1_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_1_end_0 = const()[name = tensor("xi_1_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_1_end_mask_0 = const()[name = tensor("xi_1_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_1_squeeze_mask_0 = const()[name = tensor("xi_1_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_1_cast_fp16 = slice_by_index(begin = xi_1_begin_0, end = xi_1_end_0, end_mask = xi_1_end_mask_0, squeeze_mask = xi_1_squeeze_mask_0, x = pairs_1_cast_fp16)[name = tensor("xi_1_cast_fp16")]; + tensor var_383_cast_fp16 = mul(x = xr_1_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_383_cast_fp16")]; + tensor var_384_cast_fp16 = mul(x = xi_1_cast_fp16, y = roti_cast_fp16)[name = tensor("op_384_cast_fp16")]; + tensor xor__1_cast_fp16 = sub(x = var_383_cast_fp16, y = var_384_cast_fp16)[name = tensor("xor__1_cast_fp16")]; + tensor var_387_cast_fp16 = mul(x = xr_1_cast_fp16, y = roti_cast_fp16)[name = tensor("op_387_cast_fp16")]; + tensor var_388_cast_fp16 = mul(x = xi_1_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_388_cast_fp16")]; + tensor xoi_1_cast_fp16 = add(x = var_387_cast_fp16, y = var_388_cast_fp16)[name = tensor("xoi_1_cast_fp16")]; + tensor out_1_axis_0 = const()[name = tensor("out_1_axis_0"), val = tensor(-1)]; + tensor out_1_cast_fp16 = stack(axis = out_1_axis_0, values = (xor__1_cast_fp16, xoi_1_cast_fp16))[name = tensor("out_1_cast_fp16")]; + tensor var_398 = const()[name = tensor("op_398"), val = tensor([1, 256, 16, 64])]; + tensor q_1_cast_fp16 = reshape(shape = var_398, x = out_1_cast_fp16)[name = tensor("q_1_cast_fp16")]; + tensor var_404 = const()[name = tensor("op_404"), val = tensor([256, 16, 32, 2])]; + tensor pairs_3_cast_fp16 = reshape(shape = var_404, x = x_3_cast_fp16)[name = tensor("pairs_3_cast_fp16")]; + tensor xr_3_begin_0 = const()[name = tensor("xr_3_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_3_end_0 = const()[name = tensor("xr_3_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_3_end_mask_0 = const()[name = tensor("xr_3_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_3_squeeze_mask_0 = const()[name = tensor("xr_3_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_3_cast_fp16 = slice_by_index(begin = xr_3_begin_0, end = xr_3_end_0, end_mask = xr_3_end_mask_0, squeeze_mask = xr_3_squeeze_mask_0, x = pairs_3_cast_fp16)[name = tensor("xr_3_cast_fp16")]; + tensor xi_3_begin_0 = const()[name = tensor("xi_3_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_3_end_0 = const()[name = tensor("xi_3_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_3_end_mask_0 = const()[name = tensor("xi_3_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_3_squeeze_mask_0 = const()[name = tensor("xi_3_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_3_cast_fp16 = slice_by_index(begin = xi_3_begin_0, end = xi_3_end_0, end_mask = xi_3_end_mask_0, squeeze_mask = xi_3_squeeze_mask_0, x = pairs_3_cast_fp16)[name = tensor("xi_3_cast_fp16")]; + tensor var_412_cast_fp16 = mul(x = xr_3_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_412_cast_fp16")]; + tensor var_413_cast_fp16 = mul(x = xi_3_cast_fp16, y = roti_cast_fp16)[name = tensor("op_413_cast_fp16")]; + tensor xor__3_cast_fp16 = sub(x = var_412_cast_fp16, y = var_413_cast_fp16)[name = tensor("xor__3_cast_fp16")]; + tensor var_416_cast_fp16 = mul(x = xr_3_cast_fp16, y = roti_cast_fp16)[name = tensor("op_416_cast_fp16")]; + tensor var_417_cast_fp16 = mul(x = xi_3_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_417_cast_fp16")]; + tensor xoi_3_cast_fp16 = add(x = var_416_cast_fp16, y = var_417_cast_fp16)[name = tensor("xoi_3_cast_fp16")]; + tensor out_3_axis_0 = const()[name = tensor("out_3_axis_0"), val = tensor(-1)]; + tensor out_3_cast_fp16 = stack(axis = out_3_axis_0, values = (xor__3_cast_fp16, xoi_3_cast_fp16))[name = tensor("out_3_cast_fp16")]; + tensor var_434 = const()[name = tensor("op_434"), val = tensor([256, 1024])]; + tensor var_435_cast_fp16 = reshape(shape = var_434, x = out_3_cast_fp16)[name = tensor("op_435_cast_fp16")]; + tensor written_k_1_transpose_x_1 = const()[name = tensor("written_k_1_transpose_x_1"), val = tensor(true)]; + tensor written_k_1_transpose_y_1 = const()[name = tensor("written_k_1_transpose_y_1"), val = tensor(false)]; + tensor written_k_1_cast_fp16 = matmul(transpose_x = written_k_1_transpose_x_1, transpose_y = written_k_1_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_435_cast_fp16)[name = tensor("written_k_1_cast_fp16")]; + tensor var_442 = const()[name = tensor("op_442"), val = tensor([256, 1024])]; + tensor var_443_cast_fp16 = reshape(shape = var_442, x = v_1_cast_fp16)[name = tensor("op_443_cast_fp16")]; + tensor written_v_1_transpose_x_1 = const()[name = tensor("written_v_1_transpose_x_1"), val = tensor(true)]; + tensor written_v_1_transpose_y_1 = const()[name = tensor("written_v_1_transpose_y_1"), val = tensor(false)]; + tensor written_v_1_cast_fp16 = matmul(transpose_x = written_v_1_transpose_x_1, transpose_y = written_v_1_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_443_cast_fp16)[name = tensor("written_v_1_cast_fp16")]; + tensor var_445_to_fp16 = const()[name = tensor("op_445_to_fp16"), val = tensor(0x1p+0)]; + tensor var_447_cast_fp16 = sub(x = var_445_to_fp16, y = clip_0_cast_fp16)[name = tensor("op_447_cast_fp16")]; + tensor var_452 = const()[name = tensor("op_452"), val = tensor([1, 512, 1, 1])]; + tensor keep_1_cast_fp16 = reshape(shape = var_452, x = var_447_cast_fp16)[name = tensor("keep_1_cast_fp16")]; + tensor k_cache0_to_fp16_dtype_0 = const()[name = tensor("k_cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache0_to_fp16 = cast(dtype = k_cache0_to_fp16_dtype_0, x = k_cache0)[name = tensor("cast_464")]; + tensor var_454_cast_fp16 = mul(x = k_cache0_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_454_cast_fp16")]; + tensor var_459 = const()[name = tensor("op_459"), val = tensor([1, 512, 16, 64])]; + tensor var_460_cast_fp16 = reshape(shape = var_459, x = written_k_1_cast_fp16)[name = tensor("op_460_cast_fp16")]; + tensor new_k_cache_1_cast_fp16 = add(x = var_454_cast_fp16, y = var_460_cast_fp16)[name = tensor("new_k_cache_1_cast_fp16")]; + tensor new_k_cache_1_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_1_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache0_to_fp16_dtype_0 = const()[name = tensor("v_cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache0_to_fp16 = cast(dtype = v_cache0_to_fp16_dtype_0, x = v_cache0)[name = tensor("cast_462")]; + tensor var_463_cast_fp16 = mul(x = v_cache0_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_463_cast_fp16")]; + tensor var_468 = const()[name = tensor("op_468"), val = tensor([1, 512, 16, 64])]; + tensor var_469_cast_fp16 = reshape(shape = var_468, x = written_v_1_cast_fp16)[name = tensor("op_469_cast_fp16")]; + tensor new_v_cache_1_cast_fp16 = add(x = var_463_cast_fp16, y = var_469_cast_fp16)[name = tensor("new_v_cache_1_cast_fp16")]; + tensor new_v_cache_1_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_1_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_482 = const()[name = tensor("op_482"), val = tensor([0, 2, 1, 3])]; + tensor var_493_transpose_x_0 = const()[name = tensor("op_493_transpose_x_0"), val = tensor(false)]; + tensor var_493_transpose_y_0 = const()[name = tensor("op_493_transpose_y_0"), val = tensor(false)]; + tensor transpose_69_perm_0 = const()[name = tensor("transpose_69_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_70_perm_0 = const()[name = tensor("transpose_70_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_70 = transpose(perm = transpose_70_perm_0, x = new_k_cache_1_cast_fp16)[name = tensor("transpose_204")]; + tensor transpose_69 = transpose(perm = transpose_69_perm_0, x = q_1_cast_fp16)[name = tensor("transpose_205")]; + tensor var_493_cast_fp16 = matmul(transpose_x = var_493_transpose_x_0, transpose_y = var_493_transpose_y_0, x = transpose_69, y = transpose_70)[name = tensor("op_493_cast_fp16")]; + tensor var_494_to_fp16 = const()[name = tensor("op_494_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_1_cast_fp16 = mul(x = var_493_cast_fp16, y = var_494_to_fp16)[name = tensor("scores_1_cast_fp16")]; + tensor var_500 = const()[name = tensor("op_500"), val = tensor([1, 1, 256, 512])]; + tensor var_501_cast_fp16 = reshape(shape = var_500, x = mask_bias_cast_fp16)[name = tensor("op_501_cast_fp16")]; + tensor scores_3_cast_fp16 = add(x = scores_1_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_3_cast_fp16")]; + tensor var_504 = const()[name = tensor("op_504"), val = tensor(-1)]; + tensor weights_1_cast_fp16 = softmax(axis = var_504, x = scores_3_cast_fp16)[name = tensor("weights_1_cast_fp16")]; + tensor attn_1_transpose_x_0 = const()[name = tensor("attn_1_transpose_x_0"), val = tensor(false)]; + tensor attn_1_transpose_y_0 = const()[name = tensor("attn_1_transpose_y_0"), val = tensor(false)]; + tensor values_1_cast_fp16 = transpose(perm = var_482, x = new_v_cache_1_cast_fp16)[name = tensor("transpose_206")]; + tensor attn_1_cast_fp16 = matmul(transpose_x = attn_1_transpose_x_0, transpose_y = attn_1_transpose_y_0, x = weights_1_cast_fp16, y = values_1_cast_fp16)[name = tensor("attn_1_cast_fp16")]; + tensor var_512 = const()[name = tensor("op_512"), val = tensor([0, 2, 1, 3])]; + tensor var_517 = const()[name = tensor("op_517"), val = tensor([1, 256, 1024])]; + tensor var_513_cast_fp16 = transpose(perm = var_512, x = attn_1_cast_fp16)[name = tensor("transpose_203")]; + tensor input_5_cast_fp16 = reshape(shape = var_517, x = var_513_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor attn0_out_proj_weight_to_fp16 = const()[name = tensor("attn0_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6304384)))]; + tensor linear_1_bias_0_to_fp16 = const()[name = tensor("linear_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401600)))]; + tensor linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn0_out_proj_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("linear_1_cast_fp16")]; + tensor var_526_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_526_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_7_cast_fp16 = add(x = conditioning_to_fp16, y = linear_1_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor input_9_axes_0 = const()[name = tensor("input_9_axes_0"), val = tensor([-1])]; + tensor norm0_2_weight_to_fp16 = const()[name = tensor("norm0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8403712)))]; + tensor norm0_2_bias_to_fp16 = const()[name = tensor("norm0_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8405824)))]; + tensor var_530_to_fp16 = const()[name = tensor("op_530_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_9_cast_fp16 = layer_norm(axes = input_9_axes_0, beta = norm0_2_bias_to_fp16, epsilon = var_530_to_fp16, gamma = norm0_2_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor linear0_1_weight_to_fp16 = const()[name = tensor("linear0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8407936)))]; + tensor linear_2_bias_0_to_fp16 = const()[name = tensor("linear_2_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16796608)))]; + tensor linear_2_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear0_1_weight_to_fp16, x = input_9_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor input_11_mode_0 = const()[name = tensor("input_11_mode_0"), val = tensor("EXACT")]; + tensor input_11_cast_fp16 = gelu(mode = input_11_mode_0, x = linear_2_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor linear0_2_weight_to_fp16 = const()[name = tensor("linear0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16804864)))]; + tensor linear_3_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear0_2_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("linear_3_cast_fp16")]; + tensor input_13_cast_fp16 = add(x = input_7_cast_fp16, y = linear_3_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor input_15_axes_0 = const()[name = tensor("input_15_axes_0"), val = tensor([-1])]; + tensor norm1_1_weight_to_fp16 = const()[name = tensor("norm1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25193536)))]; + tensor norm1_1_bias_to_fp16 = const()[name = tensor("norm1_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25195648)))]; + tensor var_547_to_fp16 = const()[name = tensor("op_547_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_15_cast_fp16 = layer_norm(axes = input_15_axes_0, beta = norm1_1_bias_to_fp16, epsilon = var_547_to_fp16, gamma = norm1_1_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor attn1_in_proj_weight_to_fp16 = const()[name = tensor("attn1_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25197760)))]; + tensor linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn1_in_proj_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor var_561 = const()[name = tensor("op_561"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_3_cast_fp16 = reshape(shape = var_561, x = linear_4_cast_fp16)[name = tensor("qkv_3_cast_fp16")]; + tensor x_5_begin_0 = const()[name = tensor("x_5_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_5_end_0 = const()[name = tensor("x_5_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_5_end_mask_0 = const()[name = tensor("x_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_5_squeeze_mask_0 = const()[name = tensor("x_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_5_cast_fp16 = slice_by_index(begin = x_5_begin_0, end = x_5_end_0, end_mask = x_5_end_mask_0, squeeze_mask = x_5_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("x_5_cast_fp16")]; + tensor x_7_begin_0 = const()[name = tensor("x_7_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_7_end_0 = const()[name = tensor("x_7_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_7_end_mask_0 = const()[name = tensor("x_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_7_squeeze_mask_0 = const()[name = tensor("x_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_7_cast_fp16 = slice_by_index(begin = x_7_begin_0, end = x_7_end_0, end_mask = x_7_end_mask_0, squeeze_mask = x_7_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("x_7_cast_fp16")]; + tensor v_3_begin_0 = const()[name = tensor("v_3_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_3_end_0 = const()[name = tensor("v_3_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_3_end_mask_0 = const()[name = tensor("v_3_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_3_squeeze_mask_0 = const()[name = tensor("v_3_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_3_cast_fp16 = slice_by_index(begin = v_3_begin_0, end = v_3_end_0, end_mask = v_3_end_mask_0, squeeze_mask = v_3_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("v_3_cast_fp16")]; + tensor var_606 = const()[name = tensor("op_606"), val = tensor([256, 16, 32, 2])]; + tensor pairs_5_cast_fp16 = reshape(shape = var_606, x = x_5_cast_fp16)[name = tensor("pairs_5_cast_fp16")]; + tensor xr_5_begin_0 = const()[name = tensor("xr_5_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_5_end_0 = const()[name = tensor("xr_5_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_5_end_mask_0 = const()[name = tensor("xr_5_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_5_squeeze_mask_0 = const()[name = tensor("xr_5_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_5_cast_fp16 = slice_by_index(begin = xr_5_begin_0, end = xr_5_end_0, end_mask = xr_5_end_mask_0, squeeze_mask = xr_5_squeeze_mask_0, x = pairs_5_cast_fp16)[name = tensor("xr_5_cast_fp16")]; + tensor xi_5_begin_0 = const()[name = tensor("xi_5_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_5_end_0 = const()[name = tensor("xi_5_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_5_end_mask_0 = const()[name = tensor("xi_5_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_5_squeeze_mask_0 = const()[name = tensor("xi_5_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_5_cast_fp16 = slice_by_index(begin = xi_5_begin_0, end = xi_5_end_0, end_mask = xi_5_end_mask_0, squeeze_mask = xi_5_squeeze_mask_0, x = pairs_5_cast_fp16)[name = tensor("xi_5_cast_fp16")]; + tensor var_614_cast_fp16 = mul(x = xr_5_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_614_cast_fp16")]; + tensor var_615_cast_fp16 = mul(x = xi_5_cast_fp16, y = roti_cast_fp16)[name = tensor("op_615_cast_fp16")]; + tensor xor__5_cast_fp16 = sub(x = var_614_cast_fp16, y = var_615_cast_fp16)[name = tensor("xor__5_cast_fp16")]; + tensor var_618_cast_fp16 = mul(x = xr_5_cast_fp16, y = roti_cast_fp16)[name = tensor("op_618_cast_fp16")]; + tensor var_619_cast_fp16 = mul(x = xi_5_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_619_cast_fp16")]; + tensor xoi_5_cast_fp16 = add(x = var_618_cast_fp16, y = var_619_cast_fp16)[name = tensor("xoi_5_cast_fp16")]; + tensor out_5_axis_0 = const()[name = tensor("out_5_axis_0"), val = tensor(-1)]; + tensor out_5_cast_fp16 = stack(axis = out_5_axis_0, values = (xor__5_cast_fp16, xoi_5_cast_fp16))[name = tensor("out_5_cast_fp16")]; + tensor var_629 = const()[name = tensor("op_629"), val = tensor([1, 256, 16, 64])]; + tensor q_3_cast_fp16 = reshape(shape = var_629, x = out_5_cast_fp16)[name = tensor("q_3_cast_fp16")]; + tensor var_635 = const()[name = tensor("op_635"), val = tensor([256, 16, 32, 2])]; + tensor pairs_7_cast_fp16 = reshape(shape = var_635, x = x_7_cast_fp16)[name = tensor("pairs_7_cast_fp16")]; + tensor xr_7_begin_0 = const()[name = tensor("xr_7_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_7_end_0 = const()[name = tensor("xr_7_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_7_end_mask_0 = const()[name = tensor("xr_7_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_7_squeeze_mask_0 = const()[name = tensor("xr_7_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_7_cast_fp16 = slice_by_index(begin = xr_7_begin_0, end = xr_7_end_0, end_mask = xr_7_end_mask_0, squeeze_mask = xr_7_squeeze_mask_0, x = pairs_7_cast_fp16)[name = tensor("xr_7_cast_fp16")]; + tensor xi_7_begin_0 = const()[name = tensor("xi_7_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_7_end_0 = const()[name = tensor("xi_7_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_7_end_mask_0 = const()[name = tensor("xi_7_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_7_squeeze_mask_0 = const()[name = tensor("xi_7_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_7_cast_fp16 = slice_by_index(begin = xi_7_begin_0, end = xi_7_end_0, end_mask = xi_7_end_mask_0, squeeze_mask = xi_7_squeeze_mask_0, x = pairs_7_cast_fp16)[name = tensor("xi_7_cast_fp16")]; + tensor var_643_cast_fp16 = mul(x = xr_7_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_643_cast_fp16")]; + tensor var_644_cast_fp16 = mul(x = xi_7_cast_fp16, y = roti_cast_fp16)[name = tensor("op_644_cast_fp16")]; + tensor xor__7_cast_fp16 = sub(x = var_643_cast_fp16, y = var_644_cast_fp16)[name = tensor("xor__7_cast_fp16")]; + tensor var_647_cast_fp16 = mul(x = xr_7_cast_fp16, y = roti_cast_fp16)[name = tensor("op_647_cast_fp16")]; + tensor var_648_cast_fp16 = mul(x = xi_7_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_648_cast_fp16")]; + tensor xoi_7_cast_fp16 = add(x = var_647_cast_fp16, y = var_648_cast_fp16)[name = tensor("xoi_7_cast_fp16")]; + tensor out_7_axis_0 = const()[name = tensor("out_7_axis_0"), val = tensor(-1)]; + tensor out_7_cast_fp16 = stack(axis = out_7_axis_0, values = (xor__7_cast_fp16, xoi_7_cast_fp16))[name = tensor("out_7_cast_fp16")]; + tensor var_665 = const()[name = tensor("op_665"), val = tensor([256, 1024])]; + tensor var_666_cast_fp16 = reshape(shape = var_665, x = out_7_cast_fp16)[name = tensor("op_666_cast_fp16")]; + tensor written_k_3_transpose_x_1 = const()[name = tensor("written_k_3_transpose_x_1"), val = tensor(true)]; + tensor written_k_3_transpose_y_1 = const()[name = tensor("written_k_3_transpose_y_1"), val = tensor(false)]; + tensor written_k_3_cast_fp16 = matmul(transpose_x = written_k_3_transpose_x_1, transpose_y = written_k_3_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_666_cast_fp16)[name = tensor("written_k_3_cast_fp16")]; + tensor var_673 = const()[name = tensor("op_673"), val = tensor([256, 1024])]; + tensor var_674_cast_fp16 = reshape(shape = var_673, x = v_3_cast_fp16)[name = tensor("op_674_cast_fp16")]; + tensor written_v_3_transpose_x_1 = const()[name = tensor("written_v_3_transpose_x_1"), val = tensor(true)]; + tensor written_v_3_transpose_y_1 = const()[name = tensor("written_v_3_transpose_y_1"), val = tensor(false)]; + tensor written_v_3_cast_fp16 = matmul(transpose_x = written_v_3_transpose_x_1, transpose_y = written_v_3_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_674_cast_fp16)[name = tensor("written_v_3_cast_fp16")]; + tensor k_cache1_to_fp16_dtype_0 = const()[name = tensor("k_cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache1_to_fp16 = cast(dtype = k_cache1_to_fp16_dtype_0, x = k_cache1)[name = tensor("cast_459")]; + tensor var_685_cast_fp16 = mul(x = k_cache1_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_685_cast_fp16")]; + tensor var_690 = const()[name = tensor("op_690"), val = tensor([1, 512, 16, 64])]; + tensor var_691_cast_fp16 = reshape(shape = var_690, x = written_k_3_cast_fp16)[name = tensor("op_691_cast_fp16")]; + tensor new_k_cache_3_cast_fp16 = add(x = var_685_cast_fp16, y = var_691_cast_fp16)[name = tensor("new_k_cache_3_cast_fp16")]; + tensor new_k_cache_3_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_3_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache1_to_fp16_dtype_0 = const()[name = tensor("v_cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache1_to_fp16 = cast(dtype = v_cache1_to_fp16_dtype_0, x = v_cache1)[name = tensor("cast_457")]; + tensor var_694_cast_fp16 = mul(x = v_cache1_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_694_cast_fp16")]; + tensor var_699 = const()[name = tensor("op_699"), val = tensor([1, 512, 16, 64])]; + tensor var_700_cast_fp16 = reshape(shape = var_699, x = written_v_3_cast_fp16)[name = tensor("op_700_cast_fp16")]; + tensor new_v_cache_3_cast_fp16 = add(x = var_694_cast_fp16, y = var_700_cast_fp16)[name = tensor("new_v_cache_3_cast_fp16")]; + tensor new_v_cache_3_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_3_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_713 = const()[name = tensor("op_713"), val = tensor([0, 2, 1, 3])]; + tensor var_724_transpose_x_0 = const()[name = tensor("op_724_transpose_x_0"), val = tensor(false)]; + tensor var_724_transpose_y_0 = const()[name = tensor("op_724_transpose_y_0"), val = tensor(false)]; + tensor transpose_71_perm_0 = const()[name = tensor("transpose_71_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_72_perm_0 = const()[name = tensor("transpose_72_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_72 = transpose(perm = transpose_72_perm_0, x = new_k_cache_3_cast_fp16)[name = tensor("transpose_200")]; + tensor transpose_71 = transpose(perm = transpose_71_perm_0, x = q_3_cast_fp16)[name = tensor("transpose_201")]; + tensor var_724_cast_fp16 = matmul(transpose_x = var_724_transpose_x_0, transpose_y = var_724_transpose_y_0, x = transpose_71, y = transpose_72)[name = tensor("op_724_cast_fp16")]; + tensor var_725_to_fp16 = const()[name = tensor("op_725_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_5_cast_fp16 = mul(x = var_724_cast_fp16, y = var_725_to_fp16)[name = tensor("scores_5_cast_fp16")]; + tensor scores_7_cast_fp16 = add(x = scores_5_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_7_cast_fp16")]; + tensor var_735 = const()[name = tensor("op_735"), val = tensor(-1)]; + tensor weights_3_cast_fp16 = softmax(axis = var_735, x = scores_7_cast_fp16)[name = tensor("weights_3_cast_fp16")]; + tensor attn_3_transpose_x_0 = const()[name = tensor("attn_3_transpose_x_0"), val = tensor(false)]; + tensor attn_3_transpose_y_0 = const()[name = tensor("attn_3_transpose_y_0"), val = tensor(false)]; + tensor values_3_cast_fp16 = transpose(perm = var_713, x = new_v_cache_3_cast_fp16)[name = tensor("transpose_202")]; + tensor attn_3_cast_fp16 = matmul(transpose_x = attn_3_transpose_x_0, transpose_y = attn_3_transpose_y_0, x = weights_3_cast_fp16, y = values_3_cast_fp16)[name = tensor("attn_3_cast_fp16")]; + tensor var_743 = const()[name = tensor("op_743"), val = tensor([0, 2, 1, 3])]; + tensor var_748 = const()[name = tensor("op_748"), val = tensor([1, 256, 1024])]; + tensor var_744_cast_fp16 = transpose(perm = var_743, x = attn_3_cast_fp16)[name = tensor("transpose_199")]; + tensor input_17_cast_fp16 = reshape(shape = var_748, x = var_744_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor attn1_out_proj_weight_to_fp16 = const()[name = tensor("attn1_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31489280)))]; + tensor linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn1_out_proj_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor position1_to_fp16_dtype_0 = const()[name = tensor("position1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position1_to_fp16 = cast(dtype = position1_to_fp16_dtype_0, x = position1)[name = tensor("cast_455")]; + tensor var_757_cast_fp16 = add(x = position1_to_fp16, y = valid_len_to_fp16)[name = tensor("op_757_cast_fp16")]; + tensor var_757_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_757_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_19_cast_fp16 = add(x = input_13_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor input_21_axes_0 = const()[name = tensor("input_21_axes_0"), val = tensor([-1])]; + tensor norm1_2_weight_to_fp16 = const()[name = tensor("norm1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33586496)))]; + tensor norm1_2_bias_to_fp16 = const()[name = tensor("norm1_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33588608)))]; + tensor var_761_to_fp16 = const()[name = tensor("op_761_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_21_cast_fp16 = layer_norm(axes = input_21_axes_0, beta = norm1_2_bias_to_fp16, epsilon = var_761_to_fp16, gamma = norm1_2_weight_to_fp16, x = input_19_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor linear1_1_weight_to_fp16 = const()[name = tensor("linear1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33590720)))]; + tensor linear_6_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear1_1_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor input_23_mode_0 = const()[name = tensor("input_23_mode_0"), val = tensor("EXACT")]; + tensor input_23_cast_fp16 = gelu(mode = input_23_mode_0, x = linear_6_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor linear1_2_weight_to_fp16 = const()[name = tensor("linear1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41979392)))]; + tensor linear_7_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear1_2_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_25_cast_fp16 = add(x = input_19_cast_fp16, y = linear_7_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor input_27_axes_0 = const()[name = tensor("input_27_axes_0"), val = tensor([-1])]; + tensor norm2_1_weight_to_fp16 = const()[name = tensor("norm2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50368064)))]; + tensor norm2_1_bias_to_fp16 = const()[name = tensor("norm2_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50370176)))]; + tensor var_778_to_fp16 = const()[name = tensor("op_778_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = norm2_1_bias_to_fp16, epsilon = var_778_to_fp16, gamma = norm2_1_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor attn2_in_proj_weight_to_fp16 = const()[name = tensor("attn2_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50372288)))]; + tensor linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn2_in_proj_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor var_792 = const()[name = tensor("op_792"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_5_cast_fp16 = reshape(shape = var_792, x = linear_8_cast_fp16)[name = tensor("qkv_5_cast_fp16")]; + tensor x_9_begin_0 = const()[name = tensor("x_9_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_9_end_0 = const()[name = tensor("x_9_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_9_end_mask_0 = const()[name = tensor("x_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_9_squeeze_mask_0 = const()[name = tensor("x_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_9_cast_fp16 = slice_by_index(begin = x_9_begin_0, end = x_9_end_0, end_mask = x_9_end_mask_0, squeeze_mask = x_9_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("x_9_cast_fp16")]; + tensor x_11_begin_0 = const()[name = tensor("x_11_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_11_end_0 = const()[name = tensor("x_11_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_11_end_mask_0 = const()[name = tensor("x_11_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_11_squeeze_mask_0 = const()[name = tensor("x_11_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_11_cast_fp16 = slice_by_index(begin = x_11_begin_0, end = x_11_end_0, end_mask = x_11_end_mask_0, squeeze_mask = x_11_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("x_11_cast_fp16")]; + tensor v_5_begin_0 = const()[name = tensor("v_5_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_5_end_0 = const()[name = tensor("v_5_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_5_end_mask_0 = const()[name = tensor("v_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_5_squeeze_mask_0 = const()[name = tensor("v_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_5_cast_fp16 = slice_by_index(begin = v_5_begin_0, end = v_5_end_0, end_mask = v_5_end_mask_0, squeeze_mask = v_5_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("v_5_cast_fp16")]; + tensor var_837 = const()[name = tensor("op_837"), val = tensor([256, 16, 32, 2])]; + tensor pairs_9_cast_fp16 = reshape(shape = var_837, x = x_9_cast_fp16)[name = tensor("pairs_9_cast_fp16")]; + tensor xr_9_begin_0 = const()[name = tensor("xr_9_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_9_end_0 = const()[name = tensor("xr_9_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_9_end_mask_0 = const()[name = tensor("xr_9_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_9_squeeze_mask_0 = const()[name = tensor("xr_9_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_9_cast_fp16 = slice_by_index(begin = xr_9_begin_0, end = xr_9_end_0, end_mask = xr_9_end_mask_0, squeeze_mask = xr_9_squeeze_mask_0, x = pairs_9_cast_fp16)[name = tensor("xr_9_cast_fp16")]; + tensor xi_9_begin_0 = const()[name = tensor("xi_9_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_9_end_0 = const()[name = tensor("xi_9_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_9_end_mask_0 = const()[name = tensor("xi_9_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_9_squeeze_mask_0 = const()[name = tensor("xi_9_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_9_cast_fp16 = slice_by_index(begin = xi_9_begin_0, end = xi_9_end_0, end_mask = xi_9_end_mask_0, squeeze_mask = xi_9_squeeze_mask_0, x = pairs_9_cast_fp16)[name = tensor("xi_9_cast_fp16")]; + tensor var_845_cast_fp16 = mul(x = xr_9_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_845_cast_fp16")]; + tensor var_846_cast_fp16 = mul(x = xi_9_cast_fp16, y = roti_cast_fp16)[name = tensor("op_846_cast_fp16")]; + tensor xor__9_cast_fp16 = sub(x = var_845_cast_fp16, y = var_846_cast_fp16)[name = tensor("xor__9_cast_fp16")]; + tensor var_849_cast_fp16 = mul(x = xr_9_cast_fp16, y = roti_cast_fp16)[name = tensor("op_849_cast_fp16")]; + tensor var_850_cast_fp16 = mul(x = xi_9_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_850_cast_fp16")]; + tensor xoi_9_cast_fp16 = add(x = var_849_cast_fp16, y = var_850_cast_fp16)[name = tensor("xoi_9_cast_fp16")]; + tensor out_9_axis_0 = const()[name = tensor("out_9_axis_0"), val = tensor(-1)]; + tensor out_9_cast_fp16 = stack(axis = out_9_axis_0, values = (xor__9_cast_fp16, xoi_9_cast_fp16))[name = tensor("out_9_cast_fp16")]; + tensor var_860 = const()[name = tensor("op_860"), val = tensor([1, 256, 16, 64])]; + tensor q_5_cast_fp16 = reshape(shape = var_860, x = out_9_cast_fp16)[name = tensor("q_5_cast_fp16")]; + tensor var_866 = const()[name = tensor("op_866"), val = tensor([256, 16, 32, 2])]; + tensor pairs_11_cast_fp16 = reshape(shape = var_866, x = x_11_cast_fp16)[name = tensor("pairs_11_cast_fp16")]; + tensor xr_11_begin_0 = const()[name = tensor("xr_11_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_11_end_0 = const()[name = tensor("xr_11_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_11_end_mask_0 = const()[name = tensor("xr_11_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_11_squeeze_mask_0 = const()[name = tensor("xr_11_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_11_cast_fp16 = slice_by_index(begin = xr_11_begin_0, end = xr_11_end_0, end_mask = xr_11_end_mask_0, squeeze_mask = xr_11_squeeze_mask_0, x = pairs_11_cast_fp16)[name = tensor("xr_11_cast_fp16")]; + tensor xi_11_begin_0 = const()[name = tensor("xi_11_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_11_end_0 = const()[name = tensor("xi_11_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_11_end_mask_0 = const()[name = tensor("xi_11_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_11_squeeze_mask_0 = const()[name = tensor("xi_11_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_11_cast_fp16 = slice_by_index(begin = xi_11_begin_0, end = xi_11_end_0, end_mask = xi_11_end_mask_0, squeeze_mask = xi_11_squeeze_mask_0, x = pairs_11_cast_fp16)[name = tensor("xi_11_cast_fp16")]; + tensor var_874_cast_fp16 = mul(x = xr_11_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_874_cast_fp16")]; + tensor var_875_cast_fp16 = mul(x = xi_11_cast_fp16, y = roti_cast_fp16)[name = tensor("op_875_cast_fp16")]; + tensor xor__11_cast_fp16 = sub(x = var_874_cast_fp16, y = var_875_cast_fp16)[name = tensor("xor__11_cast_fp16")]; + tensor var_878_cast_fp16 = mul(x = xr_11_cast_fp16, y = roti_cast_fp16)[name = tensor("op_878_cast_fp16")]; + tensor var_879_cast_fp16 = mul(x = xi_11_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_879_cast_fp16")]; + tensor xoi_11_cast_fp16 = add(x = var_878_cast_fp16, y = var_879_cast_fp16)[name = tensor("xoi_11_cast_fp16")]; + tensor out_11_axis_0 = const()[name = tensor("out_11_axis_0"), val = tensor(-1)]; + tensor out_11_cast_fp16 = stack(axis = out_11_axis_0, values = (xor__11_cast_fp16, xoi_11_cast_fp16))[name = tensor("out_11_cast_fp16")]; + tensor var_896 = const()[name = tensor("op_896"), val = tensor([256, 1024])]; + tensor var_897_cast_fp16 = reshape(shape = var_896, x = out_11_cast_fp16)[name = tensor("op_897_cast_fp16")]; + tensor written_k_5_transpose_x_1 = const()[name = tensor("written_k_5_transpose_x_1"), val = tensor(true)]; + tensor written_k_5_transpose_y_1 = const()[name = tensor("written_k_5_transpose_y_1"), val = tensor(false)]; + tensor written_k_5_cast_fp16 = matmul(transpose_x = written_k_5_transpose_x_1, transpose_y = written_k_5_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_897_cast_fp16)[name = tensor("written_k_5_cast_fp16")]; + tensor var_904 = const()[name = tensor("op_904"), val = tensor([256, 1024])]; + tensor var_905_cast_fp16 = reshape(shape = var_904, x = v_5_cast_fp16)[name = tensor("op_905_cast_fp16")]; + tensor written_v_5_transpose_x_1 = const()[name = tensor("written_v_5_transpose_x_1"), val = tensor(true)]; + tensor written_v_5_transpose_y_1 = const()[name = tensor("written_v_5_transpose_y_1"), val = tensor(false)]; + tensor written_v_5_cast_fp16 = matmul(transpose_x = written_v_5_transpose_x_1, transpose_y = written_v_5_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_905_cast_fp16)[name = tensor("written_v_5_cast_fp16")]; + tensor k_cache2_to_fp16_dtype_0 = const()[name = tensor("k_cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache2_to_fp16 = cast(dtype = k_cache2_to_fp16_dtype_0, x = k_cache2)[name = tensor("cast_453")]; + tensor var_916_cast_fp16 = mul(x = k_cache2_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_916_cast_fp16")]; + tensor var_921 = const()[name = tensor("op_921"), val = tensor([1, 512, 16, 64])]; + tensor var_922_cast_fp16 = reshape(shape = var_921, x = written_k_5_cast_fp16)[name = tensor("op_922_cast_fp16")]; + tensor new_k_cache_5_cast_fp16 = add(x = var_916_cast_fp16, y = var_922_cast_fp16)[name = tensor("new_k_cache_5_cast_fp16")]; + tensor new_k_cache_5_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_5_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache2_to_fp16_dtype_0 = const()[name = tensor("v_cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache2_to_fp16 = cast(dtype = v_cache2_to_fp16_dtype_0, x = v_cache2)[name = tensor("cast_451")]; + tensor var_925_cast_fp16 = mul(x = v_cache2_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_925_cast_fp16")]; + tensor var_930 = const()[name = tensor("op_930"), val = tensor([1, 512, 16, 64])]; + tensor var_931_cast_fp16 = reshape(shape = var_930, x = written_v_5_cast_fp16)[name = tensor("op_931_cast_fp16")]; + tensor new_v_cache_5_cast_fp16 = add(x = var_925_cast_fp16, y = var_931_cast_fp16)[name = tensor("new_v_cache_5_cast_fp16")]; + tensor new_v_cache_5_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_5_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_944 = const()[name = tensor("op_944"), val = tensor([0, 2, 1, 3])]; + tensor var_955_transpose_x_0 = const()[name = tensor("op_955_transpose_x_0"), val = tensor(false)]; + tensor var_955_transpose_y_0 = const()[name = tensor("op_955_transpose_y_0"), val = tensor(false)]; + tensor transpose_73_perm_0 = const()[name = tensor("transpose_73_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_74_perm_0 = const()[name = tensor("transpose_74_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_74 = transpose(perm = transpose_74_perm_0, x = new_k_cache_5_cast_fp16)[name = tensor("transpose_196")]; + tensor transpose_73 = transpose(perm = transpose_73_perm_0, x = q_5_cast_fp16)[name = tensor("transpose_197")]; + tensor var_955_cast_fp16 = matmul(transpose_x = var_955_transpose_x_0, transpose_y = var_955_transpose_y_0, x = transpose_73, y = transpose_74)[name = tensor("op_955_cast_fp16")]; + tensor var_956_to_fp16 = const()[name = tensor("op_956_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_9_cast_fp16 = mul(x = var_955_cast_fp16, y = var_956_to_fp16)[name = tensor("scores_9_cast_fp16")]; + tensor scores_11_cast_fp16 = add(x = scores_9_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_11_cast_fp16")]; + tensor var_966 = const()[name = tensor("op_966"), val = tensor(-1)]; + tensor weights_5_cast_fp16 = softmax(axis = var_966, x = scores_11_cast_fp16)[name = tensor("weights_5_cast_fp16")]; + tensor attn_5_transpose_x_0 = const()[name = tensor("attn_5_transpose_x_0"), val = tensor(false)]; + tensor attn_5_transpose_y_0 = const()[name = tensor("attn_5_transpose_y_0"), val = tensor(false)]; + tensor values_5_cast_fp16 = transpose(perm = var_944, x = new_v_cache_5_cast_fp16)[name = tensor("transpose_198")]; + tensor attn_5_cast_fp16 = matmul(transpose_x = attn_5_transpose_x_0, transpose_y = attn_5_transpose_y_0, x = weights_5_cast_fp16, y = values_5_cast_fp16)[name = tensor("attn_5_cast_fp16")]; + tensor var_974 = const()[name = tensor("op_974"), val = tensor([0, 2, 1, 3])]; + tensor var_979 = const()[name = tensor("op_979"), val = tensor([1, 256, 1024])]; + tensor var_975_cast_fp16 = transpose(perm = var_974, x = attn_5_cast_fp16)[name = tensor("transpose_195")]; + tensor input_29_cast_fp16 = reshape(shape = var_979, x = var_975_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor attn2_out_proj_weight_to_fp16 = const()[name = tensor("attn2_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(56663808)))]; + tensor linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn2_out_proj_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor position2_to_fp16_dtype_0 = const()[name = tensor("position2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position2_to_fp16 = cast(dtype = position2_to_fp16_dtype_0, x = position2)[name = tensor("cast_449")]; + tensor var_988_cast_fp16 = add(x = position2_to_fp16, y = valid_len_to_fp16)[name = tensor("op_988_cast_fp16")]; + tensor var_988_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_988_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_31_cast_fp16 = add(x = input_25_cast_fp16, y = linear_9_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; + tensor norm2_2_weight_to_fp16 = const()[name = tensor("norm2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58761024)))]; + tensor norm2_2_bias_to_fp16 = const()[name = tensor("norm2_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58763136)))]; + tensor var_992_to_fp16 = const()[name = tensor("op_992_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_33_cast_fp16 = layer_norm(axes = input_33_axes_0, beta = norm2_2_bias_to_fp16, epsilon = var_992_to_fp16, gamma = norm2_2_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor linear2_1_weight_to_fp16 = const()[name = tensor("linear2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58765248)))]; + tensor linear_10_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear2_1_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor input_35_mode_0 = const()[name = tensor("input_35_mode_0"), val = tensor("EXACT")]; + tensor input_35_cast_fp16 = gelu(mode = input_35_mode_0, x = linear_10_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor linear2_2_weight_to_fp16 = const()[name = tensor("linear2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67153920)))]; + tensor linear_11_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear2_2_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor input_37_cast_fp16 = add(x = input_31_cast_fp16, y = linear_11_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor input_39_axes_0 = const()[name = tensor("input_39_axes_0"), val = tensor([-1])]; + tensor norm3_1_weight_to_fp16 = const()[name = tensor("norm3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75542592)))]; + tensor norm3_1_bias_to_fp16 = const()[name = tensor("norm3_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75544704)))]; + tensor var_1009_to_fp16 = const()[name = tensor("op_1009_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_39_cast_fp16 = layer_norm(axes = input_39_axes_0, beta = norm3_1_bias_to_fp16, epsilon = var_1009_to_fp16, gamma = norm3_1_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor attn3_in_proj_weight_to_fp16 = const()[name = tensor("attn3_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75546816)))]; + tensor linear_12_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn3_in_proj_weight_to_fp16, x = input_39_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor var_1023 = const()[name = tensor("op_1023"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_7_cast_fp16 = reshape(shape = var_1023, x = linear_12_cast_fp16)[name = tensor("qkv_7_cast_fp16")]; + tensor x_13_begin_0 = const()[name = tensor("x_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_13_end_0 = const()[name = tensor("x_13_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_13_end_mask_0 = const()[name = tensor("x_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_13_squeeze_mask_0 = const()[name = tensor("x_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_13_cast_fp16 = slice_by_index(begin = x_13_begin_0, end = x_13_end_0, end_mask = x_13_end_mask_0, squeeze_mask = x_13_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("x_13_cast_fp16")]; + tensor x_15_begin_0 = const()[name = tensor("x_15_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_15_end_0 = const()[name = tensor("x_15_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_15_end_mask_0 = const()[name = tensor("x_15_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_15_squeeze_mask_0 = const()[name = tensor("x_15_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_15_cast_fp16 = slice_by_index(begin = x_15_begin_0, end = x_15_end_0, end_mask = x_15_end_mask_0, squeeze_mask = x_15_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("x_15_cast_fp16")]; + tensor v_7_begin_0 = const()[name = tensor("v_7_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_7_end_0 = const()[name = tensor("v_7_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_7_end_mask_0 = const()[name = tensor("v_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_7_squeeze_mask_0 = const()[name = tensor("v_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_7_cast_fp16 = slice_by_index(begin = v_7_begin_0, end = v_7_end_0, end_mask = v_7_end_mask_0, squeeze_mask = v_7_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("v_7_cast_fp16")]; + tensor var_1068 = const()[name = tensor("op_1068"), val = tensor([256, 16, 32, 2])]; + tensor pairs_13_cast_fp16 = reshape(shape = var_1068, x = x_13_cast_fp16)[name = tensor("pairs_13_cast_fp16")]; + tensor xr_13_begin_0 = const()[name = tensor("xr_13_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_13_end_0 = const()[name = tensor("xr_13_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_13_end_mask_0 = const()[name = tensor("xr_13_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_13_squeeze_mask_0 = const()[name = tensor("xr_13_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_13_cast_fp16 = slice_by_index(begin = xr_13_begin_0, end = xr_13_end_0, end_mask = xr_13_end_mask_0, squeeze_mask = xr_13_squeeze_mask_0, x = pairs_13_cast_fp16)[name = tensor("xr_13_cast_fp16")]; + tensor xi_13_begin_0 = const()[name = tensor("xi_13_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_13_end_0 = const()[name = tensor("xi_13_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_13_end_mask_0 = const()[name = tensor("xi_13_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_13_squeeze_mask_0 = const()[name = tensor("xi_13_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_13_cast_fp16 = slice_by_index(begin = xi_13_begin_0, end = xi_13_end_0, end_mask = xi_13_end_mask_0, squeeze_mask = xi_13_squeeze_mask_0, x = pairs_13_cast_fp16)[name = tensor("xi_13_cast_fp16")]; + tensor var_1076_cast_fp16 = mul(x = xr_13_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1076_cast_fp16")]; + tensor var_1077_cast_fp16 = mul(x = xi_13_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1077_cast_fp16")]; + tensor xor__13_cast_fp16 = sub(x = var_1076_cast_fp16, y = var_1077_cast_fp16)[name = tensor("xor__13_cast_fp16")]; + tensor var_1080_cast_fp16 = mul(x = xr_13_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1080_cast_fp16")]; + tensor var_1081_cast_fp16 = mul(x = xi_13_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1081_cast_fp16")]; + tensor xoi_13_cast_fp16 = add(x = var_1080_cast_fp16, y = var_1081_cast_fp16)[name = tensor("xoi_13_cast_fp16")]; + tensor out_13_axis_0 = const()[name = tensor("out_13_axis_0"), val = tensor(-1)]; + tensor out_13_cast_fp16 = stack(axis = out_13_axis_0, values = (xor__13_cast_fp16, xoi_13_cast_fp16))[name = tensor("out_13_cast_fp16")]; + tensor var_1091 = const()[name = tensor("op_1091"), val = tensor([1, 256, 16, 64])]; + tensor q_7_cast_fp16 = reshape(shape = var_1091, x = out_13_cast_fp16)[name = tensor("q_7_cast_fp16")]; + tensor var_1097 = const()[name = tensor("op_1097"), val = tensor([256, 16, 32, 2])]; + tensor pairs_15_cast_fp16 = reshape(shape = var_1097, x = x_15_cast_fp16)[name = tensor("pairs_15_cast_fp16")]; + tensor xr_15_begin_0 = const()[name = tensor("xr_15_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_15_end_0 = const()[name = tensor("xr_15_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_15_end_mask_0 = const()[name = tensor("xr_15_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_15_squeeze_mask_0 = const()[name = tensor("xr_15_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_15_cast_fp16 = slice_by_index(begin = xr_15_begin_0, end = xr_15_end_0, end_mask = xr_15_end_mask_0, squeeze_mask = xr_15_squeeze_mask_0, x = pairs_15_cast_fp16)[name = tensor("xr_15_cast_fp16")]; + tensor xi_15_begin_0 = const()[name = tensor("xi_15_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_15_end_0 = const()[name = tensor("xi_15_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_15_end_mask_0 = const()[name = tensor("xi_15_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_15_squeeze_mask_0 = const()[name = tensor("xi_15_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_15_cast_fp16 = slice_by_index(begin = xi_15_begin_0, end = xi_15_end_0, end_mask = xi_15_end_mask_0, squeeze_mask = xi_15_squeeze_mask_0, x = pairs_15_cast_fp16)[name = tensor("xi_15_cast_fp16")]; + tensor var_1105_cast_fp16 = mul(x = xr_15_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1105_cast_fp16")]; + tensor var_1106_cast_fp16 = mul(x = xi_15_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1106_cast_fp16")]; + tensor xor__15_cast_fp16 = sub(x = var_1105_cast_fp16, y = var_1106_cast_fp16)[name = tensor("xor__15_cast_fp16")]; + tensor var_1109_cast_fp16 = mul(x = xr_15_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1109_cast_fp16")]; + tensor var_1110_cast_fp16 = mul(x = xi_15_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1110_cast_fp16")]; + tensor xoi_15_cast_fp16 = add(x = var_1109_cast_fp16, y = var_1110_cast_fp16)[name = tensor("xoi_15_cast_fp16")]; + tensor out_15_axis_0 = const()[name = tensor("out_15_axis_0"), val = tensor(-1)]; + tensor out_15_cast_fp16 = stack(axis = out_15_axis_0, values = (xor__15_cast_fp16, xoi_15_cast_fp16))[name = tensor("out_15_cast_fp16")]; + tensor var_1127 = const()[name = tensor("op_1127"), val = tensor([256, 1024])]; + tensor var_1128_cast_fp16 = reshape(shape = var_1127, x = out_15_cast_fp16)[name = tensor("op_1128_cast_fp16")]; + tensor written_k_7_transpose_x_1 = const()[name = tensor("written_k_7_transpose_x_1"), val = tensor(true)]; + tensor written_k_7_transpose_y_1 = const()[name = tensor("written_k_7_transpose_y_1"), val = tensor(false)]; + tensor written_k_7_cast_fp16 = matmul(transpose_x = written_k_7_transpose_x_1, transpose_y = written_k_7_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_1128_cast_fp16)[name = tensor("written_k_7_cast_fp16")]; + tensor var_1135 = const()[name = tensor("op_1135"), val = tensor([256, 1024])]; + tensor var_1136_cast_fp16 = reshape(shape = var_1135, x = v_7_cast_fp16)[name = tensor("op_1136_cast_fp16")]; + tensor written_v_7_transpose_x_1 = const()[name = tensor("written_v_7_transpose_x_1"), val = tensor(true)]; + tensor written_v_7_transpose_y_1 = const()[name = tensor("written_v_7_transpose_y_1"), val = tensor(false)]; + tensor written_v_7_cast_fp16 = matmul(transpose_x = written_v_7_transpose_x_1, transpose_y = written_v_7_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_1136_cast_fp16)[name = tensor("written_v_7_cast_fp16")]; + tensor k_cache3_to_fp16_dtype_0 = const()[name = tensor("k_cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache3_to_fp16 = cast(dtype = k_cache3_to_fp16_dtype_0, x = k_cache3)[name = tensor("cast_447")]; + tensor var_1147_cast_fp16 = mul(x = k_cache3_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_1147_cast_fp16")]; + tensor var_1152 = const()[name = tensor("op_1152"), val = tensor([1, 512, 16, 64])]; + tensor var_1153_cast_fp16 = reshape(shape = var_1152, x = written_k_7_cast_fp16)[name = tensor("op_1153_cast_fp16")]; + tensor new_k_cache_7_cast_fp16 = add(x = var_1147_cast_fp16, y = var_1153_cast_fp16)[name = tensor("new_k_cache_7_cast_fp16")]; + tensor new_k_cache_7_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_7_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache3_to_fp16_dtype_0 = const()[name = tensor("v_cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache3_to_fp16 = cast(dtype = v_cache3_to_fp16_dtype_0, x = v_cache3)[name = tensor("cast_445")]; + tensor var_1156_cast_fp16 = mul(x = v_cache3_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_1156_cast_fp16")]; + tensor var_1161 = const()[name = tensor("op_1161"), val = tensor([1, 512, 16, 64])]; + tensor var_1162_cast_fp16 = reshape(shape = var_1161, x = written_v_7_cast_fp16)[name = tensor("op_1162_cast_fp16")]; + tensor new_v_cache_7_cast_fp16 = add(x = var_1156_cast_fp16, y = var_1162_cast_fp16)[name = tensor("new_v_cache_7_cast_fp16")]; + tensor new_v_cache_7_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_7_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_1175 = const()[name = tensor("op_1175"), val = tensor([0, 2, 1, 3])]; + tensor var_1186_transpose_x_0 = const()[name = tensor("op_1186_transpose_x_0"), val = tensor(false)]; + tensor var_1186_transpose_y_0 = const()[name = tensor("op_1186_transpose_y_0"), val = tensor(false)]; + tensor transpose_75_perm_0 = const()[name = tensor("transpose_75_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_76_perm_0 = const()[name = tensor("transpose_76_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_76 = transpose(perm = transpose_76_perm_0, x = new_k_cache_7_cast_fp16)[name = tensor("transpose_192")]; + tensor transpose_75 = transpose(perm = transpose_75_perm_0, x = q_7_cast_fp16)[name = tensor("transpose_193")]; + tensor var_1186_cast_fp16 = matmul(transpose_x = var_1186_transpose_x_0, transpose_y = var_1186_transpose_y_0, x = transpose_75, y = transpose_76)[name = tensor("op_1186_cast_fp16")]; + tensor var_1187_to_fp16 = const()[name = tensor("op_1187_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_13_cast_fp16 = mul(x = var_1186_cast_fp16, y = var_1187_to_fp16)[name = tensor("scores_13_cast_fp16")]; + tensor scores_15_cast_fp16 = add(x = scores_13_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_15_cast_fp16")]; + tensor var_1197 = const()[name = tensor("op_1197"), val = tensor(-1)]; + tensor weights_7_cast_fp16 = softmax(axis = var_1197, x = scores_15_cast_fp16)[name = tensor("weights_7_cast_fp16")]; + tensor attn_7_transpose_x_0 = const()[name = tensor("attn_7_transpose_x_0"), val = tensor(false)]; + tensor attn_7_transpose_y_0 = const()[name = tensor("attn_7_transpose_y_0"), val = tensor(false)]; + tensor values_7_cast_fp16 = transpose(perm = var_1175, x = new_v_cache_7_cast_fp16)[name = tensor("transpose_194")]; + tensor attn_7_cast_fp16 = matmul(transpose_x = attn_7_transpose_x_0, transpose_y = attn_7_transpose_y_0, x = weights_7_cast_fp16, y = values_7_cast_fp16)[name = tensor("attn_7_cast_fp16")]; + tensor var_1205 = const()[name = tensor("op_1205"), val = tensor([0, 2, 1, 3])]; + tensor var_1210 = const()[name = tensor("op_1210"), val = tensor([1, 256, 1024])]; + tensor var_1206_cast_fp16 = transpose(perm = var_1205, x = attn_7_cast_fp16)[name = tensor("transpose_191")]; + tensor input_41_cast_fp16 = reshape(shape = var_1210, x = var_1206_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor attn3_out_proj_weight_to_fp16 = const()[name = tensor("attn3_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(81838336)))]; + tensor linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn3_out_proj_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor position3_to_fp16_dtype_0 = const()[name = tensor("position3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position3_to_fp16 = cast(dtype = position3_to_fp16_dtype_0, x = position3)[name = tensor("cast_443")]; + tensor var_1219_cast_fp16 = add(x = position3_to_fp16, y = valid_len_to_fp16)[name = tensor("op_1219_cast_fp16")]; + tensor var_1219_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1219_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_43_cast_fp16 = add(x = input_37_cast_fp16, y = linear_13_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; + tensor norm3_2_weight_to_fp16 = const()[name = tensor("norm3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(83935552)))]; + tensor norm3_2_bias_to_fp16 = const()[name = tensor("norm3_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(83937664)))]; + tensor var_1223_to_fp16 = const()[name = tensor("op_1223_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_45_cast_fp16 = layer_norm(axes = input_45_axes_0, beta = norm3_2_bias_to_fp16, epsilon = var_1223_to_fp16, gamma = norm3_2_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor linear3_1_weight_to_fp16 = const()[name = tensor("linear3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(83939776)))]; + tensor linear_14_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear3_1_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor input_47_mode_0 = const()[name = tensor("input_47_mode_0"), val = tensor("EXACT")]; + tensor input_47_cast_fp16 = gelu(mode = input_47_mode_0, x = linear_14_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor linear3_2_weight_to_fp16 = const()[name = tensor("linear3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(92328448)))]; + tensor linear_15_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear3_2_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor input_49_cast_fp16 = add(x = input_43_cast_fp16, y = linear_15_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor input_51_axes_0 = const()[name = tensor("input_51_axes_0"), val = tensor([-1])]; + tensor norm4_1_weight_to_fp16 = const()[name = tensor("norm4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100717120)))]; + tensor norm4_1_bias_to_fp16 = const()[name = tensor("norm4_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100719232)))]; + tensor var_1240_to_fp16 = const()[name = tensor("op_1240_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_51_cast_fp16 = layer_norm(axes = input_51_axes_0, beta = norm4_1_bias_to_fp16, epsilon = var_1240_to_fp16, gamma = norm4_1_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor attn4_in_proj_weight_to_fp16 = const()[name = tensor("attn4_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100721344)))]; + tensor linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn4_in_proj_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor var_1254 = const()[name = tensor("op_1254"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_9_cast_fp16 = reshape(shape = var_1254, x = linear_16_cast_fp16)[name = tensor("qkv_9_cast_fp16")]; + tensor x_17_begin_0 = const()[name = tensor("x_17_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_17_end_0 = const()[name = tensor("x_17_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_17_end_mask_0 = const()[name = tensor("x_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_17_squeeze_mask_0 = const()[name = tensor("x_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_17_cast_fp16 = slice_by_index(begin = x_17_begin_0, end = x_17_end_0, end_mask = x_17_end_mask_0, squeeze_mask = x_17_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("x_17_cast_fp16")]; + tensor x_19_begin_0 = const()[name = tensor("x_19_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_19_end_0 = const()[name = tensor("x_19_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_19_end_mask_0 = const()[name = tensor("x_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_19_squeeze_mask_0 = const()[name = tensor("x_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_19_cast_fp16 = slice_by_index(begin = x_19_begin_0, end = x_19_end_0, end_mask = x_19_end_mask_0, squeeze_mask = x_19_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("x_19_cast_fp16")]; + tensor v_9_begin_0 = const()[name = tensor("v_9_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_9_end_0 = const()[name = tensor("v_9_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_9_end_mask_0 = const()[name = tensor("v_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_9_squeeze_mask_0 = const()[name = tensor("v_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_9_cast_fp16 = slice_by_index(begin = v_9_begin_0, end = v_9_end_0, end_mask = v_9_end_mask_0, squeeze_mask = v_9_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("v_9_cast_fp16")]; + tensor var_1299 = const()[name = tensor("op_1299"), val = tensor([256, 16, 32, 2])]; + tensor pairs_17_cast_fp16 = reshape(shape = var_1299, x = x_17_cast_fp16)[name = tensor("pairs_17_cast_fp16")]; + tensor xr_17_begin_0 = const()[name = tensor("xr_17_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_17_end_0 = const()[name = tensor("xr_17_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_17_end_mask_0 = const()[name = tensor("xr_17_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_17_squeeze_mask_0 = const()[name = tensor("xr_17_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_17_cast_fp16 = slice_by_index(begin = xr_17_begin_0, end = xr_17_end_0, end_mask = xr_17_end_mask_0, squeeze_mask = xr_17_squeeze_mask_0, x = pairs_17_cast_fp16)[name = tensor("xr_17_cast_fp16")]; + tensor xi_17_begin_0 = const()[name = tensor("xi_17_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_17_end_0 = const()[name = tensor("xi_17_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_17_end_mask_0 = const()[name = tensor("xi_17_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_17_squeeze_mask_0 = const()[name = tensor("xi_17_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_17_cast_fp16 = slice_by_index(begin = xi_17_begin_0, end = xi_17_end_0, end_mask = xi_17_end_mask_0, squeeze_mask = xi_17_squeeze_mask_0, x = pairs_17_cast_fp16)[name = tensor("xi_17_cast_fp16")]; + tensor var_1307_cast_fp16 = mul(x = xr_17_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1307_cast_fp16")]; + tensor var_1308_cast_fp16 = mul(x = xi_17_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1308_cast_fp16")]; + tensor xor__17_cast_fp16 = sub(x = var_1307_cast_fp16, y = var_1308_cast_fp16)[name = tensor("xor__17_cast_fp16")]; + tensor var_1311_cast_fp16 = mul(x = xr_17_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1311_cast_fp16")]; + tensor var_1312_cast_fp16 = mul(x = xi_17_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1312_cast_fp16")]; + tensor xoi_17_cast_fp16 = add(x = var_1311_cast_fp16, y = var_1312_cast_fp16)[name = tensor("xoi_17_cast_fp16")]; + tensor out_17_axis_0 = const()[name = tensor("out_17_axis_0"), val = tensor(-1)]; + tensor out_17_cast_fp16 = stack(axis = out_17_axis_0, values = (xor__17_cast_fp16, xoi_17_cast_fp16))[name = tensor("out_17_cast_fp16")]; + tensor var_1322 = const()[name = tensor("op_1322"), val = tensor([1, 256, 16, 64])]; + tensor q_9_cast_fp16 = reshape(shape = var_1322, x = out_17_cast_fp16)[name = tensor("q_9_cast_fp16")]; + tensor var_1328 = const()[name = tensor("op_1328"), val = tensor([256, 16, 32, 2])]; + tensor pairs_19_cast_fp16 = reshape(shape = var_1328, x = x_19_cast_fp16)[name = tensor("pairs_19_cast_fp16")]; + tensor xr_19_begin_0 = const()[name = tensor("xr_19_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_19_end_0 = const()[name = tensor("xr_19_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_19_end_mask_0 = const()[name = tensor("xr_19_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_19_squeeze_mask_0 = const()[name = tensor("xr_19_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_19_cast_fp16 = slice_by_index(begin = xr_19_begin_0, end = xr_19_end_0, end_mask = xr_19_end_mask_0, squeeze_mask = xr_19_squeeze_mask_0, x = pairs_19_cast_fp16)[name = tensor("xr_19_cast_fp16")]; + tensor xi_19_begin_0 = const()[name = tensor("xi_19_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_19_end_0 = const()[name = tensor("xi_19_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_19_end_mask_0 = const()[name = tensor("xi_19_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_19_squeeze_mask_0 = const()[name = tensor("xi_19_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_19_cast_fp16 = slice_by_index(begin = xi_19_begin_0, end = xi_19_end_0, end_mask = xi_19_end_mask_0, squeeze_mask = xi_19_squeeze_mask_0, x = pairs_19_cast_fp16)[name = tensor("xi_19_cast_fp16")]; + tensor var_1336_cast_fp16 = mul(x = xr_19_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1336_cast_fp16")]; + tensor var_1337_cast_fp16 = mul(x = xi_19_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1337_cast_fp16")]; + tensor xor__19_cast_fp16 = sub(x = var_1336_cast_fp16, y = var_1337_cast_fp16)[name = tensor("xor__19_cast_fp16")]; + tensor var_1340_cast_fp16 = mul(x = xr_19_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1340_cast_fp16")]; + tensor var_1341_cast_fp16 = mul(x = xi_19_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1341_cast_fp16")]; + tensor xoi_19_cast_fp16 = add(x = var_1340_cast_fp16, y = var_1341_cast_fp16)[name = tensor("xoi_19_cast_fp16")]; + tensor out_19_axis_0 = const()[name = tensor("out_19_axis_0"), val = tensor(-1)]; + tensor out_19_cast_fp16 = stack(axis = out_19_axis_0, values = (xor__19_cast_fp16, xoi_19_cast_fp16))[name = tensor("out_19_cast_fp16")]; + tensor var_1358 = const()[name = tensor("op_1358"), val = tensor([256, 1024])]; + tensor var_1359_cast_fp16 = reshape(shape = var_1358, x = out_19_cast_fp16)[name = tensor("op_1359_cast_fp16")]; + tensor written_k_9_transpose_x_1 = const()[name = tensor("written_k_9_transpose_x_1"), val = tensor(true)]; + tensor written_k_9_transpose_y_1 = const()[name = tensor("written_k_9_transpose_y_1"), val = tensor(false)]; + tensor written_k_9_cast_fp16 = matmul(transpose_x = written_k_9_transpose_x_1, transpose_y = written_k_9_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_1359_cast_fp16)[name = tensor("written_k_9_cast_fp16")]; + tensor var_1366 = const()[name = tensor("op_1366"), val = tensor([256, 1024])]; + tensor var_1367_cast_fp16 = reshape(shape = var_1366, x = v_9_cast_fp16)[name = tensor("op_1367_cast_fp16")]; + tensor written_v_9_transpose_x_1 = const()[name = tensor("written_v_9_transpose_x_1"), val = tensor(true)]; + tensor written_v_9_transpose_y_1 = const()[name = tensor("written_v_9_transpose_y_1"), val = tensor(false)]; + tensor written_v_9_cast_fp16 = matmul(transpose_x = written_v_9_transpose_x_1, transpose_y = written_v_9_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_1367_cast_fp16)[name = tensor("written_v_9_cast_fp16")]; + tensor k_cache4_to_fp16_dtype_0 = const()[name = tensor("k_cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache4_to_fp16 = cast(dtype = k_cache4_to_fp16_dtype_0, x = k_cache4)[name = tensor("cast_441")]; + tensor var_1378_cast_fp16 = mul(x = k_cache4_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_1378_cast_fp16")]; + tensor var_1383 = const()[name = tensor("op_1383"), val = tensor([1, 512, 16, 64])]; + tensor var_1384_cast_fp16 = reshape(shape = var_1383, x = written_k_9_cast_fp16)[name = tensor("op_1384_cast_fp16")]; + tensor new_k_cache_9_cast_fp16 = add(x = var_1378_cast_fp16, y = var_1384_cast_fp16)[name = tensor("new_k_cache_9_cast_fp16")]; + tensor new_k_cache_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache4_to_fp16_dtype_0 = const()[name = tensor("v_cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache4_to_fp16 = cast(dtype = v_cache4_to_fp16_dtype_0, x = v_cache4)[name = tensor("cast_439")]; + tensor var_1387_cast_fp16 = mul(x = v_cache4_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_1387_cast_fp16")]; + tensor var_1392 = const()[name = tensor("op_1392"), val = tensor([1, 512, 16, 64])]; + tensor var_1393_cast_fp16 = reshape(shape = var_1392, x = written_v_9_cast_fp16)[name = tensor("op_1393_cast_fp16")]; + tensor new_v_cache_9_cast_fp16 = add(x = var_1387_cast_fp16, y = var_1393_cast_fp16)[name = tensor("new_v_cache_9_cast_fp16")]; + tensor new_v_cache_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_1406 = const()[name = tensor("op_1406"), val = tensor([0, 2, 1, 3])]; + tensor var_1417_transpose_x_0 = const()[name = tensor("op_1417_transpose_x_0"), val = tensor(false)]; + tensor var_1417_transpose_y_0 = const()[name = tensor("op_1417_transpose_y_0"), val = tensor(false)]; + tensor transpose_77_perm_0 = const()[name = tensor("transpose_77_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_78_perm_0 = const()[name = tensor("transpose_78_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_78 = transpose(perm = transpose_78_perm_0, x = new_k_cache_9_cast_fp16)[name = tensor("transpose_188")]; + tensor transpose_77 = transpose(perm = transpose_77_perm_0, x = q_9_cast_fp16)[name = tensor("transpose_189")]; + tensor var_1417_cast_fp16 = matmul(transpose_x = var_1417_transpose_x_0, transpose_y = var_1417_transpose_y_0, x = transpose_77, y = transpose_78)[name = tensor("op_1417_cast_fp16")]; + tensor var_1418_to_fp16 = const()[name = tensor("op_1418_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_17_cast_fp16 = mul(x = var_1417_cast_fp16, y = var_1418_to_fp16)[name = tensor("scores_17_cast_fp16")]; + tensor scores_19_cast_fp16 = add(x = scores_17_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_19_cast_fp16")]; + tensor var_1428 = const()[name = tensor("op_1428"), val = tensor(-1)]; + tensor weights_9_cast_fp16 = softmax(axis = var_1428, x = scores_19_cast_fp16)[name = tensor("weights_9_cast_fp16")]; + tensor attn_9_transpose_x_0 = const()[name = tensor("attn_9_transpose_x_0"), val = tensor(false)]; + tensor attn_9_transpose_y_0 = const()[name = tensor("attn_9_transpose_y_0"), val = tensor(false)]; + tensor values_9_cast_fp16 = transpose(perm = var_1406, x = new_v_cache_9_cast_fp16)[name = tensor("transpose_190")]; + tensor attn_9_cast_fp16 = matmul(transpose_x = attn_9_transpose_x_0, transpose_y = attn_9_transpose_y_0, x = weights_9_cast_fp16, y = values_9_cast_fp16)[name = tensor("attn_9_cast_fp16")]; + tensor var_1436 = const()[name = tensor("op_1436"), val = tensor([0, 2, 1, 3])]; + tensor var_1441 = const()[name = tensor("op_1441"), val = tensor([1, 256, 1024])]; + tensor var_1437_cast_fp16 = transpose(perm = var_1436, x = attn_9_cast_fp16)[name = tensor("transpose_187")]; + tensor input_53_cast_fp16 = reshape(shape = var_1441, x = var_1437_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor attn4_out_proj_weight_to_fp16 = const()[name = tensor("attn4_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(107012864)))]; + tensor linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn4_out_proj_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor position4_to_fp16_dtype_0 = const()[name = tensor("position4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position4_to_fp16 = cast(dtype = position4_to_fp16_dtype_0, x = position4)[name = tensor("cast_437")]; + tensor var_1450_cast_fp16 = add(x = position4_to_fp16, y = valid_len_to_fp16)[name = tensor("op_1450_cast_fp16")]; + tensor var_1450_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1450_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_55_cast_fp16 = add(x = input_49_cast_fp16, y = linear_17_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor input_57_axes_0 = const()[name = tensor("input_57_axes_0"), val = tensor([-1])]; + tensor norm4_2_weight_to_fp16 = const()[name = tensor("norm4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109110080)))]; + tensor norm4_2_bias_to_fp16 = const()[name = tensor("norm4_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109112192)))]; + tensor var_1454_to_fp16 = const()[name = tensor("op_1454_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_57_cast_fp16 = layer_norm(axes = input_57_axes_0, beta = norm4_2_bias_to_fp16, epsilon = var_1454_to_fp16, gamma = norm4_2_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor linear4_1_weight_to_fp16 = const()[name = tensor("linear4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109114304)))]; + tensor linear_18_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear4_1_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor input_59_mode_0 = const()[name = tensor("input_59_mode_0"), val = tensor("EXACT")]; + tensor input_59_cast_fp16 = gelu(mode = input_59_mode_0, x = linear_18_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor linear4_2_weight_to_fp16 = const()[name = tensor("linear4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(117502976)))]; + tensor linear_19_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear4_2_weight_to_fp16, x = input_59_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_61_cast_fp16 = add(x = input_55_cast_fp16, y = linear_19_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor input_63_axes_0 = const()[name = tensor("input_63_axes_0"), val = tensor([-1])]; + tensor norm5_1_weight_to_fp16 = const()[name = tensor("norm5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125891648)))]; + tensor norm5_1_bias_to_fp16 = const()[name = tensor("norm5_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125893760)))]; + tensor var_1471_to_fp16 = const()[name = tensor("op_1471_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_63_cast_fp16 = layer_norm(axes = input_63_axes_0, beta = norm5_1_bias_to_fp16, epsilon = var_1471_to_fp16, gamma = norm5_1_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor attn5_in_proj_weight_to_fp16 = const()[name = tensor("attn5_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125895872)))]; + tensor linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn5_in_proj_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor var_1485 = const()[name = tensor("op_1485"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_11_cast_fp16 = reshape(shape = var_1485, x = linear_20_cast_fp16)[name = tensor("qkv_11_cast_fp16")]; + tensor x_21_begin_0 = const()[name = tensor("x_21_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_21_end_0 = const()[name = tensor("x_21_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_21_end_mask_0 = const()[name = tensor("x_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_21_squeeze_mask_0 = const()[name = tensor("x_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_21_cast_fp16 = slice_by_index(begin = x_21_begin_0, end = x_21_end_0, end_mask = x_21_end_mask_0, squeeze_mask = x_21_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("x_21_cast_fp16")]; + tensor x_23_begin_0 = const()[name = tensor("x_23_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_23_end_0 = const()[name = tensor("x_23_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_23_end_mask_0 = const()[name = tensor("x_23_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_23_squeeze_mask_0 = const()[name = tensor("x_23_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_23_cast_fp16 = slice_by_index(begin = x_23_begin_0, end = x_23_end_0, end_mask = x_23_end_mask_0, squeeze_mask = x_23_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("x_23_cast_fp16")]; + tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_11_cast_fp16 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("v_11_cast_fp16")]; + tensor var_1530 = const()[name = tensor("op_1530"), val = tensor([256, 16, 32, 2])]; + tensor pairs_21_cast_fp16 = reshape(shape = var_1530, x = x_21_cast_fp16)[name = tensor("pairs_21_cast_fp16")]; + tensor xr_21_begin_0 = const()[name = tensor("xr_21_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_21_end_0 = const()[name = tensor("xr_21_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_21_end_mask_0 = const()[name = tensor("xr_21_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_21_squeeze_mask_0 = const()[name = tensor("xr_21_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_21_cast_fp16 = slice_by_index(begin = xr_21_begin_0, end = xr_21_end_0, end_mask = xr_21_end_mask_0, squeeze_mask = xr_21_squeeze_mask_0, x = pairs_21_cast_fp16)[name = tensor("xr_21_cast_fp16")]; + tensor xi_21_begin_0 = const()[name = tensor("xi_21_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_21_end_0 = const()[name = tensor("xi_21_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_21_end_mask_0 = const()[name = tensor("xi_21_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_21_squeeze_mask_0 = const()[name = tensor("xi_21_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_21_cast_fp16 = slice_by_index(begin = xi_21_begin_0, end = xi_21_end_0, end_mask = xi_21_end_mask_0, squeeze_mask = xi_21_squeeze_mask_0, x = pairs_21_cast_fp16)[name = tensor("xi_21_cast_fp16")]; + tensor var_1538_cast_fp16 = mul(x = xr_21_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1538_cast_fp16")]; + tensor var_1539_cast_fp16 = mul(x = xi_21_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1539_cast_fp16")]; + tensor xor__21_cast_fp16 = sub(x = var_1538_cast_fp16, y = var_1539_cast_fp16)[name = tensor("xor__21_cast_fp16")]; + tensor var_1542_cast_fp16 = mul(x = xr_21_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1542_cast_fp16")]; + tensor var_1543_cast_fp16 = mul(x = xi_21_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1543_cast_fp16")]; + tensor xoi_21_cast_fp16 = add(x = var_1542_cast_fp16, y = var_1543_cast_fp16)[name = tensor("xoi_21_cast_fp16")]; + tensor out_21_axis_0 = const()[name = tensor("out_21_axis_0"), val = tensor(-1)]; + tensor out_21_cast_fp16 = stack(axis = out_21_axis_0, values = (xor__21_cast_fp16, xoi_21_cast_fp16))[name = tensor("out_21_cast_fp16")]; + tensor var_1553 = const()[name = tensor("op_1553"), val = tensor([1, 256, 16, 64])]; + tensor q_11_cast_fp16 = reshape(shape = var_1553, x = out_21_cast_fp16)[name = tensor("q_11_cast_fp16")]; + tensor var_1559 = const()[name = tensor("op_1559"), val = tensor([256, 16, 32, 2])]; + tensor pairs_23_cast_fp16 = reshape(shape = var_1559, x = x_23_cast_fp16)[name = tensor("pairs_23_cast_fp16")]; + tensor xr_23_begin_0 = const()[name = tensor("xr_23_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_23_end_0 = const()[name = tensor("xr_23_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_23_end_mask_0 = const()[name = tensor("xr_23_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_23_squeeze_mask_0 = const()[name = tensor("xr_23_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_23_cast_fp16 = slice_by_index(begin = xr_23_begin_0, end = xr_23_end_0, end_mask = xr_23_end_mask_0, squeeze_mask = xr_23_squeeze_mask_0, x = pairs_23_cast_fp16)[name = tensor("xr_23_cast_fp16")]; + tensor xi_23_begin_0 = const()[name = tensor("xi_23_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_23_end_0 = const()[name = tensor("xi_23_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_23_end_mask_0 = const()[name = tensor("xi_23_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_23_squeeze_mask_0 = const()[name = tensor("xi_23_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_23_cast_fp16 = slice_by_index(begin = xi_23_begin_0, end = xi_23_end_0, end_mask = xi_23_end_mask_0, squeeze_mask = xi_23_squeeze_mask_0, x = pairs_23_cast_fp16)[name = tensor("xi_23_cast_fp16")]; + tensor var_1567_cast_fp16 = mul(x = xr_23_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1567_cast_fp16")]; + tensor var_1568_cast_fp16 = mul(x = xi_23_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1568_cast_fp16")]; + tensor xor__23_cast_fp16 = sub(x = var_1567_cast_fp16, y = var_1568_cast_fp16)[name = tensor("xor__23_cast_fp16")]; + tensor var_1571_cast_fp16 = mul(x = xr_23_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1571_cast_fp16")]; + tensor var_1572_cast_fp16 = mul(x = xi_23_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1572_cast_fp16")]; + tensor xoi_23_cast_fp16 = add(x = var_1571_cast_fp16, y = var_1572_cast_fp16)[name = tensor("xoi_23_cast_fp16")]; + tensor out_23_axis_0 = const()[name = tensor("out_23_axis_0"), val = tensor(-1)]; + tensor out_23_cast_fp16 = stack(axis = out_23_axis_0, values = (xor__23_cast_fp16, xoi_23_cast_fp16))[name = tensor("out_23_cast_fp16")]; + tensor var_1589 = const()[name = tensor("op_1589"), val = tensor([256, 1024])]; + tensor var_1590_cast_fp16 = reshape(shape = var_1589, x = out_23_cast_fp16)[name = tensor("op_1590_cast_fp16")]; + tensor written_k_11_transpose_x_1 = const()[name = tensor("written_k_11_transpose_x_1"), val = tensor(true)]; + tensor written_k_11_transpose_y_1 = const()[name = tensor("written_k_11_transpose_y_1"), val = tensor(false)]; + tensor written_k_11_cast_fp16 = matmul(transpose_x = written_k_11_transpose_x_1, transpose_y = written_k_11_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_1590_cast_fp16)[name = tensor("written_k_11_cast_fp16")]; + tensor var_1597 = const()[name = tensor("op_1597"), val = tensor([256, 1024])]; + tensor var_1598_cast_fp16 = reshape(shape = var_1597, x = v_11_cast_fp16)[name = tensor("op_1598_cast_fp16")]; + tensor written_v_11_transpose_x_1 = const()[name = tensor("written_v_11_transpose_x_1"), val = tensor(true)]; + tensor written_v_11_transpose_y_1 = const()[name = tensor("written_v_11_transpose_y_1"), val = tensor(false)]; + tensor written_v_11_cast_fp16 = matmul(transpose_x = written_v_11_transpose_x_1, transpose_y = written_v_11_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_1598_cast_fp16)[name = tensor("written_v_11_cast_fp16")]; + tensor k_cache5_to_fp16_dtype_0 = const()[name = tensor("k_cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache5_to_fp16 = cast(dtype = k_cache5_to_fp16_dtype_0, x = k_cache5)[name = tensor("cast_435")]; + tensor var_1609_cast_fp16 = mul(x = k_cache5_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_1609_cast_fp16")]; + tensor var_1614 = const()[name = tensor("op_1614"), val = tensor([1, 512, 16, 64])]; + tensor var_1615_cast_fp16 = reshape(shape = var_1614, x = written_k_11_cast_fp16)[name = tensor("op_1615_cast_fp16")]; + tensor new_k_cache_11_cast_fp16 = add(x = var_1609_cast_fp16, y = var_1615_cast_fp16)[name = tensor("new_k_cache_11_cast_fp16")]; + tensor new_k_cache_11_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_11_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache5_to_fp16_dtype_0 = const()[name = tensor("v_cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache5_to_fp16 = cast(dtype = v_cache5_to_fp16_dtype_0, x = v_cache5)[name = tensor("cast_433")]; + tensor var_1618_cast_fp16 = mul(x = v_cache5_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_1618_cast_fp16")]; + tensor var_1623 = const()[name = tensor("op_1623"), val = tensor([1, 512, 16, 64])]; + tensor var_1624_cast_fp16 = reshape(shape = var_1623, x = written_v_11_cast_fp16)[name = tensor("op_1624_cast_fp16")]; + tensor new_v_cache_11_cast_fp16 = add(x = var_1618_cast_fp16, y = var_1624_cast_fp16)[name = tensor("new_v_cache_11_cast_fp16")]; + tensor new_v_cache_11_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_11_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_1637 = const()[name = tensor("op_1637"), val = tensor([0, 2, 1, 3])]; + tensor var_1648_transpose_x_0 = const()[name = tensor("op_1648_transpose_x_0"), val = tensor(false)]; + tensor var_1648_transpose_y_0 = const()[name = tensor("op_1648_transpose_y_0"), val = tensor(false)]; + tensor transpose_79_perm_0 = const()[name = tensor("transpose_79_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_80_perm_0 = const()[name = tensor("transpose_80_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_80 = transpose(perm = transpose_80_perm_0, x = new_k_cache_11_cast_fp16)[name = tensor("transpose_184")]; + tensor transpose_79 = transpose(perm = transpose_79_perm_0, x = q_11_cast_fp16)[name = tensor("transpose_185")]; + tensor var_1648_cast_fp16 = matmul(transpose_x = var_1648_transpose_x_0, transpose_y = var_1648_transpose_y_0, x = transpose_79, y = transpose_80)[name = tensor("op_1648_cast_fp16")]; + tensor var_1649_to_fp16 = const()[name = tensor("op_1649_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_21_cast_fp16 = mul(x = var_1648_cast_fp16, y = var_1649_to_fp16)[name = tensor("scores_21_cast_fp16")]; + tensor scores_23_cast_fp16 = add(x = scores_21_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_23_cast_fp16")]; + tensor var_1659 = const()[name = tensor("op_1659"), val = tensor(-1)]; + tensor weights_11_cast_fp16 = softmax(axis = var_1659, x = scores_23_cast_fp16)[name = tensor("weights_11_cast_fp16")]; + tensor attn_11_transpose_x_0 = const()[name = tensor("attn_11_transpose_x_0"), val = tensor(false)]; + tensor attn_11_transpose_y_0 = const()[name = tensor("attn_11_transpose_y_0"), val = tensor(false)]; + tensor values_11_cast_fp16 = transpose(perm = var_1637, x = new_v_cache_11_cast_fp16)[name = tensor("transpose_186")]; + tensor attn_11_cast_fp16 = matmul(transpose_x = attn_11_transpose_x_0, transpose_y = attn_11_transpose_y_0, x = weights_11_cast_fp16, y = values_11_cast_fp16)[name = tensor("attn_11_cast_fp16")]; + tensor var_1667 = const()[name = tensor("op_1667"), val = tensor([0, 2, 1, 3])]; + tensor var_1672 = const()[name = tensor("op_1672"), val = tensor([1, 256, 1024])]; + tensor var_1668_cast_fp16 = transpose(perm = var_1667, x = attn_11_cast_fp16)[name = tensor("transpose_183")]; + tensor input_65_cast_fp16 = reshape(shape = var_1672, x = var_1668_cast_fp16)[name = tensor("input_65_cast_fp16")]; + tensor attn5_out_proj_weight_to_fp16 = const()[name = tensor("attn5_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132187392)))]; + tensor linear_21_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn5_out_proj_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("linear_21_cast_fp16")]; + tensor position5_to_fp16_dtype_0 = const()[name = tensor("position5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position5_to_fp16 = cast(dtype = position5_to_fp16_dtype_0, x = position5)[name = tensor("cast_431")]; + tensor var_1681_cast_fp16 = add(x = position5_to_fp16, y = valid_len_to_fp16)[name = tensor("op_1681_cast_fp16")]; + tensor var_1681_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1681_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_67_cast_fp16 = add(x = input_61_cast_fp16, y = linear_21_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor input_69_axes_0 = const()[name = tensor("input_69_axes_0"), val = tensor([-1])]; + tensor norm5_2_weight_to_fp16 = const()[name = tensor("norm5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134284608)))]; + tensor norm5_2_bias_to_fp16 = const()[name = tensor("norm5_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134286720)))]; + tensor var_1685_to_fp16 = const()[name = tensor("op_1685_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_69_cast_fp16 = layer_norm(axes = input_69_axes_0, beta = norm5_2_bias_to_fp16, epsilon = var_1685_to_fp16, gamma = norm5_2_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("input_69_cast_fp16")]; + tensor linear5_1_weight_to_fp16 = const()[name = tensor("linear5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134288832)))]; + tensor linear_22_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear5_1_weight_to_fp16, x = input_69_cast_fp16)[name = tensor("linear_22_cast_fp16")]; + tensor input_71_mode_0 = const()[name = tensor("input_71_mode_0"), val = tensor("EXACT")]; + tensor input_71_cast_fp16 = gelu(mode = input_71_mode_0, x = linear_22_cast_fp16)[name = tensor("input_71_cast_fp16")]; + tensor linear5_2_weight_to_fp16 = const()[name = tensor("linear5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142677504)))]; + tensor linear_23_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear5_2_weight_to_fp16, x = input_71_cast_fp16)[name = tensor("linear_23_cast_fp16")]; + tensor input_73_cast_fp16 = add(x = input_67_cast_fp16, y = linear_23_cast_fp16)[name = tensor("input_73_cast_fp16")]; + tensor input_75_axes_0 = const()[name = tensor("input_75_axes_0"), val = tensor([-1])]; + tensor norm6_1_weight_to_fp16 = const()[name = tensor("norm6_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151066176)))]; + tensor norm6_1_bias_to_fp16 = const()[name = tensor("norm6_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151068288)))]; + tensor var_1702_to_fp16 = const()[name = tensor("op_1702_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_75_cast_fp16 = layer_norm(axes = input_75_axes_0, beta = norm6_1_bias_to_fp16, epsilon = var_1702_to_fp16, gamma = norm6_1_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("input_75_cast_fp16")]; + tensor attn6_in_proj_weight_to_fp16 = const()[name = tensor("attn6_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151070400)))]; + tensor linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn6_in_proj_weight_to_fp16, x = input_75_cast_fp16)[name = tensor("linear_24_cast_fp16")]; + tensor var_1716 = const()[name = tensor("op_1716"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_13_cast_fp16 = reshape(shape = var_1716, x = linear_24_cast_fp16)[name = tensor("qkv_13_cast_fp16")]; + tensor x_25_begin_0 = const()[name = tensor("x_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_25_end_0 = const()[name = tensor("x_25_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_25_end_mask_0 = const()[name = tensor("x_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_25_squeeze_mask_0 = const()[name = tensor("x_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_25_cast_fp16 = slice_by_index(begin = x_25_begin_0, end = x_25_end_0, end_mask = x_25_end_mask_0, squeeze_mask = x_25_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("x_25_cast_fp16")]; + tensor x_27_begin_0 = const()[name = tensor("x_27_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_27_end_0 = const()[name = tensor("x_27_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_27_end_mask_0 = const()[name = tensor("x_27_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_27_squeeze_mask_0 = const()[name = tensor("x_27_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_27_cast_fp16 = slice_by_index(begin = x_27_begin_0, end = x_27_end_0, end_mask = x_27_end_mask_0, squeeze_mask = x_27_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("x_27_cast_fp16")]; + tensor v_13_begin_0 = const()[name = tensor("v_13_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_13_end_0 = const()[name = tensor("v_13_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_13_end_mask_0 = const()[name = tensor("v_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_13_squeeze_mask_0 = const()[name = tensor("v_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_13_cast_fp16 = slice_by_index(begin = v_13_begin_0, end = v_13_end_0, end_mask = v_13_end_mask_0, squeeze_mask = v_13_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("v_13_cast_fp16")]; + tensor var_1761 = const()[name = tensor("op_1761"), val = tensor([256, 16, 32, 2])]; + tensor pairs_25_cast_fp16 = reshape(shape = var_1761, x = x_25_cast_fp16)[name = tensor("pairs_25_cast_fp16")]; + tensor xr_25_begin_0 = const()[name = tensor("xr_25_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_25_end_0 = const()[name = tensor("xr_25_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_25_end_mask_0 = const()[name = tensor("xr_25_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_25_squeeze_mask_0 = const()[name = tensor("xr_25_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_25_cast_fp16 = slice_by_index(begin = xr_25_begin_0, end = xr_25_end_0, end_mask = xr_25_end_mask_0, squeeze_mask = xr_25_squeeze_mask_0, x = pairs_25_cast_fp16)[name = tensor("xr_25_cast_fp16")]; + tensor xi_25_begin_0 = const()[name = tensor("xi_25_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_25_end_0 = const()[name = tensor("xi_25_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_25_end_mask_0 = const()[name = tensor("xi_25_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_25_squeeze_mask_0 = const()[name = tensor("xi_25_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_25_cast_fp16 = slice_by_index(begin = xi_25_begin_0, end = xi_25_end_0, end_mask = xi_25_end_mask_0, squeeze_mask = xi_25_squeeze_mask_0, x = pairs_25_cast_fp16)[name = tensor("xi_25_cast_fp16")]; + tensor var_1769_cast_fp16 = mul(x = xr_25_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1769_cast_fp16")]; + tensor var_1770_cast_fp16 = mul(x = xi_25_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1770_cast_fp16")]; + tensor xor__25_cast_fp16 = sub(x = var_1769_cast_fp16, y = var_1770_cast_fp16)[name = tensor("xor__25_cast_fp16")]; + tensor var_1773_cast_fp16 = mul(x = xr_25_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1773_cast_fp16")]; + tensor var_1774_cast_fp16 = mul(x = xi_25_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1774_cast_fp16")]; + tensor xoi_25_cast_fp16 = add(x = var_1773_cast_fp16, y = var_1774_cast_fp16)[name = tensor("xoi_25_cast_fp16")]; + tensor out_25_axis_0 = const()[name = tensor("out_25_axis_0"), val = tensor(-1)]; + tensor out_25_cast_fp16 = stack(axis = out_25_axis_0, values = (xor__25_cast_fp16, xoi_25_cast_fp16))[name = tensor("out_25_cast_fp16")]; + tensor var_1784 = const()[name = tensor("op_1784"), val = tensor([1, 256, 16, 64])]; + tensor q_13_cast_fp16 = reshape(shape = var_1784, x = out_25_cast_fp16)[name = tensor("q_13_cast_fp16")]; + tensor var_1790 = const()[name = tensor("op_1790"), val = tensor([256, 16, 32, 2])]; + tensor pairs_27_cast_fp16 = reshape(shape = var_1790, x = x_27_cast_fp16)[name = tensor("pairs_27_cast_fp16")]; + tensor xr_27_begin_0 = const()[name = tensor("xr_27_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_27_end_0 = const()[name = tensor("xr_27_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_27_end_mask_0 = const()[name = tensor("xr_27_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_27_squeeze_mask_0 = const()[name = tensor("xr_27_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_27_cast_fp16 = slice_by_index(begin = xr_27_begin_0, end = xr_27_end_0, end_mask = xr_27_end_mask_0, squeeze_mask = xr_27_squeeze_mask_0, x = pairs_27_cast_fp16)[name = tensor("xr_27_cast_fp16")]; + tensor xi_27_begin_0 = const()[name = tensor("xi_27_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_27_end_0 = const()[name = tensor("xi_27_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_27_end_mask_0 = const()[name = tensor("xi_27_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_27_squeeze_mask_0 = const()[name = tensor("xi_27_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_27_cast_fp16 = slice_by_index(begin = xi_27_begin_0, end = xi_27_end_0, end_mask = xi_27_end_mask_0, squeeze_mask = xi_27_squeeze_mask_0, x = pairs_27_cast_fp16)[name = tensor("xi_27_cast_fp16")]; + tensor var_1798_cast_fp16 = mul(x = xr_27_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1798_cast_fp16")]; + tensor var_1799_cast_fp16 = mul(x = xi_27_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1799_cast_fp16")]; + tensor xor__27_cast_fp16 = sub(x = var_1798_cast_fp16, y = var_1799_cast_fp16)[name = tensor("xor__27_cast_fp16")]; + tensor var_1802_cast_fp16 = mul(x = xr_27_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1802_cast_fp16")]; + tensor var_1803_cast_fp16 = mul(x = xi_27_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1803_cast_fp16")]; + tensor xoi_27_cast_fp16 = add(x = var_1802_cast_fp16, y = var_1803_cast_fp16)[name = tensor("xoi_27_cast_fp16")]; + tensor out_27_axis_0 = const()[name = tensor("out_27_axis_0"), val = tensor(-1)]; + tensor out_27_cast_fp16 = stack(axis = out_27_axis_0, values = (xor__27_cast_fp16, xoi_27_cast_fp16))[name = tensor("out_27_cast_fp16")]; + tensor var_1820 = const()[name = tensor("op_1820"), val = tensor([256, 1024])]; + tensor var_1821_cast_fp16 = reshape(shape = var_1820, x = out_27_cast_fp16)[name = tensor("op_1821_cast_fp16")]; + tensor written_k_13_transpose_x_1 = const()[name = tensor("written_k_13_transpose_x_1"), val = tensor(true)]; + tensor written_k_13_transpose_y_1 = const()[name = tensor("written_k_13_transpose_y_1"), val = tensor(false)]; + tensor written_k_13_cast_fp16 = matmul(transpose_x = written_k_13_transpose_x_1, transpose_y = written_k_13_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_1821_cast_fp16)[name = tensor("written_k_13_cast_fp16")]; + tensor var_1828 = const()[name = tensor("op_1828"), val = tensor([256, 1024])]; + tensor var_1829_cast_fp16 = reshape(shape = var_1828, x = v_13_cast_fp16)[name = tensor("op_1829_cast_fp16")]; + tensor written_v_13_transpose_x_1 = const()[name = tensor("written_v_13_transpose_x_1"), val = tensor(true)]; + tensor written_v_13_transpose_y_1 = const()[name = tensor("written_v_13_transpose_y_1"), val = tensor(false)]; + tensor written_v_13_cast_fp16 = matmul(transpose_x = written_v_13_transpose_x_1, transpose_y = written_v_13_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_1829_cast_fp16)[name = tensor("written_v_13_cast_fp16")]; + tensor k_cache6_to_fp16_dtype_0 = const()[name = tensor("k_cache6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache6_to_fp16 = cast(dtype = k_cache6_to_fp16_dtype_0, x = k_cache6)[name = tensor("cast_429")]; + tensor var_1840_cast_fp16 = mul(x = k_cache6_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_1840_cast_fp16")]; + tensor var_1845 = const()[name = tensor("op_1845"), val = tensor([1, 512, 16, 64])]; + tensor var_1846_cast_fp16 = reshape(shape = var_1845, x = written_k_13_cast_fp16)[name = tensor("op_1846_cast_fp16")]; + tensor new_k_cache_13_cast_fp16 = add(x = var_1840_cast_fp16, y = var_1846_cast_fp16)[name = tensor("new_k_cache_13_cast_fp16")]; + tensor new_k_cache_13_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_13_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache6_to_fp16_dtype_0 = const()[name = tensor("v_cache6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache6_to_fp16 = cast(dtype = v_cache6_to_fp16_dtype_0, x = v_cache6)[name = tensor("cast_427")]; + tensor var_1849_cast_fp16 = mul(x = v_cache6_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_1849_cast_fp16")]; + tensor var_1854 = const()[name = tensor("op_1854"), val = tensor([1, 512, 16, 64])]; + tensor var_1855_cast_fp16 = reshape(shape = var_1854, x = written_v_13_cast_fp16)[name = tensor("op_1855_cast_fp16")]; + tensor new_v_cache_13_cast_fp16 = add(x = var_1849_cast_fp16, y = var_1855_cast_fp16)[name = tensor("new_v_cache_13_cast_fp16")]; + tensor new_v_cache_13_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_13_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_1868 = const()[name = tensor("op_1868"), val = tensor([0, 2, 1, 3])]; + tensor var_1879_transpose_x_0 = const()[name = tensor("op_1879_transpose_x_0"), val = tensor(false)]; + tensor var_1879_transpose_y_0 = const()[name = tensor("op_1879_transpose_y_0"), val = tensor(false)]; + tensor transpose_81_perm_0 = const()[name = tensor("transpose_81_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_82_perm_0 = const()[name = tensor("transpose_82_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_82 = transpose(perm = transpose_82_perm_0, x = new_k_cache_13_cast_fp16)[name = tensor("transpose_180")]; + tensor transpose_81 = transpose(perm = transpose_81_perm_0, x = q_13_cast_fp16)[name = tensor("transpose_181")]; + tensor var_1879_cast_fp16 = matmul(transpose_x = var_1879_transpose_x_0, transpose_y = var_1879_transpose_y_0, x = transpose_81, y = transpose_82)[name = tensor("op_1879_cast_fp16")]; + tensor var_1880_to_fp16 = const()[name = tensor("op_1880_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_25_cast_fp16 = mul(x = var_1879_cast_fp16, y = var_1880_to_fp16)[name = tensor("scores_25_cast_fp16")]; + tensor scores_27_cast_fp16 = add(x = scores_25_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_27_cast_fp16")]; + tensor var_1890 = const()[name = tensor("op_1890"), val = tensor(-1)]; + tensor weights_13_cast_fp16 = softmax(axis = var_1890, x = scores_27_cast_fp16)[name = tensor("weights_13_cast_fp16")]; + tensor attn_13_transpose_x_0 = const()[name = tensor("attn_13_transpose_x_0"), val = tensor(false)]; + tensor attn_13_transpose_y_0 = const()[name = tensor("attn_13_transpose_y_0"), val = tensor(false)]; + tensor values_13_cast_fp16 = transpose(perm = var_1868, x = new_v_cache_13_cast_fp16)[name = tensor("transpose_182")]; + tensor attn_13_cast_fp16 = matmul(transpose_x = attn_13_transpose_x_0, transpose_y = attn_13_transpose_y_0, x = weights_13_cast_fp16, y = values_13_cast_fp16)[name = tensor("attn_13_cast_fp16")]; + tensor var_1898 = const()[name = tensor("op_1898"), val = tensor([0, 2, 1, 3])]; + tensor var_1903 = const()[name = tensor("op_1903"), val = tensor([1, 256, 1024])]; + tensor var_1899_cast_fp16 = transpose(perm = var_1898, x = attn_13_cast_fp16)[name = tensor("transpose_179")]; + tensor input_77_cast_fp16 = reshape(shape = var_1903, x = var_1899_cast_fp16)[name = tensor("input_77_cast_fp16")]; + tensor attn6_out_proj_weight_to_fp16 = const()[name = tensor("attn6_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(157361920)))]; + tensor linear_25_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn6_out_proj_weight_to_fp16, x = input_77_cast_fp16)[name = tensor("linear_25_cast_fp16")]; + tensor position6_to_fp16_dtype_0 = const()[name = tensor("position6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position6_to_fp16 = cast(dtype = position6_to_fp16_dtype_0, x = position6)[name = tensor("cast_425")]; + tensor var_1912_cast_fp16 = add(x = position6_to_fp16, y = valid_len_to_fp16)[name = tensor("op_1912_cast_fp16")]; + tensor var_1912_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1912_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_79_cast_fp16 = add(x = input_73_cast_fp16, y = linear_25_cast_fp16)[name = tensor("input_79_cast_fp16")]; + tensor input_81_axes_0 = const()[name = tensor("input_81_axes_0"), val = tensor([-1])]; + tensor norm6_2_weight_to_fp16 = const()[name = tensor("norm6_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159459136)))]; + tensor norm6_2_bias_to_fp16 = const()[name = tensor("norm6_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159461248)))]; + tensor var_1916_to_fp16 = const()[name = tensor("op_1916_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_81_cast_fp16 = layer_norm(axes = input_81_axes_0, beta = norm6_2_bias_to_fp16, epsilon = var_1916_to_fp16, gamma = norm6_2_weight_to_fp16, x = input_79_cast_fp16)[name = tensor("input_81_cast_fp16")]; + tensor linear6_1_weight_to_fp16 = const()[name = tensor("linear6_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159463360)))]; + tensor linear_26_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear6_1_weight_to_fp16, x = input_81_cast_fp16)[name = tensor("linear_26_cast_fp16")]; + tensor input_83_mode_0 = const()[name = tensor("input_83_mode_0"), val = tensor("EXACT")]; + tensor input_83_cast_fp16 = gelu(mode = input_83_mode_0, x = linear_26_cast_fp16)[name = tensor("input_83_cast_fp16")]; + tensor linear6_2_weight_to_fp16 = const()[name = tensor("linear6_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(167852032)))]; + tensor linear_27_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear6_2_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_27_cast_fp16")]; + tensor input_85_cast_fp16 = add(x = input_79_cast_fp16, y = linear_27_cast_fp16)[name = tensor("input_85_cast_fp16")]; + tensor input_87_axes_0 = const()[name = tensor("input_87_axes_0"), val = tensor([-1])]; + tensor norm7_1_weight_to_fp16 = const()[name = tensor("norm7_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176240704)))]; + tensor norm7_1_bias_to_fp16 = const()[name = tensor("norm7_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176242816)))]; + tensor var_1933_to_fp16 = const()[name = tensor("op_1933_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_87_cast_fp16 = layer_norm(axes = input_87_axes_0, beta = norm7_1_bias_to_fp16, epsilon = var_1933_to_fp16, gamma = norm7_1_weight_to_fp16, x = input_85_cast_fp16)[name = tensor("input_87_cast_fp16")]; + tensor attn7_in_proj_weight_to_fp16 = const()[name = tensor("attn7_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176244928)))]; + tensor linear_28_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn7_in_proj_weight_to_fp16, x = input_87_cast_fp16)[name = tensor("linear_28_cast_fp16")]; + tensor var_1947 = const()[name = tensor("op_1947"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_15_cast_fp16 = reshape(shape = var_1947, x = linear_28_cast_fp16)[name = tensor("qkv_15_cast_fp16")]; + tensor x_29_begin_0 = const()[name = tensor("x_29_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_29_end_0 = const()[name = tensor("x_29_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_29_end_mask_0 = const()[name = tensor("x_29_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_29_squeeze_mask_0 = const()[name = tensor("x_29_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_29_cast_fp16 = slice_by_index(begin = x_29_begin_0, end = x_29_end_0, end_mask = x_29_end_mask_0, squeeze_mask = x_29_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("x_29_cast_fp16")]; + tensor x_31_begin_0 = const()[name = tensor("x_31_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_31_end_0 = const()[name = tensor("x_31_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_31_end_mask_0 = const()[name = tensor("x_31_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_31_squeeze_mask_0 = const()[name = tensor("x_31_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_31_cast_fp16 = slice_by_index(begin = x_31_begin_0, end = x_31_end_0, end_mask = x_31_end_mask_0, squeeze_mask = x_31_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("x_31_cast_fp16")]; + tensor v_15_begin_0 = const()[name = tensor("v_15_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_15_end_0 = const()[name = tensor("v_15_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_15_end_mask_0 = const()[name = tensor("v_15_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_15_squeeze_mask_0 = const()[name = tensor("v_15_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_15_cast_fp16 = slice_by_index(begin = v_15_begin_0, end = v_15_end_0, end_mask = v_15_end_mask_0, squeeze_mask = v_15_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("v_15_cast_fp16")]; + tensor var_1992 = const()[name = tensor("op_1992"), val = tensor([256, 16, 32, 2])]; + tensor pairs_29_cast_fp16 = reshape(shape = var_1992, x = x_29_cast_fp16)[name = tensor("pairs_29_cast_fp16")]; + tensor xr_29_begin_0 = const()[name = tensor("xr_29_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_29_end_0 = const()[name = tensor("xr_29_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_29_end_mask_0 = const()[name = tensor("xr_29_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_29_squeeze_mask_0 = const()[name = tensor("xr_29_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_29_cast_fp16 = slice_by_index(begin = xr_29_begin_0, end = xr_29_end_0, end_mask = xr_29_end_mask_0, squeeze_mask = xr_29_squeeze_mask_0, x = pairs_29_cast_fp16)[name = tensor("xr_29_cast_fp16")]; + tensor xi_29_begin_0 = const()[name = tensor("xi_29_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_29_end_0 = const()[name = tensor("xi_29_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_29_end_mask_0 = const()[name = tensor("xi_29_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_29_squeeze_mask_0 = const()[name = tensor("xi_29_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_29_cast_fp16 = slice_by_index(begin = xi_29_begin_0, end = xi_29_end_0, end_mask = xi_29_end_mask_0, squeeze_mask = xi_29_squeeze_mask_0, x = pairs_29_cast_fp16)[name = tensor("xi_29_cast_fp16")]; + tensor var_2000_cast_fp16 = mul(x = xr_29_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2000_cast_fp16")]; + tensor var_2001_cast_fp16 = mul(x = xi_29_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2001_cast_fp16")]; + tensor xor__29_cast_fp16 = sub(x = var_2000_cast_fp16, y = var_2001_cast_fp16)[name = tensor("xor__29_cast_fp16")]; + tensor var_2004_cast_fp16 = mul(x = xr_29_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2004_cast_fp16")]; + tensor var_2005_cast_fp16 = mul(x = xi_29_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2005_cast_fp16")]; + tensor xoi_29_cast_fp16 = add(x = var_2004_cast_fp16, y = var_2005_cast_fp16)[name = tensor("xoi_29_cast_fp16")]; + tensor out_29_axis_0 = const()[name = tensor("out_29_axis_0"), val = tensor(-1)]; + tensor out_29_cast_fp16 = stack(axis = out_29_axis_0, values = (xor__29_cast_fp16, xoi_29_cast_fp16))[name = tensor("out_29_cast_fp16")]; + tensor var_2015 = const()[name = tensor("op_2015"), val = tensor([1, 256, 16, 64])]; + tensor q_15_cast_fp16 = reshape(shape = var_2015, x = out_29_cast_fp16)[name = tensor("q_15_cast_fp16")]; + tensor var_2021 = const()[name = tensor("op_2021"), val = tensor([256, 16, 32, 2])]; + tensor pairs_31_cast_fp16 = reshape(shape = var_2021, x = x_31_cast_fp16)[name = tensor("pairs_31_cast_fp16")]; + tensor xr_31_begin_0 = const()[name = tensor("xr_31_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_31_end_0 = const()[name = tensor("xr_31_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_31_end_mask_0 = const()[name = tensor("xr_31_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_31_squeeze_mask_0 = const()[name = tensor("xr_31_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_31_cast_fp16 = slice_by_index(begin = xr_31_begin_0, end = xr_31_end_0, end_mask = xr_31_end_mask_0, squeeze_mask = xr_31_squeeze_mask_0, x = pairs_31_cast_fp16)[name = tensor("xr_31_cast_fp16")]; + tensor xi_31_begin_0 = const()[name = tensor("xi_31_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_31_end_0 = const()[name = tensor("xi_31_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_31_end_mask_0 = const()[name = tensor("xi_31_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_31_squeeze_mask_0 = const()[name = tensor("xi_31_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_31_cast_fp16 = slice_by_index(begin = xi_31_begin_0, end = xi_31_end_0, end_mask = xi_31_end_mask_0, squeeze_mask = xi_31_squeeze_mask_0, x = pairs_31_cast_fp16)[name = tensor("xi_31_cast_fp16")]; + tensor var_2029_cast_fp16 = mul(x = xr_31_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2029_cast_fp16")]; + tensor var_2030_cast_fp16 = mul(x = xi_31_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2030_cast_fp16")]; + tensor xor__31_cast_fp16 = sub(x = var_2029_cast_fp16, y = var_2030_cast_fp16)[name = tensor("xor__31_cast_fp16")]; + tensor var_2033_cast_fp16 = mul(x = xr_31_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2033_cast_fp16")]; + tensor var_2034_cast_fp16 = mul(x = xi_31_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2034_cast_fp16")]; + tensor xoi_31_cast_fp16 = add(x = var_2033_cast_fp16, y = var_2034_cast_fp16)[name = tensor("xoi_31_cast_fp16")]; + tensor out_31_axis_0 = const()[name = tensor("out_31_axis_0"), val = tensor(-1)]; + tensor out_31_cast_fp16 = stack(axis = out_31_axis_0, values = (xor__31_cast_fp16, xoi_31_cast_fp16))[name = tensor("out_31_cast_fp16")]; + tensor var_2051 = const()[name = tensor("op_2051"), val = tensor([256, 1024])]; + tensor var_2052_cast_fp16 = reshape(shape = var_2051, x = out_31_cast_fp16)[name = tensor("op_2052_cast_fp16")]; + tensor written_k_15_transpose_x_1 = const()[name = tensor("written_k_15_transpose_x_1"), val = tensor(true)]; + tensor written_k_15_transpose_y_1 = const()[name = tensor("written_k_15_transpose_y_1"), val = tensor(false)]; + tensor written_k_15_cast_fp16 = matmul(transpose_x = written_k_15_transpose_x_1, transpose_y = written_k_15_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_2052_cast_fp16)[name = tensor("written_k_15_cast_fp16")]; + tensor var_2059 = const()[name = tensor("op_2059"), val = tensor([256, 1024])]; + tensor var_2060_cast_fp16 = reshape(shape = var_2059, x = v_15_cast_fp16)[name = tensor("op_2060_cast_fp16")]; + tensor written_v_15_transpose_x_1 = const()[name = tensor("written_v_15_transpose_x_1"), val = tensor(true)]; + tensor written_v_15_transpose_y_1 = const()[name = tensor("written_v_15_transpose_y_1"), val = tensor(false)]; + tensor written_v_15_cast_fp16 = matmul(transpose_x = written_v_15_transpose_x_1, transpose_y = written_v_15_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_2060_cast_fp16)[name = tensor("written_v_15_cast_fp16")]; + tensor k_cache7_to_fp16_dtype_0 = const()[name = tensor("k_cache7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache7_to_fp16 = cast(dtype = k_cache7_to_fp16_dtype_0, x = k_cache7)[name = tensor("cast_423")]; + tensor var_2071_cast_fp16 = mul(x = k_cache7_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_2071_cast_fp16")]; + tensor var_2076 = const()[name = tensor("op_2076"), val = tensor([1, 512, 16, 64])]; + tensor var_2077_cast_fp16 = reshape(shape = var_2076, x = written_k_15_cast_fp16)[name = tensor("op_2077_cast_fp16")]; + tensor new_k_cache_15_cast_fp16 = add(x = var_2071_cast_fp16, y = var_2077_cast_fp16)[name = tensor("new_k_cache_15_cast_fp16")]; + tensor new_k_cache_15_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_15_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache7_to_fp16_dtype_0 = const()[name = tensor("v_cache7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache7_to_fp16 = cast(dtype = v_cache7_to_fp16_dtype_0, x = v_cache7)[name = tensor("cast_421")]; + tensor var_2080_cast_fp16 = mul(x = v_cache7_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_2080_cast_fp16")]; + tensor var_2085 = const()[name = tensor("op_2085"), val = tensor([1, 512, 16, 64])]; + tensor var_2086_cast_fp16 = reshape(shape = var_2085, x = written_v_15_cast_fp16)[name = tensor("op_2086_cast_fp16")]; + tensor new_v_cache_15_cast_fp16 = add(x = var_2080_cast_fp16, y = var_2086_cast_fp16)[name = tensor("new_v_cache_15_cast_fp16")]; + tensor new_v_cache_15_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_15_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_2099 = const()[name = tensor("op_2099"), val = tensor([0, 2, 1, 3])]; + tensor var_2110_transpose_x_0 = const()[name = tensor("op_2110_transpose_x_0"), val = tensor(false)]; + tensor var_2110_transpose_y_0 = const()[name = tensor("op_2110_transpose_y_0"), val = tensor(false)]; + tensor transpose_83_perm_0 = const()[name = tensor("transpose_83_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_84_perm_0 = const()[name = tensor("transpose_84_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_84 = transpose(perm = transpose_84_perm_0, x = new_k_cache_15_cast_fp16)[name = tensor("transpose_176")]; + tensor transpose_83 = transpose(perm = transpose_83_perm_0, x = q_15_cast_fp16)[name = tensor("transpose_177")]; + tensor var_2110_cast_fp16 = matmul(transpose_x = var_2110_transpose_x_0, transpose_y = var_2110_transpose_y_0, x = transpose_83, y = transpose_84)[name = tensor("op_2110_cast_fp16")]; + tensor var_2111_to_fp16 = const()[name = tensor("op_2111_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_29_cast_fp16 = mul(x = var_2110_cast_fp16, y = var_2111_to_fp16)[name = tensor("scores_29_cast_fp16")]; + tensor scores_31_cast_fp16 = add(x = scores_29_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_31_cast_fp16")]; + tensor var_2121 = const()[name = tensor("op_2121"), val = tensor(-1)]; + tensor weights_15_cast_fp16 = softmax(axis = var_2121, x = scores_31_cast_fp16)[name = tensor("weights_15_cast_fp16")]; + tensor attn_15_transpose_x_0 = const()[name = tensor("attn_15_transpose_x_0"), val = tensor(false)]; + tensor attn_15_transpose_y_0 = const()[name = tensor("attn_15_transpose_y_0"), val = tensor(false)]; + tensor values_15_cast_fp16 = transpose(perm = var_2099, x = new_v_cache_15_cast_fp16)[name = tensor("transpose_178")]; + tensor attn_15_cast_fp16 = matmul(transpose_x = attn_15_transpose_x_0, transpose_y = attn_15_transpose_y_0, x = weights_15_cast_fp16, y = values_15_cast_fp16)[name = tensor("attn_15_cast_fp16")]; + tensor var_2129 = const()[name = tensor("op_2129"), val = tensor([0, 2, 1, 3])]; + tensor var_2134 = const()[name = tensor("op_2134"), val = tensor([1, 256, 1024])]; + tensor var_2130_cast_fp16 = transpose(perm = var_2129, x = attn_15_cast_fp16)[name = tensor("transpose_175")]; + tensor input_89_cast_fp16 = reshape(shape = var_2134, x = var_2130_cast_fp16)[name = tensor("input_89_cast_fp16")]; + tensor attn7_out_proj_weight_to_fp16 = const()[name = tensor("attn7_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(182536448)))]; + tensor linear_29_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn7_out_proj_weight_to_fp16, x = input_89_cast_fp16)[name = tensor("linear_29_cast_fp16")]; + tensor position7_to_fp16_dtype_0 = const()[name = tensor("position7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position7_to_fp16 = cast(dtype = position7_to_fp16_dtype_0, x = position7)[name = tensor("cast_419")]; + tensor var_2143_cast_fp16 = add(x = position7_to_fp16, y = valid_len_to_fp16)[name = tensor("op_2143_cast_fp16")]; + tensor var_2143_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2143_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_91_cast_fp16 = add(x = input_85_cast_fp16, y = linear_29_cast_fp16)[name = tensor("input_91_cast_fp16")]; + tensor input_93_axes_0 = const()[name = tensor("input_93_axes_0"), val = tensor([-1])]; + tensor norm7_2_weight_to_fp16 = const()[name = tensor("norm7_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184633664)))]; + tensor norm7_2_bias_to_fp16 = const()[name = tensor("norm7_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184635776)))]; + tensor var_2147_to_fp16 = const()[name = tensor("op_2147_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_93_cast_fp16 = layer_norm(axes = input_93_axes_0, beta = norm7_2_bias_to_fp16, epsilon = var_2147_to_fp16, gamma = norm7_2_weight_to_fp16, x = input_91_cast_fp16)[name = tensor("input_93_cast_fp16")]; + tensor linear7_1_weight_to_fp16 = const()[name = tensor("linear7_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184637888)))]; + tensor linear_30_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear7_1_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("linear_30_cast_fp16")]; + tensor input_95_mode_0 = const()[name = tensor("input_95_mode_0"), val = tensor("EXACT")]; + tensor input_95_cast_fp16 = gelu(mode = input_95_mode_0, x = linear_30_cast_fp16)[name = tensor("input_95_cast_fp16")]; + tensor linear7_2_weight_to_fp16 = const()[name = tensor("linear7_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(193026560)))]; + tensor linear_31_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear7_2_weight_to_fp16, x = input_95_cast_fp16)[name = tensor("linear_31_cast_fp16")]; + tensor input_97_cast_fp16 = add(x = input_91_cast_fp16, y = linear_31_cast_fp16)[name = tensor("input_97_cast_fp16")]; + tensor input_99_axes_0 = const()[name = tensor("input_99_axes_0"), val = tensor([-1])]; + tensor norm8_1_weight_to_fp16 = const()[name = tensor("norm8_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201415232)))]; + tensor norm8_1_bias_to_fp16 = const()[name = tensor("norm8_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201417344)))]; + tensor var_2164_to_fp16 = const()[name = tensor("op_2164_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_99_cast_fp16 = layer_norm(axes = input_99_axes_0, beta = norm8_1_bias_to_fp16, epsilon = var_2164_to_fp16, gamma = norm8_1_weight_to_fp16, x = input_97_cast_fp16)[name = tensor("input_99_cast_fp16")]; + tensor attn8_in_proj_weight_to_fp16 = const()[name = tensor("attn8_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201419456)))]; + tensor linear_32_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn8_in_proj_weight_to_fp16, x = input_99_cast_fp16)[name = tensor("linear_32_cast_fp16")]; + tensor var_2178 = const()[name = tensor("op_2178"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_17_cast_fp16 = reshape(shape = var_2178, x = linear_32_cast_fp16)[name = tensor("qkv_17_cast_fp16")]; + tensor x_33_begin_0 = const()[name = tensor("x_33_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_33_end_0 = const()[name = tensor("x_33_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_33_end_mask_0 = const()[name = tensor("x_33_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_33_squeeze_mask_0 = const()[name = tensor("x_33_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_33_cast_fp16 = slice_by_index(begin = x_33_begin_0, end = x_33_end_0, end_mask = x_33_end_mask_0, squeeze_mask = x_33_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("x_33_cast_fp16")]; + tensor x_35_begin_0 = const()[name = tensor("x_35_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_35_end_0 = const()[name = tensor("x_35_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_35_end_mask_0 = const()[name = tensor("x_35_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_35_squeeze_mask_0 = const()[name = tensor("x_35_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_35_cast_fp16 = slice_by_index(begin = x_35_begin_0, end = x_35_end_0, end_mask = x_35_end_mask_0, squeeze_mask = x_35_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("x_35_cast_fp16")]; + tensor v_17_begin_0 = const()[name = tensor("v_17_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_17_end_0 = const()[name = tensor("v_17_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_17_end_mask_0 = const()[name = tensor("v_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_17_squeeze_mask_0 = const()[name = tensor("v_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_17_cast_fp16 = slice_by_index(begin = v_17_begin_0, end = v_17_end_0, end_mask = v_17_end_mask_0, squeeze_mask = v_17_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("v_17_cast_fp16")]; + tensor var_2223 = const()[name = tensor("op_2223"), val = tensor([256, 16, 32, 2])]; + tensor pairs_33_cast_fp16 = reshape(shape = var_2223, x = x_33_cast_fp16)[name = tensor("pairs_33_cast_fp16")]; + tensor xr_33_begin_0 = const()[name = tensor("xr_33_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_33_end_0 = const()[name = tensor("xr_33_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_33_end_mask_0 = const()[name = tensor("xr_33_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_33_squeeze_mask_0 = const()[name = tensor("xr_33_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_33_cast_fp16 = slice_by_index(begin = xr_33_begin_0, end = xr_33_end_0, end_mask = xr_33_end_mask_0, squeeze_mask = xr_33_squeeze_mask_0, x = pairs_33_cast_fp16)[name = tensor("xr_33_cast_fp16")]; + tensor xi_33_begin_0 = const()[name = tensor("xi_33_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_33_end_0 = const()[name = tensor("xi_33_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_33_end_mask_0 = const()[name = tensor("xi_33_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_33_squeeze_mask_0 = const()[name = tensor("xi_33_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_33_cast_fp16 = slice_by_index(begin = xi_33_begin_0, end = xi_33_end_0, end_mask = xi_33_end_mask_0, squeeze_mask = xi_33_squeeze_mask_0, x = pairs_33_cast_fp16)[name = tensor("xi_33_cast_fp16")]; + tensor var_2231_cast_fp16 = mul(x = xr_33_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2231_cast_fp16")]; + tensor var_2232_cast_fp16 = mul(x = xi_33_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2232_cast_fp16")]; + tensor xor__33_cast_fp16 = sub(x = var_2231_cast_fp16, y = var_2232_cast_fp16)[name = tensor("xor__33_cast_fp16")]; + tensor var_2235_cast_fp16 = mul(x = xr_33_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2235_cast_fp16")]; + tensor var_2236_cast_fp16 = mul(x = xi_33_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2236_cast_fp16")]; + tensor xoi_33_cast_fp16 = add(x = var_2235_cast_fp16, y = var_2236_cast_fp16)[name = tensor("xoi_33_cast_fp16")]; + tensor out_33_axis_0 = const()[name = tensor("out_33_axis_0"), val = tensor(-1)]; + tensor out_33_cast_fp16 = stack(axis = out_33_axis_0, values = (xor__33_cast_fp16, xoi_33_cast_fp16))[name = tensor("out_33_cast_fp16")]; + tensor var_2246 = const()[name = tensor("op_2246"), val = tensor([1, 256, 16, 64])]; + tensor q_17_cast_fp16 = reshape(shape = var_2246, x = out_33_cast_fp16)[name = tensor("q_17_cast_fp16")]; + tensor var_2252 = const()[name = tensor("op_2252"), val = tensor([256, 16, 32, 2])]; + tensor pairs_35_cast_fp16 = reshape(shape = var_2252, x = x_35_cast_fp16)[name = tensor("pairs_35_cast_fp16")]; + tensor xr_35_begin_0 = const()[name = tensor("xr_35_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_35_end_0 = const()[name = tensor("xr_35_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_35_end_mask_0 = const()[name = tensor("xr_35_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_35_squeeze_mask_0 = const()[name = tensor("xr_35_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_35_cast_fp16 = slice_by_index(begin = xr_35_begin_0, end = xr_35_end_0, end_mask = xr_35_end_mask_0, squeeze_mask = xr_35_squeeze_mask_0, x = pairs_35_cast_fp16)[name = tensor("xr_35_cast_fp16")]; + tensor xi_35_begin_0 = const()[name = tensor("xi_35_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_35_end_0 = const()[name = tensor("xi_35_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_35_end_mask_0 = const()[name = tensor("xi_35_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_35_squeeze_mask_0 = const()[name = tensor("xi_35_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_35_cast_fp16 = slice_by_index(begin = xi_35_begin_0, end = xi_35_end_0, end_mask = xi_35_end_mask_0, squeeze_mask = xi_35_squeeze_mask_0, x = pairs_35_cast_fp16)[name = tensor("xi_35_cast_fp16")]; + tensor var_2260_cast_fp16 = mul(x = xr_35_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2260_cast_fp16")]; + tensor var_2261_cast_fp16 = mul(x = xi_35_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2261_cast_fp16")]; + tensor xor__35_cast_fp16 = sub(x = var_2260_cast_fp16, y = var_2261_cast_fp16)[name = tensor("xor__35_cast_fp16")]; + tensor var_2264_cast_fp16 = mul(x = xr_35_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2264_cast_fp16")]; + tensor var_2265_cast_fp16 = mul(x = xi_35_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2265_cast_fp16")]; + tensor xoi_35_cast_fp16 = add(x = var_2264_cast_fp16, y = var_2265_cast_fp16)[name = tensor("xoi_35_cast_fp16")]; + tensor out_35_axis_0 = const()[name = tensor("out_35_axis_0"), val = tensor(-1)]; + tensor out_35_cast_fp16 = stack(axis = out_35_axis_0, values = (xor__35_cast_fp16, xoi_35_cast_fp16))[name = tensor("out_35_cast_fp16")]; + tensor var_2282 = const()[name = tensor("op_2282"), val = tensor([256, 1024])]; + tensor var_2283_cast_fp16 = reshape(shape = var_2282, x = out_35_cast_fp16)[name = tensor("op_2283_cast_fp16")]; + tensor written_k_17_transpose_x_1 = const()[name = tensor("written_k_17_transpose_x_1"), val = tensor(true)]; + tensor written_k_17_transpose_y_1 = const()[name = tensor("written_k_17_transpose_y_1"), val = tensor(false)]; + tensor written_k_17_cast_fp16 = matmul(transpose_x = written_k_17_transpose_x_1, transpose_y = written_k_17_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_2283_cast_fp16)[name = tensor("written_k_17_cast_fp16")]; + tensor var_2290 = const()[name = tensor("op_2290"), val = tensor([256, 1024])]; + tensor var_2291_cast_fp16 = reshape(shape = var_2290, x = v_17_cast_fp16)[name = tensor("op_2291_cast_fp16")]; + tensor written_v_17_transpose_x_1 = const()[name = tensor("written_v_17_transpose_x_1"), val = tensor(true)]; + tensor written_v_17_transpose_y_1 = const()[name = tensor("written_v_17_transpose_y_1"), val = tensor(false)]; + tensor written_v_17_cast_fp16 = matmul(transpose_x = written_v_17_transpose_x_1, transpose_y = written_v_17_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_2291_cast_fp16)[name = tensor("written_v_17_cast_fp16")]; + tensor k_cache8_to_fp16_dtype_0 = const()[name = tensor("k_cache8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache8_to_fp16 = cast(dtype = k_cache8_to_fp16_dtype_0, x = k_cache8)[name = tensor("cast_417")]; + tensor var_2302_cast_fp16 = mul(x = k_cache8_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_2302_cast_fp16")]; + tensor var_2307 = const()[name = tensor("op_2307"), val = tensor([1, 512, 16, 64])]; + tensor var_2308_cast_fp16 = reshape(shape = var_2307, x = written_k_17_cast_fp16)[name = tensor("op_2308_cast_fp16")]; + tensor new_k_cache_17_cast_fp16 = add(x = var_2302_cast_fp16, y = var_2308_cast_fp16)[name = tensor("new_k_cache_17_cast_fp16")]; + tensor new_k_cache_17_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_17_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache8_to_fp16_dtype_0 = const()[name = tensor("v_cache8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache8_to_fp16 = cast(dtype = v_cache8_to_fp16_dtype_0, x = v_cache8)[name = tensor("cast_415")]; + tensor var_2311_cast_fp16 = mul(x = v_cache8_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_2311_cast_fp16")]; + tensor var_2316 = const()[name = tensor("op_2316"), val = tensor([1, 512, 16, 64])]; + tensor var_2317_cast_fp16 = reshape(shape = var_2316, x = written_v_17_cast_fp16)[name = tensor("op_2317_cast_fp16")]; + tensor new_v_cache_17_cast_fp16 = add(x = var_2311_cast_fp16, y = var_2317_cast_fp16)[name = tensor("new_v_cache_17_cast_fp16")]; + tensor new_v_cache_17_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_17_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_2330 = const()[name = tensor("op_2330"), val = tensor([0, 2, 1, 3])]; + tensor var_2341_transpose_x_0 = const()[name = tensor("op_2341_transpose_x_0"), val = tensor(false)]; + tensor var_2341_transpose_y_0 = const()[name = tensor("op_2341_transpose_y_0"), val = tensor(false)]; + tensor transpose_85_perm_0 = const()[name = tensor("transpose_85_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_86_perm_0 = const()[name = tensor("transpose_86_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_86 = transpose(perm = transpose_86_perm_0, x = new_k_cache_17_cast_fp16)[name = tensor("transpose_172")]; + tensor transpose_85 = transpose(perm = transpose_85_perm_0, x = q_17_cast_fp16)[name = tensor("transpose_173")]; + tensor var_2341_cast_fp16 = matmul(transpose_x = var_2341_transpose_x_0, transpose_y = var_2341_transpose_y_0, x = transpose_85, y = transpose_86)[name = tensor("op_2341_cast_fp16")]; + tensor var_2342_to_fp16 = const()[name = tensor("op_2342_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_33_cast_fp16 = mul(x = var_2341_cast_fp16, y = var_2342_to_fp16)[name = tensor("scores_33_cast_fp16")]; + tensor scores_35_cast_fp16 = add(x = scores_33_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_35_cast_fp16")]; + tensor var_2352 = const()[name = tensor("op_2352"), val = tensor(-1)]; + tensor weights_17_cast_fp16 = softmax(axis = var_2352, x = scores_35_cast_fp16)[name = tensor("weights_17_cast_fp16")]; + tensor attn_17_transpose_x_0 = const()[name = tensor("attn_17_transpose_x_0"), val = tensor(false)]; + tensor attn_17_transpose_y_0 = const()[name = tensor("attn_17_transpose_y_0"), val = tensor(false)]; + tensor values_17_cast_fp16 = transpose(perm = var_2330, x = new_v_cache_17_cast_fp16)[name = tensor("transpose_174")]; + tensor attn_17_cast_fp16 = matmul(transpose_x = attn_17_transpose_x_0, transpose_y = attn_17_transpose_y_0, x = weights_17_cast_fp16, y = values_17_cast_fp16)[name = tensor("attn_17_cast_fp16")]; + tensor var_2360 = const()[name = tensor("op_2360"), val = tensor([0, 2, 1, 3])]; + tensor var_2365 = const()[name = tensor("op_2365"), val = tensor([1, 256, 1024])]; + tensor var_2361_cast_fp16 = transpose(perm = var_2360, x = attn_17_cast_fp16)[name = tensor("transpose_171")]; + tensor input_101_cast_fp16 = reshape(shape = var_2365, x = var_2361_cast_fp16)[name = tensor("input_101_cast_fp16")]; + tensor attn8_out_proj_weight_to_fp16 = const()[name = tensor("attn8_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(207710976)))]; + tensor linear_33_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn8_out_proj_weight_to_fp16, x = input_101_cast_fp16)[name = tensor("linear_33_cast_fp16")]; + tensor position8_to_fp16_dtype_0 = const()[name = tensor("position8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position8_to_fp16 = cast(dtype = position8_to_fp16_dtype_0, x = position8)[name = tensor("cast_413")]; + tensor var_2374_cast_fp16 = add(x = position8_to_fp16, y = valid_len_to_fp16)[name = tensor("op_2374_cast_fp16")]; + tensor var_2374_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2374_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_103_cast_fp16 = add(x = input_97_cast_fp16, y = linear_33_cast_fp16)[name = tensor("input_103_cast_fp16")]; + tensor input_105_axes_0 = const()[name = tensor("input_105_axes_0"), val = tensor([-1])]; + tensor norm8_2_weight_to_fp16 = const()[name = tensor("norm8_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209808192)))]; + tensor norm8_2_bias_to_fp16 = const()[name = tensor("norm8_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209810304)))]; + tensor var_2378_to_fp16 = const()[name = tensor("op_2378_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_105_cast_fp16 = layer_norm(axes = input_105_axes_0, beta = norm8_2_bias_to_fp16, epsilon = var_2378_to_fp16, gamma = norm8_2_weight_to_fp16, x = input_103_cast_fp16)[name = tensor("input_105_cast_fp16")]; + tensor linear8_1_weight_to_fp16 = const()[name = tensor("linear8_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209812416)))]; + tensor linear_34_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear8_1_weight_to_fp16, x = input_105_cast_fp16)[name = tensor("linear_34_cast_fp16")]; + tensor input_107_mode_0 = const()[name = tensor("input_107_mode_0"), val = tensor("EXACT")]; + tensor input_107_cast_fp16 = gelu(mode = input_107_mode_0, x = linear_34_cast_fp16)[name = tensor("input_107_cast_fp16")]; + tensor linear8_2_weight_to_fp16 = const()[name = tensor("linear8_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(218201088)))]; + tensor linear_35_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear8_2_weight_to_fp16, x = input_107_cast_fp16)[name = tensor("linear_35_cast_fp16")]; + tensor input_109_cast_fp16 = add(x = input_103_cast_fp16, y = linear_35_cast_fp16)[name = tensor("input_109_cast_fp16")]; + tensor input_111_axes_0 = const()[name = tensor("input_111_axes_0"), val = tensor([-1])]; + tensor norm9_1_weight_to_fp16 = const()[name = tensor("norm9_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226589760)))]; + tensor norm9_1_bias_to_fp16 = const()[name = tensor("norm9_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226591872)))]; + tensor var_2395_to_fp16 = const()[name = tensor("op_2395_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_111_cast_fp16 = layer_norm(axes = input_111_axes_0, beta = norm9_1_bias_to_fp16, epsilon = var_2395_to_fp16, gamma = norm9_1_weight_to_fp16, x = input_109_cast_fp16)[name = tensor("input_111_cast_fp16")]; + tensor attn9_in_proj_weight_to_fp16 = const()[name = tensor("attn9_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226593984)))]; + tensor linear_36_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn9_in_proj_weight_to_fp16, x = input_111_cast_fp16)[name = tensor("linear_36_cast_fp16")]; + tensor var_2409 = const()[name = tensor("op_2409"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_19_cast_fp16 = reshape(shape = var_2409, x = linear_36_cast_fp16)[name = tensor("qkv_19_cast_fp16")]; + tensor x_37_begin_0 = const()[name = tensor("x_37_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_37_end_0 = const()[name = tensor("x_37_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_37_end_mask_0 = const()[name = tensor("x_37_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_37_squeeze_mask_0 = const()[name = tensor("x_37_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_37_cast_fp16 = slice_by_index(begin = x_37_begin_0, end = x_37_end_0, end_mask = x_37_end_mask_0, squeeze_mask = x_37_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("x_37_cast_fp16")]; + tensor x_39_begin_0 = const()[name = tensor("x_39_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_39_end_0 = const()[name = tensor("x_39_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_39_end_mask_0 = const()[name = tensor("x_39_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_39_squeeze_mask_0 = const()[name = tensor("x_39_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_39_cast_fp16 = slice_by_index(begin = x_39_begin_0, end = x_39_end_0, end_mask = x_39_end_mask_0, squeeze_mask = x_39_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("x_39_cast_fp16")]; + tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_19_cast_fp16 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("v_19_cast_fp16")]; + tensor var_2454 = const()[name = tensor("op_2454"), val = tensor([256, 16, 32, 2])]; + tensor pairs_37_cast_fp16 = reshape(shape = var_2454, x = x_37_cast_fp16)[name = tensor("pairs_37_cast_fp16")]; + tensor xr_37_begin_0 = const()[name = tensor("xr_37_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_37_end_0 = const()[name = tensor("xr_37_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_37_end_mask_0 = const()[name = tensor("xr_37_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_37_squeeze_mask_0 = const()[name = tensor("xr_37_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_37_cast_fp16 = slice_by_index(begin = xr_37_begin_0, end = xr_37_end_0, end_mask = xr_37_end_mask_0, squeeze_mask = xr_37_squeeze_mask_0, x = pairs_37_cast_fp16)[name = tensor("xr_37_cast_fp16")]; + tensor xi_37_begin_0 = const()[name = tensor("xi_37_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_37_end_0 = const()[name = tensor("xi_37_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_37_end_mask_0 = const()[name = tensor("xi_37_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_37_squeeze_mask_0 = const()[name = tensor("xi_37_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_37_cast_fp16 = slice_by_index(begin = xi_37_begin_0, end = xi_37_end_0, end_mask = xi_37_end_mask_0, squeeze_mask = xi_37_squeeze_mask_0, x = pairs_37_cast_fp16)[name = tensor("xi_37_cast_fp16")]; + tensor var_2462_cast_fp16 = mul(x = xr_37_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2462_cast_fp16")]; + tensor var_2463_cast_fp16 = mul(x = xi_37_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2463_cast_fp16")]; + tensor xor__37_cast_fp16 = sub(x = var_2462_cast_fp16, y = var_2463_cast_fp16)[name = tensor("xor__37_cast_fp16")]; + tensor var_2466_cast_fp16 = mul(x = xr_37_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2466_cast_fp16")]; + tensor var_2467_cast_fp16 = mul(x = xi_37_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2467_cast_fp16")]; + tensor xoi_37_cast_fp16 = add(x = var_2466_cast_fp16, y = var_2467_cast_fp16)[name = tensor("xoi_37_cast_fp16")]; + tensor out_37_axis_0 = const()[name = tensor("out_37_axis_0"), val = tensor(-1)]; + tensor out_37_cast_fp16 = stack(axis = out_37_axis_0, values = (xor__37_cast_fp16, xoi_37_cast_fp16))[name = tensor("out_37_cast_fp16")]; + tensor var_2477 = const()[name = tensor("op_2477"), val = tensor([1, 256, 16, 64])]; + tensor q_19_cast_fp16 = reshape(shape = var_2477, x = out_37_cast_fp16)[name = tensor("q_19_cast_fp16")]; + tensor var_2483 = const()[name = tensor("op_2483"), val = tensor([256, 16, 32, 2])]; + tensor pairs_39_cast_fp16 = reshape(shape = var_2483, x = x_39_cast_fp16)[name = tensor("pairs_39_cast_fp16")]; + tensor xr_39_begin_0 = const()[name = tensor("xr_39_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_39_end_0 = const()[name = tensor("xr_39_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_39_end_mask_0 = const()[name = tensor("xr_39_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_39_squeeze_mask_0 = const()[name = tensor("xr_39_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_39_cast_fp16 = slice_by_index(begin = xr_39_begin_0, end = xr_39_end_0, end_mask = xr_39_end_mask_0, squeeze_mask = xr_39_squeeze_mask_0, x = pairs_39_cast_fp16)[name = tensor("xr_39_cast_fp16")]; + tensor xi_39_begin_0 = const()[name = tensor("xi_39_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_39_end_0 = const()[name = tensor("xi_39_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_39_end_mask_0 = const()[name = tensor("xi_39_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_39_squeeze_mask_0 = const()[name = tensor("xi_39_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_39_cast_fp16 = slice_by_index(begin = xi_39_begin_0, end = xi_39_end_0, end_mask = xi_39_end_mask_0, squeeze_mask = xi_39_squeeze_mask_0, x = pairs_39_cast_fp16)[name = tensor("xi_39_cast_fp16")]; + tensor var_2491_cast_fp16 = mul(x = xr_39_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2491_cast_fp16")]; + tensor var_2492_cast_fp16 = mul(x = xi_39_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2492_cast_fp16")]; + tensor xor__39_cast_fp16 = sub(x = var_2491_cast_fp16, y = var_2492_cast_fp16)[name = tensor("xor__39_cast_fp16")]; + tensor var_2495_cast_fp16 = mul(x = xr_39_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2495_cast_fp16")]; + tensor var_2496_cast_fp16 = mul(x = xi_39_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2496_cast_fp16")]; + tensor xoi_39_cast_fp16 = add(x = var_2495_cast_fp16, y = var_2496_cast_fp16)[name = tensor("xoi_39_cast_fp16")]; + tensor out_39_axis_0 = const()[name = tensor("out_39_axis_0"), val = tensor(-1)]; + tensor out_39_cast_fp16 = stack(axis = out_39_axis_0, values = (xor__39_cast_fp16, xoi_39_cast_fp16))[name = tensor("out_39_cast_fp16")]; + tensor var_2513 = const()[name = tensor("op_2513"), val = tensor([256, 1024])]; + tensor var_2514_cast_fp16 = reshape(shape = var_2513, x = out_39_cast_fp16)[name = tensor("op_2514_cast_fp16")]; + tensor written_k_19_transpose_x_1 = const()[name = tensor("written_k_19_transpose_x_1"), val = tensor(true)]; + tensor written_k_19_transpose_y_1 = const()[name = tensor("written_k_19_transpose_y_1"), val = tensor(false)]; + tensor written_k_19_cast_fp16 = matmul(transpose_x = written_k_19_transpose_x_1, transpose_y = written_k_19_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_2514_cast_fp16)[name = tensor("written_k_19_cast_fp16")]; + tensor var_2521 = const()[name = tensor("op_2521"), val = tensor([256, 1024])]; + tensor var_2522_cast_fp16 = reshape(shape = var_2521, x = v_19_cast_fp16)[name = tensor("op_2522_cast_fp16")]; + tensor written_v_19_transpose_x_1 = const()[name = tensor("written_v_19_transpose_x_1"), val = tensor(true)]; + tensor written_v_19_transpose_y_1 = const()[name = tensor("written_v_19_transpose_y_1"), val = tensor(false)]; + tensor written_v_19_cast_fp16 = matmul(transpose_x = written_v_19_transpose_x_1, transpose_y = written_v_19_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_2522_cast_fp16)[name = tensor("written_v_19_cast_fp16")]; + tensor k_cache9_to_fp16_dtype_0 = const()[name = tensor("k_cache9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache9_to_fp16 = cast(dtype = k_cache9_to_fp16_dtype_0, x = k_cache9)[name = tensor("cast_411")]; + tensor var_2533_cast_fp16 = mul(x = k_cache9_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_2533_cast_fp16")]; + tensor var_2538 = const()[name = tensor("op_2538"), val = tensor([1, 512, 16, 64])]; + tensor var_2539_cast_fp16 = reshape(shape = var_2538, x = written_k_19_cast_fp16)[name = tensor("op_2539_cast_fp16")]; + tensor new_k_cache_19_cast_fp16 = add(x = var_2533_cast_fp16, y = var_2539_cast_fp16)[name = tensor("new_k_cache_19_cast_fp16")]; + tensor new_k_cache_19_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_19_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache9_to_fp16_dtype_0 = const()[name = tensor("v_cache9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache9_to_fp16 = cast(dtype = v_cache9_to_fp16_dtype_0, x = v_cache9)[name = tensor("cast_409")]; + tensor var_2542_cast_fp16 = mul(x = v_cache9_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_2542_cast_fp16")]; + tensor var_2547 = const()[name = tensor("op_2547"), val = tensor([1, 512, 16, 64])]; + tensor var_2548_cast_fp16 = reshape(shape = var_2547, x = written_v_19_cast_fp16)[name = tensor("op_2548_cast_fp16")]; + tensor new_v_cache_19_cast_fp16 = add(x = var_2542_cast_fp16, y = var_2548_cast_fp16)[name = tensor("new_v_cache_19_cast_fp16")]; + tensor new_v_cache_19_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_19_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_2561 = const()[name = tensor("op_2561"), val = tensor([0, 2, 1, 3])]; + tensor var_2572_transpose_x_0 = const()[name = tensor("op_2572_transpose_x_0"), val = tensor(false)]; + tensor var_2572_transpose_y_0 = const()[name = tensor("op_2572_transpose_y_0"), val = tensor(false)]; + tensor transpose_87_perm_0 = const()[name = tensor("transpose_87_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_88_perm_0 = const()[name = tensor("transpose_88_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_88 = transpose(perm = transpose_88_perm_0, x = new_k_cache_19_cast_fp16)[name = tensor("transpose_168")]; + tensor transpose_87 = transpose(perm = transpose_87_perm_0, x = q_19_cast_fp16)[name = tensor("transpose_169")]; + tensor var_2572_cast_fp16 = matmul(transpose_x = var_2572_transpose_x_0, transpose_y = var_2572_transpose_y_0, x = transpose_87, y = transpose_88)[name = tensor("op_2572_cast_fp16")]; + tensor var_2573_to_fp16 = const()[name = tensor("op_2573_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_37_cast_fp16 = mul(x = var_2572_cast_fp16, y = var_2573_to_fp16)[name = tensor("scores_37_cast_fp16")]; + tensor scores_39_cast_fp16 = add(x = scores_37_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_39_cast_fp16")]; + tensor var_2583 = const()[name = tensor("op_2583"), val = tensor(-1)]; + tensor weights_19_cast_fp16 = softmax(axis = var_2583, x = scores_39_cast_fp16)[name = tensor("weights_19_cast_fp16")]; + tensor attn_19_transpose_x_0 = const()[name = tensor("attn_19_transpose_x_0"), val = tensor(false)]; + tensor attn_19_transpose_y_0 = const()[name = tensor("attn_19_transpose_y_0"), val = tensor(false)]; + tensor values_19_cast_fp16 = transpose(perm = var_2561, x = new_v_cache_19_cast_fp16)[name = tensor("transpose_170")]; + tensor attn_19_cast_fp16 = matmul(transpose_x = attn_19_transpose_x_0, transpose_y = attn_19_transpose_y_0, x = weights_19_cast_fp16, y = values_19_cast_fp16)[name = tensor("attn_19_cast_fp16")]; + tensor var_2591 = const()[name = tensor("op_2591"), val = tensor([0, 2, 1, 3])]; + tensor var_2596 = const()[name = tensor("op_2596"), val = tensor([1, 256, 1024])]; + tensor var_2592_cast_fp16 = transpose(perm = var_2591, x = attn_19_cast_fp16)[name = tensor("transpose_167")]; + tensor input_113_cast_fp16 = reshape(shape = var_2596, x = var_2592_cast_fp16)[name = tensor("input_113_cast_fp16")]; + tensor attn9_out_proj_weight_to_fp16 = const()[name = tensor("attn9_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(232885504)))]; + tensor linear_37_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn9_out_proj_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("linear_37_cast_fp16")]; + tensor position9_to_fp16_dtype_0 = const()[name = tensor("position9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position9_to_fp16 = cast(dtype = position9_to_fp16_dtype_0, x = position9)[name = tensor("cast_407")]; + tensor var_2605_cast_fp16 = add(x = position9_to_fp16, y = valid_len_to_fp16)[name = tensor("op_2605_cast_fp16")]; + tensor var_2605_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2605_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_115_cast_fp16 = add(x = input_109_cast_fp16, y = linear_37_cast_fp16)[name = tensor("input_115_cast_fp16")]; + tensor input_117_axes_0 = const()[name = tensor("input_117_axes_0"), val = tensor([-1])]; + tensor norm9_2_weight_to_fp16 = const()[name = tensor("norm9_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(234982720)))]; + tensor norm9_2_bias_to_fp16 = const()[name = tensor("norm9_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(234984832)))]; + tensor var_2609_to_fp16 = const()[name = tensor("op_2609_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_117_cast_fp16 = layer_norm(axes = input_117_axes_0, beta = norm9_2_bias_to_fp16, epsilon = var_2609_to_fp16, gamma = norm9_2_weight_to_fp16, x = input_115_cast_fp16)[name = tensor("input_117_cast_fp16")]; + tensor linear9_1_weight_to_fp16 = const()[name = tensor("linear9_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(234986944)))]; + tensor linear_38_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear9_1_weight_to_fp16, x = input_117_cast_fp16)[name = tensor("linear_38_cast_fp16")]; + tensor input_119_mode_0 = const()[name = tensor("input_119_mode_0"), val = tensor("EXACT")]; + tensor input_119_cast_fp16 = gelu(mode = input_119_mode_0, x = linear_38_cast_fp16)[name = tensor("input_119_cast_fp16")]; + tensor linear9_2_weight_to_fp16 = const()[name = tensor("linear9_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(243375616)))]; + tensor linear_39_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear9_2_weight_to_fp16, x = input_119_cast_fp16)[name = tensor("linear_39_cast_fp16")]; + tensor input_121_cast_fp16 = add(x = input_115_cast_fp16, y = linear_39_cast_fp16)[name = tensor("input_121_cast_fp16")]; + tensor input_123_axes_0 = const()[name = tensor("input_123_axes_0"), val = tensor([-1])]; + tensor norm10_1_weight_to_fp16 = const()[name = tensor("norm10_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(251764288)))]; + tensor norm10_1_bias_to_fp16 = const()[name = tensor("norm10_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(251766400)))]; + tensor var_2626_to_fp16 = const()[name = tensor("op_2626_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_123_cast_fp16 = layer_norm(axes = input_123_axes_0, beta = norm10_1_bias_to_fp16, epsilon = var_2626_to_fp16, gamma = norm10_1_weight_to_fp16, x = input_121_cast_fp16)[name = tensor("input_123_cast_fp16")]; + tensor attn10_in_proj_weight_to_fp16 = const()[name = tensor("attn10_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(251768512)))]; + tensor linear_40_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn10_in_proj_weight_to_fp16, x = input_123_cast_fp16)[name = tensor("linear_40_cast_fp16")]; + tensor var_2640 = const()[name = tensor("op_2640"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_21_cast_fp16 = reshape(shape = var_2640, x = linear_40_cast_fp16)[name = tensor("qkv_21_cast_fp16")]; + tensor x_41_begin_0 = const()[name = tensor("x_41_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_41_end_0 = const()[name = tensor("x_41_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_41_end_mask_0 = const()[name = tensor("x_41_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_41_squeeze_mask_0 = const()[name = tensor("x_41_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_41_cast_fp16 = slice_by_index(begin = x_41_begin_0, end = x_41_end_0, end_mask = x_41_end_mask_0, squeeze_mask = x_41_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("x_41_cast_fp16")]; + tensor x_43_begin_0 = const()[name = tensor("x_43_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_43_end_0 = const()[name = tensor("x_43_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_43_end_mask_0 = const()[name = tensor("x_43_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_43_squeeze_mask_0 = const()[name = tensor("x_43_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_43_cast_fp16 = slice_by_index(begin = x_43_begin_0, end = x_43_end_0, end_mask = x_43_end_mask_0, squeeze_mask = x_43_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("x_43_cast_fp16")]; + tensor v_21_begin_0 = const()[name = tensor("v_21_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_21_end_0 = const()[name = tensor("v_21_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_21_end_mask_0 = const()[name = tensor("v_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_21_squeeze_mask_0 = const()[name = tensor("v_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_21_cast_fp16 = slice_by_index(begin = v_21_begin_0, end = v_21_end_0, end_mask = v_21_end_mask_0, squeeze_mask = v_21_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("v_21_cast_fp16")]; + tensor var_2685 = const()[name = tensor("op_2685"), val = tensor([256, 16, 32, 2])]; + tensor pairs_41_cast_fp16 = reshape(shape = var_2685, x = x_41_cast_fp16)[name = tensor("pairs_41_cast_fp16")]; + tensor xr_41_begin_0 = const()[name = tensor("xr_41_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_41_end_0 = const()[name = tensor("xr_41_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_41_end_mask_0 = const()[name = tensor("xr_41_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_41_squeeze_mask_0 = const()[name = tensor("xr_41_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_41_cast_fp16 = slice_by_index(begin = xr_41_begin_0, end = xr_41_end_0, end_mask = xr_41_end_mask_0, squeeze_mask = xr_41_squeeze_mask_0, x = pairs_41_cast_fp16)[name = tensor("xr_41_cast_fp16")]; + tensor xi_41_begin_0 = const()[name = tensor("xi_41_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_41_end_0 = const()[name = tensor("xi_41_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_41_end_mask_0 = const()[name = tensor("xi_41_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_41_squeeze_mask_0 = const()[name = tensor("xi_41_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_41_cast_fp16 = slice_by_index(begin = xi_41_begin_0, end = xi_41_end_0, end_mask = xi_41_end_mask_0, squeeze_mask = xi_41_squeeze_mask_0, x = pairs_41_cast_fp16)[name = tensor("xi_41_cast_fp16")]; + tensor var_2693_cast_fp16 = mul(x = xr_41_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2693_cast_fp16")]; + tensor var_2694_cast_fp16 = mul(x = xi_41_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2694_cast_fp16")]; + tensor xor__41_cast_fp16 = sub(x = var_2693_cast_fp16, y = var_2694_cast_fp16)[name = tensor("xor__41_cast_fp16")]; + tensor var_2697_cast_fp16 = mul(x = xr_41_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2697_cast_fp16")]; + tensor var_2698_cast_fp16 = mul(x = xi_41_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2698_cast_fp16")]; + tensor xoi_41_cast_fp16 = add(x = var_2697_cast_fp16, y = var_2698_cast_fp16)[name = tensor("xoi_41_cast_fp16")]; + tensor out_41_axis_0 = const()[name = tensor("out_41_axis_0"), val = tensor(-1)]; + tensor out_41_cast_fp16 = stack(axis = out_41_axis_0, values = (xor__41_cast_fp16, xoi_41_cast_fp16))[name = tensor("out_41_cast_fp16")]; + tensor var_2708 = const()[name = tensor("op_2708"), val = tensor([1, 256, 16, 64])]; + tensor q_21_cast_fp16 = reshape(shape = var_2708, x = out_41_cast_fp16)[name = tensor("q_21_cast_fp16")]; + tensor var_2714 = const()[name = tensor("op_2714"), val = tensor([256, 16, 32, 2])]; + tensor pairs_43_cast_fp16 = reshape(shape = var_2714, x = x_43_cast_fp16)[name = tensor("pairs_43_cast_fp16")]; + tensor xr_43_begin_0 = const()[name = tensor("xr_43_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_43_end_0 = const()[name = tensor("xr_43_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_43_end_mask_0 = const()[name = tensor("xr_43_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_43_squeeze_mask_0 = const()[name = tensor("xr_43_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_43_cast_fp16 = slice_by_index(begin = xr_43_begin_0, end = xr_43_end_0, end_mask = xr_43_end_mask_0, squeeze_mask = xr_43_squeeze_mask_0, x = pairs_43_cast_fp16)[name = tensor("xr_43_cast_fp16")]; + tensor xi_43_begin_0 = const()[name = tensor("xi_43_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_43_end_0 = const()[name = tensor("xi_43_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_43_end_mask_0 = const()[name = tensor("xi_43_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_43_squeeze_mask_0 = const()[name = tensor("xi_43_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_43_cast_fp16 = slice_by_index(begin = xi_43_begin_0, end = xi_43_end_0, end_mask = xi_43_end_mask_0, squeeze_mask = xi_43_squeeze_mask_0, x = pairs_43_cast_fp16)[name = tensor("xi_43_cast_fp16")]; + tensor var_2722_cast_fp16 = mul(x = xr_43_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2722_cast_fp16")]; + tensor var_2723_cast_fp16 = mul(x = xi_43_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2723_cast_fp16")]; + tensor xor__43_cast_fp16 = sub(x = var_2722_cast_fp16, y = var_2723_cast_fp16)[name = tensor("xor__43_cast_fp16")]; + tensor var_2726_cast_fp16 = mul(x = xr_43_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2726_cast_fp16")]; + tensor var_2727_cast_fp16 = mul(x = xi_43_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2727_cast_fp16")]; + tensor xoi_43_cast_fp16 = add(x = var_2726_cast_fp16, y = var_2727_cast_fp16)[name = tensor("xoi_43_cast_fp16")]; + tensor out_43_axis_0 = const()[name = tensor("out_43_axis_0"), val = tensor(-1)]; + tensor out_43_cast_fp16 = stack(axis = out_43_axis_0, values = (xor__43_cast_fp16, xoi_43_cast_fp16))[name = tensor("out_43_cast_fp16")]; + tensor var_2744 = const()[name = tensor("op_2744"), val = tensor([256, 1024])]; + tensor var_2745_cast_fp16 = reshape(shape = var_2744, x = out_43_cast_fp16)[name = tensor("op_2745_cast_fp16")]; + tensor written_k_21_transpose_x_1 = const()[name = tensor("written_k_21_transpose_x_1"), val = tensor(true)]; + tensor written_k_21_transpose_y_1 = const()[name = tensor("written_k_21_transpose_y_1"), val = tensor(false)]; + tensor written_k_21_cast_fp16 = matmul(transpose_x = written_k_21_transpose_x_1, transpose_y = written_k_21_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_2745_cast_fp16)[name = tensor("written_k_21_cast_fp16")]; + tensor var_2752 = const()[name = tensor("op_2752"), val = tensor([256, 1024])]; + tensor var_2753_cast_fp16 = reshape(shape = var_2752, x = v_21_cast_fp16)[name = tensor("op_2753_cast_fp16")]; + tensor written_v_21_transpose_x_1 = const()[name = tensor("written_v_21_transpose_x_1"), val = tensor(true)]; + tensor written_v_21_transpose_y_1 = const()[name = tensor("written_v_21_transpose_y_1"), val = tensor(false)]; + tensor written_v_21_cast_fp16 = matmul(transpose_x = written_v_21_transpose_x_1, transpose_y = written_v_21_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_2753_cast_fp16)[name = tensor("written_v_21_cast_fp16")]; + tensor k_cache10_to_fp16_dtype_0 = const()[name = tensor("k_cache10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache10_to_fp16 = cast(dtype = k_cache10_to_fp16_dtype_0, x = k_cache10)[name = tensor("cast_405")]; + tensor var_2764_cast_fp16 = mul(x = k_cache10_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_2764_cast_fp16")]; + tensor var_2769 = const()[name = tensor("op_2769"), val = tensor([1, 512, 16, 64])]; + tensor var_2770_cast_fp16 = reshape(shape = var_2769, x = written_k_21_cast_fp16)[name = tensor("op_2770_cast_fp16")]; + tensor new_k_cache_21_cast_fp16 = add(x = var_2764_cast_fp16, y = var_2770_cast_fp16)[name = tensor("new_k_cache_21_cast_fp16")]; + tensor new_k_cache_21_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_21_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache10_to_fp16_dtype_0 = const()[name = tensor("v_cache10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache10_to_fp16 = cast(dtype = v_cache10_to_fp16_dtype_0, x = v_cache10)[name = tensor("cast_403")]; + tensor var_2773_cast_fp16 = mul(x = v_cache10_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_2773_cast_fp16")]; + tensor var_2778 = const()[name = tensor("op_2778"), val = tensor([1, 512, 16, 64])]; + tensor var_2779_cast_fp16 = reshape(shape = var_2778, x = written_v_21_cast_fp16)[name = tensor("op_2779_cast_fp16")]; + tensor new_v_cache_21_cast_fp16 = add(x = var_2773_cast_fp16, y = var_2779_cast_fp16)[name = tensor("new_v_cache_21_cast_fp16")]; + tensor new_v_cache_21_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_21_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_2792 = const()[name = tensor("op_2792"), val = tensor([0, 2, 1, 3])]; + tensor var_2803_transpose_x_0 = const()[name = tensor("op_2803_transpose_x_0"), val = tensor(false)]; + tensor var_2803_transpose_y_0 = const()[name = tensor("op_2803_transpose_y_0"), val = tensor(false)]; + tensor transpose_89_perm_0 = const()[name = tensor("transpose_89_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_90_perm_0 = const()[name = tensor("transpose_90_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_90 = transpose(perm = transpose_90_perm_0, x = new_k_cache_21_cast_fp16)[name = tensor("transpose_164")]; + tensor transpose_89 = transpose(perm = transpose_89_perm_0, x = q_21_cast_fp16)[name = tensor("transpose_165")]; + tensor var_2803_cast_fp16 = matmul(transpose_x = var_2803_transpose_x_0, transpose_y = var_2803_transpose_y_0, x = transpose_89, y = transpose_90)[name = tensor("op_2803_cast_fp16")]; + tensor var_2804_to_fp16 = const()[name = tensor("op_2804_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_41_cast_fp16 = mul(x = var_2803_cast_fp16, y = var_2804_to_fp16)[name = tensor("scores_41_cast_fp16")]; + tensor scores_43_cast_fp16 = add(x = scores_41_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_43_cast_fp16")]; + tensor var_2814 = const()[name = tensor("op_2814"), val = tensor(-1)]; + tensor weights_21_cast_fp16 = softmax(axis = var_2814, x = scores_43_cast_fp16)[name = tensor("weights_21_cast_fp16")]; + tensor attn_21_transpose_x_0 = const()[name = tensor("attn_21_transpose_x_0"), val = tensor(false)]; + tensor attn_21_transpose_y_0 = const()[name = tensor("attn_21_transpose_y_0"), val = tensor(false)]; + tensor values_21_cast_fp16 = transpose(perm = var_2792, x = new_v_cache_21_cast_fp16)[name = tensor("transpose_166")]; + tensor attn_21_cast_fp16 = matmul(transpose_x = attn_21_transpose_x_0, transpose_y = attn_21_transpose_y_0, x = weights_21_cast_fp16, y = values_21_cast_fp16)[name = tensor("attn_21_cast_fp16")]; + tensor var_2822 = const()[name = tensor("op_2822"), val = tensor([0, 2, 1, 3])]; + tensor var_2827 = const()[name = tensor("op_2827"), val = tensor([1, 256, 1024])]; + tensor var_2823_cast_fp16 = transpose(perm = var_2822, x = attn_21_cast_fp16)[name = tensor("transpose_163")]; + tensor input_125_cast_fp16 = reshape(shape = var_2827, x = var_2823_cast_fp16)[name = tensor("input_125_cast_fp16")]; + tensor attn10_out_proj_weight_to_fp16 = const()[name = tensor("attn10_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(258060032)))]; + tensor linear_41_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn10_out_proj_weight_to_fp16, x = input_125_cast_fp16)[name = tensor("linear_41_cast_fp16")]; + tensor position10_to_fp16_dtype_0 = const()[name = tensor("position10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position10_to_fp16 = cast(dtype = position10_to_fp16_dtype_0, x = position10)[name = tensor("cast_401")]; + tensor var_2836_cast_fp16 = add(x = position10_to_fp16, y = valid_len_to_fp16)[name = tensor("op_2836_cast_fp16")]; + tensor var_2836_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2836_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_127_cast_fp16 = add(x = input_121_cast_fp16, y = linear_41_cast_fp16)[name = tensor("input_127_cast_fp16")]; + tensor input_129_axes_0 = const()[name = tensor("input_129_axes_0"), val = tensor([-1])]; + tensor norm10_2_weight_to_fp16 = const()[name = tensor("norm10_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260157248)))]; + tensor norm10_2_bias_to_fp16 = const()[name = tensor("norm10_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260159360)))]; + tensor var_2840_to_fp16 = const()[name = tensor("op_2840_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_129_cast_fp16 = layer_norm(axes = input_129_axes_0, beta = norm10_2_bias_to_fp16, epsilon = var_2840_to_fp16, gamma = norm10_2_weight_to_fp16, x = input_127_cast_fp16)[name = tensor("input_129_cast_fp16")]; + tensor linear10_1_weight_to_fp16 = const()[name = tensor("linear10_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260161472)))]; + tensor linear_42_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear10_1_weight_to_fp16, x = input_129_cast_fp16)[name = tensor("linear_42_cast_fp16")]; + tensor input_131_mode_0 = const()[name = tensor("input_131_mode_0"), val = tensor("EXACT")]; + tensor input_131_cast_fp16 = gelu(mode = input_131_mode_0, x = linear_42_cast_fp16)[name = tensor("input_131_cast_fp16")]; + tensor linear10_2_weight_to_fp16 = const()[name = tensor("linear10_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(268550144)))]; + tensor linear_43_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear10_2_weight_to_fp16, x = input_131_cast_fp16)[name = tensor("linear_43_cast_fp16")]; + tensor input_133_cast_fp16 = add(x = input_127_cast_fp16, y = linear_43_cast_fp16)[name = tensor("input_133_cast_fp16")]; + tensor input_135_axes_0 = const()[name = tensor("input_135_axes_0"), val = tensor([-1])]; + tensor norm11_1_weight_to_fp16 = const()[name = tensor("norm11_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(276938816)))]; + tensor norm11_1_bias_to_fp16 = const()[name = tensor("norm11_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(276940928)))]; + tensor var_2857_to_fp16 = const()[name = tensor("op_2857_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_135_cast_fp16 = layer_norm(axes = input_135_axes_0, beta = norm11_1_bias_to_fp16, epsilon = var_2857_to_fp16, gamma = norm11_1_weight_to_fp16, x = input_133_cast_fp16)[name = tensor("input_135_cast_fp16")]; + tensor attn11_in_proj_weight_to_fp16 = const()[name = tensor("attn11_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(276943040)))]; + tensor linear_44_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn11_in_proj_weight_to_fp16, x = input_135_cast_fp16)[name = tensor("linear_44_cast_fp16")]; + tensor var_2871 = const()[name = tensor("op_2871"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_23_cast_fp16 = reshape(shape = var_2871, x = linear_44_cast_fp16)[name = tensor("qkv_23_cast_fp16")]; + tensor x_45_begin_0 = const()[name = tensor("x_45_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_45_end_0 = const()[name = tensor("x_45_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_45_end_mask_0 = const()[name = tensor("x_45_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_45_squeeze_mask_0 = const()[name = tensor("x_45_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_45_cast_fp16 = slice_by_index(begin = x_45_begin_0, end = x_45_end_0, end_mask = x_45_end_mask_0, squeeze_mask = x_45_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("x_45_cast_fp16")]; + tensor x_47_begin_0 = const()[name = tensor("x_47_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_47_end_0 = const()[name = tensor("x_47_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_47_end_mask_0 = const()[name = tensor("x_47_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_47_squeeze_mask_0 = const()[name = tensor("x_47_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_47_cast_fp16 = slice_by_index(begin = x_47_begin_0, end = x_47_end_0, end_mask = x_47_end_mask_0, squeeze_mask = x_47_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("x_47_cast_fp16")]; + tensor v_23_begin_0 = const()[name = tensor("v_23_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_23_end_0 = const()[name = tensor("v_23_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_23_end_mask_0 = const()[name = tensor("v_23_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_23_squeeze_mask_0 = const()[name = tensor("v_23_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_23_cast_fp16 = slice_by_index(begin = v_23_begin_0, end = v_23_end_0, end_mask = v_23_end_mask_0, squeeze_mask = v_23_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("v_23_cast_fp16")]; + tensor var_2916 = const()[name = tensor("op_2916"), val = tensor([256, 16, 32, 2])]; + tensor pairs_45_cast_fp16 = reshape(shape = var_2916, x = x_45_cast_fp16)[name = tensor("pairs_45_cast_fp16")]; + tensor xr_45_begin_0 = const()[name = tensor("xr_45_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_45_end_0 = const()[name = tensor("xr_45_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_45_end_mask_0 = const()[name = tensor("xr_45_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_45_squeeze_mask_0 = const()[name = tensor("xr_45_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_45_cast_fp16 = slice_by_index(begin = xr_45_begin_0, end = xr_45_end_0, end_mask = xr_45_end_mask_0, squeeze_mask = xr_45_squeeze_mask_0, x = pairs_45_cast_fp16)[name = tensor("xr_45_cast_fp16")]; + tensor xi_45_begin_0 = const()[name = tensor("xi_45_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_45_end_0 = const()[name = tensor("xi_45_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_45_end_mask_0 = const()[name = tensor("xi_45_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_45_squeeze_mask_0 = const()[name = tensor("xi_45_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_45_cast_fp16 = slice_by_index(begin = xi_45_begin_0, end = xi_45_end_0, end_mask = xi_45_end_mask_0, squeeze_mask = xi_45_squeeze_mask_0, x = pairs_45_cast_fp16)[name = tensor("xi_45_cast_fp16")]; + tensor var_2924_cast_fp16 = mul(x = xr_45_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2924_cast_fp16")]; + tensor var_2925_cast_fp16 = mul(x = xi_45_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2925_cast_fp16")]; + tensor xor__45_cast_fp16 = sub(x = var_2924_cast_fp16, y = var_2925_cast_fp16)[name = tensor("xor__45_cast_fp16")]; + tensor var_2928_cast_fp16 = mul(x = xr_45_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2928_cast_fp16")]; + tensor var_2929_cast_fp16 = mul(x = xi_45_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2929_cast_fp16")]; + tensor xoi_45_cast_fp16 = add(x = var_2928_cast_fp16, y = var_2929_cast_fp16)[name = tensor("xoi_45_cast_fp16")]; + tensor out_45_axis_0 = const()[name = tensor("out_45_axis_0"), val = tensor(-1)]; + tensor out_45_cast_fp16 = stack(axis = out_45_axis_0, values = (xor__45_cast_fp16, xoi_45_cast_fp16))[name = tensor("out_45_cast_fp16")]; + tensor var_2939 = const()[name = tensor("op_2939"), val = tensor([1, 256, 16, 64])]; + tensor q_23_cast_fp16 = reshape(shape = var_2939, x = out_45_cast_fp16)[name = tensor("q_23_cast_fp16")]; + tensor var_2945 = const()[name = tensor("op_2945"), val = tensor([256, 16, 32, 2])]; + tensor pairs_47_cast_fp16 = reshape(shape = var_2945, x = x_47_cast_fp16)[name = tensor("pairs_47_cast_fp16")]; + tensor xr_47_begin_0 = const()[name = tensor("xr_47_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_47_end_0 = const()[name = tensor("xr_47_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_47_end_mask_0 = const()[name = tensor("xr_47_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_47_squeeze_mask_0 = const()[name = tensor("xr_47_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_47_cast_fp16 = slice_by_index(begin = xr_47_begin_0, end = xr_47_end_0, end_mask = xr_47_end_mask_0, squeeze_mask = xr_47_squeeze_mask_0, x = pairs_47_cast_fp16)[name = tensor("xr_47_cast_fp16")]; + tensor xi_47_begin_0 = const()[name = tensor("xi_47_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_47_end_0 = const()[name = tensor("xi_47_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_47_end_mask_0 = const()[name = tensor("xi_47_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_47_squeeze_mask_0 = const()[name = tensor("xi_47_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_47_cast_fp16 = slice_by_index(begin = xi_47_begin_0, end = xi_47_end_0, end_mask = xi_47_end_mask_0, squeeze_mask = xi_47_squeeze_mask_0, x = pairs_47_cast_fp16)[name = tensor("xi_47_cast_fp16")]; + tensor var_2953_cast_fp16 = mul(x = xr_47_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2953_cast_fp16")]; + tensor var_2954_cast_fp16 = mul(x = xi_47_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2954_cast_fp16")]; + tensor xor__47_cast_fp16 = sub(x = var_2953_cast_fp16, y = var_2954_cast_fp16)[name = tensor("xor__47_cast_fp16")]; + tensor var_2957_cast_fp16 = mul(x = xr_47_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2957_cast_fp16")]; + tensor var_2958_cast_fp16 = mul(x = xi_47_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2958_cast_fp16")]; + tensor xoi_47_cast_fp16 = add(x = var_2957_cast_fp16, y = var_2958_cast_fp16)[name = tensor("xoi_47_cast_fp16")]; + tensor out_47_axis_0 = const()[name = tensor("out_47_axis_0"), val = tensor(-1)]; + tensor out_47_cast_fp16 = stack(axis = out_47_axis_0, values = (xor__47_cast_fp16, xoi_47_cast_fp16))[name = tensor("out_47_cast_fp16")]; + tensor var_2975 = const()[name = tensor("op_2975"), val = tensor([256, 1024])]; + tensor var_2976_cast_fp16 = reshape(shape = var_2975, x = out_47_cast_fp16)[name = tensor("op_2976_cast_fp16")]; + tensor written_k_23_transpose_x_1 = const()[name = tensor("written_k_23_transpose_x_1"), val = tensor(true)]; + tensor written_k_23_transpose_y_1 = const()[name = tensor("written_k_23_transpose_y_1"), val = tensor(false)]; + tensor written_k_23_cast_fp16 = matmul(transpose_x = written_k_23_transpose_x_1, transpose_y = written_k_23_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_2976_cast_fp16)[name = tensor("written_k_23_cast_fp16")]; + tensor var_2983 = const()[name = tensor("op_2983"), val = tensor([256, 1024])]; + tensor var_2984_cast_fp16 = reshape(shape = var_2983, x = v_23_cast_fp16)[name = tensor("op_2984_cast_fp16")]; + tensor written_v_23_transpose_x_1 = const()[name = tensor("written_v_23_transpose_x_1"), val = tensor(true)]; + tensor written_v_23_transpose_y_1 = const()[name = tensor("written_v_23_transpose_y_1"), val = tensor(false)]; + tensor written_v_23_cast_fp16 = matmul(transpose_x = written_v_23_transpose_x_1, transpose_y = written_v_23_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_2984_cast_fp16)[name = tensor("written_v_23_cast_fp16")]; + tensor k_cache11_to_fp16_dtype_0 = const()[name = tensor("k_cache11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache11_to_fp16 = cast(dtype = k_cache11_to_fp16_dtype_0, x = k_cache11)[name = tensor("cast_399")]; + tensor var_2995_cast_fp16 = mul(x = k_cache11_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_2995_cast_fp16")]; + tensor var_3000 = const()[name = tensor("op_3000"), val = tensor([1, 512, 16, 64])]; + tensor var_3001_cast_fp16 = reshape(shape = var_3000, x = written_k_23_cast_fp16)[name = tensor("op_3001_cast_fp16")]; + tensor new_k_cache_23_cast_fp16 = add(x = var_2995_cast_fp16, y = var_3001_cast_fp16)[name = tensor("new_k_cache_23_cast_fp16")]; + tensor new_k_cache_23_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_23_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache11_to_fp16_dtype_0 = const()[name = tensor("v_cache11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache11_to_fp16 = cast(dtype = v_cache11_to_fp16_dtype_0, x = v_cache11)[name = tensor("cast_397")]; + tensor var_3004_cast_fp16 = mul(x = v_cache11_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_3004_cast_fp16")]; + tensor var_3009 = const()[name = tensor("op_3009"), val = tensor([1, 512, 16, 64])]; + tensor var_3010_cast_fp16 = reshape(shape = var_3009, x = written_v_23_cast_fp16)[name = tensor("op_3010_cast_fp16")]; + tensor new_v_cache_23_cast_fp16 = add(x = var_3004_cast_fp16, y = var_3010_cast_fp16)[name = tensor("new_v_cache_23_cast_fp16")]; + tensor new_v_cache_23_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_23_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_3023 = const()[name = tensor("op_3023"), val = tensor([0, 2, 1, 3])]; + tensor var_3034_transpose_x_0 = const()[name = tensor("op_3034_transpose_x_0"), val = tensor(false)]; + tensor var_3034_transpose_y_0 = const()[name = tensor("op_3034_transpose_y_0"), val = tensor(false)]; + tensor transpose_91_perm_0 = const()[name = tensor("transpose_91_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_92_perm_0 = const()[name = tensor("transpose_92_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_92 = transpose(perm = transpose_92_perm_0, x = new_k_cache_23_cast_fp16)[name = tensor("transpose_160")]; + tensor transpose_91 = transpose(perm = transpose_91_perm_0, x = q_23_cast_fp16)[name = tensor("transpose_161")]; + tensor var_3034_cast_fp16 = matmul(transpose_x = var_3034_transpose_x_0, transpose_y = var_3034_transpose_y_0, x = transpose_91, y = transpose_92)[name = tensor("op_3034_cast_fp16")]; + tensor var_3035_to_fp16 = const()[name = tensor("op_3035_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_45_cast_fp16 = mul(x = var_3034_cast_fp16, y = var_3035_to_fp16)[name = tensor("scores_45_cast_fp16")]; + tensor scores_47_cast_fp16 = add(x = scores_45_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_47_cast_fp16")]; + tensor var_3045 = const()[name = tensor("op_3045"), val = tensor(-1)]; + tensor weights_23_cast_fp16 = softmax(axis = var_3045, x = scores_47_cast_fp16)[name = tensor("weights_23_cast_fp16")]; + tensor attn_23_transpose_x_0 = const()[name = tensor("attn_23_transpose_x_0"), val = tensor(false)]; + tensor attn_23_transpose_y_0 = const()[name = tensor("attn_23_transpose_y_0"), val = tensor(false)]; + tensor values_23_cast_fp16 = transpose(perm = var_3023, x = new_v_cache_23_cast_fp16)[name = tensor("transpose_162")]; + tensor attn_23_cast_fp16 = matmul(transpose_x = attn_23_transpose_x_0, transpose_y = attn_23_transpose_y_0, x = weights_23_cast_fp16, y = values_23_cast_fp16)[name = tensor("attn_23_cast_fp16")]; + tensor var_3053 = const()[name = tensor("op_3053"), val = tensor([0, 2, 1, 3])]; + tensor var_3058 = const()[name = tensor("op_3058"), val = tensor([1, 256, 1024])]; + tensor var_3054_cast_fp16 = transpose(perm = var_3053, x = attn_23_cast_fp16)[name = tensor("transpose_159")]; + tensor input_137_cast_fp16 = reshape(shape = var_3058, x = var_3054_cast_fp16)[name = tensor("input_137_cast_fp16")]; + tensor attn11_out_proj_weight_to_fp16 = const()[name = tensor("attn11_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(283234560)))]; + tensor linear_45_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn11_out_proj_weight_to_fp16, x = input_137_cast_fp16)[name = tensor("linear_45_cast_fp16")]; + tensor position11_to_fp16_dtype_0 = const()[name = tensor("position11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position11_to_fp16 = cast(dtype = position11_to_fp16_dtype_0, x = position11)[name = tensor("cast_395")]; + tensor var_3067_cast_fp16 = add(x = position11_to_fp16, y = valid_len_to_fp16)[name = tensor("op_3067_cast_fp16")]; + tensor var_3067_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3067_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_139_cast_fp16 = add(x = input_133_cast_fp16, y = linear_45_cast_fp16)[name = tensor("input_139_cast_fp16")]; + tensor input_141_axes_0 = const()[name = tensor("input_141_axes_0"), val = tensor([-1])]; + tensor norm11_2_weight_to_fp16 = const()[name = tensor("norm11_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285331776)))]; + tensor norm11_2_bias_to_fp16 = const()[name = tensor("norm11_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285333888)))]; + tensor var_3071_to_fp16 = const()[name = tensor("op_3071_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_141_cast_fp16 = layer_norm(axes = input_141_axes_0, beta = norm11_2_bias_to_fp16, epsilon = var_3071_to_fp16, gamma = norm11_2_weight_to_fp16, x = input_139_cast_fp16)[name = tensor("input_141_cast_fp16")]; + tensor linear11_1_weight_to_fp16 = const()[name = tensor("linear11_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285336000)))]; + tensor linear_46_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear11_1_weight_to_fp16, x = input_141_cast_fp16)[name = tensor("linear_46_cast_fp16")]; + tensor input_143_mode_0 = const()[name = tensor("input_143_mode_0"), val = tensor("EXACT")]; + tensor input_143_cast_fp16 = gelu(mode = input_143_mode_0, x = linear_46_cast_fp16)[name = tensor("input_143_cast_fp16")]; + tensor linear11_2_weight_to_fp16 = const()[name = tensor("linear11_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293724672)))]; + tensor linear_47_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear11_2_weight_to_fp16, x = input_143_cast_fp16)[name = tensor("linear_47_cast_fp16")]; + tensor input_145_cast_fp16 = add(x = input_139_cast_fp16, y = linear_47_cast_fp16)[name = tensor("input_145_cast_fp16")]; + tensor input_147_axes_0 = const()[name = tensor("input_147_axes_0"), val = tensor([-1])]; + tensor norm12_1_weight_to_fp16 = const()[name = tensor("norm12_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302113344)))]; + tensor norm12_1_bias_to_fp16 = const()[name = tensor("norm12_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302115456)))]; + tensor var_3088_to_fp16 = const()[name = tensor("op_3088_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_147_cast_fp16 = layer_norm(axes = input_147_axes_0, beta = norm12_1_bias_to_fp16, epsilon = var_3088_to_fp16, gamma = norm12_1_weight_to_fp16, x = input_145_cast_fp16)[name = tensor("input_147_cast_fp16")]; + tensor attn12_in_proj_weight_to_fp16 = const()[name = tensor("attn12_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302117568)))]; + tensor linear_48_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn12_in_proj_weight_to_fp16, x = input_147_cast_fp16)[name = tensor("linear_48_cast_fp16")]; + tensor var_3102 = const()[name = tensor("op_3102"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_25_cast_fp16 = reshape(shape = var_3102, x = linear_48_cast_fp16)[name = tensor("qkv_25_cast_fp16")]; + tensor x_49_begin_0 = const()[name = tensor("x_49_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_49_end_0 = const()[name = tensor("x_49_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_49_end_mask_0 = const()[name = tensor("x_49_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_49_squeeze_mask_0 = const()[name = tensor("x_49_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_49_cast_fp16 = slice_by_index(begin = x_49_begin_0, end = x_49_end_0, end_mask = x_49_end_mask_0, squeeze_mask = x_49_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("x_49_cast_fp16")]; + tensor x_51_begin_0 = const()[name = tensor("x_51_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_51_end_0 = const()[name = tensor("x_51_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_51_end_mask_0 = const()[name = tensor("x_51_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_51_squeeze_mask_0 = const()[name = tensor("x_51_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_51_cast_fp16 = slice_by_index(begin = x_51_begin_0, end = x_51_end_0, end_mask = x_51_end_mask_0, squeeze_mask = x_51_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("x_51_cast_fp16")]; + tensor v_25_begin_0 = const()[name = tensor("v_25_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_25_end_0 = const()[name = tensor("v_25_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_25_end_mask_0 = const()[name = tensor("v_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_25_squeeze_mask_0 = const()[name = tensor("v_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_25_cast_fp16 = slice_by_index(begin = v_25_begin_0, end = v_25_end_0, end_mask = v_25_end_mask_0, squeeze_mask = v_25_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("v_25_cast_fp16")]; + tensor var_3147 = const()[name = tensor("op_3147"), val = tensor([256, 16, 32, 2])]; + tensor pairs_49_cast_fp16 = reshape(shape = var_3147, x = x_49_cast_fp16)[name = tensor("pairs_49_cast_fp16")]; + tensor xr_49_begin_0 = const()[name = tensor("xr_49_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_49_end_0 = const()[name = tensor("xr_49_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_49_end_mask_0 = const()[name = tensor("xr_49_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_49_squeeze_mask_0 = const()[name = tensor("xr_49_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_49_cast_fp16 = slice_by_index(begin = xr_49_begin_0, end = xr_49_end_0, end_mask = xr_49_end_mask_0, squeeze_mask = xr_49_squeeze_mask_0, x = pairs_49_cast_fp16)[name = tensor("xr_49_cast_fp16")]; + tensor xi_49_begin_0 = const()[name = tensor("xi_49_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_49_end_0 = const()[name = tensor("xi_49_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_49_end_mask_0 = const()[name = tensor("xi_49_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_49_squeeze_mask_0 = const()[name = tensor("xi_49_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_49_cast_fp16 = slice_by_index(begin = xi_49_begin_0, end = xi_49_end_0, end_mask = xi_49_end_mask_0, squeeze_mask = xi_49_squeeze_mask_0, x = pairs_49_cast_fp16)[name = tensor("xi_49_cast_fp16")]; + tensor var_3155_cast_fp16 = mul(x = xr_49_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3155_cast_fp16")]; + tensor var_3156_cast_fp16 = mul(x = xi_49_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3156_cast_fp16")]; + tensor xor__49_cast_fp16 = sub(x = var_3155_cast_fp16, y = var_3156_cast_fp16)[name = tensor("xor__49_cast_fp16")]; + tensor var_3159_cast_fp16 = mul(x = xr_49_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3159_cast_fp16")]; + tensor var_3160_cast_fp16 = mul(x = xi_49_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3160_cast_fp16")]; + tensor xoi_49_cast_fp16 = add(x = var_3159_cast_fp16, y = var_3160_cast_fp16)[name = tensor("xoi_49_cast_fp16")]; + tensor out_49_axis_0 = const()[name = tensor("out_49_axis_0"), val = tensor(-1)]; + tensor out_49_cast_fp16 = stack(axis = out_49_axis_0, values = (xor__49_cast_fp16, xoi_49_cast_fp16))[name = tensor("out_49_cast_fp16")]; + tensor var_3170 = const()[name = tensor("op_3170"), val = tensor([1, 256, 16, 64])]; + tensor q_25_cast_fp16 = reshape(shape = var_3170, x = out_49_cast_fp16)[name = tensor("q_25_cast_fp16")]; + tensor var_3176 = const()[name = tensor("op_3176"), val = tensor([256, 16, 32, 2])]; + tensor pairs_51_cast_fp16 = reshape(shape = var_3176, x = x_51_cast_fp16)[name = tensor("pairs_51_cast_fp16")]; + tensor xr_51_begin_0 = const()[name = tensor("xr_51_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_51_end_0 = const()[name = tensor("xr_51_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_51_end_mask_0 = const()[name = tensor("xr_51_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_51_squeeze_mask_0 = const()[name = tensor("xr_51_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_51_cast_fp16 = slice_by_index(begin = xr_51_begin_0, end = xr_51_end_0, end_mask = xr_51_end_mask_0, squeeze_mask = xr_51_squeeze_mask_0, x = pairs_51_cast_fp16)[name = tensor("xr_51_cast_fp16")]; + tensor xi_51_begin_0 = const()[name = tensor("xi_51_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_51_end_0 = const()[name = tensor("xi_51_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_51_end_mask_0 = const()[name = tensor("xi_51_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_51_squeeze_mask_0 = const()[name = tensor("xi_51_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_51_cast_fp16 = slice_by_index(begin = xi_51_begin_0, end = xi_51_end_0, end_mask = xi_51_end_mask_0, squeeze_mask = xi_51_squeeze_mask_0, x = pairs_51_cast_fp16)[name = tensor("xi_51_cast_fp16")]; + tensor var_3184_cast_fp16 = mul(x = xr_51_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3184_cast_fp16")]; + tensor var_3185_cast_fp16 = mul(x = xi_51_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3185_cast_fp16")]; + tensor xor__51_cast_fp16 = sub(x = var_3184_cast_fp16, y = var_3185_cast_fp16)[name = tensor("xor__51_cast_fp16")]; + tensor var_3188_cast_fp16 = mul(x = xr_51_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3188_cast_fp16")]; + tensor var_3189_cast_fp16 = mul(x = xi_51_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3189_cast_fp16")]; + tensor xoi_51_cast_fp16 = add(x = var_3188_cast_fp16, y = var_3189_cast_fp16)[name = tensor("xoi_51_cast_fp16")]; + tensor out_51_axis_0 = const()[name = tensor("out_51_axis_0"), val = tensor(-1)]; + tensor out_51_cast_fp16 = stack(axis = out_51_axis_0, values = (xor__51_cast_fp16, xoi_51_cast_fp16))[name = tensor("out_51_cast_fp16")]; + tensor var_3206 = const()[name = tensor("op_3206"), val = tensor([256, 1024])]; + tensor var_3207_cast_fp16 = reshape(shape = var_3206, x = out_51_cast_fp16)[name = tensor("op_3207_cast_fp16")]; + tensor written_k_25_transpose_x_1 = const()[name = tensor("written_k_25_transpose_x_1"), val = tensor(true)]; + tensor written_k_25_transpose_y_1 = const()[name = tensor("written_k_25_transpose_y_1"), val = tensor(false)]; + tensor written_k_25_cast_fp16 = matmul(transpose_x = written_k_25_transpose_x_1, transpose_y = written_k_25_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_3207_cast_fp16)[name = tensor("written_k_25_cast_fp16")]; + tensor var_3214 = const()[name = tensor("op_3214"), val = tensor([256, 1024])]; + tensor var_3215_cast_fp16 = reshape(shape = var_3214, x = v_25_cast_fp16)[name = tensor("op_3215_cast_fp16")]; + tensor written_v_25_transpose_x_1 = const()[name = tensor("written_v_25_transpose_x_1"), val = tensor(true)]; + tensor written_v_25_transpose_y_1 = const()[name = tensor("written_v_25_transpose_y_1"), val = tensor(false)]; + tensor written_v_25_cast_fp16 = matmul(transpose_x = written_v_25_transpose_x_1, transpose_y = written_v_25_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_3215_cast_fp16)[name = tensor("written_v_25_cast_fp16")]; + tensor k_cache12_to_fp16_dtype_0 = const()[name = tensor("k_cache12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache12_to_fp16 = cast(dtype = k_cache12_to_fp16_dtype_0, x = k_cache12)[name = tensor("cast_393")]; + tensor var_3226_cast_fp16 = mul(x = k_cache12_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_3226_cast_fp16")]; + tensor var_3231 = const()[name = tensor("op_3231"), val = tensor([1, 512, 16, 64])]; + tensor var_3232_cast_fp16 = reshape(shape = var_3231, x = written_k_25_cast_fp16)[name = tensor("op_3232_cast_fp16")]; + tensor new_k_cache_25_cast_fp16 = add(x = var_3226_cast_fp16, y = var_3232_cast_fp16)[name = tensor("new_k_cache_25_cast_fp16")]; + tensor new_k_cache_25_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_25_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache12_to_fp16_dtype_0 = const()[name = tensor("v_cache12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache12_to_fp16 = cast(dtype = v_cache12_to_fp16_dtype_0, x = v_cache12)[name = tensor("cast_391")]; + tensor var_3235_cast_fp16 = mul(x = v_cache12_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_3235_cast_fp16")]; + tensor var_3240 = const()[name = tensor("op_3240"), val = tensor([1, 512, 16, 64])]; + tensor var_3241_cast_fp16 = reshape(shape = var_3240, x = written_v_25_cast_fp16)[name = tensor("op_3241_cast_fp16")]; + tensor new_v_cache_25_cast_fp16 = add(x = var_3235_cast_fp16, y = var_3241_cast_fp16)[name = tensor("new_v_cache_25_cast_fp16")]; + tensor new_v_cache_25_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_25_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_3254 = const()[name = tensor("op_3254"), val = tensor([0, 2, 1, 3])]; + tensor var_3265_transpose_x_0 = const()[name = tensor("op_3265_transpose_x_0"), val = tensor(false)]; + tensor var_3265_transpose_y_0 = const()[name = tensor("op_3265_transpose_y_0"), val = tensor(false)]; + tensor transpose_93_perm_0 = const()[name = tensor("transpose_93_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_94_perm_0 = const()[name = tensor("transpose_94_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_94 = transpose(perm = transpose_94_perm_0, x = new_k_cache_25_cast_fp16)[name = tensor("transpose_156")]; + tensor transpose_93 = transpose(perm = transpose_93_perm_0, x = q_25_cast_fp16)[name = tensor("transpose_157")]; + tensor var_3265_cast_fp16 = matmul(transpose_x = var_3265_transpose_x_0, transpose_y = var_3265_transpose_y_0, x = transpose_93, y = transpose_94)[name = tensor("op_3265_cast_fp16")]; + tensor var_3266_to_fp16 = const()[name = tensor("op_3266_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_49_cast_fp16 = mul(x = var_3265_cast_fp16, y = var_3266_to_fp16)[name = tensor("scores_49_cast_fp16")]; + tensor scores_51_cast_fp16 = add(x = scores_49_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_51_cast_fp16")]; + tensor var_3276 = const()[name = tensor("op_3276"), val = tensor(-1)]; + tensor weights_25_cast_fp16 = softmax(axis = var_3276, x = scores_51_cast_fp16)[name = tensor("weights_25_cast_fp16")]; + tensor attn_25_transpose_x_0 = const()[name = tensor("attn_25_transpose_x_0"), val = tensor(false)]; + tensor attn_25_transpose_y_0 = const()[name = tensor("attn_25_transpose_y_0"), val = tensor(false)]; + tensor values_25_cast_fp16 = transpose(perm = var_3254, x = new_v_cache_25_cast_fp16)[name = tensor("transpose_158")]; + tensor attn_25_cast_fp16 = matmul(transpose_x = attn_25_transpose_x_0, transpose_y = attn_25_transpose_y_0, x = weights_25_cast_fp16, y = values_25_cast_fp16)[name = tensor("attn_25_cast_fp16")]; + tensor var_3284 = const()[name = tensor("op_3284"), val = tensor([0, 2, 1, 3])]; + tensor var_3289 = const()[name = tensor("op_3289"), val = tensor([1, 256, 1024])]; + tensor var_3285_cast_fp16 = transpose(perm = var_3284, x = attn_25_cast_fp16)[name = tensor("transpose_155")]; + tensor input_149_cast_fp16 = reshape(shape = var_3289, x = var_3285_cast_fp16)[name = tensor("input_149_cast_fp16")]; + tensor attn12_out_proj_weight_to_fp16 = const()[name = tensor("attn12_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(308409088)))]; + tensor linear_49_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn12_out_proj_weight_to_fp16, x = input_149_cast_fp16)[name = tensor("linear_49_cast_fp16")]; + tensor position12_to_fp16_dtype_0 = const()[name = tensor("position12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position12_to_fp16 = cast(dtype = position12_to_fp16_dtype_0, x = position12)[name = tensor("cast_389")]; + tensor var_3298_cast_fp16 = add(x = position12_to_fp16, y = valid_len_to_fp16)[name = tensor("op_3298_cast_fp16")]; + tensor var_3298_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3298_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_151_cast_fp16 = add(x = input_145_cast_fp16, y = linear_49_cast_fp16)[name = tensor("input_151_cast_fp16")]; + tensor input_153_axes_0 = const()[name = tensor("input_153_axes_0"), val = tensor([-1])]; + tensor norm12_2_weight_to_fp16 = const()[name = tensor("norm12_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(310506304)))]; + tensor norm12_2_bias_to_fp16 = const()[name = tensor("norm12_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(310508416)))]; + tensor var_3302_to_fp16 = const()[name = tensor("op_3302_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_153_cast_fp16 = layer_norm(axes = input_153_axes_0, beta = norm12_2_bias_to_fp16, epsilon = var_3302_to_fp16, gamma = norm12_2_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("input_153_cast_fp16")]; + tensor linear12_1_weight_to_fp16 = const()[name = tensor("linear12_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(310510528)))]; + tensor linear_50_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear12_1_weight_to_fp16, x = input_153_cast_fp16)[name = tensor("linear_50_cast_fp16")]; + tensor input_155_mode_0 = const()[name = tensor("input_155_mode_0"), val = tensor("EXACT")]; + tensor input_155_cast_fp16 = gelu(mode = input_155_mode_0, x = linear_50_cast_fp16)[name = tensor("input_155_cast_fp16")]; + tensor linear12_2_weight_to_fp16 = const()[name = tensor("linear12_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(318899200)))]; + tensor linear_51_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear12_2_weight_to_fp16, x = input_155_cast_fp16)[name = tensor("linear_51_cast_fp16")]; + tensor input_157_cast_fp16 = add(x = input_151_cast_fp16, y = linear_51_cast_fp16)[name = tensor("input_157_cast_fp16")]; + tensor input_159_axes_0 = const()[name = tensor("input_159_axes_0"), val = tensor([-1])]; + tensor norm13_1_weight_to_fp16 = const()[name = tensor("norm13_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(327287872)))]; + tensor norm13_1_bias_to_fp16 = const()[name = tensor("norm13_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(327289984)))]; + tensor var_3319_to_fp16 = const()[name = tensor("op_3319_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_159_cast_fp16 = layer_norm(axes = input_159_axes_0, beta = norm13_1_bias_to_fp16, epsilon = var_3319_to_fp16, gamma = norm13_1_weight_to_fp16, x = input_157_cast_fp16)[name = tensor("input_159_cast_fp16")]; + tensor attn13_in_proj_weight_to_fp16 = const()[name = tensor("attn13_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(327292096)))]; + tensor linear_52_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn13_in_proj_weight_to_fp16, x = input_159_cast_fp16)[name = tensor("linear_52_cast_fp16")]; + tensor var_3333 = const()[name = tensor("op_3333"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_27_cast_fp16 = reshape(shape = var_3333, x = linear_52_cast_fp16)[name = tensor("qkv_27_cast_fp16")]; + tensor x_53_begin_0 = const()[name = tensor("x_53_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_53_end_0 = const()[name = tensor("x_53_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_53_end_mask_0 = const()[name = tensor("x_53_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_53_squeeze_mask_0 = const()[name = tensor("x_53_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_53_cast_fp16 = slice_by_index(begin = x_53_begin_0, end = x_53_end_0, end_mask = x_53_end_mask_0, squeeze_mask = x_53_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("x_53_cast_fp16")]; + tensor x_55_begin_0 = const()[name = tensor("x_55_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_55_end_0 = const()[name = tensor("x_55_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_55_end_mask_0 = const()[name = tensor("x_55_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_55_squeeze_mask_0 = const()[name = tensor("x_55_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_55_cast_fp16 = slice_by_index(begin = x_55_begin_0, end = x_55_end_0, end_mask = x_55_end_mask_0, squeeze_mask = x_55_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("x_55_cast_fp16")]; + tensor v_27_begin_0 = const()[name = tensor("v_27_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_27_end_0 = const()[name = tensor("v_27_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_27_end_mask_0 = const()[name = tensor("v_27_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_27_squeeze_mask_0 = const()[name = tensor("v_27_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_27_cast_fp16 = slice_by_index(begin = v_27_begin_0, end = v_27_end_0, end_mask = v_27_end_mask_0, squeeze_mask = v_27_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("v_27_cast_fp16")]; + tensor var_3378 = const()[name = tensor("op_3378"), val = tensor([256, 16, 32, 2])]; + tensor pairs_53_cast_fp16 = reshape(shape = var_3378, x = x_53_cast_fp16)[name = tensor("pairs_53_cast_fp16")]; + tensor xr_53_begin_0 = const()[name = tensor("xr_53_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_53_end_0 = const()[name = tensor("xr_53_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_53_end_mask_0 = const()[name = tensor("xr_53_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_53_squeeze_mask_0 = const()[name = tensor("xr_53_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_53_cast_fp16 = slice_by_index(begin = xr_53_begin_0, end = xr_53_end_0, end_mask = xr_53_end_mask_0, squeeze_mask = xr_53_squeeze_mask_0, x = pairs_53_cast_fp16)[name = tensor("xr_53_cast_fp16")]; + tensor xi_53_begin_0 = const()[name = tensor("xi_53_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_53_end_0 = const()[name = tensor("xi_53_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_53_end_mask_0 = const()[name = tensor("xi_53_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_53_squeeze_mask_0 = const()[name = tensor("xi_53_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_53_cast_fp16 = slice_by_index(begin = xi_53_begin_0, end = xi_53_end_0, end_mask = xi_53_end_mask_0, squeeze_mask = xi_53_squeeze_mask_0, x = pairs_53_cast_fp16)[name = tensor("xi_53_cast_fp16")]; + tensor var_3386_cast_fp16 = mul(x = xr_53_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3386_cast_fp16")]; + tensor var_3387_cast_fp16 = mul(x = xi_53_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3387_cast_fp16")]; + tensor xor__53_cast_fp16 = sub(x = var_3386_cast_fp16, y = var_3387_cast_fp16)[name = tensor("xor__53_cast_fp16")]; + tensor var_3390_cast_fp16 = mul(x = xr_53_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3390_cast_fp16")]; + tensor var_3391_cast_fp16 = mul(x = xi_53_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3391_cast_fp16")]; + tensor xoi_53_cast_fp16 = add(x = var_3390_cast_fp16, y = var_3391_cast_fp16)[name = tensor("xoi_53_cast_fp16")]; + tensor out_53_axis_0 = const()[name = tensor("out_53_axis_0"), val = tensor(-1)]; + tensor out_53_cast_fp16 = stack(axis = out_53_axis_0, values = (xor__53_cast_fp16, xoi_53_cast_fp16))[name = tensor("out_53_cast_fp16")]; + tensor var_3401 = const()[name = tensor("op_3401"), val = tensor([1, 256, 16, 64])]; + tensor q_27_cast_fp16 = reshape(shape = var_3401, x = out_53_cast_fp16)[name = tensor("q_27_cast_fp16")]; + tensor var_3407 = const()[name = tensor("op_3407"), val = tensor([256, 16, 32, 2])]; + tensor pairs_55_cast_fp16 = reshape(shape = var_3407, x = x_55_cast_fp16)[name = tensor("pairs_55_cast_fp16")]; + tensor xr_55_begin_0 = const()[name = tensor("xr_55_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_55_end_0 = const()[name = tensor("xr_55_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_55_end_mask_0 = const()[name = tensor("xr_55_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_55_squeeze_mask_0 = const()[name = tensor("xr_55_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_55_cast_fp16 = slice_by_index(begin = xr_55_begin_0, end = xr_55_end_0, end_mask = xr_55_end_mask_0, squeeze_mask = xr_55_squeeze_mask_0, x = pairs_55_cast_fp16)[name = tensor("xr_55_cast_fp16")]; + tensor xi_55_begin_0 = const()[name = tensor("xi_55_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_55_end_0 = const()[name = tensor("xi_55_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_55_end_mask_0 = const()[name = tensor("xi_55_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_55_squeeze_mask_0 = const()[name = tensor("xi_55_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_55_cast_fp16 = slice_by_index(begin = xi_55_begin_0, end = xi_55_end_0, end_mask = xi_55_end_mask_0, squeeze_mask = xi_55_squeeze_mask_0, x = pairs_55_cast_fp16)[name = tensor("xi_55_cast_fp16")]; + tensor var_3415_cast_fp16 = mul(x = xr_55_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3415_cast_fp16")]; + tensor var_3416_cast_fp16 = mul(x = xi_55_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3416_cast_fp16")]; + tensor xor__55_cast_fp16 = sub(x = var_3415_cast_fp16, y = var_3416_cast_fp16)[name = tensor("xor__55_cast_fp16")]; + tensor var_3419_cast_fp16 = mul(x = xr_55_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3419_cast_fp16")]; + tensor var_3420_cast_fp16 = mul(x = xi_55_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3420_cast_fp16")]; + tensor xoi_55_cast_fp16 = add(x = var_3419_cast_fp16, y = var_3420_cast_fp16)[name = tensor("xoi_55_cast_fp16")]; + tensor out_55_axis_0 = const()[name = tensor("out_55_axis_0"), val = tensor(-1)]; + tensor out_55_cast_fp16 = stack(axis = out_55_axis_0, values = (xor__55_cast_fp16, xoi_55_cast_fp16))[name = tensor("out_55_cast_fp16")]; + tensor var_3437 = const()[name = tensor("op_3437"), val = tensor([256, 1024])]; + tensor var_3438_cast_fp16 = reshape(shape = var_3437, x = out_55_cast_fp16)[name = tensor("op_3438_cast_fp16")]; + tensor written_k_27_transpose_x_1 = const()[name = tensor("written_k_27_transpose_x_1"), val = tensor(true)]; + tensor written_k_27_transpose_y_1 = const()[name = tensor("written_k_27_transpose_y_1"), val = tensor(false)]; + tensor written_k_27_cast_fp16 = matmul(transpose_x = written_k_27_transpose_x_1, transpose_y = written_k_27_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_3438_cast_fp16)[name = tensor("written_k_27_cast_fp16")]; + tensor var_3445 = const()[name = tensor("op_3445"), val = tensor([256, 1024])]; + tensor var_3446_cast_fp16 = reshape(shape = var_3445, x = v_27_cast_fp16)[name = tensor("op_3446_cast_fp16")]; + tensor written_v_27_transpose_x_1 = const()[name = tensor("written_v_27_transpose_x_1"), val = tensor(true)]; + tensor written_v_27_transpose_y_1 = const()[name = tensor("written_v_27_transpose_y_1"), val = tensor(false)]; + tensor written_v_27_cast_fp16 = matmul(transpose_x = written_v_27_transpose_x_1, transpose_y = written_v_27_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_3446_cast_fp16)[name = tensor("written_v_27_cast_fp16")]; + tensor k_cache13_to_fp16_dtype_0 = const()[name = tensor("k_cache13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache13_to_fp16 = cast(dtype = k_cache13_to_fp16_dtype_0, x = k_cache13)[name = tensor("cast_387")]; + tensor var_3457_cast_fp16 = mul(x = k_cache13_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_3457_cast_fp16")]; + tensor var_3462 = const()[name = tensor("op_3462"), val = tensor([1, 512, 16, 64])]; + tensor var_3463_cast_fp16 = reshape(shape = var_3462, x = written_k_27_cast_fp16)[name = tensor("op_3463_cast_fp16")]; + tensor new_k_cache_27_cast_fp16 = add(x = var_3457_cast_fp16, y = var_3463_cast_fp16)[name = tensor("new_k_cache_27_cast_fp16")]; + tensor new_k_cache_27_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_27_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache13_to_fp16_dtype_0 = const()[name = tensor("v_cache13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache13_to_fp16 = cast(dtype = v_cache13_to_fp16_dtype_0, x = v_cache13)[name = tensor("cast_385")]; + tensor var_3466_cast_fp16 = mul(x = v_cache13_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_3466_cast_fp16")]; + tensor var_3471 = const()[name = tensor("op_3471"), val = tensor([1, 512, 16, 64])]; + tensor var_3472_cast_fp16 = reshape(shape = var_3471, x = written_v_27_cast_fp16)[name = tensor("op_3472_cast_fp16")]; + tensor new_v_cache_27_cast_fp16 = add(x = var_3466_cast_fp16, y = var_3472_cast_fp16)[name = tensor("new_v_cache_27_cast_fp16")]; + tensor new_v_cache_27_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_27_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_3485 = const()[name = tensor("op_3485"), val = tensor([0, 2, 1, 3])]; + tensor var_3496_transpose_x_0 = const()[name = tensor("op_3496_transpose_x_0"), val = tensor(false)]; + tensor var_3496_transpose_y_0 = const()[name = tensor("op_3496_transpose_y_0"), val = tensor(false)]; + tensor transpose_95_perm_0 = const()[name = tensor("transpose_95_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_96_perm_0 = const()[name = tensor("transpose_96_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_96 = transpose(perm = transpose_96_perm_0, x = new_k_cache_27_cast_fp16)[name = tensor("transpose_152")]; + tensor transpose_95 = transpose(perm = transpose_95_perm_0, x = q_27_cast_fp16)[name = tensor("transpose_153")]; + tensor var_3496_cast_fp16 = matmul(transpose_x = var_3496_transpose_x_0, transpose_y = var_3496_transpose_y_0, x = transpose_95, y = transpose_96)[name = tensor("op_3496_cast_fp16")]; + tensor var_3497_to_fp16 = const()[name = tensor("op_3497_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_53_cast_fp16 = mul(x = var_3496_cast_fp16, y = var_3497_to_fp16)[name = tensor("scores_53_cast_fp16")]; + tensor scores_55_cast_fp16 = add(x = scores_53_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_55_cast_fp16")]; + tensor var_3507 = const()[name = tensor("op_3507"), val = tensor(-1)]; + tensor weights_27_cast_fp16 = softmax(axis = var_3507, x = scores_55_cast_fp16)[name = tensor("weights_27_cast_fp16")]; + tensor attn_27_transpose_x_0 = const()[name = tensor("attn_27_transpose_x_0"), val = tensor(false)]; + tensor attn_27_transpose_y_0 = const()[name = tensor("attn_27_transpose_y_0"), val = tensor(false)]; + tensor values_27_cast_fp16 = transpose(perm = var_3485, x = new_v_cache_27_cast_fp16)[name = tensor("transpose_154")]; + tensor attn_27_cast_fp16 = matmul(transpose_x = attn_27_transpose_x_0, transpose_y = attn_27_transpose_y_0, x = weights_27_cast_fp16, y = values_27_cast_fp16)[name = tensor("attn_27_cast_fp16")]; + tensor var_3515 = const()[name = tensor("op_3515"), val = tensor([0, 2, 1, 3])]; + tensor var_3520 = const()[name = tensor("op_3520"), val = tensor([1, 256, 1024])]; + tensor var_3516_cast_fp16 = transpose(perm = var_3515, x = attn_27_cast_fp16)[name = tensor("transpose_151")]; + tensor input_161_cast_fp16 = reshape(shape = var_3520, x = var_3516_cast_fp16)[name = tensor("input_161_cast_fp16")]; + tensor attn13_out_proj_weight_to_fp16 = const()[name = tensor("attn13_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(333583616)))]; + tensor linear_53_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn13_out_proj_weight_to_fp16, x = input_161_cast_fp16)[name = tensor("linear_53_cast_fp16")]; + tensor position13_to_fp16_dtype_0 = const()[name = tensor("position13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position13_to_fp16 = cast(dtype = position13_to_fp16_dtype_0, x = position13)[name = tensor("cast_383")]; + tensor var_3529_cast_fp16 = add(x = position13_to_fp16, y = valid_len_to_fp16)[name = tensor("op_3529_cast_fp16")]; + tensor var_3529_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3529_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_163_cast_fp16 = add(x = input_157_cast_fp16, y = linear_53_cast_fp16)[name = tensor("input_163_cast_fp16")]; + tensor input_165_axes_0 = const()[name = tensor("input_165_axes_0"), val = tensor([-1])]; + tensor norm13_2_weight_to_fp16 = const()[name = tensor("norm13_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335680832)))]; + tensor norm13_2_bias_to_fp16 = const()[name = tensor("norm13_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335682944)))]; + tensor var_3533_to_fp16 = const()[name = tensor("op_3533_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_165_cast_fp16 = layer_norm(axes = input_165_axes_0, beta = norm13_2_bias_to_fp16, epsilon = var_3533_to_fp16, gamma = norm13_2_weight_to_fp16, x = input_163_cast_fp16)[name = tensor("input_165_cast_fp16")]; + tensor linear13_1_weight_to_fp16 = const()[name = tensor("linear13_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335685056)))]; + tensor linear_54_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear13_1_weight_to_fp16, x = input_165_cast_fp16)[name = tensor("linear_54_cast_fp16")]; + tensor input_167_mode_0 = const()[name = tensor("input_167_mode_0"), val = tensor("EXACT")]; + tensor input_167_cast_fp16 = gelu(mode = input_167_mode_0, x = linear_54_cast_fp16)[name = tensor("input_167_cast_fp16")]; + tensor linear13_2_weight_to_fp16 = const()[name = tensor("linear13_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(344073728)))]; + tensor linear_55_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear13_2_weight_to_fp16, x = input_167_cast_fp16)[name = tensor("linear_55_cast_fp16")]; + tensor input_169_cast_fp16 = add(x = input_163_cast_fp16, y = linear_55_cast_fp16)[name = tensor("input_169_cast_fp16")]; + tensor input_171_axes_0 = const()[name = tensor("input_171_axes_0"), val = tensor([-1])]; + tensor norm14_1_weight_to_fp16 = const()[name = tensor("norm14_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(352462400)))]; + tensor norm14_1_bias_to_fp16 = const()[name = tensor("norm14_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(352464512)))]; + tensor var_3550_to_fp16 = const()[name = tensor("op_3550_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_171_cast_fp16 = layer_norm(axes = input_171_axes_0, beta = norm14_1_bias_to_fp16, epsilon = var_3550_to_fp16, gamma = norm14_1_weight_to_fp16, x = input_169_cast_fp16)[name = tensor("input_171_cast_fp16")]; + tensor attn14_in_proj_weight_to_fp16 = const()[name = tensor("attn14_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(352466624)))]; + tensor linear_56_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn14_in_proj_weight_to_fp16, x = input_171_cast_fp16)[name = tensor("linear_56_cast_fp16")]; + tensor var_3564 = const()[name = tensor("op_3564"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_29_cast_fp16 = reshape(shape = var_3564, x = linear_56_cast_fp16)[name = tensor("qkv_29_cast_fp16")]; + tensor x_57_begin_0 = const()[name = tensor("x_57_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_57_end_0 = const()[name = tensor("x_57_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_57_end_mask_0 = const()[name = tensor("x_57_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_57_squeeze_mask_0 = const()[name = tensor("x_57_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_57_cast_fp16 = slice_by_index(begin = x_57_begin_0, end = x_57_end_0, end_mask = x_57_end_mask_0, squeeze_mask = x_57_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("x_57_cast_fp16")]; + tensor x_59_begin_0 = const()[name = tensor("x_59_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_59_end_0 = const()[name = tensor("x_59_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_59_end_mask_0 = const()[name = tensor("x_59_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_59_squeeze_mask_0 = const()[name = tensor("x_59_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_59_cast_fp16 = slice_by_index(begin = x_59_begin_0, end = x_59_end_0, end_mask = x_59_end_mask_0, squeeze_mask = x_59_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("x_59_cast_fp16")]; + tensor v_29_begin_0 = const()[name = tensor("v_29_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_29_end_0 = const()[name = tensor("v_29_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_29_end_mask_0 = const()[name = tensor("v_29_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_29_squeeze_mask_0 = const()[name = tensor("v_29_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_29_cast_fp16 = slice_by_index(begin = v_29_begin_0, end = v_29_end_0, end_mask = v_29_end_mask_0, squeeze_mask = v_29_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("v_29_cast_fp16")]; + tensor var_3609 = const()[name = tensor("op_3609"), val = tensor([256, 16, 32, 2])]; + tensor pairs_57_cast_fp16 = reshape(shape = var_3609, x = x_57_cast_fp16)[name = tensor("pairs_57_cast_fp16")]; + tensor xr_57_begin_0 = const()[name = tensor("xr_57_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_57_end_0 = const()[name = tensor("xr_57_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_57_end_mask_0 = const()[name = tensor("xr_57_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_57_squeeze_mask_0 = const()[name = tensor("xr_57_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_57_cast_fp16 = slice_by_index(begin = xr_57_begin_0, end = xr_57_end_0, end_mask = xr_57_end_mask_0, squeeze_mask = xr_57_squeeze_mask_0, x = pairs_57_cast_fp16)[name = tensor("xr_57_cast_fp16")]; + tensor xi_57_begin_0 = const()[name = tensor("xi_57_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_57_end_0 = const()[name = tensor("xi_57_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_57_end_mask_0 = const()[name = tensor("xi_57_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_57_squeeze_mask_0 = const()[name = tensor("xi_57_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_57_cast_fp16 = slice_by_index(begin = xi_57_begin_0, end = xi_57_end_0, end_mask = xi_57_end_mask_0, squeeze_mask = xi_57_squeeze_mask_0, x = pairs_57_cast_fp16)[name = tensor("xi_57_cast_fp16")]; + tensor var_3617_cast_fp16 = mul(x = xr_57_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3617_cast_fp16")]; + tensor var_3618_cast_fp16 = mul(x = xi_57_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3618_cast_fp16")]; + tensor xor__57_cast_fp16 = sub(x = var_3617_cast_fp16, y = var_3618_cast_fp16)[name = tensor("xor__57_cast_fp16")]; + tensor var_3621_cast_fp16 = mul(x = xr_57_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3621_cast_fp16")]; + tensor var_3622_cast_fp16 = mul(x = xi_57_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3622_cast_fp16")]; + tensor xoi_57_cast_fp16 = add(x = var_3621_cast_fp16, y = var_3622_cast_fp16)[name = tensor("xoi_57_cast_fp16")]; + tensor out_57_axis_0 = const()[name = tensor("out_57_axis_0"), val = tensor(-1)]; + tensor out_57_cast_fp16 = stack(axis = out_57_axis_0, values = (xor__57_cast_fp16, xoi_57_cast_fp16))[name = tensor("out_57_cast_fp16")]; + tensor var_3632 = const()[name = tensor("op_3632"), val = tensor([1, 256, 16, 64])]; + tensor q_29_cast_fp16 = reshape(shape = var_3632, x = out_57_cast_fp16)[name = tensor("q_29_cast_fp16")]; + tensor var_3638 = const()[name = tensor("op_3638"), val = tensor([256, 16, 32, 2])]; + tensor pairs_59_cast_fp16 = reshape(shape = var_3638, x = x_59_cast_fp16)[name = tensor("pairs_59_cast_fp16")]; + tensor xr_59_begin_0 = const()[name = tensor("xr_59_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_59_end_0 = const()[name = tensor("xr_59_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_59_end_mask_0 = const()[name = tensor("xr_59_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_59_squeeze_mask_0 = const()[name = tensor("xr_59_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_59_cast_fp16 = slice_by_index(begin = xr_59_begin_0, end = xr_59_end_0, end_mask = xr_59_end_mask_0, squeeze_mask = xr_59_squeeze_mask_0, x = pairs_59_cast_fp16)[name = tensor("xr_59_cast_fp16")]; + tensor xi_59_begin_0 = const()[name = tensor("xi_59_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_59_end_0 = const()[name = tensor("xi_59_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_59_end_mask_0 = const()[name = tensor("xi_59_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_59_squeeze_mask_0 = const()[name = tensor("xi_59_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_59_cast_fp16 = slice_by_index(begin = xi_59_begin_0, end = xi_59_end_0, end_mask = xi_59_end_mask_0, squeeze_mask = xi_59_squeeze_mask_0, x = pairs_59_cast_fp16)[name = tensor("xi_59_cast_fp16")]; + tensor var_3646_cast_fp16 = mul(x = xr_59_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3646_cast_fp16")]; + tensor var_3647_cast_fp16 = mul(x = xi_59_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3647_cast_fp16")]; + tensor xor__59_cast_fp16 = sub(x = var_3646_cast_fp16, y = var_3647_cast_fp16)[name = tensor("xor__59_cast_fp16")]; + tensor var_3650_cast_fp16 = mul(x = xr_59_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3650_cast_fp16")]; + tensor var_3651_cast_fp16 = mul(x = xi_59_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3651_cast_fp16")]; + tensor xoi_59_cast_fp16 = add(x = var_3650_cast_fp16, y = var_3651_cast_fp16)[name = tensor("xoi_59_cast_fp16")]; + tensor out_59_axis_0 = const()[name = tensor("out_59_axis_0"), val = tensor(-1)]; + tensor out_59_cast_fp16 = stack(axis = out_59_axis_0, values = (xor__59_cast_fp16, xoi_59_cast_fp16))[name = tensor("out_59_cast_fp16")]; + tensor var_3668 = const()[name = tensor("op_3668"), val = tensor([256, 1024])]; + tensor var_3669_cast_fp16 = reshape(shape = var_3668, x = out_59_cast_fp16)[name = tensor("op_3669_cast_fp16")]; + tensor written_k_29_transpose_x_1 = const()[name = tensor("written_k_29_transpose_x_1"), val = tensor(true)]; + tensor written_k_29_transpose_y_1 = const()[name = tensor("written_k_29_transpose_y_1"), val = tensor(false)]; + tensor written_k_29_cast_fp16 = matmul(transpose_x = written_k_29_transpose_x_1, transpose_y = written_k_29_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_3669_cast_fp16)[name = tensor("written_k_29_cast_fp16")]; + tensor var_3676 = const()[name = tensor("op_3676"), val = tensor([256, 1024])]; + tensor var_3677_cast_fp16 = reshape(shape = var_3676, x = v_29_cast_fp16)[name = tensor("op_3677_cast_fp16")]; + tensor written_v_29_transpose_x_1 = const()[name = tensor("written_v_29_transpose_x_1"), val = tensor(true)]; + tensor written_v_29_transpose_y_1 = const()[name = tensor("written_v_29_transpose_y_1"), val = tensor(false)]; + tensor written_v_29_cast_fp16 = matmul(transpose_x = written_v_29_transpose_x_1, transpose_y = written_v_29_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_3677_cast_fp16)[name = tensor("written_v_29_cast_fp16")]; + tensor k_cache14_to_fp16_dtype_0 = const()[name = tensor("k_cache14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache14_to_fp16 = cast(dtype = k_cache14_to_fp16_dtype_0, x = k_cache14)[name = tensor("cast_381")]; + tensor var_3688_cast_fp16 = mul(x = k_cache14_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_3688_cast_fp16")]; + tensor var_3693 = const()[name = tensor("op_3693"), val = tensor([1, 512, 16, 64])]; + tensor var_3694_cast_fp16 = reshape(shape = var_3693, x = written_k_29_cast_fp16)[name = tensor("op_3694_cast_fp16")]; + tensor new_k_cache_29_cast_fp16 = add(x = var_3688_cast_fp16, y = var_3694_cast_fp16)[name = tensor("new_k_cache_29_cast_fp16")]; + tensor new_k_cache_29_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_29_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache14_to_fp16_dtype_0 = const()[name = tensor("v_cache14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache14_to_fp16 = cast(dtype = v_cache14_to_fp16_dtype_0, x = v_cache14)[name = tensor("cast_379")]; + tensor var_3697_cast_fp16 = mul(x = v_cache14_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_3697_cast_fp16")]; + tensor var_3702 = const()[name = tensor("op_3702"), val = tensor([1, 512, 16, 64])]; + tensor var_3703_cast_fp16 = reshape(shape = var_3702, x = written_v_29_cast_fp16)[name = tensor("op_3703_cast_fp16")]; + tensor new_v_cache_29_cast_fp16 = add(x = var_3697_cast_fp16, y = var_3703_cast_fp16)[name = tensor("new_v_cache_29_cast_fp16")]; + tensor new_v_cache_29_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_29_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_3716 = const()[name = tensor("op_3716"), val = tensor([0, 2, 1, 3])]; + tensor var_3727_transpose_x_0 = const()[name = tensor("op_3727_transpose_x_0"), val = tensor(false)]; + tensor var_3727_transpose_y_0 = const()[name = tensor("op_3727_transpose_y_0"), val = tensor(false)]; + tensor transpose_97_perm_0 = const()[name = tensor("transpose_97_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_98_perm_0 = const()[name = tensor("transpose_98_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_98 = transpose(perm = transpose_98_perm_0, x = new_k_cache_29_cast_fp16)[name = tensor("transpose_148")]; + tensor transpose_97 = transpose(perm = transpose_97_perm_0, x = q_29_cast_fp16)[name = tensor("transpose_149")]; + tensor var_3727_cast_fp16 = matmul(transpose_x = var_3727_transpose_x_0, transpose_y = var_3727_transpose_y_0, x = transpose_97, y = transpose_98)[name = tensor("op_3727_cast_fp16")]; + tensor var_3728_to_fp16 = const()[name = tensor("op_3728_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_57_cast_fp16 = mul(x = var_3727_cast_fp16, y = var_3728_to_fp16)[name = tensor("scores_57_cast_fp16")]; + tensor scores_59_cast_fp16 = add(x = scores_57_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_59_cast_fp16")]; + tensor var_3738 = const()[name = tensor("op_3738"), val = tensor(-1)]; + tensor weights_29_cast_fp16 = softmax(axis = var_3738, x = scores_59_cast_fp16)[name = tensor("weights_29_cast_fp16")]; + tensor attn_29_transpose_x_0 = const()[name = tensor("attn_29_transpose_x_0"), val = tensor(false)]; + tensor attn_29_transpose_y_0 = const()[name = tensor("attn_29_transpose_y_0"), val = tensor(false)]; + tensor values_29_cast_fp16 = transpose(perm = var_3716, x = new_v_cache_29_cast_fp16)[name = tensor("transpose_150")]; + tensor attn_29_cast_fp16 = matmul(transpose_x = attn_29_transpose_x_0, transpose_y = attn_29_transpose_y_0, x = weights_29_cast_fp16, y = values_29_cast_fp16)[name = tensor("attn_29_cast_fp16")]; + tensor var_3746 = const()[name = tensor("op_3746"), val = tensor([0, 2, 1, 3])]; + tensor var_3751 = const()[name = tensor("op_3751"), val = tensor([1, 256, 1024])]; + tensor var_3747_cast_fp16 = transpose(perm = var_3746, x = attn_29_cast_fp16)[name = tensor("transpose_147")]; + tensor input_173_cast_fp16 = reshape(shape = var_3751, x = var_3747_cast_fp16)[name = tensor("input_173_cast_fp16")]; + tensor attn14_out_proj_weight_to_fp16 = const()[name = tensor("attn14_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(358758144)))]; + tensor linear_57_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn14_out_proj_weight_to_fp16, x = input_173_cast_fp16)[name = tensor("linear_57_cast_fp16")]; + tensor position14_to_fp16_dtype_0 = const()[name = tensor("position14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position14_to_fp16 = cast(dtype = position14_to_fp16_dtype_0, x = position14)[name = tensor("cast_377")]; + tensor var_3760_cast_fp16 = add(x = position14_to_fp16, y = valid_len_to_fp16)[name = tensor("op_3760_cast_fp16")]; + tensor var_3760_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3760_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_175_cast_fp16 = add(x = input_169_cast_fp16, y = linear_57_cast_fp16)[name = tensor("input_175_cast_fp16")]; + tensor input_177_axes_0 = const()[name = tensor("input_177_axes_0"), val = tensor([-1])]; + tensor norm14_2_weight_to_fp16 = const()[name = tensor("norm14_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(360855360)))]; + tensor norm14_2_bias_to_fp16 = const()[name = tensor("norm14_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(360857472)))]; + tensor var_3764_to_fp16 = const()[name = tensor("op_3764_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_177_cast_fp16 = layer_norm(axes = input_177_axes_0, beta = norm14_2_bias_to_fp16, epsilon = var_3764_to_fp16, gamma = norm14_2_weight_to_fp16, x = input_175_cast_fp16)[name = tensor("input_177_cast_fp16")]; + tensor linear14_1_weight_to_fp16 = const()[name = tensor("linear14_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(360859584)))]; + tensor linear_58_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear14_1_weight_to_fp16, x = input_177_cast_fp16)[name = tensor("linear_58_cast_fp16")]; + tensor input_179_mode_0 = const()[name = tensor("input_179_mode_0"), val = tensor("EXACT")]; + tensor input_179_cast_fp16 = gelu(mode = input_179_mode_0, x = linear_58_cast_fp16)[name = tensor("input_179_cast_fp16")]; + tensor linear14_2_weight_to_fp16 = const()[name = tensor("linear14_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(369248256)))]; + tensor linear_59_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear14_2_weight_to_fp16, x = input_179_cast_fp16)[name = tensor("linear_59_cast_fp16")]; + tensor input_181_cast_fp16 = add(x = input_175_cast_fp16, y = linear_59_cast_fp16)[name = tensor("input_181_cast_fp16")]; + tensor input_183_axes_0 = const()[name = tensor("input_183_axes_0"), val = tensor([-1])]; + tensor norm15_1_weight_to_fp16 = const()[name = tensor("norm15_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(377636928)))]; + tensor norm15_1_bias_to_fp16 = const()[name = tensor("norm15_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(377639040)))]; + tensor var_3781_to_fp16 = const()[name = tensor("op_3781_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_183_cast_fp16 = layer_norm(axes = input_183_axes_0, beta = norm15_1_bias_to_fp16, epsilon = var_3781_to_fp16, gamma = norm15_1_weight_to_fp16, x = input_181_cast_fp16)[name = tensor("input_183_cast_fp16")]; + tensor attn15_in_proj_weight_to_fp16 = const()[name = tensor("attn15_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(377641152)))]; + tensor linear_60_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn15_in_proj_weight_to_fp16, x = input_183_cast_fp16)[name = tensor("linear_60_cast_fp16")]; + tensor var_3795 = const()[name = tensor("op_3795"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_31_cast_fp16 = reshape(shape = var_3795, x = linear_60_cast_fp16)[name = tensor("qkv_31_cast_fp16")]; + tensor x_61_begin_0 = const()[name = tensor("x_61_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_61_end_0 = const()[name = tensor("x_61_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_61_end_mask_0 = const()[name = tensor("x_61_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_61_squeeze_mask_0 = const()[name = tensor("x_61_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_61_cast_fp16 = slice_by_index(begin = x_61_begin_0, end = x_61_end_0, end_mask = x_61_end_mask_0, squeeze_mask = x_61_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("x_61_cast_fp16")]; + tensor x_63_begin_0 = const()[name = tensor("x_63_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_63_end_0 = const()[name = tensor("x_63_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_63_end_mask_0 = const()[name = tensor("x_63_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_63_squeeze_mask_0 = const()[name = tensor("x_63_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_63_cast_fp16 = slice_by_index(begin = x_63_begin_0, end = x_63_end_0, end_mask = x_63_end_mask_0, squeeze_mask = x_63_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("x_63_cast_fp16")]; + tensor v_31_begin_0 = const()[name = tensor("v_31_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_31_end_0 = const()[name = tensor("v_31_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_31_end_mask_0 = const()[name = tensor("v_31_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_31_squeeze_mask_0 = const()[name = tensor("v_31_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_31_cast_fp16 = slice_by_index(begin = v_31_begin_0, end = v_31_end_0, end_mask = v_31_end_mask_0, squeeze_mask = v_31_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("v_31_cast_fp16")]; + tensor var_3840 = const()[name = tensor("op_3840"), val = tensor([256, 16, 32, 2])]; + tensor pairs_61_cast_fp16 = reshape(shape = var_3840, x = x_61_cast_fp16)[name = tensor("pairs_61_cast_fp16")]; + tensor xr_61_begin_0 = const()[name = tensor("xr_61_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_61_end_0 = const()[name = tensor("xr_61_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_61_end_mask_0 = const()[name = tensor("xr_61_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_61_squeeze_mask_0 = const()[name = tensor("xr_61_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_61_cast_fp16 = slice_by_index(begin = xr_61_begin_0, end = xr_61_end_0, end_mask = xr_61_end_mask_0, squeeze_mask = xr_61_squeeze_mask_0, x = pairs_61_cast_fp16)[name = tensor("xr_61_cast_fp16")]; + tensor xi_61_begin_0 = const()[name = tensor("xi_61_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_61_end_0 = const()[name = tensor("xi_61_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_61_end_mask_0 = const()[name = tensor("xi_61_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_61_squeeze_mask_0 = const()[name = tensor("xi_61_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_61_cast_fp16 = slice_by_index(begin = xi_61_begin_0, end = xi_61_end_0, end_mask = xi_61_end_mask_0, squeeze_mask = xi_61_squeeze_mask_0, x = pairs_61_cast_fp16)[name = tensor("xi_61_cast_fp16")]; + tensor var_3848_cast_fp16 = mul(x = xr_61_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3848_cast_fp16")]; + tensor var_3849_cast_fp16 = mul(x = xi_61_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3849_cast_fp16")]; + tensor xor__61_cast_fp16 = sub(x = var_3848_cast_fp16, y = var_3849_cast_fp16)[name = tensor("xor__61_cast_fp16")]; + tensor var_3852_cast_fp16 = mul(x = xr_61_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3852_cast_fp16")]; + tensor var_3853_cast_fp16 = mul(x = xi_61_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3853_cast_fp16")]; + tensor xoi_61_cast_fp16 = add(x = var_3852_cast_fp16, y = var_3853_cast_fp16)[name = tensor("xoi_61_cast_fp16")]; + tensor out_61_axis_0 = const()[name = tensor("out_61_axis_0"), val = tensor(-1)]; + tensor out_61_cast_fp16 = stack(axis = out_61_axis_0, values = (xor__61_cast_fp16, xoi_61_cast_fp16))[name = tensor("out_61_cast_fp16")]; + tensor var_3863 = const()[name = tensor("op_3863"), val = tensor([1, 256, 16, 64])]; + tensor q_31_cast_fp16 = reshape(shape = var_3863, x = out_61_cast_fp16)[name = tensor("q_31_cast_fp16")]; + tensor var_3869 = const()[name = tensor("op_3869"), val = tensor([256, 16, 32, 2])]; + tensor pairs_63_cast_fp16 = reshape(shape = var_3869, x = x_63_cast_fp16)[name = tensor("pairs_63_cast_fp16")]; + tensor xr_63_begin_0 = const()[name = tensor("xr_63_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_63_end_0 = const()[name = tensor("xr_63_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_63_end_mask_0 = const()[name = tensor("xr_63_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_63_squeeze_mask_0 = const()[name = tensor("xr_63_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_63_cast_fp16 = slice_by_index(begin = xr_63_begin_0, end = xr_63_end_0, end_mask = xr_63_end_mask_0, squeeze_mask = xr_63_squeeze_mask_0, x = pairs_63_cast_fp16)[name = tensor("xr_63_cast_fp16")]; + tensor xi_63_begin_0 = const()[name = tensor("xi_63_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_63_end_0 = const()[name = tensor("xi_63_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_63_end_mask_0 = const()[name = tensor("xi_63_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_63_squeeze_mask_0 = const()[name = tensor("xi_63_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_63_cast_fp16 = slice_by_index(begin = xi_63_begin_0, end = xi_63_end_0, end_mask = xi_63_end_mask_0, squeeze_mask = xi_63_squeeze_mask_0, x = pairs_63_cast_fp16)[name = tensor("xi_63_cast_fp16")]; + tensor var_3877_cast_fp16 = mul(x = xr_63_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3877_cast_fp16")]; + tensor var_3878_cast_fp16 = mul(x = xi_63_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3878_cast_fp16")]; + tensor xor__63_cast_fp16 = sub(x = var_3877_cast_fp16, y = var_3878_cast_fp16)[name = tensor("xor__63_cast_fp16")]; + tensor var_3881_cast_fp16 = mul(x = xr_63_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3881_cast_fp16")]; + tensor var_3882_cast_fp16 = mul(x = xi_63_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3882_cast_fp16")]; + tensor xoi_63_cast_fp16 = add(x = var_3881_cast_fp16, y = var_3882_cast_fp16)[name = tensor("xoi_63_cast_fp16")]; + tensor out_63_axis_0 = const()[name = tensor("out_63_axis_0"), val = tensor(-1)]; + tensor out_63_cast_fp16 = stack(axis = out_63_axis_0, values = (xor__63_cast_fp16, xoi_63_cast_fp16))[name = tensor("out_63_cast_fp16")]; + tensor var_3899 = const()[name = tensor("op_3899"), val = tensor([256, 1024])]; + tensor var_3900_cast_fp16 = reshape(shape = var_3899, x = out_63_cast_fp16)[name = tensor("op_3900_cast_fp16")]; + tensor written_k_31_transpose_x_1 = const()[name = tensor("written_k_31_transpose_x_1"), val = tensor(true)]; + tensor written_k_31_transpose_y_1 = const()[name = tensor("written_k_31_transpose_y_1"), val = tensor(false)]; + tensor written_k_31_cast_fp16 = matmul(transpose_x = written_k_31_transpose_x_1, transpose_y = written_k_31_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_3900_cast_fp16)[name = tensor("written_k_31_cast_fp16")]; + tensor var_3907 = const()[name = tensor("op_3907"), val = tensor([256, 1024])]; + tensor var_3908_cast_fp16 = reshape(shape = var_3907, x = v_31_cast_fp16)[name = tensor("op_3908_cast_fp16")]; + tensor written_v_31_transpose_x_1 = const()[name = tensor("written_v_31_transpose_x_1"), val = tensor(true)]; + tensor written_v_31_transpose_y_1 = const()[name = tensor("written_v_31_transpose_y_1"), val = tensor(false)]; + tensor written_v_31_cast_fp16 = matmul(transpose_x = written_v_31_transpose_x_1, transpose_y = written_v_31_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_3908_cast_fp16)[name = tensor("written_v_31_cast_fp16")]; + tensor k_cache15_to_fp16_dtype_0 = const()[name = tensor("k_cache15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache15_to_fp16 = cast(dtype = k_cache15_to_fp16_dtype_0, x = k_cache15)[name = tensor("cast_375")]; + tensor var_3919_cast_fp16 = mul(x = k_cache15_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_3919_cast_fp16")]; + tensor var_3924 = const()[name = tensor("op_3924"), val = tensor([1, 512, 16, 64])]; + tensor var_3925_cast_fp16 = reshape(shape = var_3924, x = written_k_31_cast_fp16)[name = tensor("op_3925_cast_fp16")]; + tensor new_k_cache_31_cast_fp16 = add(x = var_3919_cast_fp16, y = var_3925_cast_fp16)[name = tensor("new_k_cache_31_cast_fp16")]; + tensor new_k_cache_31_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_31_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache15_to_fp16_dtype_0 = const()[name = tensor("v_cache15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache15_to_fp16 = cast(dtype = v_cache15_to_fp16_dtype_0, x = v_cache15)[name = tensor("cast_373")]; + tensor var_3928_cast_fp16 = mul(x = v_cache15_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_3928_cast_fp16")]; + tensor var_3933 = const()[name = tensor("op_3933"), val = tensor([1, 512, 16, 64])]; + tensor var_3934_cast_fp16 = reshape(shape = var_3933, x = written_v_31_cast_fp16)[name = tensor("op_3934_cast_fp16")]; + tensor new_v_cache_31_cast_fp16 = add(x = var_3928_cast_fp16, y = var_3934_cast_fp16)[name = tensor("new_v_cache_31_cast_fp16")]; + tensor new_v_cache_31_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_31_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_3947 = const()[name = tensor("op_3947"), val = tensor([0, 2, 1, 3])]; + tensor var_3958_transpose_x_0 = const()[name = tensor("op_3958_transpose_x_0"), val = tensor(false)]; + tensor var_3958_transpose_y_0 = const()[name = tensor("op_3958_transpose_y_0"), val = tensor(false)]; + tensor transpose_99_perm_0 = const()[name = tensor("transpose_99_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_100_perm_0 = const()[name = tensor("transpose_100_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_100 = transpose(perm = transpose_100_perm_0, x = new_k_cache_31_cast_fp16)[name = tensor("transpose_144")]; + tensor transpose_99 = transpose(perm = transpose_99_perm_0, x = q_31_cast_fp16)[name = tensor("transpose_145")]; + tensor var_3958_cast_fp16 = matmul(transpose_x = var_3958_transpose_x_0, transpose_y = var_3958_transpose_y_0, x = transpose_99, y = transpose_100)[name = tensor("op_3958_cast_fp16")]; + tensor var_3959_to_fp16 = const()[name = tensor("op_3959_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_61_cast_fp16 = mul(x = var_3958_cast_fp16, y = var_3959_to_fp16)[name = tensor("scores_61_cast_fp16")]; + tensor scores_63_cast_fp16 = add(x = scores_61_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_63_cast_fp16")]; + tensor var_3969 = const()[name = tensor("op_3969"), val = tensor(-1)]; + tensor weights_31_cast_fp16 = softmax(axis = var_3969, x = scores_63_cast_fp16)[name = tensor("weights_31_cast_fp16")]; + tensor attn_31_transpose_x_0 = const()[name = tensor("attn_31_transpose_x_0"), val = tensor(false)]; + tensor attn_31_transpose_y_0 = const()[name = tensor("attn_31_transpose_y_0"), val = tensor(false)]; + tensor values_31_cast_fp16 = transpose(perm = var_3947, x = new_v_cache_31_cast_fp16)[name = tensor("transpose_146")]; + tensor attn_31_cast_fp16 = matmul(transpose_x = attn_31_transpose_x_0, transpose_y = attn_31_transpose_y_0, x = weights_31_cast_fp16, y = values_31_cast_fp16)[name = tensor("attn_31_cast_fp16")]; + tensor var_3977 = const()[name = tensor("op_3977"), val = tensor([0, 2, 1, 3])]; + tensor var_3982 = const()[name = tensor("op_3982"), val = tensor([1, 256, 1024])]; + tensor var_3978_cast_fp16 = transpose(perm = var_3977, x = attn_31_cast_fp16)[name = tensor("transpose_143")]; + tensor input_185_cast_fp16 = reshape(shape = var_3982, x = var_3978_cast_fp16)[name = tensor("input_185_cast_fp16")]; + tensor attn15_out_proj_weight_to_fp16 = const()[name = tensor("attn15_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(383932672)))]; + tensor linear_61_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn15_out_proj_weight_to_fp16, x = input_185_cast_fp16)[name = tensor("linear_61_cast_fp16")]; + tensor position15_to_fp16_dtype_0 = const()[name = tensor("position15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position15_to_fp16 = cast(dtype = position15_to_fp16_dtype_0, x = position15)[name = tensor("cast_371")]; + tensor var_3991_cast_fp16 = add(x = position15_to_fp16, y = valid_len_to_fp16)[name = tensor("op_3991_cast_fp16")]; + tensor var_3991_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3991_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_187_cast_fp16 = add(x = input_181_cast_fp16, y = linear_61_cast_fp16)[name = tensor("input_187_cast_fp16")]; + tensor input_189_axes_0 = const()[name = tensor("input_189_axes_0"), val = tensor([-1])]; + tensor norm15_2_weight_to_fp16 = const()[name = tensor("norm15_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(386029888)))]; + tensor norm15_2_bias_to_fp16 = const()[name = tensor("norm15_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(386032000)))]; + tensor var_3995_to_fp16 = const()[name = tensor("op_3995_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_189_cast_fp16 = layer_norm(axes = input_189_axes_0, beta = norm15_2_bias_to_fp16, epsilon = var_3995_to_fp16, gamma = norm15_2_weight_to_fp16, x = input_187_cast_fp16)[name = tensor("input_189_cast_fp16")]; + tensor linear15_1_weight_to_fp16 = const()[name = tensor("linear15_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(386034112)))]; + tensor linear_62_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear15_1_weight_to_fp16, x = input_189_cast_fp16)[name = tensor("linear_62_cast_fp16")]; + tensor input_191_mode_0 = const()[name = tensor("input_191_mode_0"), val = tensor("EXACT")]; + tensor input_191_cast_fp16 = gelu(mode = input_191_mode_0, x = linear_62_cast_fp16)[name = tensor("input_191_cast_fp16")]; + tensor linear15_2_weight_to_fp16 = const()[name = tensor("linear15_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(394422784)))]; + tensor linear_63_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear15_2_weight_to_fp16, x = input_191_cast_fp16)[name = tensor("linear_63_cast_fp16")]; + tensor input_193_cast_fp16 = add(x = input_187_cast_fp16, y = linear_63_cast_fp16)[name = tensor("input_193_cast_fp16")]; + tensor input_195_axes_0 = const()[name = tensor("input_195_axes_0"), val = tensor([-1])]; + tensor norm16_1_weight_to_fp16 = const()[name = tensor("norm16_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(402811456)))]; + tensor norm16_1_bias_to_fp16 = const()[name = tensor("norm16_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(402813568)))]; + tensor var_4012_to_fp16 = const()[name = tensor("op_4012_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_195_cast_fp16 = layer_norm(axes = input_195_axes_0, beta = norm16_1_bias_to_fp16, epsilon = var_4012_to_fp16, gamma = norm16_1_weight_to_fp16, x = input_193_cast_fp16)[name = tensor("input_195_cast_fp16")]; + tensor attn16_in_proj_weight_to_fp16 = const()[name = tensor("attn16_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(402815680)))]; + tensor linear_64_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn16_in_proj_weight_to_fp16, x = input_195_cast_fp16)[name = tensor("linear_64_cast_fp16")]; + tensor var_4026 = const()[name = tensor("op_4026"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_33_cast_fp16 = reshape(shape = var_4026, x = linear_64_cast_fp16)[name = tensor("qkv_33_cast_fp16")]; + tensor x_65_begin_0 = const()[name = tensor("x_65_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_65_end_0 = const()[name = tensor("x_65_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_65_end_mask_0 = const()[name = tensor("x_65_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_65_squeeze_mask_0 = const()[name = tensor("x_65_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_65_cast_fp16 = slice_by_index(begin = x_65_begin_0, end = x_65_end_0, end_mask = x_65_end_mask_0, squeeze_mask = x_65_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("x_65_cast_fp16")]; + tensor x_67_begin_0 = const()[name = tensor("x_67_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_67_end_0 = const()[name = tensor("x_67_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_67_end_mask_0 = const()[name = tensor("x_67_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_67_squeeze_mask_0 = const()[name = tensor("x_67_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_67_cast_fp16 = slice_by_index(begin = x_67_begin_0, end = x_67_end_0, end_mask = x_67_end_mask_0, squeeze_mask = x_67_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("x_67_cast_fp16")]; + tensor v_33_begin_0 = const()[name = tensor("v_33_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_33_end_0 = const()[name = tensor("v_33_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_33_end_mask_0 = const()[name = tensor("v_33_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_33_squeeze_mask_0 = const()[name = tensor("v_33_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_33_cast_fp16 = slice_by_index(begin = v_33_begin_0, end = v_33_end_0, end_mask = v_33_end_mask_0, squeeze_mask = v_33_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("v_33_cast_fp16")]; + tensor var_4071 = const()[name = tensor("op_4071"), val = tensor([256, 16, 32, 2])]; + tensor pairs_65_cast_fp16 = reshape(shape = var_4071, x = x_65_cast_fp16)[name = tensor("pairs_65_cast_fp16")]; + tensor xr_65_begin_0 = const()[name = tensor("xr_65_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_65_end_0 = const()[name = tensor("xr_65_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_65_end_mask_0 = const()[name = tensor("xr_65_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_65_squeeze_mask_0 = const()[name = tensor("xr_65_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_65_cast_fp16 = slice_by_index(begin = xr_65_begin_0, end = xr_65_end_0, end_mask = xr_65_end_mask_0, squeeze_mask = xr_65_squeeze_mask_0, x = pairs_65_cast_fp16)[name = tensor("xr_65_cast_fp16")]; + tensor xi_65_begin_0 = const()[name = tensor("xi_65_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_65_end_0 = const()[name = tensor("xi_65_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_65_end_mask_0 = const()[name = tensor("xi_65_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_65_squeeze_mask_0 = const()[name = tensor("xi_65_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_65_cast_fp16 = slice_by_index(begin = xi_65_begin_0, end = xi_65_end_0, end_mask = xi_65_end_mask_0, squeeze_mask = xi_65_squeeze_mask_0, x = pairs_65_cast_fp16)[name = tensor("xi_65_cast_fp16")]; + tensor var_4079_cast_fp16 = mul(x = xr_65_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4079_cast_fp16")]; + tensor var_4080_cast_fp16 = mul(x = xi_65_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4080_cast_fp16")]; + tensor xor__65_cast_fp16 = sub(x = var_4079_cast_fp16, y = var_4080_cast_fp16)[name = tensor("xor__65_cast_fp16")]; + tensor var_4083_cast_fp16 = mul(x = xr_65_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4083_cast_fp16")]; + tensor var_4084_cast_fp16 = mul(x = xi_65_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4084_cast_fp16")]; + tensor xoi_65_cast_fp16 = add(x = var_4083_cast_fp16, y = var_4084_cast_fp16)[name = tensor("xoi_65_cast_fp16")]; + tensor out_65_axis_0 = const()[name = tensor("out_65_axis_0"), val = tensor(-1)]; + tensor out_65_cast_fp16 = stack(axis = out_65_axis_0, values = (xor__65_cast_fp16, xoi_65_cast_fp16))[name = tensor("out_65_cast_fp16")]; + tensor var_4094 = const()[name = tensor("op_4094"), val = tensor([1, 256, 16, 64])]; + tensor q_33_cast_fp16 = reshape(shape = var_4094, x = out_65_cast_fp16)[name = tensor("q_33_cast_fp16")]; + tensor var_4100 = const()[name = tensor("op_4100"), val = tensor([256, 16, 32, 2])]; + tensor pairs_67_cast_fp16 = reshape(shape = var_4100, x = x_67_cast_fp16)[name = tensor("pairs_67_cast_fp16")]; + tensor xr_67_begin_0 = const()[name = tensor("xr_67_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_67_end_0 = const()[name = tensor("xr_67_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_67_end_mask_0 = const()[name = tensor("xr_67_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_67_squeeze_mask_0 = const()[name = tensor("xr_67_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_67_cast_fp16 = slice_by_index(begin = xr_67_begin_0, end = xr_67_end_0, end_mask = xr_67_end_mask_0, squeeze_mask = xr_67_squeeze_mask_0, x = pairs_67_cast_fp16)[name = tensor("xr_67_cast_fp16")]; + tensor xi_67_begin_0 = const()[name = tensor("xi_67_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_67_end_0 = const()[name = tensor("xi_67_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_67_end_mask_0 = const()[name = tensor("xi_67_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_67_squeeze_mask_0 = const()[name = tensor("xi_67_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_67_cast_fp16 = slice_by_index(begin = xi_67_begin_0, end = xi_67_end_0, end_mask = xi_67_end_mask_0, squeeze_mask = xi_67_squeeze_mask_0, x = pairs_67_cast_fp16)[name = tensor("xi_67_cast_fp16")]; + tensor var_4108_cast_fp16 = mul(x = xr_67_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4108_cast_fp16")]; + tensor var_4109_cast_fp16 = mul(x = xi_67_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4109_cast_fp16")]; + tensor xor__67_cast_fp16 = sub(x = var_4108_cast_fp16, y = var_4109_cast_fp16)[name = tensor("xor__67_cast_fp16")]; + tensor var_4112_cast_fp16 = mul(x = xr_67_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4112_cast_fp16")]; + tensor var_4113_cast_fp16 = mul(x = xi_67_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4113_cast_fp16")]; + tensor xoi_67_cast_fp16 = add(x = var_4112_cast_fp16, y = var_4113_cast_fp16)[name = tensor("xoi_67_cast_fp16")]; + tensor out_67_axis_0 = const()[name = tensor("out_67_axis_0"), val = tensor(-1)]; + tensor out_67_cast_fp16 = stack(axis = out_67_axis_0, values = (xor__67_cast_fp16, xoi_67_cast_fp16))[name = tensor("out_67_cast_fp16")]; + tensor var_4130 = const()[name = tensor("op_4130"), val = tensor([256, 1024])]; + tensor var_4131_cast_fp16 = reshape(shape = var_4130, x = out_67_cast_fp16)[name = tensor("op_4131_cast_fp16")]; + tensor written_k_33_transpose_x_1 = const()[name = tensor("written_k_33_transpose_x_1"), val = tensor(true)]; + tensor written_k_33_transpose_y_1 = const()[name = tensor("written_k_33_transpose_y_1"), val = tensor(false)]; + tensor written_k_33_cast_fp16 = matmul(transpose_x = written_k_33_transpose_x_1, transpose_y = written_k_33_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_4131_cast_fp16)[name = tensor("written_k_33_cast_fp16")]; + tensor var_4138 = const()[name = tensor("op_4138"), val = tensor([256, 1024])]; + tensor var_4139_cast_fp16 = reshape(shape = var_4138, x = v_33_cast_fp16)[name = tensor("op_4139_cast_fp16")]; + tensor written_v_33_transpose_x_1 = const()[name = tensor("written_v_33_transpose_x_1"), val = tensor(true)]; + tensor written_v_33_transpose_y_1 = const()[name = tensor("written_v_33_transpose_y_1"), val = tensor(false)]; + tensor written_v_33_cast_fp16 = matmul(transpose_x = written_v_33_transpose_x_1, transpose_y = written_v_33_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_4139_cast_fp16)[name = tensor("written_v_33_cast_fp16")]; + tensor k_cache16_to_fp16_dtype_0 = const()[name = tensor("k_cache16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache16_to_fp16 = cast(dtype = k_cache16_to_fp16_dtype_0, x = k_cache16)[name = tensor("cast_369")]; + tensor var_4150_cast_fp16 = mul(x = k_cache16_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_4150_cast_fp16")]; + tensor var_4155 = const()[name = tensor("op_4155"), val = tensor([1, 512, 16, 64])]; + tensor var_4156_cast_fp16 = reshape(shape = var_4155, x = written_k_33_cast_fp16)[name = tensor("op_4156_cast_fp16")]; + tensor new_k_cache_33_cast_fp16 = add(x = var_4150_cast_fp16, y = var_4156_cast_fp16)[name = tensor("new_k_cache_33_cast_fp16")]; + tensor new_k_cache_33_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_33_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache16_to_fp16_dtype_0 = const()[name = tensor("v_cache16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache16_to_fp16 = cast(dtype = v_cache16_to_fp16_dtype_0, x = v_cache16)[name = tensor("cast_367")]; + tensor var_4159_cast_fp16 = mul(x = v_cache16_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_4159_cast_fp16")]; + tensor var_4164 = const()[name = tensor("op_4164"), val = tensor([1, 512, 16, 64])]; + tensor var_4165_cast_fp16 = reshape(shape = var_4164, x = written_v_33_cast_fp16)[name = tensor("op_4165_cast_fp16")]; + tensor new_v_cache_33_cast_fp16 = add(x = var_4159_cast_fp16, y = var_4165_cast_fp16)[name = tensor("new_v_cache_33_cast_fp16")]; + tensor new_v_cache_33_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_33_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_4178 = const()[name = tensor("op_4178"), val = tensor([0, 2, 1, 3])]; + tensor var_4189_transpose_x_0 = const()[name = tensor("op_4189_transpose_x_0"), val = tensor(false)]; + tensor var_4189_transpose_y_0 = const()[name = tensor("op_4189_transpose_y_0"), val = tensor(false)]; + tensor transpose_101_perm_0 = const()[name = tensor("transpose_101_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_102_perm_0 = const()[name = tensor("transpose_102_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_102 = transpose(perm = transpose_102_perm_0, x = new_k_cache_33_cast_fp16)[name = tensor("transpose_140")]; + tensor transpose_101 = transpose(perm = transpose_101_perm_0, x = q_33_cast_fp16)[name = tensor("transpose_141")]; + tensor var_4189_cast_fp16 = matmul(transpose_x = var_4189_transpose_x_0, transpose_y = var_4189_transpose_y_0, x = transpose_101, y = transpose_102)[name = tensor("op_4189_cast_fp16")]; + tensor var_4190_to_fp16 = const()[name = tensor("op_4190_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_65_cast_fp16 = mul(x = var_4189_cast_fp16, y = var_4190_to_fp16)[name = tensor("scores_65_cast_fp16")]; + tensor scores_67_cast_fp16 = add(x = scores_65_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_67_cast_fp16")]; + tensor var_4200 = const()[name = tensor("op_4200"), val = tensor(-1)]; + tensor weights_33_cast_fp16 = softmax(axis = var_4200, x = scores_67_cast_fp16)[name = tensor("weights_33_cast_fp16")]; + tensor attn_33_transpose_x_0 = const()[name = tensor("attn_33_transpose_x_0"), val = tensor(false)]; + tensor attn_33_transpose_y_0 = const()[name = tensor("attn_33_transpose_y_0"), val = tensor(false)]; + tensor values_33_cast_fp16 = transpose(perm = var_4178, x = new_v_cache_33_cast_fp16)[name = tensor("transpose_142")]; + tensor attn_33_cast_fp16 = matmul(transpose_x = attn_33_transpose_x_0, transpose_y = attn_33_transpose_y_0, x = weights_33_cast_fp16, y = values_33_cast_fp16)[name = tensor("attn_33_cast_fp16")]; + tensor var_4208 = const()[name = tensor("op_4208"), val = tensor([0, 2, 1, 3])]; + tensor var_4213 = const()[name = tensor("op_4213"), val = tensor([1, 256, 1024])]; + tensor var_4209_cast_fp16 = transpose(perm = var_4208, x = attn_33_cast_fp16)[name = tensor("transpose_139")]; + tensor input_197_cast_fp16 = reshape(shape = var_4213, x = var_4209_cast_fp16)[name = tensor("input_197_cast_fp16")]; + tensor attn16_out_proj_weight_to_fp16 = const()[name = tensor("attn16_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(409107200)))]; + tensor linear_65_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn16_out_proj_weight_to_fp16, x = input_197_cast_fp16)[name = tensor("linear_65_cast_fp16")]; + tensor position16_to_fp16_dtype_0 = const()[name = tensor("position16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position16_to_fp16 = cast(dtype = position16_to_fp16_dtype_0, x = position16)[name = tensor("cast_365")]; + tensor var_4222_cast_fp16 = add(x = position16_to_fp16, y = valid_len_to_fp16)[name = tensor("op_4222_cast_fp16")]; + tensor var_4222_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4222_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_199_cast_fp16 = add(x = input_193_cast_fp16, y = linear_65_cast_fp16)[name = tensor("input_199_cast_fp16")]; + tensor input_201_axes_0 = const()[name = tensor("input_201_axes_0"), val = tensor([-1])]; + tensor norm16_2_weight_to_fp16 = const()[name = tensor("norm16_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411204416)))]; + tensor norm16_2_bias_to_fp16 = const()[name = tensor("norm16_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411206528)))]; + tensor var_4226_to_fp16 = const()[name = tensor("op_4226_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_201_cast_fp16 = layer_norm(axes = input_201_axes_0, beta = norm16_2_bias_to_fp16, epsilon = var_4226_to_fp16, gamma = norm16_2_weight_to_fp16, x = input_199_cast_fp16)[name = tensor("input_201_cast_fp16")]; + tensor linear16_1_weight_to_fp16 = const()[name = tensor("linear16_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411208640)))]; + tensor linear_66_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear16_1_weight_to_fp16, x = input_201_cast_fp16)[name = tensor("linear_66_cast_fp16")]; + tensor input_203_mode_0 = const()[name = tensor("input_203_mode_0"), val = tensor("EXACT")]; + tensor input_203_cast_fp16 = gelu(mode = input_203_mode_0, x = linear_66_cast_fp16)[name = tensor("input_203_cast_fp16")]; + tensor linear16_2_weight_to_fp16 = const()[name = tensor("linear16_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(419597312)))]; + tensor linear_67_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear16_2_weight_to_fp16, x = input_203_cast_fp16)[name = tensor("linear_67_cast_fp16")]; + tensor input_205_cast_fp16 = add(x = input_199_cast_fp16, y = linear_67_cast_fp16)[name = tensor("input_205_cast_fp16")]; + tensor input_207_axes_0 = const()[name = tensor("input_207_axes_0"), val = tensor([-1])]; + tensor norm17_1_weight_to_fp16 = const()[name = tensor("norm17_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(427985984)))]; + tensor norm17_1_bias_to_fp16 = const()[name = tensor("norm17_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(427988096)))]; + tensor var_4243_to_fp16 = const()[name = tensor("op_4243_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_207_cast_fp16 = layer_norm(axes = input_207_axes_0, beta = norm17_1_bias_to_fp16, epsilon = var_4243_to_fp16, gamma = norm17_1_weight_to_fp16, x = input_205_cast_fp16)[name = tensor("input_207_cast_fp16")]; + tensor attn17_in_proj_weight_to_fp16 = const()[name = tensor("attn17_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(427990208)))]; + tensor linear_68_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn17_in_proj_weight_to_fp16, x = input_207_cast_fp16)[name = tensor("linear_68_cast_fp16")]; + tensor var_4257 = const()[name = tensor("op_4257"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_35_cast_fp16 = reshape(shape = var_4257, x = linear_68_cast_fp16)[name = tensor("qkv_35_cast_fp16")]; + tensor x_69_begin_0 = const()[name = tensor("x_69_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_69_end_0 = const()[name = tensor("x_69_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_69_end_mask_0 = const()[name = tensor("x_69_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_69_squeeze_mask_0 = const()[name = tensor("x_69_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_69_cast_fp16 = slice_by_index(begin = x_69_begin_0, end = x_69_end_0, end_mask = x_69_end_mask_0, squeeze_mask = x_69_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("x_69_cast_fp16")]; + tensor x_71_begin_0 = const()[name = tensor("x_71_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_71_end_0 = const()[name = tensor("x_71_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_71_end_mask_0 = const()[name = tensor("x_71_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_71_squeeze_mask_0 = const()[name = tensor("x_71_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_71_cast_fp16 = slice_by_index(begin = x_71_begin_0, end = x_71_end_0, end_mask = x_71_end_mask_0, squeeze_mask = x_71_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("x_71_cast_fp16")]; + tensor v_35_begin_0 = const()[name = tensor("v_35_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_35_end_0 = const()[name = tensor("v_35_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_35_end_mask_0 = const()[name = tensor("v_35_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_35_squeeze_mask_0 = const()[name = tensor("v_35_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_35_cast_fp16 = slice_by_index(begin = v_35_begin_0, end = v_35_end_0, end_mask = v_35_end_mask_0, squeeze_mask = v_35_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("v_35_cast_fp16")]; + tensor var_4302 = const()[name = tensor("op_4302"), val = tensor([256, 16, 32, 2])]; + tensor pairs_69_cast_fp16 = reshape(shape = var_4302, x = x_69_cast_fp16)[name = tensor("pairs_69_cast_fp16")]; + tensor xr_69_begin_0 = const()[name = tensor("xr_69_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_69_end_0 = const()[name = tensor("xr_69_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_69_end_mask_0 = const()[name = tensor("xr_69_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_69_squeeze_mask_0 = const()[name = tensor("xr_69_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_69_cast_fp16 = slice_by_index(begin = xr_69_begin_0, end = xr_69_end_0, end_mask = xr_69_end_mask_0, squeeze_mask = xr_69_squeeze_mask_0, x = pairs_69_cast_fp16)[name = tensor("xr_69_cast_fp16")]; + tensor xi_69_begin_0 = const()[name = tensor("xi_69_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_69_end_0 = const()[name = tensor("xi_69_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_69_end_mask_0 = const()[name = tensor("xi_69_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_69_squeeze_mask_0 = const()[name = tensor("xi_69_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_69_cast_fp16 = slice_by_index(begin = xi_69_begin_0, end = xi_69_end_0, end_mask = xi_69_end_mask_0, squeeze_mask = xi_69_squeeze_mask_0, x = pairs_69_cast_fp16)[name = tensor("xi_69_cast_fp16")]; + tensor var_4310_cast_fp16 = mul(x = xr_69_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4310_cast_fp16")]; + tensor var_4311_cast_fp16 = mul(x = xi_69_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4311_cast_fp16")]; + tensor xor__69_cast_fp16 = sub(x = var_4310_cast_fp16, y = var_4311_cast_fp16)[name = tensor("xor__69_cast_fp16")]; + tensor var_4314_cast_fp16 = mul(x = xr_69_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4314_cast_fp16")]; + tensor var_4315_cast_fp16 = mul(x = xi_69_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4315_cast_fp16")]; + tensor xoi_69_cast_fp16 = add(x = var_4314_cast_fp16, y = var_4315_cast_fp16)[name = tensor("xoi_69_cast_fp16")]; + tensor out_69_axis_0 = const()[name = tensor("out_69_axis_0"), val = tensor(-1)]; + tensor out_69_cast_fp16 = stack(axis = out_69_axis_0, values = (xor__69_cast_fp16, xoi_69_cast_fp16))[name = tensor("out_69_cast_fp16")]; + tensor var_4325 = const()[name = tensor("op_4325"), val = tensor([1, 256, 16, 64])]; + tensor q_35_cast_fp16 = reshape(shape = var_4325, x = out_69_cast_fp16)[name = tensor("q_35_cast_fp16")]; + tensor var_4331 = const()[name = tensor("op_4331"), val = tensor([256, 16, 32, 2])]; + tensor pairs_71_cast_fp16 = reshape(shape = var_4331, x = x_71_cast_fp16)[name = tensor("pairs_71_cast_fp16")]; + tensor xr_71_begin_0 = const()[name = tensor("xr_71_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_71_end_0 = const()[name = tensor("xr_71_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_71_end_mask_0 = const()[name = tensor("xr_71_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_71_squeeze_mask_0 = const()[name = tensor("xr_71_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_71_cast_fp16 = slice_by_index(begin = xr_71_begin_0, end = xr_71_end_0, end_mask = xr_71_end_mask_0, squeeze_mask = xr_71_squeeze_mask_0, x = pairs_71_cast_fp16)[name = tensor("xr_71_cast_fp16")]; + tensor xi_71_begin_0 = const()[name = tensor("xi_71_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_71_end_0 = const()[name = tensor("xi_71_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_71_end_mask_0 = const()[name = tensor("xi_71_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_71_squeeze_mask_0 = const()[name = tensor("xi_71_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_71_cast_fp16 = slice_by_index(begin = xi_71_begin_0, end = xi_71_end_0, end_mask = xi_71_end_mask_0, squeeze_mask = xi_71_squeeze_mask_0, x = pairs_71_cast_fp16)[name = tensor("xi_71_cast_fp16")]; + tensor var_4339_cast_fp16 = mul(x = xr_71_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4339_cast_fp16")]; + tensor var_4340_cast_fp16 = mul(x = xi_71_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4340_cast_fp16")]; + tensor xor__71_cast_fp16 = sub(x = var_4339_cast_fp16, y = var_4340_cast_fp16)[name = tensor("xor__71_cast_fp16")]; + tensor var_4343_cast_fp16 = mul(x = xr_71_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4343_cast_fp16")]; + tensor var_4344_cast_fp16 = mul(x = xi_71_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4344_cast_fp16")]; + tensor xoi_71_cast_fp16 = add(x = var_4343_cast_fp16, y = var_4344_cast_fp16)[name = tensor("xoi_71_cast_fp16")]; + tensor out_71_axis_0 = const()[name = tensor("out_71_axis_0"), val = tensor(-1)]; + tensor out_71_cast_fp16 = stack(axis = out_71_axis_0, values = (xor__71_cast_fp16, xoi_71_cast_fp16))[name = tensor("out_71_cast_fp16")]; + tensor var_4361 = const()[name = tensor("op_4361"), val = tensor([256, 1024])]; + tensor var_4362_cast_fp16 = reshape(shape = var_4361, x = out_71_cast_fp16)[name = tensor("op_4362_cast_fp16")]; + tensor written_k_35_transpose_x_1 = const()[name = tensor("written_k_35_transpose_x_1"), val = tensor(true)]; + tensor written_k_35_transpose_y_1 = const()[name = tensor("written_k_35_transpose_y_1"), val = tensor(false)]; + tensor written_k_35_cast_fp16 = matmul(transpose_x = written_k_35_transpose_x_1, transpose_y = written_k_35_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_4362_cast_fp16)[name = tensor("written_k_35_cast_fp16")]; + tensor var_4369 = const()[name = tensor("op_4369"), val = tensor([256, 1024])]; + tensor var_4370_cast_fp16 = reshape(shape = var_4369, x = v_35_cast_fp16)[name = tensor("op_4370_cast_fp16")]; + tensor written_v_35_transpose_x_1 = const()[name = tensor("written_v_35_transpose_x_1"), val = tensor(true)]; + tensor written_v_35_transpose_y_1 = const()[name = tensor("written_v_35_transpose_y_1"), val = tensor(false)]; + tensor written_v_35_cast_fp16 = matmul(transpose_x = written_v_35_transpose_x_1, transpose_y = written_v_35_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_4370_cast_fp16)[name = tensor("written_v_35_cast_fp16")]; + tensor k_cache17_to_fp16_dtype_0 = const()[name = tensor("k_cache17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache17_to_fp16 = cast(dtype = k_cache17_to_fp16_dtype_0, x = k_cache17)[name = tensor("cast_363")]; + tensor var_4381_cast_fp16 = mul(x = k_cache17_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_4381_cast_fp16")]; + tensor var_4386 = const()[name = tensor("op_4386"), val = tensor([1, 512, 16, 64])]; + tensor var_4387_cast_fp16 = reshape(shape = var_4386, x = written_k_35_cast_fp16)[name = tensor("op_4387_cast_fp16")]; + tensor new_k_cache_35_cast_fp16 = add(x = var_4381_cast_fp16, y = var_4387_cast_fp16)[name = tensor("new_k_cache_35_cast_fp16")]; + tensor new_k_cache_35_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_35_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache17_to_fp16_dtype_0 = const()[name = tensor("v_cache17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache17_to_fp16 = cast(dtype = v_cache17_to_fp16_dtype_0, x = v_cache17)[name = tensor("cast_361")]; + tensor var_4390_cast_fp16 = mul(x = v_cache17_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_4390_cast_fp16")]; + tensor var_4395 = const()[name = tensor("op_4395"), val = tensor([1, 512, 16, 64])]; + tensor var_4396_cast_fp16 = reshape(shape = var_4395, x = written_v_35_cast_fp16)[name = tensor("op_4396_cast_fp16")]; + tensor new_v_cache_35_cast_fp16 = add(x = var_4390_cast_fp16, y = var_4396_cast_fp16)[name = tensor("new_v_cache_35_cast_fp16")]; + tensor new_v_cache_35_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_35_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_4409 = const()[name = tensor("op_4409"), val = tensor([0, 2, 1, 3])]; + tensor var_4420_transpose_x_0 = const()[name = tensor("op_4420_transpose_x_0"), val = tensor(false)]; + tensor var_4420_transpose_y_0 = const()[name = tensor("op_4420_transpose_y_0"), val = tensor(false)]; + tensor transpose_103_perm_0 = const()[name = tensor("transpose_103_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_104_perm_0 = const()[name = tensor("transpose_104_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_104 = transpose(perm = transpose_104_perm_0, x = new_k_cache_35_cast_fp16)[name = tensor("transpose_136")]; + tensor transpose_103 = transpose(perm = transpose_103_perm_0, x = q_35_cast_fp16)[name = tensor("transpose_137")]; + tensor var_4420_cast_fp16 = matmul(transpose_x = var_4420_transpose_x_0, transpose_y = var_4420_transpose_y_0, x = transpose_103, y = transpose_104)[name = tensor("op_4420_cast_fp16")]; + tensor var_4421_to_fp16 = const()[name = tensor("op_4421_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_69_cast_fp16 = mul(x = var_4420_cast_fp16, y = var_4421_to_fp16)[name = tensor("scores_69_cast_fp16")]; + tensor scores_71_cast_fp16 = add(x = scores_69_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_71_cast_fp16")]; + tensor var_4431 = const()[name = tensor("op_4431"), val = tensor(-1)]; + tensor weights_35_cast_fp16 = softmax(axis = var_4431, x = scores_71_cast_fp16)[name = tensor("weights_35_cast_fp16")]; + tensor attn_35_transpose_x_0 = const()[name = tensor("attn_35_transpose_x_0"), val = tensor(false)]; + tensor attn_35_transpose_y_0 = const()[name = tensor("attn_35_transpose_y_0"), val = tensor(false)]; + tensor values_35_cast_fp16 = transpose(perm = var_4409, x = new_v_cache_35_cast_fp16)[name = tensor("transpose_138")]; + tensor attn_35_cast_fp16 = matmul(transpose_x = attn_35_transpose_x_0, transpose_y = attn_35_transpose_y_0, x = weights_35_cast_fp16, y = values_35_cast_fp16)[name = tensor("attn_35_cast_fp16")]; + tensor var_4439 = const()[name = tensor("op_4439"), val = tensor([0, 2, 1, 3])]; + tensor var_4444 = const()[name = tensor("op_4444"), val = tensor([1, 256, 1024])]; + tensor var_4440_cast_fp16 = transpose(perm = var_4439, x = attn_35_cast_fp16)[name = tensor("transpose_135")]; + tensor input_209_cast_fp16 = reshape(shape = var_4444, x = var_4440_cast_fp16)[name = tensor("input_209_cast_fp16")]; + tensor attn17_out_proj_weight_to_fp16 = const()[name = tensor("attn17_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(434281728)))]; + tensor linear_69_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn17_out_proj_weight_to_fp16, x = input_209_cast_fp16)[name = tensor("linear_69_cast_fp16")]; + tensor position17_to_fp16_dtype_0 = const()[name = tensor("position17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position17_to_fp16 = cast(dtype = position17_to_fp16_dtype_0, x = position17)[name = tensor("cast_359")]; + tensor var_4453_cast_fp16 = add(x = position17_to_fp16, y = valid_len_to_fp16)[name = tensor("op_4453_cast_fp16")]; + tensor var_4453_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4453_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_211_cast_fp16 = add(x = input_205_cast_fp16, y = linear_69_cast_fp16)[name = tensor("input_211_cast_fp16")]; + tensor input_213_axes_0 = const()[name = tensor("input_213_axes_0"), val = tensor([-1])]; + tensor norm17_2_weight_to_fp16 = const()[name = tensor("norm17_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(436378944)))]; + tensor norm17_2_bias_to_fp16 = const()[name = tensor("norm17_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(436381056)))]; + tensor var_4457_to_fp16 = const()[name = tensor("op_4457_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_213_cast_fp16 = layer_norm(axes = input_213_axes_0, beta = norm17_2_bias_to_fp16, epsilon = var_4457_to_fp16, gamma = norm17_2_weight_to_fp16, x = input_211_cast_fp16)[name = tensor("input_213_cast_fp16")]; + tensor linear17_1_weight_to_fp16 = const()[name = tensor("linear17_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(436383168)))]; + tensor linear_70_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear17_1_weight_to_fp16, x = input_213_cast_fp16)[name = tensor("linear_70_cast_fp16")]; + tensor input_215_mode_0 = const()[name = tensor("input_215_mode_0"), val = tensor("EXACT")]; + tensor input_215_cast_fp16 = gelu(mode = input_215_mode_0, x = linear_70_cast_fp16)[name = tensor("input_215_cast_fp16")]; + tensor linear17_2_weight_to_fp16 = const()[name = tensor("linear17_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(444771840)))]; + tensor linear_71_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear17_2_weight_to_fp16, x = input_215_cast_fp16)[name = tensor("linear_71_cast_fp16")]; + tensor input_217_cast_fp16 = add(x = input_211_cast_fp16, y = linear_71_cast_fp16)[name = tensor("input_217_cast_fp16")]; + tensor input_219_axes_0 = const()[name = tensor("input_219_axes_0"), val = tensor([-1])]; + tensor norm18_1_weight_to_fp16 = const()[name = tensor("norm18_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(453160512)))]; + tensor norm18_1_bias_to_fp16 = const()[name = tensor("norm18_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(453162624)))]; + tensor var_4474_to_fp16 = const()[name = tensor("op_4474_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_219_cast_fp16 = layer_norm(axes = input_219_axes_0, beta = norm18_1_bias_to_fp16, epsilon = var_4474_to_fp16, gamma = norm18_1_weight_to_fp16, x = input_217_cast_fp16)[name = tensor("input_219_cast_fp16")]; + tensor attn18_in_proj_weight_to_fp16 = const()[name = tensor("attn18_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(453164736)))]; + tensor linear_72_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn18_in_proj_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_72_cast_fp16")]; + tensor var_4488 = const()[name = tensor("op_4488"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_37_cast_fp16 = reshape(shape = var_4488, x = linear_72_cast_fp16)[name = tensor("qkv_37_cast_fp16")]; + tensor x_73_begin_0 = const()[name = tensor("x_73_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_73_end_0 = const()[name = tensor("x_73_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_73_end_mask_0 = const()[name = tensor("x_73_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_73_squeeze_mask_0 = const()[name = tensor("x_73_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_73_cast_fp16 = slice_by_index(begin = x_73_begin_0, end = x_73_end_0, end_mask = x_73_end_mask_0, squeeze_mask = x_73_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("x_73_cast_fp16")]; + tensor x_75_begin_0 = const()[name = tensor("x_75_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_75_end_0 = const()[name = tensor("x_75_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_75_end_mask_0 = const()[name = tensor("x_75_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_75_squeeze_mask_0 = const()[name = tensor("x_75_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_75_cast_fp16 = slice_by_index(begin = x_75_begin_0, end = x_75_end_0, end_mask = x_75_end_mask_0, squeeze_mask = x_75_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("x_75_cast_fp16")]; + tensor v_37_begin_0 = const()[name = tensor("v_37_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_37_end_0 = const()[name = tensor("v_37_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_37_end_mask_0 = const()[name = tensor("v_37_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_37_squeeze_mask_0 = const()[name = tensor("v_37_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_37_cast_fp16 = slice_by_index(begin = v_37_begin_0, end = v_37_end_0, end_mask = v_37_end_mask_0, squeeze_mask = v_37_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("v_37_cast_fp16")]; + tensor var_4533 = const()[name = tensor("op_4533"), val = tensor([256, 16, 32, 2])]; + tensor pairs_73_cast_fp16 = reshape(shape = var_4533, x = x_73_cast_fp16)[name = tensor("pairs_73_cast_fp16")]; + tensor xr_73_begin_0 = const()[name = tensor("xr_73_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_73_end_0 = const()[name = tensor("xr_73_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_73_end_mask_0 = const()[name = tensor("xr_73_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_73_squeeze_mask_0 = const()[name = tensor("xr_73_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_73_cast_fp16 = slice_by_index(begin = xr_73_begin_0, end = xr_73_end_0, end_mask = xr_73_end_mask_0, squeeze_mask = xr_73_squeeze_mask_0, x = pairs_73_cast_fp16)[name = tensor("xr_73_cast_fp16")]; + tensor xi_73_begin_0 = const()[name = tensor("xi_73_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_73_end_0 = const()[name = tensor("xi_73_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_73_end_mask_0 = const()[name = tensor("xi_73_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_73_squeeze_mask_0 = const()[name = tensor("xi_73_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_73_cast_fp16 = slice_by_index(begin = xi_73_begin_0, end = xi_73_end_0, end_mask = xi_73_end_mask_0, squeeze_mask = xi_73_squeeze_mask_0, x = pairs_73_cast_fp16)[name = tensor("xi_73_cast_fp16")]; + tensor var_4541_cast_fp16 = mul(x = xr_73_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4541_cast_fp16")]; + tensor var_4542_cast_fp16 = mul(x = xi_73_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4542_cast_fp16")]; + tensor xor__73_cast_fp16 = sub(x = var_4541_cast_fp16, y = var_4542_cast_fp16)[name = tensor("xor__73_cast_fp16")]; + tensor var_4545_cast_fp16 = mul(x = xr_73_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4545_cast_fp16")]; + tensor var_4546_cast_fp16 = mul(x = xi_73_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4546_cast_fp16")]; + tensor xoi_73_cast_fp16 = add(x = var_4545_cast_fp16, y = var_4546_cast_fp16)[name = tensor("xoi_73_cast_fp16")]; + tensor out_73_axis_0 = const()[name = tensor("out_73_axis_0"), val = tensor(-1)]; + tensor out_73_cast_fp16 = stack(axis = out_73_axis_0, values = (xor__73_cast_fp16, xoi_73_cast_fp16))[name = tensor("out_73_cast_fp16")]; + tensor var_4556 = const()[name = tensor("op_4556"), val = tensor([1, 256, 16, 64])]; + tensor q_37_cast_fp16 = reshape(shape = var_4556, x = out_73_cast_fp16)[name = tensor("q_37_cast_fp16")]; + tensor var_4562 = const()[name = tensor("op_4562"), val = tensor([256, 16, 32, 2])]; + tensor pairs_75_cast_fp16 = reshape(shape = var_4562, x = x_75_cast_fp16)[name = tensor("pairs_75_cast_fp16")]; + tensor xr_75_begin_0 = const()[name = tensor("xr_75_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_75_end_0 = const()[name = tensor("xr_75_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_75_end_mask_0 = const()[name = tensor("xr_75_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_75_squeeze_mask_0 = const()[name = tensor("xr_75_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_75_cast_fp16 = slice_by_index(begin = xr_75_begin_0, end = xr_75_end_0, end_mask = xr_75_end_mask_0, squeeze_mask = xr_75_squeeze_mask_0, x = pairs_75_cast_fp16)[name = tensor("xr_75_cast_fp16")]; + tensor xi_75_begin_0 = const()[name = tensor("xi_75_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_75_end_0 = const()[name = tensor("xi_75_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_75_end_mask_0 = const()[name = tensor("xi_75_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_75_squeeze_mask_0 = const()[name = tensor("xi_75_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_75_cast_fp16 = slice_by_index(begin = xi_75_begin_0, end = xi_75_end_0, end_mask = xi_75_end_mask_0, squeeze_mask = xi_75_squeeze_mask_0, x = pairs_75_cast_fp16)[name = tensor("xi_75_cast_fp16")]; + tensor var_4570_cast_fp16 = mul(x = xr_75_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4570_cast_fp16")]; + tensor var_4571_cast_fp16 = mul(x = xi_75_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4571_cast_fp16")]; + tensor xor__75_cast_fp16 = sub(x = var_4570_cast_fp16, y = var_4571_cast_fp16)[name = tensor("xor__75_cast_fp16")]; + tensor var_4574_cast_fp16 = mul(x = xr_75_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4574_cast_fp16")]; + tensor var_4575_cast_fp16 = mul(x = xi_75_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4575_cast_fp16")]; + tensor xoi_75_cast_fp16 = add(x = var_4574_cast_fp16, y = var_4575_cast_fp16)[name = tensor("xoi_75_cast_fp16")]; + tensor out_75_axis_0 = const()[name = tensor("out_75_axis_0"), val = tensor(-1)]; + tensor out_75_cast_fp16 = stack(axis = out_75_axis_0, values = (xor__75_cast_fp16, xoi_75_cast_fp16))[name = tensor("out_75_cast_fp16")]; + tensor var_4592 = const()[name = tensor("op_4592"), val = tensor([256, 1024])]; + tensor var_4593_cast_fp16 = reshape(shape = var_4592, x = out_75_cast_fp16)[name = tensor("op_4593_cast_fp16")]; + tensor written_k_37_transpose_x_1 = const()[name = tensor("written_k_37_transpose_x_1"), val = tensor(true)]; + tensor written_k_37_transpose_y_1 = const()[name = tensor("written_k_37_transpose_y_1"), val = tensor(false)]; + tensor written_k_37_cast_fp16 = matmul(transpose_x = written_k_37_transpose_x_1, transpose_y = written_k_37_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_4593_cast_fp16)[name = tensor("written_k_37_cast_fp16")]; + tensor var_4600 = const()[name = tensor("op_4600"), val = tensor([256, 1024])]; + tensor var_4601_cast_fp16 = reshape(shape = var_4600, x = v_37_cast_fp16)[name = tensor("op_4601_cast_fp16")]; + tensor written_v_37_transpose_x_1 = const()[name = tensor("written_v_37_transpose_x_1"), val = tensor(true)]; + tensor written_v_37_transpose_y_1 = const()[name = tensor("written_v_37_transpose_y_1"), val = tensor(false)]; + tensor written_v_37_cast_fp16 = matmul(transpose_x = written_v_37_transpose_x_1, transpose_y = written_v_37_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_4601_cast_fp16)[name = tensor("written_v_37_cast_fp16")]; + tensor k_cache18_to_fp16_dtype_0 = const()[name = tensor("k_cache18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache18_to_fp16 = cast(dtype = k_cache18_to_fp16_dtype_0, x = k_cache18)[name = tensor("cast_357")]; + tensor var_4612_cast_fp16 = mul(x = k_cache18_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_4612_cast_fp16")]; + tensor var_4617 = const()[name = tensor("op_4617"), val = tensor([1, 512, 16, 64])]; + tensor var_4618_cast_fp16 = reshape(shape = var_4617, x = written_k_37_cast_fp16)[name = tensor("op_4618_cast_fp16")]; + tensor new_k_cache_37_cast_fp16 = add(x = var_4612_cast_fp16, y = var_4618_cast_fp16)[name = tensor("new_k_cache_37_cast_fp16")]; + tensor new_k_cache_37_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_37_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache18_to_fp16_dtype_0 = const()[name = tensor("v_cache18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache18_to_fp16 = cast(dtype = v_cache18_to_fp16_dtype_0, x = v_cache18)[name = tensor("cast_355")]; + tensor var_4621_cast_fp16 = mul(x = v_cache18_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_4621_cast_fp16")]; + tensor var_4626 = const()[name = tensor("op_4626"), val = tensor([1, 512, 16, 64])]; + tensor var_4627_cast_fp16 = reshape(shape = var_4626, x = written_v_37_cast_fp16)[name = tensor("op_4627_cast_fp16")]; + tensor new_v_cache_37_cast_fp16 = add(x = var_4621_cast_fp16, y = var_4627_cast_fp16)[name = tensor("new_v_cache_37_cast_fp16")]; + tensor new_v_cache_37_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_37_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_4640 = const()[name = tensor("op_4640"), val = tensor([0, 2, 1, 3])]; + tensor var_4651_transpose_x_0 = const()[name = tensor("op_4651_transpose_x_0"), val = tensor(false)]; + tensor var_4651_transpose_y_0 = const()[name = tensor("op_4651_transpose_y_0"), val = tensor(false)]; + tensor transpose_105_perm_0 = const()[name = tensor("transpose_105_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_106_perm_0 = const()[name = tensor("transpose_106_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_106 = transpose(perm = transpose_106_perm_0, x = new_k_cache_37_cast_fp16)[name = tensor("transpose_132")]; + tensor transpose_105 = transpose(perm = transpose_105_perm_0, x = q_37_cast_fp16)[name = tensor("transpose_133")]; + tensor var_4651_cast_fp16 = matmul(transpose_x = var_4651_transpose_x_0, transpose_y = var_4651_transpose_y_0, x = transpose_105, y = transpose_106)[name = tensor("op_4651_cast_fp16")]; + tensor var_4652_to_fp16 = const()[name = tensor("op_4652_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_73_cast_fp16 = mul(x = var_4651_cast_fp16, y = var_4652_to_fp16)[name = tensor("scores_73_cast_fp16")]; + tensor scores_75_cast_fp16 = add(x = scores_73_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_75_cast_fp16")]; + tensor var_4662 = const()[name = tensor("op_4662"), val = tensor(-1)]; + tensor weights_37_cast_fp16 = softmax(axis = var_4662, x = scores_75_cast_fp16)[name = tensor("weights_37_cast_fp16")]; + tensor attn_37_transpose_x_0 = const()[name = tensor("attn_37_transpose_x_0"), val = tensor(false)]; + tensor attn_37_transpose_y_0 = const()[name = tensor("attn_37_transpose_y_0"), val = tensor(false)]; + tensor values_37_cast_fp16 = transpose(perm = var_4640, x = new_v_cache_37_cast_fp16)[name = tensor("transpose_134")]; + tensor attn_37_cast_fp16 = matmul(transpose_x = attn_37_transpose_x_0, transpose_y = attn_37_transpose_y_0, x = weights_37_cast_fp16, y = values_37_cast_fp16)[name = tensor("attn_37_cast_fp16")]; + tensor var_4670 = const()[name = tensor("op_4670"), val = tensor([0, 2, 1, 3])]; + tensor var_4675 = const()[name = tensor("op_4675"), val = tensor([1, 256, 1024])]; + tensor var_4671_cast_fp16 = transpose(perm = var_4670, x = attn_37_cast_fp16)[name = tensor("transpose_131")]; + tensor input_221_cast_fp16 = reshape(shape = var_4675, x = var_4671_cast_fp16)[name = tensor("input_221_cast_fp16")]; + tensor attn18_out_proj_weight_to_fp16 = const()[name = tensor("attn18_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(459456256)))]; + tensor linear_73_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn18_out_proj_weight_to_fp16, x = input_221_cast_fp16)[name = tensor("linear_73_cast_fp16")]; + tensor position18_to_fp16_dtype_0 = const()[name = tensor("position18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position18_to_fp16 = cast(dtype = position18_to_fp16_dtype_0, x = position18)[name = tensor("cast_353")]; + tensor var_4684_cast_fp16 = add(x = position18_to_fp16, y = valid_len_to_fp16)[name = tensor("op_4684_cast_fp16")]; + tensor var_4684_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4684_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_223_cast_fp16 = add(x = input_217_cast_fp16, y = linear_73_cast_fp16)[name = tensor("input_223_cast_fp16")]; + tensor input_225_axes_0 = const()[name = tensor("input_225_axes_0"), val = tensor([-1])]; + tensor norm18_2_weight_to_fp16 = const()[name = tensor("norm18_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(461553472)))]; + tensor norm18_2_bias_to_fp16 = const()[name = tensor("norm18_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(461555584)))]; + tensor var_4688_to_fp16 = const()[name = tensor("op_4688_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_225_cast_fp16 = layer_norm(axes = input_225_axes_0, beta = norm18_2_bias_to_fp16, epsilon = var_4688_to_fp16, gamma = norm18_2_weight_to_fp16, x = input_223_cast_fp16)[name = tensor("input_225_cast_fp16")]; + tensor linear18_1_weight_to_fp16 = const()[name = tensor("linear18_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(461557696)))]; + tensor linear_74_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear18_1_weight_to_fp16, x = input_225_cast_fp16)[name = tensor("linear_74_cast_fp16")]; + tensor input_227_mode_0 = const()[name = tensor("input_227_mode_0"), val = tensor("EXACT")]; + tensor input_227_cast_fp16 = gelu(mode = input_227_mode_0, x = linear_74_cast_fp16)[name = tensor("input_227_cast_fp16")]; + tensor linear18_2_weight_to_fp16 = const()[name = tensor("linear18_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(469946368)))]; + tensor linear_75_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear18_2_weight_to_fp16, x = input_227_cast_fp16)[name = tensor("linear_75_cast_fp16")]; + tensor input_229_cast_fp16 = add(x = input_223_cast_fp16, y = linear_75_cast_fp16)[name = tensor("input_229_cast_fp16")]; + tensor input_231_axes_0 = const()[name = tensor("input_231_axes_0"), val = tensor([-1])]; + tensor norm19_1_weight_to_fp16 = const()[name = tensor("norm19_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(478335040)))]; + tensor norm19_1_bias_to_fp16 = const()[name = tensor("norm19_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(478337152)))]; + tensor var_4705_to_fp16 = const()[name = tensor("op_4705_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_231_cast_fp16 = layer_norm(axes = input_231_axes_0, beta = norm19_1_bias_to_fp16, epsilon = var_4705_to_fp16, gamma = norm19_1_weight_to_fp16, x = input_229_cast_fp16)[name = tensor("input_231_cast_fp16")]; + tensor attn19_in_proj_weight_to_fp16 = const()[name = tensor("attn19_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(478339264)))]; + tensor linear_76_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn19_in_proj_weight_to_fp16, x = input_231_cast_fp16)[name = tensor("linear_76_cast_fp16")]; + tensor var_4719 = const()[name = tensor("op_4719"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_39_cast_fp16 = reshape(shape = var_4719, x = linear_76_cast_fp16)[name = tensor("qkv_39_cast_fp16")]; + tensor x_77_begin_0 = const()[name = tensor("x_77_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_77_end_0 = const()[name = tensor("x_77_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_77_end_mask_0 = const()[name = tensor("x_77_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_77_squeeze_mask_0 = const()[name = tensor("x_77_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_77_cast_fp16 = slice_by_index(begin = x_77_begin_0, end = x_77_end_0, end_mask = x_77_end_mask_0, squeeze_mask = x_77_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("x_77_cast_fp16")]; + tensor x_79_begin_0 = const()[name = tensor("x_79_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_79_end_0 = const()[name = tensor("x_79_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_79_end_mask_0 = const()[name = tensor("x_79_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_79_squeeze_mask_0 = const()[name = tensor("x_79_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_79_cast_fp16 = slice_by_index(begin = x_79_begin_0, end = x_79_end_0, end_mask = x_79_end_mask_0, squeeze_mask = x_79_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("x_79_cast_fp16")]; + tensor v_39_begin_0 = const()[name = tensor("v_39_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_39_end_0 = const()[name = tensor("v_39_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_39_end_mask_0 = const()[name = tensor("v_39_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_39_squeeze_mask_0 = const()[name = tensor("v_39_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_39_cast_fp16 = slice_by_index(begin = v_39_begin_0, end = v_39_end_0, end_mask = v_39_end_mask_0, squeeze_mask = v_39_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("v_39_cast_fp16")]; + tensor var_4764 = const()[name = tensor("op_4764"), val = tensor([256, 16, 32, 2])]; + tensor pairs_77_cast_fp16 = reshape(shape = var_4764, x = x_77_cast_fp16)[name = tensor("pairs_77_cast_fp16")]; + tensor xr_77_begin_0 = const()[name = tensor("xr_77_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_77_end_0 = const()[name = tensor("xr_77_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_77_end_mask_0 = const()[name = tensor("xr_77_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_77_squeeze_mask_0 = const()[name = tensor("xr_77_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_77_cast_fp16 = slice_by_index(begin = xr_77_begin_0, end = xr_77_end_0, end_mask = xr_77_end_mask_0, squeeze_mask = xr_77_squeeze_mask_0, x = pairs_77_cast_fp16)[name = tensor("xr_77_cast_fp16")]; + tensor xi_77_begin_0 = const()[name = tensor("xi_77_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_77_end_0 = const()[name = tensor("xi_77_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_77_end_mask_0 = const()[name = tensor("xi_77_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_77_squeeze_mask_0 = const()[name = tensor("xi_77_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_77_cast_fp16 = slice_by_index(begin = xi_77_begin_0, end = xi_77_end_0, end_mask = xi_77_end_mask_0, squeeze_mask = xi_77_squeeze_mask_0, x = pairs_77_cast_fp16)[name = tensor("xi_77_cast_fp16")]; + tensor var_4772_cast_fp16 = mul(x = xr_77_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4772_cast_fp16")]; + tensor var_4773_cast_fp16 = mul(x = xi_77_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4773_cast_fp16")]; + tensor xor__77_cast_fp16 = sub(x = var_4772_cast_fp16, y = var_4773_cast_fp16)[name = tensor("xor__77_cast_fp16")]; + tensor var_4776_cast_fp16 = mul(x = xr_77_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4776_cast_fp16")]; + tensor var_4777_cast_fp16 = mul(x = xi_77_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4777_cast_fp16")]; + tensor xoi_77_cast_fp16 = add(x = var_4776_cast_fp16, y = var_4777_cast_fp16)[name = tensor("xoi_77_cast_fp16")]; + tensor out_77_axis_0 = const()[name = tensor("out_77_axis_0"), val = tensor(-1)]; + tensor out_77_cast_fp16 = stack(axis = out_77_axis_0, values = (xor__77_cast_fp16, xoi_77_cast_fp16))[name = tensor("out_77_cast_fp16")]; + tensor var_4787 = const()[name = tensor("op_4787"), val = tensor([1, 256, 16, 64])]; + tensor q_39_cast_fp16 = reshape(shape = var_4787, x = out_77_cast_fp16)[name = tensor("q_39_cast_fp16")]; + tensor var_4793 = const()[name = tensor("op_4793"), val = tensor([256, 16, 32, 2])]; + tensor pairs_79_cast_fp16 = reshape(shape = var_4793, x = x_79_cast_fp16)[name = tensor("pairs_79_cast_fp16")]; + tensor xr_79_begin_0 = const()[name = tensor("xr_79_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_79_end_0 = const()[name = tensor("xr_79_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_79_end_mask_0 = const()[name = tensor("xr_79_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_79_squeeze_mask_0 = const()[name = tensor("xr_79_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_79_cast_fp16 = slice_by_index(begin = xr_79_begin_0, end = xr_79_end_0, end_mask = xr_79_end_mask_0, squeeze_mask = xr_79_squeeze_mask_0, x = pairs_79_cast_fp16)[name = tensor("xr_79_cast_fp16")]; + tensor xi_79_begin_0 = const()[name = tensor("xi_79_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_79_end_0 = const()[name = tensor("xi_79_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_79_end_mask_0 = const()[name = tensor("xi_79_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_79_squeeze_mask_0 = const()[name = tensor("xi_79_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_79_cast_fp16 = slice_by_index(begin = xi_79_begin_0, end = xi_79_end_0, end_mask = xi_79_end_mask_0, squeeze_mask = xi_79_squeeze_mask_0, x = pairs_79_cast_fp16)[name = tensor("xi_79_cast_fp16")]; + tensor var_4801_cast_fp16 = mul(x = xr_79_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4801_cast_fp16")]; + tensor var_4802_cast_fp16 = mul(x = xi_79_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4802_cast_fp16")]; + tensor xor__79_cast_fp16 = sub(x = var_4801_cast_fp16, y = var_4802_cast_fp16)[name = tensor("xor__79_cast_fp16")]; + tensor var_4805_cast_fp16 = mul(x = xr_79_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4805_cast_fp16")]; + tensor var_4806_cast_fp16 = mul(x = xi_79_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4806_cast_fp16")]; + tensor xoi_79_cast_fp16 = add(x = var_4805_cast_fp16, y = var_4806_cast_fp16)[name = tensor("xoi_79_cast_fp16")]; + tensor out_79_axis_0 = const()[name = tensor("out_79_axis_0"), val = tensor(-1)]; + tensor out_79_cast_fp16 = stack(axis = out_79_axis_0, values = (xor__79_cast_fp16, xoi_79_cast_fp16))[name = tensor("out_79_cast_fp16")]; + tensor var_4823 = const()[name = tensor("op_4823"), val = tensor([256, 1024])]; + tensor var_4824_cast_fp16 = reshape(shape = var_4823, x = out_79_cast_fp16)[name = tensor("op_4824_cast_fp16")]; + tensor written_k_39_transpose_x_1 = const()[name = tensor("written_k_39_transpose_x_1"), val = tensor(true)]; + tensor written_k_39_transpose_y_1 = const()[name = tensor("written_k_39_transpose_y_1"), val = tensor(false)]; + tensor written_k_39_cast_fp16 = matmul(transpose_x = written_k_39_transpose_x_1, transpose_y = written_k_39_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_4824_cast_fp16)[name = tensor("written_k_39_cast_fp16")]; + tensor var_4831 = const()[name = tensor("op_4831"), val = tensor([256, 1024])]; + tensor var_4832_cast_fp16 = reshape(shape = var_4831, x = v_39_cast_fp16)[name = tensor("op_4832_cast_fp16")]; + tensor written_v_39_transpose_x_1 = const()[name = tensor("written_v_39_transpose_x_1"), val = tensor(true)]; + tensor written_v_39_transpose_y_1 = const()[name = tensor("written_v_39_transpose_y_1"), val = tensor(false)]; + tensor written_v_39_cast_fp16 = matmul(transpose_x = written_v_39_transpose_x_1, transpose_y = written_v_39_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_4832_cast_fp16)[name = tensor("written_v_39_cast_fp16")]; + tensor k_cache19_to_fp16_dtype_0 = const()[name = tensor("k_cache19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache19_to_fp16 = cast(dtype = k_cache19_to_fp16_dtype_0, x = k_cache19)[name = tensor("cast_351")]; + tensor var_4843_cast_fp16 = mul(x = k_cache19_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_4843_cast_fp16")]; + tensor var_4848 = const()[name = tensor("op_4848"), val = tensor([1, 512, 16, 64])]; + tensor var_4849_cast_fp16 = reshape(shape = var_4848, x = written_k_39_cast_fp16)[name = tensor("op_4849_cast_fp16")]; + tensor new_k_cache_39_cast_fp16 = add(x = var_4843_cast_fp16, y = var_4849_cast_fp16)[name = tensor("new_k_cache_39_cast_fp16")]; + tensor new_k_cache_39_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_39_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache19_to_fp16_dtype_0 = const()[name = tensor("v_cache19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache19_to_fp16 = cast(dtype = v_cache19_to_fp16_dtype_0, x = v_cache19)[name = tensor("cast_349")]; + tensor var_4852_cast_fp16 = mul(x = v_cache19_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_4852_cast_fp16")]; + tensor var_4857 = const()[name = tensor("op_4857"), val = tensor([1, 512, 16, 64])]; + tensor var_4858_cast_fp16 = reshape(shape = var_4857, x = written_v_39_cast_fp16)[name = tensor("op_4858_cast_fp16")]; + tensor new_v_cache_39_cast_fp16 = add(x = var_4852_cast_fp16, y = var_4858_cast_fp16)[name = tensor("new_v_cache_39_cast_fp16")]; + tensor new_v_cache_39_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_39_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_4871 = const()[name = tensor("op_4871"), val = tensor([0, 2, 1, 3])]; + tensor var_4882_transpose_x_0 = const()[name = tensor("op_4882_transpose_x_0"), val = tensor(false)]; + tensor var_4882_transpose_y_0 = const()[name = tensor("op_4882_transpose_y_0"), val = tensor(false)]; + tensor transpose_107_perm_0 = const()[name = tensor("transpose_107_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_108_perm_0 = const()[name = tensor("transpose_108_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_108 = transpose(perm = transpose_108_perm_0, x = new_k_cache_39_cast_fp16)[name = tensor("transpose_128")]; + tensor transpose_107 = transpose(perm = transpose_107_perm_0, x = q_39_cast_fp16)[name = tensor("transpose_129")]; + tensor var_4882_cast_fp16 = matmul(transpose_x = var_4882_transpose_x_0, transpose_y = var_4882_transpose_y_0, x = transpose_107, y = transpose_108)[name = tensor("op_4882_cast_fp16")]; + tensor var_4883_to_fp16 = const()[name = tensor("op_4883_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_77_cast_fp16 = mul(x = var_4882_cast_fp16, y = var_4883_to_fp16)[name = tensor("scores_77_cast_fp16")]; + tensor scores_79_cast_fp16 = add(x = scores_77_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_79_cast_fp16")]; + tensor var_4893 = const()[name = tensor("op_4893"), val = tensor(-1)]; + tensor weights_39_cast_fp16 = softmax(axis = var_4893, x = scores_79_cast_fp16)[name = tensor("weights_39_cast_fp16")]; + tensor attn_39_transpose_x_0 = const()[name = tensor("attn_39_transpose_x_0"), val = tensor(false)]; + tensor attn_39_transpose_y_0 = const()[name = tensor("attn_39_transpose_y_0"), val = tensor(false)]; + tensor values_39_cast_fp16 = transpose(perm = var_4871, x = new_v_cache_39_cast_fp16)[name = tensor("transpose_130")]; + tensor attn_39_cast_fp16 = matmul(transpose_x = attn_39_transpose_x_0, transpose_y = attn_39_transpose_y_0, x = weights_39_cast_fp16, y = values_39_cast_fp16)[name = tensor("attn_39_cast_fp16")]; + tensor var_4901 = const()[name = tensor("op_4901"), val = tensor([0, 2, 1, 3])]; + tensor var_4906 = const()[name = tensor("op_4906"), val = tensor([1, 256, 1024])]; + tensor var_4902_cast_fp16 = transpose(perm = var_4901, x = attn_39_cast_fp16)[name = tensor("transpose_127")]; + tensor input_233_cast_fp16 = reshape(shape = var_4906, x = var_4902_cast_fp16)[name = tensor("input_233_cast_fp16")]; + tensor attn19_out_proj_weight_to_fp16 = const()[name = tensor("attn19_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(484630784)))]; + tensor linear_77_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn19_out_proj_weight_to_fp16, x = input_233_cast_fp16)[name = tensor("linear_77_cast_fp16")]; + tensor position19_to_fp16_dtype_0 = const()[name = tensor("position19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position19_to_fp16 = cast(dtype = position19_to_fp16_dtype_0, x = position19)[name = tensor("cast_347")]; + tensor var_4915_cast_fp16 = add(x = position19_to_fp16, y = valid_len_to_fp16)[name = tensor("op_4915_cast_fp16")]; + tensor var_4915_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4915_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_235_cast_fp16 = add(x = input_229_cast_fp16, y = linear_77_cast_fp16)[name = tensor("input_235_cast_fp16")]; + tensor input_237_axes_0 = const()[name = tensor("input_237_axes_0"), val = tensor([-1])]; + tensor norm19_2_weight_to_fp16 = const()[name = tensor("norm19_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486728000)))]; + tensor norm19_2_bias_to_fp16 = const()[name = tensor("norm19_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486730112)))]; + tensor var_4919_to_fp16 = const()[name = tensor("op_4919_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_237_cast_fp16 = layer_norm(axes = input_237_axes_0, beta = norm19_2_bias_to_fp16, epsilon = var_4919_to_fp16, gamma = norm19_2_weight_to_fp16, x = input_235_cast_fp16)[name = tensor("input_237_cast_fp16")]; + tensor linear19_1_weight_to_fp16 = const()[name = tensor("linear19_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486732224)))]; + tensor linear_78_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear19_1_weight_to_fp16, x = input_237_cast_fp16)[name = tensor("linear_78_cast_fp16")]; + tensor input_239_mode_0 = const()[name = tensor("input_239_mode_0"), val = tensor("EXACT")]; + tensor input_239_cast_fp16 = gelu(mode = input_239_mode_0, x = linear_78_cast_fp16)[name = tensor("input_239_cast_fp16")]; + tensor linear19_2_weight_to_fp16 = const()[name = tensor("linear19_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(495120896)))]; + tensor linear_79_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear19_2_weight_to_fp16, x = input_239_cast_fp16)[name = tensor("linear_79_cast_fp16")]; + tensor input_241_cast_fp16 = add(x = input_235_cast_fp16, y = linear_79_cast_fp16)[name = tensor("input_241_cast_fp16")]; + tensor input_243_axes_0 = const()[name = tensor("input_243_axes_0"), val = tensor([-1])]; + tensor norm20_1_weight_to_fp16 = const()[name = tensor("norm20_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(503509568)))]; + tensor norm20_1_bias_to_fp16 = const()[name = tensor("norm20_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(503511680)))]; + tensor var_4936_to_fp16 = const()[name = tensor("op_4936_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_243_cast_fp16 = layer_norm(axes = input_243_axes_0, beta = norm20_1_bias_to_fp16, epsilon = var_4936_to_fp16, gamma = norm20_1_weight_to_fp16, x = input_241_cast_fp16)[name = tensor("input_243_cast_fp16")]; + tensor attn20_in_proj_weight_to_fp16 = const()[name = tensor("attn20_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(503513792)))]; + tensor linear_80_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn20_in_proj_weight_to_fp16, x = input_243_cast_fp16)[name = tensor("linear_80_cast_fp16")]; + tensor var_4950 = const()[name = tensor("op_4950"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_41_cast_fp16 = reshape(shape = var_4950, x = linear_80_cast_fp16)[name = tensor("qkv_41_cast_fp16")]; + tensor x_81_begin_0 = const()[name = tensor("x_81_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_81_end_0 = const()[name = tensor("x_81_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_81_end_mask_0 = const()[name = tensor("x_81_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_81_squeeze_mask_0 = const()[name = tensor("x_81_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_81_cast_fp16 = slice_by_index(begin = x_81_begin_0, end = x_81_end_0, end_mask = x_81_end_mask_0, squeeze_mask = x_81_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("x_81_cast_fp16")]; + tensor x_83_begin_0 = const()[name = tensor("x_83_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_83_end_0 = const()[name = tensor("x_83_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_83_end_mask_0 = const()[name = tensor("x_83_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_83_squeeze_mask_0 = const()[name = tensor("x_83_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_83_cast_fp16 = slice_by_index(begin = x_83_begin_0, end = x_83_end_0, end_mask = x_83_end_mask_0, squeeze_mask = x_83_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("x_83_cast_fp16")]; + tensor v_41_begin_0 = const()[name = tensor("v_41_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_41_end_0 = const()[name = tensor("v_41_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_41_end_mask_0 = const()[name = tensor("v_41_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_41_squeeze_mask_0 = const()[name = tensor("v_41_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_41_cast_fp16 = slice_by_index(begin = v_41_begin_0, end = v_41_end_0, end_mask = v_41_end_mask_0, squeeze_mask = v_41_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("v_41_cast_fp16")]; + tensor var_4995 = const()[name = tensor("op_4995"), val = tensor([256, 16, 32, 2])]; + tensor pairs_81_cast_fp16 = reshape(shape = var_4995, x = x_81_cast_fp16)[name = tensor("pairs_81_cast_fp16")]; + tensor xr_81_begin_0 = const()[name = tensor("xr_81_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_81_end_0 = const()[name = tensor("xr_81_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_81_end_mask_0 = const()[name = tensor("xr_81_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_81_squeeze_mask_0 = const()[name = tensor("xr_81_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_81_cast_fp16 = slice_by_index(begin = xr_81_begin_0, end = xr_81_end_0, end_mask = xr_81_end_mask_0, squeeze_mask = xr_81_squeeze_mask_0, x = pairs_81_cast_fp16)[name = tensor("xr_81_cast_fp16")]; + tensor xi_81_begin_0 = const()[name = tensor("xi_81_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_81_end_0 = const()[name = tensor("xi_81_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_81_end_mask_0 = const()[name = tensor("xi_81_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_81_squeeze_mask_0 = const()[name = tensor("xi_81_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_81_cast_fp16 = slice_by_index(begin = xi_81_begin_0, end = xi_81_end_0, end_mask = xi_81_end_mask_0, squeeze_mask = xi_81_squeeze_mask_0, x = pairs_81_cast_fp16)[name = tensor("xi_81_cast_fp16")]; + tensor var_5003_cast_fp16 = mul(x = xr_81_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5003_cast_fp16")]; + tensor var_5004_cast_fp16 = mul(x = xi_81_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5004_cast_fp16")]; + tensor xor__81_cast_fp16 = sub(x = var_5003_cast_fp16, y = var_5004_cast_fp16)[name = tensor("xor__81_cast_fp16")]; + tensor var_5007_cast_fp16 = mul(x = xr_81_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5007_cast_fp16")]; + tensor var_5008_cast_fp16 = mul(x = xi_81_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5008_cast_fp16")]; + tensor xoi_81_cast_fp16 = add(x = var_5007_cast_fp16, y = var_5008_cast_fp16)[name = tensor("xoi_81_cast_fp16")]; + tensor out_81_axis_0 = const()[name = tensor("out_81_axis_0"), val = tensor(-1)]; + tensor out_81_cast_fp16 = stack(axis = out_81_axis_0, values = (xor__81_cast_fp16, xoi_81_cast_fp16))[name = tensor("out_81_cast_fp16")]; + tensor var_5018 = const()[name = tensor("op_5018"), val = tensor([1, 256, 16, 64])]; + tensor q_41_cast_fp16 = reshape(shape = var_5018, x = out_81_cast_fp16)[name = tensor("q_41_cast_fp16")]; + tensor var_5024 = const()[name = tensor("op_5024"), val = tensor([256, 16, 32, 2])]; + tensor pairs_83_cast_fp16 = reshape(shape = var_5024, x = x_83_cast_fp16)[name = tensor("pairs_83_cast_fp16")]; + tensor xr_83_begin_0 = const()[name = tensor("xr_83_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_83_end_0 = const()[name = tensor("xr_83_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_83_end_mask_0 = const()[name = tensor("xr_83_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_83_squeeze_mask_0 = const()[name = tensor("xr_83_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_83_cast_fp16 = slice_by_index(begin = xr_83_begin_0, end = xr_83_end_0, end_mask = xr_83_end_mask_0, squeeze_mask = xr_83_squeeze_mask_0, x = pairs_83_cast_fp16)[name = tensor("xr_83_cast_fp16")]; + tensor xi_83_begin_0 = const()[name = tensor("xi_83_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_83_end_0 = const()[name = tensor("xi_83_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_83_end_mask_0 = const()[name = tensor("xi_83_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_83_squeeze_mask_0 = const()[name = tensor("xi_83_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_83_cast_fp16 = slice_by_index(begin = xi_83_begin_0, end = xi_83_end_0, end_mask = xi_83_end_mask_0, squeeze_mask = xi_83_squeeze_mask_0, x = pairs_83_cast_fp16)[name = tensor("xi_83_cast_fp16")]; + tensor var_5032_cast_fp16 = mul(x = xr_83_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5032_cast_fp16")]; + tensor var_5033_cast_fp16 = mul(x = xi_83_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5033_cast_fp16")]; + tensor xor__83_cast_fp16 = sub(x = var_5032_cast_fp16, y = var_5033_cast_fp16)[name = tensor("xor__83_cast_fp16")]; + tensor var_5036_cast_fp16 = mul(x = xr_83_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5036_cast_fp16")]; + tensor var_5037_cast_fp16 = mul(x = xi_83_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5037_cast_fp16")]; + tensor xoi_83_cast_fp16 = add(x = var_5036_cast_fp16, y = var_5037_cast_fp16)[name = tensor("xoi_83_cast_fp16")]; + tensor out_83_axis_0 = const()[name = tensor("out_83_axis_0"), val = tensor(-1)]; + tensor out_83_cast_fp16 = stack(axis = out_83_axis_0, values = (xor__83_cast_fp16, xoi_83_cast_fp16))[name = tensor("out_83_cast_fp16")]; + tensor var_5054 = const()[name = tensor("op_5054"), val = tensor([256, 1024])]; + tensor var_5055_cast_fp16 = reshape(shape = var_5054, x = out_83_cast_fp16)[name = tensor("op_5055_cast_fp16")]; + tensor written_k_41_transpose_x_1 = const()[name = tensor("written_k_41_transpose_x_1"), val = tensor(true)]; + tensor written_k_41_transpose_y_1 = const()[name = tensor("written_k_41_transpose_y_1"), val = tensor(false)]; + tensor written_k_41_cast_fp16 = matmul(transpose_x = written_k_41_transpose_x_1, transpose_y = written_k_41_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_5055_cast_fp16)[name = tensor("written_k_41_cast_fp16")]; + tensor var_5062 = const()[name = tensor("op_5062"), val = tensor([256, 1024])]; + tensor var_5063_cast_fp16 = reshape(shape = var_5062, x = v_41_cast_fp16)[name = tensor("op_5063_cast_fp16")]; + tensor written_v_41_transpose_x_1 = const()[name = tensor("written_v_41_transpose_x_1"), val = tensor(true)]; + tensor written_v_41_transpose_y_1 = const()[name = tensor("written_v_41_transpose_y_1"), val = tensor(false)]; + tensor written_v_41_cast_fp16 = matmul(transpose_x = written_v_41_transpose_x_1, transpose_y = written_v_41_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_5063_cast_fp16)[name = tensor("written_v_41_cast_fp16")]; + tensor k_cache20_to_fp16_dtype_0 = const()[name = tensor("k_cache20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache20_to_fp16 = cast(dtype = k_cache20_to_fp16_dtype_0, x = k_cache20)[name = tensor("cast_345")]; + tensor var_5074_cast_fp16 = mul(x = k_cache20_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_5074_cast_fp16")]; + tensor var_5079 = const()[name = tensor("op_5079"), val = tensor([1, 512, 16, 64])]; + tensor var_5080_cast_fp16 = reshape(shape = var_5079, x = written_k_41_cast_fp16)[name = tensor("op_5080_cast_fp16")]; + tensor new_k_cache_41_cast_fp16 = add(x = var_5074_cast_fp16, y = var_5080_cast_fp16)[name = tensor("new_k_cache_41_cast_fp16")]; + tensor new_k_cache_41_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_41_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache20_to_fp16_dtype_0 = const()[name = tensor("v_cache20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache20_to_fp16 = cast(dtype = v_cache20_to_fp16_dtype_0, x = v_cache20)[name = tensor("cast_343")]; + tensor var_5083_cast_fp16 = mul(x = v_cache20_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_5083_cast_fp16")]; + tensor var_5088 = const()[name = tensor("op_5088"), val = tensor([1, 512, 16, 64])]; + tensor var_5089_cast_fp16 = reshape(shape = var_5088, x = written_v_41_cast_fp16)[name = tensor("op_5089_cast_fp16")]; + tensor new_v_cache_41_cast_fp16 = add(x = var_5083_cast_fp16, y = var_5089_cast_fp16)[name = tensor("new_v_cache_41_cast_fp16")]; + tensor new_v_cache_41_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_41_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_5102 = const()[name = tensor("op_5102"), val = tensor([0, 2, 1, 3])]; + tensor var_5113_transpose_x_0 = const()[name = tensor("op_5113_transpose_x_0"), val = tensor(false)]; + tensor var_5113_transpose_y_0 = const()[name = tensor("op_5113_transpose_y_0"), val = tensor(false)]; + tensor transpose_109_perm_0 = const()[name = tensor("transpose_109_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_110_perm_0 = const()[name = tensor("transpose_110_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_110 = transpose(perm = transpose_110_perm_0, x = new_k_cache_41_cast_fp16)[name = tensor("transpose_124")]; + tensor transpose_109 = transpose(perm = transpose_109_perm_0, x = q_41_cast_fp16)[name = tensor("transpose_125")]; + tensor var_5113_cast_fp16 = matmul(transpose_x = var_5113_transpose_x_0, transpose_y = var_5113_transpose_y_0, x = transpose_109, y = transpose_110)[name = tensor("op_5113_cast_fp16")]; + tensor var_5114_to_fp16 = const()[name = tensor("op_5114_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_81_cast_fp16 = mul(x = var_5113_cast_fp16, y = var_5114_to_fp16)[name = tensor("scores_81_cast_fp16")]; + tensor scores_83_cast_fp16 = add(x = scores_81_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_83_cast_fp16")]; + tensor var_5124 = const()[name = tensor("op_5124"), val = tensor(-1)]; + tensor weights_41_cast_fp16 = softmax(axis = var_5124, x = scores_83_cast_fp16)[name = tensor("weights_41_cast_fp16")]; + tensor attn_41_transpose_x_0 = const()[name = tensor("attn_41_transpose_x_0"), val = tensor(false)]; + tensor attn_41_transpose_y_0 = const()[name = tensor("attn_41_transpose_y_0"), val = tensor(false)]; + tensor values_41_cast_fp16 = transpose(perm = var_5102, x = new_v_cache_41_cast_fp16)[name = tensor("transpose_126")]; + tensor attn_41_cast_fp16 = matmul(transpose_x = attn_41_transpose_x_0, transpose_y = attn_41_transpose_y_0, x = weights_41_cast_fp16, y = values_41_cast_fp16)[name = tensor("attn_41_cast_fp16")]; + tensor var_5132 = const()[name = tensor("op_5132"), val = tensor([0, 2, 1, 3])]; + tensor var_5137 = const()[name = tensor("op_5137"), val = tensor([1, 256, 1024])]; + tensor var_5133_cast_fp16 = transpose(perm = var_5132, x = attn_41_cast_fp16)[name = tensor("transpose_123")]; + tensor input_245_cast_fp16 = reshape(shape = var_5137, x = var_5133_cast_fp16)[name = tensor("input_245_cast_fp16")]; + tensor attn20_out_proj_weight_to_fp16 = const()[name = tensor("attn20_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(509805312)))]; + tensor linear_81_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn20_out_proj_weight_to_fp16, x = input_245_cast_fp16)[name = tensor("linear_81_cast_fp16")]; + tensor position20_to_fp16_dtype_0 = const()[name = tensor("position20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position20_to_fp16 = cast(dtype = position20_to_fp16_dtype_0, x = position20)[name = tensor("cast_341")]; + tensor var_5146_cast_fp16 = add(x = position20_to_fp16, y = valid_len_to_fp16)[name = tensor("op_5146_cast_fp16")]; + tensor var_5146_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5146_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_247_cast_fp16 = add(x = input_241_cast_fp16, y = linear_81_cast_fp16)[name = tensor("input_247_cast_fp16")]; + tensor input_249_axes_0 = const()[name = tensor("input_249_axes_0"), val = tensor([-1])]; + tensor norm20_2_weight_to_fp16 = const()[name = tensor("norm20_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(511902528)))]; + tensor norm20_2_bias_to_fp16 = const()[name = tensor("norm20_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(511904640)))]; + tensor var_5150_to_fp16 = const()[name = tensor("op_5150_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_249_cast_fp16 = layer_norm(axes = input_249_axes_0, beta = norm20_2_bias_to_fp16, epsilon = var_5150_to_fp16, gamma = norm20_2_weight_to_fp16, x = input_247_cast_fp16)[name = tensor("input_249_cast_fp16")]; + tensor linear20_1_weight_to_fp16 = const()[name = tensor("linear20_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(511906752)))]; + tensor linear_82_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear20_1_weight_to_fp16, x = input_249_cast_fp16)[name = tensor("linear_82_cast_fp16")]; + tensor input_251_mode_0 = const()[name = tensor("input_251_mode_0"), val = tensor("EXACT")]; + tensor input_251_cast_fp16 = gelu(mode = input_251_mode_0, x = linear_82_cast_fp16)[name = tensor("input_251_cast_fp16")]; + tensor linear20_2_weight_to_fp16 = const()[name = tensor("linear20_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(520295424)))]; + tensor linear_83_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear20_2_weight_to_fp16, x = input_251_cast_fp16)[name = tensor("linear_83_cast_fp16")]; + tensor input_253_cast_fp16 = add(x = input_247_cast_fp16, y = linear_83_cast_fp16)[name = tensor("input_253_cast_fp16")]; + tensor input_255_axes_0 = const()[name = tensor("input_255_axes_0"), val = tensor([-1])]; + tensor norm21_1_weight_to_fp16 = const()[name = tensor("norm21_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(528684096)))]; + tensor norm21_1_bias_to_fp16 = const()[name = tensor("norm21_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(528686208)))]; + tensor var_5167_to_fp16 = const()[name = tensor("op_5167_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_255_cast_fp16 = layer_norm(axes = input_255_axes_0, beta = norm21_1_bias_to_fp16, epsilon = var_5167_to_fp16, gamma = norm21_1_weight_to_fp16, x = input_253_cast_fp16)[name = tensor("input_255_cast_fp16")]; + tensor attn21_in_proj_weight_to_fp16 = const()[name = tensor("attn21_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(528688320)))]; + tensor linear_84_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn21_in_proj_weight_to_fp16, x = input_255_cast_fp16)[name = tensor("linear_84_cast_fp16")]; + tensor var_5181 = const()[name = tensor("op_5181"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_43_cast_fp16 = reshape(shape = var_5181, x = linear_84_cast_fp16)[name = tensor("qkv_43_cast_fp16")]; + tensor x_85_begin_0 = const()[name = tensor("x_85_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_85_end_0 = const()[name = tensor("x_85_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_85_end_mask_0 = const()[name = tensor("x_85_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_85_squeeze_mask_0 = const()[name = tensor("x_85_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_85_cast_fp16 = slice_by_index(begin = x_85_begin_0, end = x_85_end_0, end_mask = x_85_end_mask_0, squeeze_mask = x_85_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("x_85_cast_fp16")]; + tensor x_87_begin_0 = const()[name = tensor("x_87_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_87_end_0 = const()[name = tensor("x_87_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_87_end_mask_0 = const()[name = tensor("x_87_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_87_squeeze_mask_0 = const()[name = tensor("x_87_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_87_cast_fp16 = slice_by_index(begin = x_87_begin_0, end = x_87_end_0, end_mask = x_87_end_mask_0, squeeze_mask = x_87_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("x_87_cast_fp16")]; + tensor v_43_begin_0 = const()[name = tensor("v_43_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_43_end_0 = const()[name = tensor("v_43_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_43_end_mask_0 = const()[name = tensor("v_43_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_43_squeeze_mask_0 = const()[name = tensor("v_43_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_43_cast_fp16 = slice_by_index(begin = v_43_begin_0, end = v_43_end_0, end_mask = v_43_end_mask_0, squeeze_mask = v_43_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("v_43_cast_fp16")]; + tensor var_5226 = const()[name = tensor("op_5226"), val = tensor([256, 16, 32, 2])]; + tensor pairs_85_cast_fp16 = reshape(shape = var_5226, x = x_85_cast_fp16)[name = tensor("pairs_85_cast_fp16")]; + tensor xr_85_begin_0 = const()[name = tensor("xr_85_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_85_end_0 = const()[name = tensor("xr_85_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_85_end_mask_0 = const()[name = tensor("xr_85_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_85_squeeze_mask_0 = const()[name = tensor("xr_85_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_85_cast_fp16 = slice_by_index(begin = xr_85_begin_0, end = xr_85_end_0, end_mask = xr_85_end_mask_0, squeeze_mask = xr_85_squeeze_mask_0, x = pairs_85_cast_fp16)[name = tensor("xr_85_cast_fp16")]; + tensor xi_85_begin_0 = const()[name = tensor("xi_85_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_85_end_0 = const()[name = tensor("xi_85_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_85_end_mask_0 = const()[name = tensor("xi_85_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_85_squeeze_mask_0 = const()[name = tensor("xi_85_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_85_cast_fp16 = slice_by_index(begin = xi_85_begin_0, end = xi_85_end_0, end_mask = xi_85_end_mask_0, squeeze_mask = xi_85_squeeze_mask_0, x = pairs_85_cast_fp16)[name = tensor("xi_85_cast_fp16")]; + tensor var_5234_cast_fp16 = mul(x = xr_85_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5234_cast_fp16")]; + tensor var_5235_cast_fp16 = mul(x = xi_85_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5235_cast_fp16")]; + tensor xor__85_cast_fp16 = sub(x = var_5234_cast_fp16, y = var_5235_cast_fp16)[name = tensor("xor__85_cast_fp16")]; + tensor var_5238_cast_fp16 = mul(x = xr_85_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5238_cast_fp16")]; + tensor var_5239_cast_fp16 = mul(x = xi_85_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5239_cast_fp16")]; + tensor xoi_85_cast_fp16 = add(x = var_5238_cast_fp16, y = var_5239_cast_fp16)[name = tensor("xoi_85_cast_fp16")]; + tensor out_85_axis_0 = const()[name = tensor("out_85_axis_0"), val = tensor(-1)]; + tensor out_85_cast_fp16 = stack(axis = out_85_axis_0, values = (xor__85_cast_fp16, xoi_85_cast_fp16))[name = tensor("out_85_cast_fp16")]; + tensor var_5249 = const()[name = tensor("op_5249"), val = tensor([1, 256, 16, 64])]; + tensor q_43_cast_fp16 = reshape(shape = var_5249, x = out_85_cast_fp16)[name = tensor("q_43_cast_fp16")]; + tensor var_5255 = const()[name = tensor("op_5255"), val = tensor([256, 16, 32, 2])]; + tensor pairs_87_cast_fp16 = reshape(shape = var_5255, x = x_87_cast_fp16)[name = tensor("pairs_87_cast_fp16")]; + tensor xr_87_begin_0 = const()[name = tensor("xr_87_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_87_end_0 = const()[name = tensor("xr_87_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_87_end_mask_0 = const()[name = tensor("xr_87_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_87_squeeze_mask_0 = const()[name = tensor("xr_87_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_87_cast_fp16 = slice_by_index(begin = xr_87_begin_0, end = xr_87_end_0, end_mask = xr_87_end_mask_0, squeeze_mask = xr_87_squeeze_mask_0, x = pairs_87_cast_fp16)[name = tensor("xr_87_cast_fp16")]; + tensor xi_87_begin_0 = const()[name = tensor("xi_87_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_87_end_0 = const()[name = tensor("xi_87_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_87_end_mask_0 = const()[name = tensor("xi_87_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_87_squeeze_mask_0 = const()[name = tensor("xi_87_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_87_cast_fp16 = slice_by_index(begin = xi_87_begin_0, end = xi_87_end_0, end_mask = xi_87_end_mask_0, squeeze_mask = xi_87_squeeze_mask_0, x = pairs_87_cast_fp16)[name = tensor("xi_87_cast_fp16")]; + tensor var_5263_cast_fp16 = mul(x = xr_87_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5263_cast_fp16")]; + tensor var_5264_cast_fp16 = mul(x = xi_87_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5264_cast_fp16")]; + tensor xor__87_cast_fp16 = sub(x = var_5263_cast_fp16, y = var_5264_cast_fp16)[name = tensor("xor__87_cast_fp16")]; + tensor var_5267_cast_fp16 = mul(x = xr_87_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5267_cast_fp16")]; + tensor var_5268_cast_fp16 = mul(x = xi_87_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5268_cast_fp16")]; + tensor xoi_87_cast_fp16 = add(x = var_5267_cast_fp16, y = var_5268_cast_fp16)[name = tensor("xoi_87_cast_fp16")]; + tensor out_87_axis_0 = const()[name = tensor("out_87_axis_0"), val = tensor(-1)]; + tensor out_87_cast_fp16 = stack(axis = out_87_axis_0, values = (xor__87_cast_fp16, xoi_87_cast_fp16))[name = tensor("out_87_cast_fp16")]; + tensor var_5285 = const()[name = tensor("op_5285"), val = tensor([256, 1024])]; + tensor var_5286_cast_fp16 = reshape(shape = var_5285, x = out_87_cast_fp16)[name = tensor("op_5286_cast_fp16")]; + tensor written_k_43_transpose_x_1 = const()[name = tensor("written_k_43_transpose_x_1"), val = tensor(true)]; + tensor written_k_43_transpose_y_1 = const()[name = tensor("written_k_43_transpose_y_1"), val = tensor(false)]; + tensor written_k_43_cast_fp16 = matmul(transpose_x = written_k_43_transpose_x_1, transpose_y = written_k_43_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_5286_cast_fp16)[name = tensor("written_k_43_cast_fp16")]; + tensor var_5293 = const()[name = tensor("op_5293"), val = tensor([256, 1024])]; + tensor var_5294_cast_fp16 = reshape(shape = var_5293, x = v_43_cast_fp16)[name = tensor("op_5294_cast_fp16")]; + tensor written_v_43_transpose_x_1 = const()[name = tensor("written_v_43_transpose_x_1"), val = tensor(true)]; + tensor written_v_43_transpose_y_1 = const()[name = tensor("written_v_43_transpose_y_1"), val = tensor(false)]; + tensor written_v_43_cast_fp16 = matmul(transpose_x = written_v_43_transpose_x_1, transpose_y = written_v_43_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_5294_cast_fp16)[name = tensor("written_v_43_cast_fp16")]; + tensor k_cache21_to_fp16_dtype_0 = const()[name = tensor("k_cache21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache21_to_fp16 = cast(dtype = k_cache21_to_fp16_dtype_0, x = k_cache21)[name = tensor("cast_339")]; + tensor var_5305_cast_fp16 = mul(x = k_cache21_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_5305_cast_fp16")]; + tensor var_5310 = const()[name = tensor("op_5310"), val = tensor([1, 512, 16, 64])]; + tensor var_5311_cast_fp16 = reshape(shape = var_5310, x = written_k_43_cast_fp16)[name = tensor("op_5311_cast_fp16")]; + tensor new_k_cache_43_cast_fp16 = add(x = var_5305_cast_fp16, y = var_5311_cast_fp16)[name = tensor("new_k_cache_43_cast_fp16")]; + tensor new_k_cache_43_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_43_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache21_to_fp16_dtype_0 = const()[name = tensor("v_cache21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache21_to_fp16 = cast(dtype = v_cache21_to_fp16_dtype_0, x = v_cache21)[name = tensor("cast_337")]; + tensor var_5314_cast_fp16 = mul(x = v_cache21_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_5314_cast_fp16")]; + tensor var_5319 = const()[name = tensor("op_5319"), val = tensor([1, 512, 16, 64])]; + tensor var_5320_cast_fp16 = reshape(shape = var_5319, x = written_v_43_cast_fp16)[name = tensor("op_5320_cast_fp16")]; + tensor new_v_cache_43_cast_fp16 = add(x = var_5314_cast_fp16, y = var_5320_cast_fp16)[name = tensor("new_v_cache_43_cast_fp16")]; + tensor new_v_cache_43_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_43_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_5333 = const()[name = tensor("op_5333"), val = tensor([0, 2, 1, 3])]; + tensor var_5344_transpose_x_0 = const()[name = tensor("op_5344_transpose_x_0"), val = tensor(false)]; + tensor var_5344_transpose_y_0 = const()[name = tensor("op_5344_transpose_y_0"), val = tensor(false)]; + tensor transpose_111_perm_0 = const()[name = tensor("transpose_111_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_112_perm_0 = const()[name = tensor("transpose_112_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_112 = transpose(perm = transpose_112_perm_0, x = new_k_cache_43_cast_fp16)[name = tensor("transpose_120")]; + tensor transpose_111 = transpose(perm = transpose_111_perm_0, x = q_43_cast_fp16)[name = tensor("transpose_121")]; + tensor var_5344_cast_fp16 = matmul(transpose_x = var_5344_transpose_x_0, transpose_y = var_5344_transpose_y_0, x = transpose_111, y = transpose_112)[name = tensor("op_5344_cast_fp16")]; + tensor var_5345_to_fp16 = const()[name = tensor("op_5345_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_85_cast_fp16 = mul(x = var_5344_cast_fp16, y = var_5345_to_fp16)[name = tensor("scores_85_cast_fp16")]; + tensor scores_87_cast_fp16 = add(x = scores_85_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_87_cast_fp16")]; + tensor var_5355 = const()[name = tensor("op_5355"), val = tensor(-1)]; + tensor weights_43_cast_fp16 = softmax(axis = var_5355, x = scores_87_cast_fp16)[name = tensor("weights_43_cast_fp16")]; + tensor attn_43_transpose_x_0 = const()[name = tensor("attn_43_transpose_x_0"), val = tensor(false)]; + tensor attn_43_transpose_y_0 = const()[name = tensor("attn_43_transpose_y_0"), val = tensor(false)]; + tensor values_43_cast_fp16 = transpose(perm = var_5333, x = new_v_cache_43_cast_fp16)[name = tensor("transpose_122")]; + tensor attn_43_cast_fp16 = matmul(transpose_x = attn_43_transpose_x_0, transpose_y = attn_43_transpose_y_0, x = weights_43_cast_fp16, y = values_43_cast_fp16)[name = tensor("attn_43_cast_fp16")]; + tensor var_5363 = const()[name = tensor("op_5363"), val = tensor([0, 2, 1, 3])]; + tensor var_5368 = const()[name = tensor("op_5368"), val = tensor([1, 256, 1024])]; + tensor var_5364_cast_fp16 = transpose(perm = var_5363, x = attn_43_cast_fp16)[name = tensor("transpose_119")]; + tensor input_257_cast_fp16 = reshape(shape = var_5368, x = var_5364_cast_fp16)[name = tensor("input_257_cast_fp16")]; + tensor attn21_out_proj_weight_to_fp16 = const()[name = tensor("attn21_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(534979840)))]; + tensor linear_85_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn21_out_proj_weight_to_fp16, x = input_257_cast_fp16)[name = tensor("linear_85_cast_fp16")]; + tensor position21_to_fp16_dtype_0 = const()[name = tensor("position21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position21_to_fp16 = cast(dtype = position21_to_fp16_dtype_0, x = position21)[name = tensor("cast_335")]; + tensor var_5377_cast_fp16 = add(x = position21_to_fp16, y = valid_len_to_fp16)[name = tensor("op_5377_cast_fp16")]; + tensor var_5377_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5377_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_259_cast_fp16 = add(x = input_253_cast_fp16, y = linear_85_cast_fp16)[name = tensor("input_259_cast_fp16")]; + tensor input_261_axes_0 = const()[name = tensor("input_261_axes_0"), val = tensor([-1])]; + tensor norm21_2_weight_to_fp16 = const()[name = tensor("norm21_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(537077056)))]; + tensor norm21_2_bias_to_fp16 = const()[name = tensor("norm21_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(537079168)))]; + tensor var_5381_to_fp16 = const()[name = tensor("op_5381_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_261_cast_fp16 = layer_norm(axes = input_261_axes_0, beta = norm21_2_bias_to_fp16, epsilon = var_5381_to_fp16, gamma = norm21_2_weight_to_fp16, x = input_259_cast_fp16)[name = tensor("input_261_cast_fp16")]; + tensor linear21_1_weight_to_fp16 = const()[name = tensor("linear21_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(537081280)))]; + tensor linear_86_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear21_1_weight_to_fp16, x = input_261_cast_fp16)[name = tensor("linear_86_cast_fp16")]; + tensor input_263_mode_0 = const()[name = tensor("input_263_mode_0"), val = tensor("EXACT")]; + tensor input_263_cast_fp16 = gelu(mode = input_263_mode_0, x = linear_86_cast_fp16)[name = tensor("input_263_cast_fp16")]; + tensor linear21_2_weight_to_fp16 = const()[name = tensor("linear21_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(545469952)))]; + tensor linear_87_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear21_2_weight_to_fp16, x = input_263_cast_fp16)[name = tensor("linear_87_cast_fp16")]; + tensor input_265_cast_fp16 = add(x = input_259_cast_fp16, y = linear_87_cast_fp16)[name = tensor("input_265_cast_fp16")]; + tensor input_267_axes_0 = const()[name = tensor("input_267_axes_0"), val = tensor([-1])]; + tensor norm22_1_weight_to_fp16 = const()[name = tensor("norm22_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(553858624)))]; + tensor norm22_1_bias_to_fp16 = const()[name = tensor("norm22_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(553860736)))]; + tensor var_5398_to_fp16 = const()[name = tensor("op_5398_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_267_cast_fp16 = layer_norm(axes = input_267_axes_0, beta = norm22_1_bias_to_fp16, epsilon = var_5398_to_fp16, gamma = norm22_1_weight_to_fp16, x = input_265_cast_fp16)[name = tensor("input_267_cast_fp16")]; + tensor attn22_in_proj_weight_to_fp16 = const()[name = tensor("attn22_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(553862848)))]; + tensor linear_88_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn22_in_proj_weight_to_fp16, x = input_267_cast_fp16)[name = tensor("linear_88_cast_fp16")]; + tensor var_5412 = const()[name = tensor("op_5412"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_45_cast_fp16 = reshape(shape = var_5412, x = linear_88_cast_fp16)[name = tensor("qkv_45_cast_fp16")]; + tensor x_89_begin_0 = const()[name = tensor("x_89_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_89_end_0 = const()[name = tensor("x_89_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_89_end_mask_0 = const()[name = tensor("x_89_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_89_squeeze_mask_0 = const()[name = tensor("x_89_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_89_cast_fp16 = slice_by_index(begin = x_89_begin_0, end = x_89_end_0, end_mask = x_89_end_mask_0, squeeze_mask = x_89_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("x_89_cast_fp16")]; + tensor x_91_begin_0 = const()[name = tensor("x_91_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_91_end_0 = const()[name = tensor("x_91_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_91_end_mask_0 = const()[name = tensor("x_91_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_91_squeeze_mask_0 = const()[name = tensor("x_91_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_91_cast_fp16 = slice_by_index(begin = x_91_begin_0, end = x_91_end_0, end_mask = x_91_end_mask_0, squeeze_mask = x_91_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("x_91_cast_fp16")]; + tensor v_45_begin_0 = const()[name = tensor("v_45_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_45_end_0 = const()[name = tensor("v_45_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_45_end_mask_0 = const()[name = tensor("v_45_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_45_squeeze_mask_0 = const()[name = tensor("v_45_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_45_cast_fp16 = slice_by_index(begin = v_45_begin_0, end = v_45_end_0, end_mask = v_45_end_mask_0, squeeze_mask = v_45_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("v_45_cast_fp16")]; + tensor var_5457 = const()[name = tensor("op_5457"), val = tensor([256, 16, 32, 2])]; + tensor pairs_89_cast_fp16 = reshape(shape = var_5457, x = x_89_cast_fp16)[name = tensor("pairs_89_cast_fp16")]; + tensor xr_89_begin_0 = const()[name = tensor("xr_89_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_89_end_0 = const()[name = tensor("xr_89_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_89_end_mask_0 = const()[name = tensor("xr_89_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_89_squeeze_mask_0 = const()[name = tensor("xr_89_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_89_cast_fp16 = slice_by_index(begin = xr_89_begin_0, end = xr_89_end_0, end_mask = xr_89_end_mask_0, squeeze_mask = xr_89_squeeze_mask_0, x = pairs_89_cast_fp16)[name = tensor("xr_89_cast_fp16")]; + tensor xi_89_begin_0 = const()[name = tensor("xi_89_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_89_end_0 = const()[name = tensor("xi_89_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_89_end_mask_0 = const()[name = tensor("xi_89_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_89_squeeze_mask_0 = const()[name = tensor("xi_89_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_89_cast_fp16 = slice_by_index(begin = xi_89_begin_0, end = xi_89_end_0, end_mask = xi_89_end_mask_0, squeeze_mask = xi_89_squeeze_mask_0, x = pairs_89_cast_fp16)[name = tensor("xi_89_cast_fp16")]; + tensor var_5465_cast_fp16 = mul(x = xr_89_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5465_cast_fp16")]; + tensor var_5466_cast_fp16 = mul(x = xi_89_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5466_cast_fp16")]; + tensor xor__89_cast_fp16 = sub(x = var_5465_cast_fp16, y = var_5466_cast_fp16)[name = tensor("xor__89_cast_fp16")]; + tensor var_5469_cast_fp16 = mul(x = xr_89_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5469_cast_fp16")]; + tensor var_5470_cast_fp16 = mul(x = xi_89_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5470_cast_fp16")]; + tensor xoi_89_cast_fp16 = add(x = var_5469_cast_fp16, y = var_5470_cast_fp16)[name = tensor("xoi_89_cast_fp16")]; + tensor out_89_axis_0 = const()[name = tensor("out_89_axis_0"), val = tensor(-1)]; + tensor out_89_cast_fp16 = stack(axis = out_89_axis_0, values = (xor__89_cast_fp16, xoi_89_cast_fp16))[name = tensor("out_89_cast_fp16")]; + tensor var_5480 = const()[name = tensor("op_5480"), val = tensor([1, 256, 16, 64])]; + tensor q_45_cast_fp16 = reshape(shape = var_5480, x = out_89_cast_fp16)[name = tensor("q_45_cast_fp16")]; + tensor var_5486 = const()[name = tensor("op_5486"), val = tensor([256, 16, 32, 2])]; + tensor pairs_91_cast_fp16 = reshape(shape = var_5486, x = x_91_cast_fp16)[name = tensor("pairs_91_cast_fp16")]; + tensor xr_91_begin_0 = const()[name = tensor("xr_91_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_91_end_0 = const()[name = tensor("xr_91_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_91_end_mask_0 = const()[name = tensor("xr_91_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_91_squeeze_mask_0 = const()[name = tensor("xr_91_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_91_cast_fp16 = slice_by_index(begin = xr_91_begin_0, end = xr_91_end_0, end_mask = xr_91_end_mask_0, squeeze_mask = xr_91_squeeze_mask_0, x = pairs_91_cast_fp16)[name = tensor("xr_91_cast_fp16")]; + tensor xi_91_begin_0 = const()[name = tensor("xi_91_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_91_end_0 = const()[name = tensor("xi_91_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_91_end_mask_0 = const()[name = tensor("xi_91_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_91_squeeze_mask_0 = const()[name = tensor("xi_91_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_91_cast_fp16 = slice_by_index(begin = xi_91_begin_0, end = xi_91_end_0, end_mask = xi_91_end_mask_0, squeeze_mask = xi_91_squeeze_mask_0, x = pairs_91_cast_fp16)[name = tensor("xi_91_cast_fp16")]; + tensor var_5494_cast_fp16 = mul(x = xr_91_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5494_cast_fp16")]; + tensor var_5495_cast_fp16 = mul(x = xi_91_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5495_cast_fp16")]; + tensor xor__91_cast_fp16 = sub(x = var_5494_cast_fp16, y = var_5495_cast_fp16)[name = tensor("xor__91_cast_fp16")]; + tensor var_5498_cast_fp16 = mul(x = xr_91_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5498_cast_fp16")]; + tensor var_5499_cast_fp16 = mul(x = xi_91_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5499_cast_fp16")]; + tensor xoi_91_cast_fp16 = add(x = var_5498_cast_fp16, y = var_5499_cast_fp16)[name = tensor("xoi_91_cast_fp16")]; + tensor out_91_axis_0 = const()[name = tensor("out_91_axis_0"), val = tensor(-1)]; + tensor out_91_cast_fp16 = stack(axis = out_91_axis_0, values = (xor__91_cast_fp16, xoi_91_cast_fp16))[name = tensor("out_91_cast_fp16")]; + tensor var_5516 = const()[name = tensor("op_5516"), val = tensor([256, 1024])]; + tensor var_5517_cast_fp16 = reshape(shape = var_5516, x = out_91_cast_fp16)[name = tensor("op_5517_cast_fp16")]; + tensor written_k_45_transpose_x_1 = const()[name = tensor("written_k_45_transpose_x_1"), val = tensor(true)]; + tensor written_k_45_transpose_y_1 = const()[name = tensor("written_k_45_transpose_y_1"), val = tensor(false)]; + tensor written_k_45_cast_fp16 = matmul(transpose_x = written_k_45_transpose_x_1, transpose_y = written_k_45_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_5517_cast_fp16)[name = tensor("written_k_45_cast_fp16")]; + tensor var_5524 = const()[name = tensor("op_5524"), val = tensor([256, 1024])]; + tensor var_5525_cast_fp16 = reshape(shape = var_5524, x = v_45_cast_fp16)[name = tensor("op_5525_cast_fp16")]; + tensor written_v_45_transpose_x_1 = const()[name = tensor("written_v_45_transpose_x_1"), val = tensor(true)]; + tensor written_v_45_transpose_y_1 = const()[name = tensor("written_v_45_transpose_y_1"), val = tensor(false)]; + tensor written_v_45_cast_fp16 = matmul(transpose_x = written_v_45_transpose_x_1, transpose_y = written_v_45_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_5525_cast_fp16)[name = tensor("written_v_45_cast_fp16")]; + tensor k_cache22_to_fp16_dtype_0 = const()[name = tensor("k_cache22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache22_to_fp16 = cast(dtype = k_cache22_to_fp16_dtype_0, x = k_cache22)[name = tensor("cast_333")]; + tensor var_5536_cast_fp16 = mul(x = k_cache22_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_5536_cast_fp16")]; + tensor var_5541 = const()[name = tensor("op_5541"), val = tensor([1, 512, 16, 64])]; + tensor var_5542_cast_fp16 = reshape(shape = var_5541, x = written_k_45_cast_fp16)[name = tensor("op_5542_cast_fp16")]; + tensor new_k_cache_45_cast_fp16 = add(x = var_5536_cast_fp16, y = var_5542_cast_fp16)[name = tensor("new_k_cache_45_cast_fp16")]; + tensor new_k_cache_45_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_45_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache22_to_fp16_dtype_0 = const()[name = tensor("v_cache22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache22_to_fp16 = cast(dtype = v_cache22_to_fp16_dtype_0, x = v_cache22)[name = tensor("cast_331")]; + tensor var_5545_cast_fp16 = mul(x = v_cache22_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_5545_cast_fp16")]; + tensor var_5550 = const()[name = tensor("op_5550"), val = tensor([1, 512, 16, 64])]; + tensor var_5551_cast_fp16 = reshape(shape = var_5550, x = written_v_45_cast_fp16)[name = tensor("op_5551_cast_fp16")]; + tensor new_v_cache_45_cast_fp16 = add(x = var_5545_cast_fp16, y = var_5551_cast_fp16)[name = tensor("new_v_cache_45_cast_fp16")]; + tensor new_v_cache_45_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_45_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_5564 = const()[name = tensor("op_5564"), val = tensor([0, 2, 1, 3])]; + tensor var_5575_transpose_x_0 = const()[name = tensor("op_5575_transpose_x_0"), val = tensor(false)]; + tensor var_5575_transpose_y_0 = const()[name = tensor("op_5575_transpose_y_0"), val = tensor(false)]; + tensor transpose_113_perm_0 = const()[name = tensor("transpose_113_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_114_perm_0 = const()[name = tensor("transpose_114_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_114 = transpose(perm = transpose_114_perm_0, x = new_k_cache_45_cast_fp16)[name = tensor("transpose_116")]; + tensor transpose_113 = transpose(perm = transpose_113_perm_0, x = q_45_cast_fp16)[name = tensor("transpose_117")]; + tensor var_5575_cast_fp16 = matmul(transpose_x = var_5575_transpose_x_0, transpose_y = var_5575_transpose_y_0, x = transpose_113, y = transpose_114)[name = tensor("op_5575_cast_fp16")]; + tensor var_5576_to_fp16 = const()[name = tensor("op_5576_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_89_cast_fp16 = mul(x = var_5575_cast_fp16, y = var_5576_to_fp16)[name = tensor("scores_89_cast_fp16")]; + tensor scores_91_cast_fp16 = add(x = scores_89_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_91_cast_fp16")]; + tensor var_5586 = const()[name = tensor("op_5586"), val = tensor(-1)]; + tensor weights_45_cast_fp16 = softmax(axis = var_5586, x = scores_91_cast_fp16)[name = tensor("weights_45_cast_fp16")]; + tensor attn_45_transpose_x_0 = const()[name = tensor("attn_45_transpose_x_0"), val = tensor(false)]; + tensor attn_45_transpose_y_0 = const()[name = tensor("attn_45_transpose_y_0"), val = tensor(false)]; + tensor values_45_cast_fp16 = transpose(perm = var_5564, x = new_v_cache_45_cast_fp16)[name = tensor("transpose_118")]; + tensor attn_45_cast_fp16 = matmul(transpose_x = attn_45_transpose_x_0, transpose_y = attn_45_transpose_y_0, x = weights_45_cast_fp16, y = values_45_cast_fp16)[name = tensor("attn_45_cast_fp16")]; + tensor var_5594 = const()[name = tensor("op_5594"), val = tensor([0, 2, 1, 3])]; + tensor var_5599 = const()[name = tensor("op_5599"), val = tensor([1, 256, 1024])]; + tensor var_5595_cast_fp16 = transpose(perm = var_5594, x = attn_45_cast_fp16)[name = tensor("transpose_115")]; + tensor input_269_cast_fp16 = reshape(shape = var_5599, x = var_5595_cast_fp16)[name = tensor("input_269_cast_fp16")]; + tensor attn22_out_proj_weight_to_fp16 = const()[name = tensor("attn22_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(560154368)))]; + tensor linear_89_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn22_out_proj_weight_to_fp16, x = input_269_cast_fp16)[name = tensor("linear_89_cast_fp16")]; + tensor position22_to_fp16_dtype_0 = const()[name = tensor("position22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position22_to_fp16 = cast(dtype = position22_to_fp16_dtype_0, x = position22)[name = tensor("cast_329")]; + tensor var_5608_cast_fp16 = add(x = position22_to_fp16, y = valid_len_to_fp16)[name = tensor("op_5608_cast_fp16")]; + tensor var_5608_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5608_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_271_cast_fp16 = add(x = input_265_cast_fp16, y = linear_89_cast_fp16)[name = tensor("input_271_cast_fp16")]; + tensor input_273_axes_0 = const()[name = tensor("input_273_axes_0"), val = tensor([-1])]; + tensor norm22_2_weight_to_fp16 = const()[name = tensor("norm22_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(562251584)))]; + tensor norm22_2_bias_to_fp16 = const()[name = tensor("norm22_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(562253696)))]; + tensor var_5612_to_fp16 = const()[name = tensor("op_5612_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_273_cast_fp16 = layer_norm(axes = input_273_axes_0, beta = norm22_2_bias_to_fp16, epsilon = var_5612_to_fp16, gamma = norm22_2_weight_to_fp16, x = input_271_cast_fp16)[name = tensor("input_273_cast_fp16")]; + tensor linear22_1_weight_to_fp16 = const()[name = tensor("linear22_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(562255808)))]; + tensor linear_90_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear22_1_weight_to_fp16, x = input_273_cast_fp16)[name = tensor("linear_90_cast_fp16")]; + tensor input_275_mode_0 = const()[name = tensor("input_275_mode_0"), val = tensor("EXACT")]; + tensor input_275_cast_fp16 = gelu(mode = input_275_mode_0, x = linear_90_cast_fp16)[name = tensor("input_275_cast_fp16")]; + tensor linear22_2_weight_to_fp16 = const()[name = tensor("linear22_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(570644480)))]; + tensor linear_91_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear22_2_weight_to_fp16, x = input_275_cast_fp16)[name = tensor("linear_91_cast_fp16")]; + tensor input_277_cast_fp16 = add(x = input_271_cast_fp16, y = linear_91_cast_fp16)[name = tensor("input_277_cast_fp16")]; + tensor input_279_axes_0 = const()[name = tensor("input_279_axes_0"), val = tensor([-1])]; + tensor norm23_1_weight_to_fp16 = const()[name = tensor("norm23_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(579033152)))]; + tensor norm23_1_bias_to_fp16 = const()[name = tensor("norm23_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(579035264)))]; + tensor var_5629_to_fp16 = const()[name = tensor("op_5629_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_279_cast_fp16 = layer_norm(axes = input_279_axes_0, beta = norm23_1_bias_to_fp16, epsilon = var_5629_to_fp16, gamma = norm23_1_weight_to_fp16, x = input_277_cast_fp16)[name = tensor("input_279_cast_fp16")]; + tensor attn23_in_proj_weight_to_fp16 = const()[name = tensor("attn23_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(579037376)))]; + tensor linear_92_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn23_in_proj_weight_to_fp16, x = input_279_cast_fp16)[name = tensor("linear_92_cast_fp16")]; + tensor var_5643 = const()[name = tensor("op_5643"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_cast_fp16 = reshape(shape = var_5643, x = linear_92_cast_fp16)[name = tensor("qkv_cast_fp16")]; + tensor x_begin_0 = const()[name = tensor("x_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_end_0 = const()[name = tensor("x_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_end_mask_0 = const()[name = tensor("x_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_squeeze_mask_0 = const()[name = tensor("x_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_cast_fp16 = slice_by_index(begin = x_begin_0, end = x_end_0, end_mask = x_end_mask_0, squeeze_mask = x_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("x_cast_fp16")]; + tensor v_begin_0 = const()[name = tensor("v_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_end_0 = const()[name = tensor("v_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_end_mask_0 = const()[name = tensor("v_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_squeeze_mask_0 = const()[name = tensor("v_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_cast_fp16 = slice_by_index(begin = v_begin_0, end = v_end_0, end_mask = v_end_mask_0, squeeze_mask = v_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("v_cast_fp16")]; + tensor var_5675 = const()[name = tensor("op_5675"), val = tensor([256, 16, 32, 2])]; + tensor pairs_cast_fp16 = reshape(shape = var_5675, x = x_cast_fp16)[name = tensor("pairs_cast_fp16")]; + tensor xr_begin_0 = const()[name = tensor("xr_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_end_0 = const()[name = tensor("xr_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_end_mask_0 = const()[name = tensor("xr_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_squeeze_mask_0 = const()[name = tensor("xr_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_cast_fp16 = slice_by_index(begin = xr_begin_0, end = xr_end_0, end_mask = xr_end_mask_0, squeeze_mask = xr_squeeze_mask_0, x = pairs_cast_fp16)[name = tensor("xr_cast_fp16")]; + tensor xi_begin_0 = const()[name = tensor("xi_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_end_0 = const()[name = tensor("xi_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_end_mask_0 = const()[name = tensor("xi_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_squeeze_mask_0 = const()[name = tensor("xi_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_cast_fp16 = slice_by_index(begin = xi_begin_0, end = xi_end_0, end_mask = xi_end_mask_0, squeeze_mask = xi_squeeze_mask_0, x = pairs_cast_fp16)[name = tensor("xi_cast_fp16")]; + tensor var_5683_cast_fp16 = mul(x = xr_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5683_cast_fp16")]; + tensor var_5684_cast_fp16 = mul(x = xi_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5684_cast_fp16")]; + tensor xor__cast_fp16 = sub(x = var_5683_cast_fp16, y = var_5684_cast_fp16)[name = tensor("xor__cast_fp16")]; + tensor var_5687_cast_fp16 = mul(x = xr_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5687_cast_fp16")]; + tensor var_5688_cast_fp16 = mul(x = xi_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5688_cast_fp16")]; + tensor xoi_cast_fp16 = add(x = var_5687_cast_fp16, y = var_5688_cast_fp16)[name = tensor("xoi_cast_fp16")]; + tensor out_axis_0 = const()[name = tensor("out_axis_0"), val = tensor(-1)]; + tensor out_cast_fp16 = stack(axis = out_axis_0, values = (xor__cast_fp16, xoi_cast_fp16))[name = tensor("out_cast_fp16")]; + tensor var_5705 = const()[name = tensor("op_5705"), val = tensor([256, 1024])]; + tensor var_5706_cast_fp16 = reshape(shape = var_5705, x = out_cast_fp16)[name = tensor("op_5706_cast_fp16")]; + tensor written_k_transpose_x_1 = const()[name = tensor("written_k_transpose_x_1"), val = tensor(true)]; + tensor written_k_transpose_y_1 = const()[name = tensor("written_k_transpose_y_1"), val = tensor(false)]; + tensor written_k_cast_fp16 = matmul(transpose_x = written_k_transpose_x_1, transpose_y = written_k_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_5706_cast_fp16)[name = tensor("written_k_cast_fp16")]; + tensor var_5713 = const()[name = tensor("op_5713"), val = tensor([256, 1024])]; + tensor var_5714_cast_fp16 = reshape(shape = var_5713, x = v_cast_fp16)[name = tensor("op_5714_cast_fp16")]; + tensor written_v_transpose_x_1 = const()[name = tensor("written_v_transpose_x_1"), val = tensor(true)]; + tensor written_v_transpose_y_1 = const()[name = tensor("written_v_transpose_y_1"), val = tensor(false)]; + tensor written_v_cast_fp16 = matmul(transpose_x = written_v_transpose_x_1, transpose_y = written_v_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_5714_cast_fp16)[name = tensor("written_v_cast_fp16")]; + tensor k_cache23_to_fp16_dtype_0 = const()[name = tensor("k_cache23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache23_to_fp16 = cast(dtype = k_cache23_to_fp16_dtype_0, x = k_cache23)[name = tensor("cast_327")]; + tensor var_5725_cast_fp16 = mul(x = k_cache23_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_5725_cast_fp16")]; + tensor var_5730 = const()[name = tensor("op_5730"), val = tensor([1, 512, 16, 64])]; + tensor var_5731_cast_fp16 = reshape(shape = var_5730, x = written_k_cast_fp16)[name = tensor("op_5731_cast_fp16")]; + tensor new_k_cache_cast_fp16 = add(x = var_5725_cast_fp16, y = var_5731_cast_fp16)[name = tensor("new_k_cache_cast_fp16")]; + tensor new_k_cache_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache23_to_fp16_dtype_0 = const()[name = tensor("v_cache23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache23_to_fp16 = cast(dtype = v_cache23_to_fp16_dtype_0, x = v_cache23)[name = tensor("cast_325")]; + tensor var_5734_cast_fp16 = mul(x = v_cache23_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_5734_cast_fp16")]; + tensor var_5739 = const()[name = tensor("op_5739"), val = tensor([1, 512, 16, 64])]; + tensor var_5740_cast_fp16 = reshape(shape = var_5739, x = written_v_cast_fp16)[name = tensor("op_5740_cast_fp16")]; + tensor new_v_cache_cast_fp16 = add(x = var_5734_cast_fp16, y = var_5740_cast_fp16)[name = tensor("new_v_cache_cast_fp16")]; + tensor new_v_cache_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor position23_to_fp16_dtype_0 = const()[name = tensor("position23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position23_to_fp16 = cast(dtype = position23_to_fp16_dtype_0, x = position23)[name = tensor("cast_323")]; + tensor var_5747_cast_fp16 = add(x = position23_to_fp16, y = valid_len_to_fp16)[name = tensor("op_5747_cast_fp16")]; + tensor var_5747_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5747_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor new_position23 = cast(dtype = var_5747_cast_fp16_to_fp32_dtype_0, x = var_5747_cast_fp16)[name = tensor("cast_322")]; + tensor new_v_cache23 = cast(dtype = new_v_cache_cast_fp16_to_fp32_dtype_0, x = new_v_cache_cast_fp16)[name = tensor("cast_324")]; + tensor new_k_cache23 = cast(dtype = new_k_cache_cast_fp16_to_fp32_dtype_0, x = new_k_cache_cast_fp16)[name = tensor("cast_326")]; + tensor new_position22 = cast(dtype = var_5608_cast_fp16_to_fp32_dtype_0, x = var_5608_cast_fp16)[name = tensor("cast_328")]; + tensor new_v_cache22 = cast(dtype = new_v_cache_45_cast_fp16_to_fp32_dtype_0, x = new_v_cache_45_cast_fp16)[name = tensor("cast_330")]; + tensor new_k_cache22 = cast(dtype = new_k_cache_45_cast_fp16_to_fp32_dtype_0, x = new_k_cache_45_cast_fp16)[name = tensor("cast_332")]; + tensor new_position21 = cast(dtype = var_5377_cast_fp16_to_fp32_dtype_0, x = var_5377_cast_fp16)[name = tensor("cast_334")]; + tensor new_v_cache21 = cast(dtype = new_v_cache_43_cast_fp16_to_fp32_dtype_0, x = new_v_cache_43_cast_fp16)[name = tensor("cast_336")]; + tensor new_k_cache21 = cast(dtype = new_k_cache_43_cast_fp16_to_fp32_dtype_0, x = new_k_cache_43_cast_fp16)[name = tensor("cast_338")]; + tensor new_position20 = cast(dtype = var_5146_cast_fp16_to_fp32_dtype_0, x = var_5146_cast_fp16)[name = tensor("cast_340")]; + tensor new_v_cache20 = cast(dtype = new_v_cache_41_cast_fp16_to_fp32_dtype_0, x = new_v_cache_41_cast_fp16)[name = tensor("cast_342")]; + tensor new_k_cache20 = cast(dtype = new_k_cache_41_cast_fp16_to_fp32_dtype_0, x = new_k_cache_41_cast_fp16)[name = tensor("cast_344")]; + tensor new_position19 = cast(dtype = var_4915_cast_fp16_to_fp32_dtype_0, x = var_4915_cast_fp16)[name = tensor("cast_346")]; + tensor new_v_cache19 = cast(dtype = new_v_cache_39_cast_fp16_to_fp32_dtype_0, x = new_v_cache_39_cast_fp16)[name = tensor("cast_348")]; + tensor new_k_cache19 = cast(dtype = new_k_cache_39_cast_fp16_to_fp32_dtype_0, x = new_k_cache_39_cast_fp16)[name = tensor("cast_350")]; + tensor new_position18 = cast(dtype = var_4684_cast_fp16_to_fp32_dtype_0, x = var_4684_cast_fp16)[name = tensor("cast_352")]; + tensor new_v_cache18 = cast(dtype = new_v_cache_37_cast_fp16_to_fp32_dtype_0, x = new_v_cache_37_cast_fp16)[name = tensor("cast_354")]; + tensor new_k_cache18 = cast(dtype = new_k_cache_37_cast_fp16_to_fp32_dtype_0, x = new_k_cache_37_cast_fp16)[name = tensor("cast_356")]; + tensor new_position17 = cast(dtype = var_4453_cast_fp16_to_fp32_dtype_0, x = var_4453_cast_fp16)[name = tensor("cast_358")]; + tensor new_v_cache17 = cast(dtype = new_v_cache_35_cast_fp16_to_fp32_dtype_0, x = new_v_cache_35_cast_fp16)[name = tensor("cast_360")]; + tensor new_k_cache17 = cast(dtype = new_k_cache_35_cast_fp16_to_fp32_dtype_0, x = new_k_cache_35_cast_fp16)[name = tensor("cast_362")]; + tensor new_position16 = cast(dtype = var_4222_cast_fp16_to_fp32_dtype_0, x = var_4222_cast_fp16)[name = tensor("cast_364")]; + tensor new_v_cache16 = cast(dtype = new_v_cache_33_cast_fp16_to_fp32_dtype_0, x = new_v_cache_33_cast_fp16)[name = tensor("cast_366")]; + tensor new_k_cache16 = cast(dtype = new_k_cache_33_cast_fp16_to_fp32_dtype_0, x = new_k_cache_33_cast_fp16)[name = tensor("cast_368")]; + tensor new_position15 = cast(dtype = var_3991_cast_fp16_to_fp32_dtype_0, x = var_3991_cast_fp16)[name = tensor("cast_370")]; + tensor new_v_cache15 = cast(dtype = new_v_cache_31_cast_fp16_to_fp32_dtype_0, x = new_v_cache_31_cast_fp16)[name = tensor("cast_372")]; + tensor new_k_cache15 = cast(dtype = new_k_cache_31_cast_fp16_to_fp32_dtype_0, x = new_k_cache_31_cast_fp16)[name = tensor("cast_374")]; + tensor new_position14 = cast(dtype = var_3760_cast_fp16_to_fp32_dtype_0, x = var_3760_cast_fp16)[name = tensor("cast_376")]; + tensor new_v_cache14 = cast(dtype = new_v_cache_29_cast_fp16_to_fp32_dtype_0, x = new_v_cache_29_cast_fp16)[name = tensor("cast_378")]; + tensor new_k_cache14 = cast(dtype = new_k_cache_29_cast_fp16_to_fp32_dtype_0, x = new_k_cache_29_cast_fp16)[name = tensor("cast_380")]; + tensor new_position13 = cast(dtype = var_3529_cast_fp16_to_fp32_dtype_0, x = var_3529_cast_fp16)[name = tensor("cast_382")]; + tensor new_v_cache13 = cast(dtype = new_v_cache_27_cast_fp16_to_fp32_dtype_0, x = new_v_cache_27_cast_fp16)[name = tensor("cast_384")]; + tensor new_k_cache13 = cast(dtype = new_k_cache_27_cast_fp16_to_fp32_dtype_0, x = new_k_cache_27_cast_fp16)[name = tensor("cast_386")]; + tensor new_position12 = cast(dtype = var_3298_cast_fp16_to_fp32_dtype_0, x = var_3298_cast_fp16)[name = tensor("cast_388")]; + tensor new_v_cache12 = cast(dtype = new_v_cache_25_cast_fp16_to_fp32_dtype_0, x = new_v_cache_25_cast_fp16)[name = tensor("cast_390")]; + tensor new_k_cache12 = cast(dtype = new_k_cache_25_cast_fp16_to_fp32_dtype_0, x = new_k_cache_25_cast_fp16)[name = tensor("cast_392")]; + tensor new_position11 = cast(dtype = var_3067_cast_fp16_to_fp32_dtype_0, x = var_3067_cast_fp16)[name = tensor("cast_394")]; + tensor new_v_cache11 = cast(dtype = new_v_cache_23_cast_fp16_to_fp32_dtype_0, x = new_v_cache_23_cast_fp16)[name = tensor("cast_396")]; + tensor new_k_cache11 = cast(dtype = new_k_cache_23_cast_fp16_to_fp32_dtype_0, x = new_k_cache_23_cast_fp16)[name = tensor("cast_398")]; + tensor new_position10 = cast(dtype = var_2836_cast_fp16_to_fp32_dtype_0, x = var_2836_cast_fp16)[name = tensor("cast_400")]; + tensor new_v_cache10 = cast(dtype = new_v_cache_21_cast_fp16_to_fp32_dtype_0, x = new_v_cache_21_cast_fp16)[name = tensor("cast_402")]; + tensor new_k_cache10 = cast(dtype = new_k_cache_21_cast_fp16_to_fp32_dtype_0, x = new_k_cache_21_cast_fp16)[name = tensor("cast_404")]; + tensor new_position9 = cast(dtype = var_2605_cast_fp16_to_fp32_dtype_0, x = var_2605_cast_fp16)[name = tensor("cast_406")]; + tensor new_v_cache9 = cast(dtype = new_v_cache_19_cast_fp16_to_fp32_dtype_0, x = new_v_cache_19_cast_fp16)[name = tensor("cast_408")]; + tensor new_k_cache9 = cast(dtype = new_k_cache_19_cast_fp16_to_fp32_dtype_0, x = new_k_cache_19_cast_fp16)[name = tensor("cast_410")]; + tensor new_position8 = cast(dtype = var_2374_cast_fp16_to_fp32_dtype_0, x = var_2374_cast_fp16)[name = tensor("cast_412")]; + tensor new_v_cache8 = cast(dtype = new_v_cache_17_cast_fp16_to_fp32_dtype_0, x = new_v_cache_17_cast_fp16)[name = tensor("cast_414")]; + tensor new_k_cache8 = cast(dtype = new_k_cache_17_cast_fp16_to_fp32_dtype_0, x = new_k_cache_17_cast_fp16)[name = tensor("cast_416")]; + tensor new_position7 = cast(dtype = var_2143_cast_fp16_to_fp32_dtype_0, x = var_2143_cast_fp16)[name = tensor("cast_418")]; + tensor new_v_cache7 = cast(dtype = new_v_cache_15_cast_fp16_to_fp32_dtype_0, x = new_v_cache_15_cast_fp16)[name = tensor("cast_420")]; + tensor new_k_cache7 = cast(dtype = new_k_cache_15_cast_fp16_to_fp32_dtype_0, x = new_k_cache_15_cast_fp16)[name = tensor("cast_422")]; + tensor new_position6 = cast(dtype = var_1912_cast_fp16_to_fp32_dtype_0, x = var_1912_cast_fp16)[name = tensor("cast_424")]; + tensor new_v_cache6 = cast(dtype = new_v_cache_13_cast_fp16_to_fp32_dtype_0, x = new_v_cache_13_cast_fp16)[name = tensor("cast_426")]; + tensor new_k_cache6 = cast(dtype = new_k_cache_13_cast_fp16_to_fp32_dtype_0, x = new_k_cache_13_cast_fp16)[name = tensor("cast_428")]; + tensor new_position5 = cast(dtype = var_1681_cast_fp16_to_fp32_dtype_0, x = var_1681_cast_fp16)[name = tensor("cast_430")]; + tensor new_v_cache5 = cast(dtype = new_v_cache_11_cast_fp16_to_fp32_dtype_0, x = new_v_cache_11_cast_fp16)[name = tensor("cast_432")]; + tensor new_k_cache5 = cast(dtype = new_k_cache_11_cast_fp16_to_fp32_dtype_0, x = new_k_cache_11_cast_fp16)[name = tensor("cast_434")]; + tensor new_position4 = cast(dtype = var_1450_cast_fp16_to_fp32_dtype_0, x = var_1450_cast_fp16)[name = tensor("cast_436")]; + tensor new_v_cache4 = cast(dtype = new_v_cache_9_cast_fp16_to_fp32_dtype_0, x = new_v_cache_9_cast_fp16)[name = tensor("cast_438")]; + tensor new_k_cache4 = cast(dtype = new_k_cache_9_cast_fp16_to_fp32_dtype_0, x = new_k_cache_9_cast_fp16)[name = tensor("cast_440")]; + tensor new_position3 = cast(dtype = var_1219_cast_fp16_to_fp32_dtype_0, x = var_1219_cast_fp16)[name = tensor("cast_442")]; + tensor new_v_cache3 = cast(dtype = new_v_cache_7_cast_fp16_to_fp32_dtype_0, x = new_v_cache_7_cast_fp16)[name = tensor("cast_444")]; + tensor new_k_cache3 = cast(dtype = new_k_cache_7_cast_fp16_to_fp32_dtype_0, x = new_k_cache_7_cast_fp16)[name = tensor("cast_446")]; + tensor new_position2 = cast(dtype = var_988_cast_fp16_to_fp32_dtype_0, x = var_988_cast_fp16)[name = tensor("cast_448")]; + tensor new_v_cache2 = cast(dtype = new_v_cache_5_cast_fp16_to_fp32_dtype_0, x = new_v_cache_5_cast_fp16)[name = tensor("cast_450")]; + tensor new_k_cache2 = cast(dtype = new_k_cache_5_cast_fp16_to_fp32_dtype_0, x = new_k_cache_5_cast_fp16)[name = tensor("cast_452")]; + tensor new_position1 = cast(dtype = var_757_cast_fp16_to_fp32_dtype_0, x = var_757_cast_fp16)[name = tensor("cast_454")]; + tensor new_v_cache1 = cast(dtype = new_v_cache_3_cast_fp16_to_fp32_dtype_0, x = new_v_cache_3_cast_fp16)[name = tensor("cast_456")]; + tensor new_k_cache1 = cast(dtype = new_k_cache_3_cast_fp16_to_fp32_dtype_0, x = new_k_cache_3_cast_fp16)[name = tensor("cast_458")]; + tensor new_position0 = cast(dtype = var_526_cast_fp16_to_fp32_dtype_0, x = var_297_cast_fp16)[name = tensor("cast_460")]; + tensor new_v_cache0 = cast(dtype = new_v_cache_1_cast_fp16_to_fp32_dtype_0, x = new_v_cache_1_cast_fp16)[name = tensor("cast_461")]; + tensor new_k_cache0 = cast(dtype = new_k_cache_1_cast_fp16_to_fp32_dtype_0, x = new_k_cache_1_cast_fp16)[name = tensor("cast_463")]; + } -> (new_k_cache0, new_v_cache0, new_position0, new_k_cache1, new_v_cache1, new_position1, new_k_cache2, new_v_cache2, new_position2, new_k_cache3, new_v_cache3, new_position3, new_k_cache4, new_v_cache4, new_position4, new_k_cache5, new_v_cache5, new_position5, new_k_cache6, new_v_cache6, new_position6, new_k_cache7, new_v_cache7, new_position7, new_k_cache8, new_v_cache8, new_position8, new_k_cache9, new_v_cache9, new_position9, new_k_cache10, new_v_cache10, new_position10, new_k_cache11, new_v_cache11, new_position11, new_k_cache12, new_v_cache12, new_position12, new_k_cache13, new_v_cache13, new_position13, new_k_cache14, new_v_cache14, new_position14, new_k_cache15, new_v_cache15, new_position15, new_k_cache16, new_v_cache16, new_position16, new_k_cache17, new_v_cache17, new_position17, new_k_cache18, new_v_cache18, new_position18, new_k_cache19, new_v_cache19, new_position19, new_k_cache20, new_v_cache20, new_position20, new_k_cache21, new_v_cache21, new_position21, new_k_cache22, new_v_cache22, new_position22, new_k_cache23, new_v_cache23, new_position23); +} \ No newline at end of file diff --git a/v2.1/french_24l/cond_prefill_ane.mlmodelc/weights/weight.bin b/v2.1/french_24l/cond_prefill_ane.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..ee1012eb0432419c41e9f5b56b2a0e85022b2742 --- /dev/null +++ b/v2.1/french_24l/cond_prefill_ane.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ac7fcaa3e59effe50cf24838f2c101f44025eb377fc80d1a31cdc4f4b6c91414 +size 585328896 diff --git a/v2.1/french_24l/cond_prefill_ane.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/french_24l/cond_prefill_ane.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..fa36e18629c97bdd9ae507d4ec7f5ccb7e294762 --- /dev/null +++ b/v2.1/french_24l/cond_prefill_ane.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:28a0aeec6c1cf574b2af67a46a41fdd3a0afe8c242a4a7dfd5a193a9099ecde2 +size 483717 diff --git a/v2.1/french_24l/cond_prefill_ane.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/french_24l/cond_prefill_ane.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..ee1012eb0432419c41e9f5b56b2a0e85022b2742 --- /dev/null +++ b/v2.1/french_24l/cond_prefill_ane.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ac7fcaa3e59effe50cf24838f2c101f44025eb377fc80d1a31cdc4f4b6c91414 +size 585328896 diff --git a/v2.1/french_24l/cond_prefill_ane.mlpackage/Manifest.json b/v2.1/french_24l/cond_prefill_ane.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..cab6d07bca4c9612b4a820e74d86778fc8f3ee6d --- /dev/null +++ b/v2.1/french_24l/cond_prefill_ane.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "16CB2E73-068A-4A55-83BE-26CA21ECA245": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "FAA33DD1-8377-41CB-B1B9-D200B192AB46": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "FAA33DD1-8377-41CB-B1B9-D200B192AB46" +} diff --git a/v2.1/french_24l/flowlm_step_ane.mlmodelc/analytics/coremldata.bin b/v2.1/french_24l/flowlm_step_ane.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..da88fba01ed1c6b9cf0769fab1ceb14a40556562 --- /dev/null +++ b/v2.1/french_24l/flowlm_step_ane.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1eb6aa4b3af494b35c8efdaafcb887f78f07a11585973309edb4d5a51a4a2fc1 +size 243 diff --git a/v2.1/french_24l/flowlm_step_ane.mlmodelc/coremldata.bin b/v2.1/french_24l/flowlm_step_ane.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..b88ab092d303f342c311d6df5cd0ed12be002f54 --- /dev/null +++ b/v2.1/french_24l/flowlm_step_ane.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ac58bb3dc9e618b3b0ff1fb442fb4409e23a929d343fd1330a578cad4bc32e5b +size 4522 diff --git a/v2.1/french_24l/flowlm_step_ane.mlmodelc/model.mil b/v2.1/french_24l/flowlm_step_ane.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..505e9f1f543fd3e9e3fa4d558ffb04e1256fb146 --- /dev/null +++ b/v2.1/french_24l/flowlm_step_ane.mlmodelc/model.mil @@ -0,0 +1,3699 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.9.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor k_cache0, tensor k_cache1, tensor k_cache10, tensor k_cache11, tensor k_cache12, tensor k_cache13, tensor k_cache14, tensor k_cache15, tensor k_cache16, tensor k_cache17, tensor k_cache18, tensor k_cache19, tensor k_cache2, tensor k_cache20, tensor k_cache21, tensor k_cache22, tensor k_cache23, tensor k_cache3, tensor k_cache4, tensor k_cache5, tensor k_cache6, tensor k_cache7, tensor k_cache8, tensor k_cache9, tensor position0, tensor position1, tensor position10, tensor position11, tensor position12, tensor position13, tensor position14, tensor position15, tensor position16, tensor position17, tensor position18, tensor position19, tensor position2, tensor position20, tensor position21, tensor position22, tensor position23, tensor position3, tensor position4, tensor position5, tensor position6, tensor position7, tensor position8, tensor position9, tensor sequence, tensor v_cache0, tensor v_cache1, tensor v_cache10, tensor v_cache11, tensor v_cache12, tensor v_cache13, tensor v_cache14, tensor v_cache15, tensor v_cache16, tensor v_cache17, tensor v_cache18, tensor v_cache19, tensor v_cache2, tensor v_cache20, tensor v_cache21, tensor v_cache22, tensor v_cache23, tensor v_cache3, tensor v_cache4, tensor v_cache5, tensor v_cache6, tensor v_cache7, tensor v_cache8, tensor v_cache9) { + tensor sequence_to_fp16_dtype_0 = const()[name = tensor("sequence_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor input_linear_weight_to_fp16 = const()[name = tensor("input_linear_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(65664)))]; + tensor sequence_to_fp16 = cast(dtype = sequence_to_fp16_dtype_0, x = sequence)[name = tensor("cast_578")]; + tensor linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = input_linear_weight_to_fp16, x = sequence_to_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; + tensor norm0_1_weight_to_fp16 = const()[name = tensor("norm0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67776)))]; + tensor norm0_1_bias_to_fp16 = const()[name = tensor("norm0_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(69888)))]; + tensor var_227_to_fp16 = const()[name = tensor("op_227_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_5_cast_fp16 = layer_norm(axes = input_5_axes_0, beta = norm0_1_bias_to_fp16, epsilon = var_227_to_fp16, gamma = norm0_1_weight_to_fp16, x = linear_0_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor attn0_in_proj_weight_to_fp16 = const()[name = tensor("attn0_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(72000)))]; + tensor linear_1_bias_0_to_fp16 = const()[name = tensor("linear_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6363520)))]; + tensor linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn0_in_proj_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("linear_1_cast_fp16")]; + tensor var_240 = const()[name = tensor("op_240"), val = tensor([1, 3, 16, 64])]; + tensor qkv_1_cast_fp16 = reshape(shape = var_240, x = linear_1_cast_fp16)[name = tensor("qkv_1_cast_fp16")]; + tensor x_1_begin_0 = const()[name = tensor("x_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_1_end_0 = const()[name = tensor("x_1_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_1_end_mask_0 = const()[name = tensor("x_1_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_1_squeeze_mask_0 = const()[name = tensor("x_1_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_1_cast_fp16 = slice_by_index(begin = x_1_begin_0, end = x_1_end_0, end_mask = x_1_end_mask_0, squeeze_mask = x_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("x_1_cast_fp16")]; + tensor x_3_begin_0 = const()[name = tensor("x_3_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_3_end_0 = const()[name = tensor("x_3_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_3_end_mask_0 = const()[name = tensor("x_3_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_3_squeeze_mask_0 = const()[name = tensor("x_3_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_3_cast_fp16 = slice_by_index(begin = x_3_begin_0, end = x_3_end_0, end_mask = x_3_end_mask_0, squeeze_mask = x_3_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("x_3_cast_fp16")]; + tensor v_1_begin_0 = const()[name = tensor("v_1_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_1_end_0 = const()[name = tensor("v_1_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_1_end_mask_0 = const()[name = tensor("v_1_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_1_squeeze_mask_0 = const()[name = tensor("v_1_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_1_cast_fp16 = slice_by_index(begin = v_1_begin_0, end = v_1_end_0, end_mask = v_1_end_mask_0, squeeze_mask = v_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("v_1_cast_fp16")]; + tensor var_269 = const()[name = tensor("op_269"), val = tensor([1, 1, 1])]; + tensor position0_to_fp16_dtype_0 = const()[name = tensor("position0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position0_to_fp16 = cast(dtype = position0_to_fp16_dtype_0, x = position0)[name = tensor("cast_577")]; + tensor var_270_cast_fp16 = reshape(shape = var_269, x = position0_to_fp16)[name = tensor("op_270_cast_fp16")]; + tensor rope_freqs_to_fp16 = const()[name = tensor("rope_freqs_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6369728)))]; + tensor angles_1_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_270_cast_fp16)[name = tensor("angles_1_cast_fp16")]; + tensor rotr_1_cast_fp16 = cos(x = angles_1_cast_fp16)[name = tensor("rotr_1_cast_fp16")]; + tensor roti_1_cast_fp16 = sin(x = angles_1_cast_fp16)[name = tensor("roti_1_cast_fp16")]; + tensor var_278 = const()[name = tensor("op_278"), val = tensor([1, 16, 32, 2])]; + tensor pairs_1_cast_fp16 = reshape(shape = var_278, x = x_1_cast_fp16)[name = tensor("pairs_1_cast_fp16")]; + tensor xr_1_begin_0 = const()[name = tensor("xr_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_1_end_0 = const()[name = tensor("xr_1_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_1_end_mask_0 = const()[name = tensor("xr_1_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_1_squeeze_mask_0 = const()[name = tensor("xr_1_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_1_cast_fp16 = slice_by_index(begin = xr_1_begin_0, end = xr_1_end_0, end_mask = xr_1_end_mask_0, squeeze_mask = xr_1_squeeze_mask_0, x = pairs_1_cast_fp16)[name = tensor("xr_1_cast_fp16")]; + tensor xi_1_begin_0 = const()[name = tensor("xi_1_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_1_end_0 = const()[name = tensor("xi_1_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_1_end_mask_0 = const()[name = tensor("xi_1_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_1_squeeze_mask_0 = const()[name = tensor("xi_1_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_1_cast_fp16 = slice_by_index(begin = xi_1_begin_0, end = xi_1_end_0, end_mask = xi_1_end_mask_0, squeeze_mask = xi_1_squeeze_mask_0, x = pairs_1_cast_fp16)[name = tensor("xi_1_cast_fp16")]; + tensor var_286_cast_fp16 = mul(x = xr_1_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_286_cast_fp16")]; + tensor var_287_cast_fp16 = mul(x = xi_1_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_287_cast_fp16")]; + tensor xor__1_cast_fp16 = sub(x = var_286_cast_fp16, y = var_287_cast_fp16)[name = tensor("xor__1_cast_fp16")]; + tensor var_290_cast_fp16 = mul(x = xr_1_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_290_cast_fp16")]; + tensor var_291_cast_fp16 = mul(x = xi_1_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_291_cast_fp16")]; + tensor xoi_1_cast_fp16 = add(x = var_290_cast_fp16, y = var_291_cast_fp16)[name = tensor("xoi_1_cast_fp16")]; + tensor out_1_axis_0 = const()[name = tensor("out_1_axis_0"), val = tensor(-1)]; + tensor out_1_cast_fp16 = stack(axis = out_1_axis_0, values = (xor__1_cast_fp16, xoi_1_cast_fp16))[name = tensor("out_1_cast_fp16")]; + tensor var_314 = const()[name = tensor("op_314"), val = tensor([1, 16, 32, 2])]; + tensor pairs_3_cast_fp16 = reshape(shape = var_314, x = x_3_cast_fp16)[name = tensor("pairs_3_cast_fp16")]; + tensor xr_3_begin_0 = const()[name = tensor("xr_3_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_3_end_0 = const()[name = tensor("xr_3_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_3_end_mask_0 = const()[name = tensor("xr_3_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_3_squeeze_mask_0 = const()[name = tensor("xr_3_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_3_cast_fp16 = slice_by_index(begin = xr_3_begin_0, end = xr_3_end_0, end_mask = xr_3_end_mask_0, squeeze_mask = xr_3_squeeze_mask_0, x = pairs_3_cast_fp16)[name = tensor("xr_3_cast_fp16")]; + tensor xi_3_begin_0 = const()[name = tensor("xi_3_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_3_end_0 = const()[name = tensor("xi_3_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_3_end_mask_0 = const()[name = tensor("xi_3_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_3_squeeze_mask_0 = const()[name = tensor("xi_3_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_3_cast_fp16 = slice_by_index(begin = xi_3_begin_0, end = xi_3_end_0, end_mask = xi_3_end_mask_0, squeeze_mask = xi_3_squeeze_mask_0, x = pairs_3_cast_fp16)[name = tensor("xi_3_cast_fp16")]; + tensor var_322_cast_fp16 = mul(x = xr_3_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_322_cast_fp16")]; + tensor var_323_cast_fp16 = mul(x = xi_3_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_323_cast_fp16")]; + tensor xor__3_cast_fp16 = sub(x = var_322_cast_fp16, y = var_323_cast_fp16)[name = tensor("xor__3_cast_fp16")]; + tensor var_326_cast_fp16 = mul(x = xr_3_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_326_cast_fp16")]; + tensor var_327_cast_fp16 = mul(x = xi_3_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_327_cast_fp16")]; + tensor xoi_3_cast_fp16 = add(x = var_326_cast_fp16, y = var_327_cast_fp16)[name = tensor("xoi_3_cast_fp16")]; + tensor out_3_axis_0 = const()[name = tensor("out_3_axis_0"), val = tensor(-1)]; + tensor out_3_cast_fp16 = stack(axis = out_3_axis_0, values = (xor__3_cast_fp16, xoi_3_cast_fp16))[name = tensor("out_3_cast_fp16")]; + tensor _inversed_339_y_0_to_fp16 = const()[name = tensor("_inversed_339_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_339_cast_fp16 = mul(x = position0_to_fp16, y = _inversed_339_y_0_to_fp16)[name = tensor("_inversed_339_cast_fp16")]; + tensor var_340_cast_fp16 = floor(x = _inversed_339_cast_fp16)[name = tensor("op_340_cast_fp16")]; + tensor var_341_to_fp16 = const()[name = tensor("op_341_to_fp16"), val = tensor(0x1p+9)]; + tensor var_342_cast_fp16 = mul(x = var_340_cast_fp16, y = var_341_to_fp16)[name = tensor("op_342_cast_fp16")]; + tensor write_idx_1_cast_fp16 = sub(x = position0_to_fp16, y = var_342_cast_fp16)[name = tensor("write_idx_1_cast_fp16")]; + tensor slot_indices_to_fp16 = const()[name = tensor("slot_indices_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6369856)))]; + tensor var_345_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_1_cast_fp16)[name = tensor("op_345_cast_fp16")]; + tensor var_355 = const()[name = tensor("op_355"), val = tensor([1, 512, 1, 1])]; + tensor onehot_1_to_fp16_dtype_0 = const()[name = tensor("onehot_1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_345_cast_fp16_to_fp16 = cast(dtype = onehot_1_to_fp16_dtype_0, x = var_345_cast_fp16)[name = tensor("cast_576")]; + tensor onehot_3_cast_fp16 = reshape(shape = var_355, x = var_345_cast_fp16_to_fp16)[name = tensor("onehot_3_cast_fp16")]; + tensor var_357_to_fp16 = const()[name = tensor("op_357_to_fp16"), val = tensor(0x1p+0)]; + tensor var_359_cast_fp16 = sub(x = var_357_to_fp16, y = onehot_3_cast_fp16)[name = tensor("op_359_cast_fp16")]; + tensor k_cache0_to_fp16_dtype_0 = const()[name = tensor("k_cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache0_to_fp16 = cast(dtype = k_cache0_to_fp16_dtype_0, x = k_cache0)[name = tensor("cast_575")]; + tensor var_360_cast_fp16 = mul(x = k_cache0_to_fp16, y = var_359_cast_fp16)[name = tensor("op_360_cast_fp16")]; + tensor var_365 = const()[name = tensor("op_365"), val = tensor([1, 1, 16, 64])]; + tensor var_366_cast_fp16 = reshape(shape = var_365, x = out_3_cast_fp16)[name = tensor("op_366_cast_fp16")]; + tensor var_367_cast_fp16 = mul(x = var_366_cast_fp16, y = onehot_3_cast_fp16)[name = tensor("op_367_cast_fp16")]; + tensor new_k_cache_1_cast_fp16 = add(x = var_360_cast_fp16, y = var_367_cast_fp16)[name = tensor("new_k_cache_1_cast_fp16")]; + tensor new_k_cache_1_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_1_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache0_to_fp16_dtype_0 = const()[name = tensor("v_cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache0_to_fp16 = cast(dtype = v_cache0_to_fp16_dtype_0, x = v_cache0)[name = tensor("cast_573")]; + tensor var_373_cast_fp16 = mul(x = v_cache0_to_fp16, y = var_359_cast_fp16)[name = tensor("op_373_cast_fp16")]; + tensor var_378 = const()[name = tensor("op_378"), val = tensor([1, 1, 16, 64])]; + tensor var_379_cast_fp16 = reshape(shape = var_378, x = v_1_cast_fp16)[name = tensor("op_379_cast_fp16")]; + tensor var_380_cast_fp16 = mul(x = var_379_cast_fp16, y = onehot_3_cast_fp16)[name = tensor("op_380_cast_fp16")]; + tensor new_v_cache_1_cast_fp16 = add(x = var_373_cast_fp16, y = var_380_cast_fp16)[name = tensor("new_v_cache_1_cast_fp16")]; + tensor new_v_cache_1_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_1_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_0_perm_0 = const()[name = tensor("transpose_0_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_393 = const()[name = tensor("op_393"), val = tensor([0, 2, 1, 3])]; + tensor var_399 = const()[name = tensor("op_399"), val = tensor([1, 16, 1, 64])]; + tensor q4_1_cast_fp16 = reshape(shape = var_399, x = out_1_cast_fp16)[name = tensor("q4_1_cast_fp16")]; + tensor var_404_transpose_x_0 = const()[name = tensor("op_404_transpose_x_0"), val = tensor(false)]; + tensor var_404_transpose_y_0 = const()[name = tensor("op_404_transpose_y_0"), val = tensor(false)]; + tensor transpose_0_cast_fp16 = transpose(perm = transpose_0_perm_0, x = new_k_cache_1_cast_fp16)[name = tensor("transpose_95")]; + tensor var_404_cast_fp16 = matmul(transpose_x = var_404_transpose_x_0, transpose_y = var_404_transpose_y_0, x = q4_1_cast_fp16, y = transpose_0_cast_fp16)[name = tensor("op_404_cast_fp16")]; + tensor var_405_to_fp16 = const()[name = tensor("op_405_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_1_cast_fp16 = mul(x = var_404_cast_fp16, y = var_405_to_fp16)[name = tensor("scores_1_cast_fp16")]; + tensor var_409 = const()[name = tensor("op_409"), val = tensor([1, 1])]; + tensor var_410_cast_fp16 = reshape(shape = var_409, x = position0_to_fp16)[name = tensor("op_410_cast_fp16")]; + tensor var_411_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_410_cast_fp16)[name = tensor("op_411_cast_fp16")]; + tensor mask_1_to_fp16_dtype_0 = const()[name = tensor("mask_1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_418_to_fp16 = const()[name = tensor("op_418_to_fp16"), val = tensor(0x1p+0)]; + tensor var_411_cast_fp16_to_fp16 = cast(dtype = mask_1_to_fp16_dtype_0, x = var_411_cast_fp16)[name = tensor("cast_571")]; + tensor var_419_cast_fp16 = sub(x = var_411_cast_fp16_to_fp16, y = var_418_to_fp16)[name = tensor("op_419_cast_fp16")]; + tensor var_424 = const()[name = tensor("op_424"), val = tensor([1, 1, 1, 512])]; + tensor var_425_cast_fp16 = reshape(shape = var_424, x = var_419_cast_fp16)[name = tensor("op_425_cast_fp16")]; + tensor var_426_to_fp16 = const()[name = tensor("op_426_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_427_cast_fp16 = mul(x = var_425_cast_fp16, y = var_426_to_fp16)[name = tensor("op_427_cast_fp16")]; + tensor scores_3_cast_fp16 = add(x = scores_1_cast_fp16, y = var_427_cast_fp16)[name = tensor("scores_3_cast_fp16")]; + tensor var_430 = const()[name = tensor("op_430"), val = tensor(-1)]; + tensor weights_1_cast_fp16 = softmax(axis = var_430, x = scores_3_cast_fp16)[name = tensor("weights_1_cast_fp16")]; + tensor attn_1_transpose_x_0 = const()[name = tensor("attn_1_transpose_x_0"), val = tensor(false)]; + tensor attn_1_transpose_y_0 = const()[name = tensor("attn_1_transpose_y_0"), val = tensor(false)]; + tensor values_1_cast_fp16 = transpose(perm = var_393, x = new_v_cache_1_cast_fp16)[name = tensor("transpose_94")]; + tensor attn_1_cast_fp16 = matmul(transpose_x = attn_1_transpose_x_0, transpose_y = attn_1_transpose_y_0, x = weights_1_cast_fp16, y = values_1_cast_fp16)[name = tensor("attn_1_cast_fp16")]; + tensor var_438 = const()[name = tensor("op_438"), val = tensor([0, 2, 1, 3])]; + tensor var_443 = const()[name = tensor("op_443"), val = tensor([1, 1, 1024])]; + tensor var_439_cast_fp16 = transpose(perm = var_438, x = attn_1_cast_fp16)[name = tensor("transpose_93")]; + tensor input_7_cast_fp16 = reshape(shape = var_443, x = var_439_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor attn0_out_proj_weight_to_fp16 = const()[name = tensor("attn0_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6370944)))]; + tensor linear_2_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn0_out_proj_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor var_449_to_fp16 = const()[name = tensor("op_449_to_fp16"), val = tensor(0x1p+0)]; + tensor var_450_cast_fp16 = add(x = position0_to_fp16, y = var_449_to_fp16)[name = tensor("op_450_cast_fp16")]; + tensor var_450_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_450_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_9_cast_fp16 = add(x = linear_0_cast_fp16, y = linear_2_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor input_11_axes_0 = const()[name = tensor("input_11_axes_0"), val = tensor([-1])]; + tensor norm0_2_weight_to_fp16 = const()[name = tensor("norm0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8468160)))]; + tensor norm0_2_bias_to_fp16 = const()[name = tensor("norm0_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8470272)))]; + tensor var_454_to_fp16 = const()[name = tensor("op_454_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_11_cast_fp16 = layer_norm(axes = input_11_axes_0, beta = norm0_2_bias_to_fp16, epsilon = var_454_to_fp16, gamma = norm0_2_weight_to_fp16, x = input_9_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor linear0_1_weight_to_fp16 = const()[name = tensor("linear0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8472384)))]; + tensor linear_3_bias_0_to_fp16 = const()[name = tensor("linear_3_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16861056)))]; + tensor linear_3_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear0_1_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("linear_3_cast_fp16")]; + tensor input_13_mode_0 = const()[name = tensor("input_13_mode_0"), val = tensor("EXACT")]; + tensor input_13_cast_fp16 = gelu(mode = input_13_mode_0, x = linear_3_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor linear0_2_weight_to_fp16 = const()[name = tensor("linear0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16869312)))]; + tensor linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear0_2_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor input_15_cast_fp16 = add(x = input_9_cast_fp16, y = linear_4_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor input_17_axes_0 = const()[name = tensor("input_17_axes_0"), val = tensor([-1])]; + tensor norm1_1_weight_to_fp16 = const()[name = tensor("norm1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25257984)))]; + tensor norm1_1_bias_to_fp16 = const()[name = tensor("norm1_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25260096)))]; + tensor var_471_to_fp16 = const()[name = tensor("op_471_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_17_cast_fp16 = layer_norm(axes = input_17_axes_0, beta = norm1_1_bias_to_fp16, epsilon = var_471_to_fp16, gamma = norm1_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor attn1_in_proj_weight_to_fp16 = const()[name = tensor("attn1_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25262208)))]; + tensor linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn1_in_proj_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor var_484 = const()[name = tensor("op_484"), val = tensor([1, 3, 16, 64])]; + tensor qkv_3_cast_fp16 = reshape(shape = var_484, x = linear_5_cast_fp16)[name = tensor("qkv_3_cast_fp16")]; + tensor x_5_begin_0 = const()[name = tensor("x_5_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_5_end_0 = const()[name = tensor("x_5_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_5_end_mask_0 = const()[name = tensor("x_5_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_5_squeeze_mask_0 = const()[name = tensor("x_5_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_5_cast_fp16 = slice_by_index(begin = x_5_begin_0, end = x_5_end_0, end_mask = x_5_end_mask_0, squeeze_mask = x_5_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("x_5_cast_fp16")]; + tensor x_7_begin_0 = const()[name = tensor("x_7_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_7_end_0 = const()[name = tensor("x_7_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_7_end_mask_0 = const()[name = tensor("x_7_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_7_squeeze_mask_0 = const()[name = tensor("x_7_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_7_cast_fp16 = slice_by_index(begin = x_7_begin_0, end = x_7_end_0, end_mask = x_7_end_mask_0, squeeze_mask = x_7_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("x_7_cast_fp16")]; + tensor v_3_begin_0 = const()[name = tensor("v_3_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_3_end_0 = const()[name = tensor("v_3_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_3_end_mask_0 = const()[name = tensor("v_3_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_3_squeeze_mask_0 = const()[name = tensor("v_3_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_3_cast_fp16 = slice_by_index(begin = v_3_begin_0, end = v_3_end_0, end_mask = v_3_end_mask_0, squeeze_mask = v_3_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("v_3_cast_fp16")]; + tensor var_513 = const()[name = tensor("op_513"), val = tensor([1, 1, 1])]; + tensor position1_to_fp16_dtype_0 = const()[name = tensor("position1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position1_to_fp16 = cast(dtype = position1_to_fp16_dtype_0, x = position1)[name = tensor("cast_569")]; + tensor var_514_cast_fp16 = reshape(shape = var_513, x = position1_to_fp16)[name = tensor("op_514_cast_fp16")]; + tensor angles_5_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_514_cast_fp16)[name = tensor("angles_5_cast_fp16")]; + tensor rotr_5_cast_fp16 = cos(x = angles_5_cast_fp16)[name = tensor("rotr_5_cast_fp16")]; + tensor roti_5_cast_fp16 = sin(x = angles_5_cast_fp16)[name = tensor("roti_5_cast_fp16")]; + tensor var_522 = const()[name = tensor("op_522"), val = tensor([1, 16, 32, 2])]; + tensor pairs_5_cast_fp16 = reshape(shape = var_522, x = x_5_cast_fp16)[name = tensor("pairs_5_cast_fp16")]; + tensor xr_5_begin_0 = const()[name = tensor("xr_5_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_5_end_0 = const()[name = tensor("xr_5_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_5_end_mask_0 = const()[name = tensor("xr_5_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_5_squeeze_mask_0 = const()[name = tensor("xr_5_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_5_cast_fp16 = slice_by_index(begin = xr_5_begin_0, end = xr_5_end_0, end_mask = xr_5_end_mask_0, squeeze_mask = xr_5_squeeze_mask_0, x = pairs_5_cast_fp16)[name = tensor("xr_5_cast_fp16")]; + tensor xi_5_begin_0 = const()[name = tensor("xi_5_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_5_end_0 = const()[name = tensor("xi_5_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_5_end_mask_0 = const()[name = tensor("xi_5_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_5_squeeze_mask_0 = const()[name = tensor("xi_5_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_5_cast_fp16 = slice_by_index(begin = xi_5_begin_0, end = xi_5_end_0, end_mask = xi_5_end_mask_0, squeeze_mask = xi_5_squeeze_mask_0, x = pairs_5_cast_fp16)[name = tensor("xi_5_cast_fp16")]; + tensor var_530_cast_fp16 = mul(x = xr_5_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_530_cast_fp16")]; + tensor var_531_cast_fp16 = mul(x = xi_5_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_531_cast_fp16")]; + tensor xor__5_cast_fp16 = sub(x = var_530_cast_fp16, y = var_531_cast_fp16)[name = tensor("xor__5_cast_fp16")]; + tensor var_534_cast_fp16 = mul(x = xr_5_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_534_cast_fp16")]; + tensor var_535_cast_fp16 = mul(x = xi_5_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_535_cast_fp16")]; + tensor xoi_5_cast_fp16 = add(x = var_534_cast_fp16, y = var_535_cast_fp16)[name = tensor("xoi_5_cast_fp16")]; + tensor out_5_axis_0 = const()[name = tensor("out_5_axis_0"), val = tensor(-1)]; + tensor out_5_cast_fp16 = stack(axis = out_5_axis_0, values = (xor__5_cast_fp16, xoi_5_cast_fp16))[name = tensor("out_5_cast_fp16")]; + tensor var_558 = const()[name = tensor("op_558"), val = tensor([1, 16, 32, 2])]; + tensor pairs_7_cast_fp16 = reshape(shape = var_558, x = x_7_cast_fp16)[name = tensor("pairs_7_cast_fp16")]; + tensor xr_7_begin_0 = const()[name = tensor("xr_7_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_7_end_0 = const()[name = tensor("xr_7_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_7_end_mask_0 = const()[name = tensor("xr_7_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_7_squeeze_mask_0 = const()[name = tensor("xr_7_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_7_cast_fp16 = slice_by_index(begin = xr_7_begin_0, end = xr_7_end_0, end_mask = xr_7_end_mask_0, squeeze_mask = xr_7_squeeze_mask_0, x = pairs_7_cast_fp16)[name = tensor("xr_7_cast_fp16")]; + tensor xi_7_begin_0 = const()[name = tensor("xi_7_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_7_end_0 = const()[name = tensor("xi_7_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_7_end_mask_0 = const()[name = tensor("xi_7_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_7_squeeze_mask_0 = const()[name = tensor("xi_7_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_7_cast_fp16 = slice_by_index(begin = xi_7_begin_0, end = xi_7_end_0, end_mask = xi_7_end_mask_0, squeeze_mask = xi_7_squeeze_mask_0, x = pairs_7_cast_fp16)[name = tensor("xi_7_cast_fp16")]; + tensor var_566_cast_fp16 = mul(x = xr_7_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_566_cast_fp16")]; + tensor var_567_cast_fp16 = mul(x = xi_7_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_567_cast_fp16")]; + tensor xor__7_cast_fp16 = sub(x = var_566_cast_fp16, y = var_567_cast_fp16)[name = tensor("xor__7_cast_fp16")]; + tensor var_570_cast_fp16 = mul(x = xr_7_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_570_cast_fp16")]; + tensor var_571_cast_fp16 = mul(x = xi_7_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_571_cast_fp16")]; + tensor xoi_7_cast_fp16 = add(x = var_570_cast_fp16, y = var_571_cast_fp16)[name = tensor("xoi_7_cast_fp16")]; + tensor out_7_axis_0 = const()[name = tensor("out_7_axis_0"), val = tensor(-1)]; + tensor out_7_cast_fp16 = stack(axis = out_7_axis_0, values = (xor__7_cast_fp16, xoi_7_cast_fp16))[name = tensor("out_7_cast_fp16")]; + tensor _inversed_583_y_0_to_fp16 = const()[name = tensor("_inversed_583_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_583_cast_fp16 = mul(x = position1_to_fp16, y = _inversed_583_y_0_to_fp16)[name = tensor("_inversed_583_cast_fp16")]; + tensor var_584_cast_fp16 = floor(x = _inversed_583_cast_fp16)[name = tensor("op_584_cast_fp16")]; + tensor var_585_to_fp16 = const()[name = tensor("op_585_to_fp16"), val = tensor(0x1p+9)]; + tensor var_586_cast_fp16 = mul(x = var_584_cast_fp16, y = var_585_to_fp16)[name = tensor("op_586_cast_fp16")]; + tensor write_idx_3_cast_fp16 = sub(x = position1_to_fp16, y = var_586_cast_fp16)[name = tensor("write_idx_3_cast_fp16")]; + tensor var_589_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_3_cast_fp16)[name = tensor("op_589_cast_fp16")]; + tensor var_599 = const()[name = tensor("op_599"), val = tensor([1, 512, 1, 1])]; + tensor onehot_5_to_fp16_dtype_0 = const()[name = tensor("onehot_5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_589_cast_fp16_to_fp16 = cast(dtype = onehot_5_to_fp16_dtype_0, x = var_589_cast_fp16)[name = tensor("cast_568")]; + tensor onehot_7_cast_fp16 = reshape(shape = var_599, x = var_589_cast_fp16_to_fp16)[name = tensor("onehot_7_cast_fp16")]; + tensor var_601_to_fp16 = const()[name = tensor("op_601_to_fp16"), val = tensor(0x1p+0)]; + tensor var_603_cast_fp16 = sub(x = var_601_to_fp16, y = onehot_7_cast_fp16)[name = tensor("op_603_cast_fp16")]; + tensor k_cache1_to_fp16_dtype_0 = const()[name = tensor("k_cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache1_to_fp16 = cast(dtype = k_cache1_to_fp16_dtype_0, x = k_cache1)[name = tensor("cast_567")]; + tensor var_604_cast_fp16 = mul(x = k_cache1_to_fp16, y = var_603_cast_fp16)[name = tensor("op_604_cast_fp16")]; + tensor var_609 = const()[name = tensor("op_609"), val = tensor([1, 1, 16, 64])]; + tensor var_610_cast_fp16 = reshape(shape = var_609, x = out_7_cast_fp16)[name = tensor("op_610_cast_fp16")]; + tensor var_611_cast_fp16 = mul(x = var_610_cast_fp16, y = onehot_7_cast_fp16)[name = tensor("op_611_cast_fp16")]; + tensor new_k_cache_3_cast_fp16 = add(x = var_604_cast_fp16, y = var_611_cast_fp16)[name = tensor("new_k_cache_3_cast_fp16")]; + tensor new_k_cache_3_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_3_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache1_to_fp16_dtype_0 = const()[name = tensor("v_cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache1_to_fp16 = cast(dtype = v_cache1_to_fp16_dtype_0, x = v_cache1)[name = tensor("cast_565")]; + tensor var_617_cast_fp16 = mul(x = v_cache1_to_fp16, y = var_603_cast_fp16)[name = tensor("op_617_cast_fp16")]; + tensor var_622 = const()[name = tensor("op_622"), val = tensor([1, 1, 16, 64])]; + tensor var_623_cast_fp16 = reshape(shape = var_622, x = v_3_cast_fp16)[name = tensor("op_623_cast_fp16")]; + tensor var_624_cast_fp16 = mul(x = var_623_cast_fp16, y = onehot_7_cast_fp16)[name = tensor("op_624_cast_fp16")]; + tensor new_v_cache_3_cast_fp16 = add(x = var_617_cast_fp16, y = var_624_cast_fp16)[name = tensor("new_v_cache_3_cast_fp16")]; + tensor new_v_cache_3_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_3_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_1_perm_0 = const()[name = tensor("transpose_1_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_637 = const()[name = tensor("op_637"), val = tensor([0, 2, 1, 3])]; + tensor var_643 = const()[name = tensor("op_643"), val = tensor([1, 16, 1, 64])]; + tensor q4_3_cast_fp16 = reshape(shape = var_643, x = out_5_cast_fp16)[name = tensor("q4_3_cast_fp16")]; + tensor var_648_transpose_x_0 = const()[name = tensor("op_648_transpose_x_0"), val = tensor(false)]; + tensor var_648_transpose_y_0 = const()[name = tensor("op_648_transpose_y_0"), val = tensor(false)]; + tensor transpose_1_cast_fp16 = transpose(perm = transpose_1_perm_0, x = new_k_cache_3_cast_fp16)[name = tensor("transpose_92")]; + tensor var_648_cast_fp16 = matmul(transpose_x = var_648_transpose_x_0, transpose_y = var_648_transpose_y_0, x = q4_3_cast_fp16, y = transpose_1_cast_fp16)[name = tensor("op_648_cast_fp16")]; + tensor var_649_to_fp16 = const()[name = tensor("op_649_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_5_cast_fp16 = mul(x = var_648_cast_fp16, y = var_649_to_fp16)[name = tensor("scores_5_cast_fp16")]; + tensor var_653 = const()[name = tensor("op_653"), val = tensor([1, 1])]; + tensor var_654_cast_fp16 = reshape(shape = var_653, x = position1_to_fp16)[name = tensor("op_654_cast_fp16")]; + tensor var_655_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_654_cast_fp16)[name = tensor("op_655_cast_fp16")]; + tensor mask_3_to_fp16_dtype_0 = const()[name = tensor("mask_3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_662_to_fp16 = const()[name = tensor("op_662_to_fp16"), val = tensor(0x1p+0)]; + tensor var_655_cast_fp16_to_fp16 = cast(dtype = mask_3_to_fp16_dtype_0, x = var_655_cast_fp16)[name = tensor("cast_563")]; + tensor var_663_cast_fp16 = sub(x = var_655_cast_fp16_to_fp16, y = var_662_to_fp16)[name = tensor("op_663_cast_fp16")]; + tensor var_668 = const()[name = tensor("op_668"), val = tensor([1, 1, 1, 512])]; + tensor var_669_cast_fp16 = reshape(shape = var_668, x = var_663_cast_fp16)[name = tensor("op_669_cast_fp16")]; + tensor var_670_to_fp16 = const()[name = tensor("op_670_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_671_cast_fp16 = mul(x = var_669_cast_fp16, y = var_670_to_fp16)[name = tensor("op_671_cast_fp16")]; + tensor scores_7_cast_fp16 = add(x = scores_5_cast_fp16, y = var_671_cast_fp16)[name = tensor("scores_7_cast_fp16")]; + tensor var_674 = const()[name = tensor("op_674"), val = tensor(-1)]; + tensor weights_3_cast_fp16 = softmax(axis = var_674, x = scores_7_cast_fp16)[name = tensor("weights_3_cast_fp16")]; + tensor attn_3_transpose_x_0 = const()[name = tensor("attn_3_transpose_x_0"), val = tensor(false)]; + tensor attn_3_transpose_y_0 = const()[name = tensor("attn_3_transpose_y_0"), val = tensor(false)]; + tensor values_3_cast_fp16 = transpose(perm = var_637, x = new_v_cache_3_cast_fp16)[name = tensor("transpose_91")]; + tensor attn_3_cast_fp16 = matmul(transpose_x = attn_3_transpose_x_0, transpose_y = attn_3_transpose_y_0, x = weights_3_cast_fp16, y = values_3_cast_fp16)[name = tensor("attn_3_cast_fp16")]; + tensor var_682 = const()[name = tensor("op_682"), val = tensor([0, 2, 1, 3])]; + tensor var_687 = const()[name = tensor("op_687"), val = tensor([1, 1, 1024])]; + tensor var_683_cast_fp16 = transpose(perm = var_682, x = attn_3_cast_fp16)[name = tensor("transpose_90")]; + tensor input_19_cast_fp16 = reshape(shape = var_687, x = var_683_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor attn1_out_proj_weight_to_fp16 = const()[name = tensor("attn1_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31553728)))]; + tensor linear_6_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn1_out_proj_weight_to_fp16, x = input_19_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor var_693_to_fp16 = const()[name = tensor("op_693_to_fp16"), val = tensor(0x1p+0)]; + tensor var_694_cast_fp16 = add(x = position1_to_fp16, y = var_693_to_fp16)[name = tensor("op_694_cast_fp16")]; + tensor var_694_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_694_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_21_cast_fp16 = add(x = input_15_cast_fp16, y = linear_6_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor input_23_axes_0 = const()[name = tensor("input_23_axes_0"), val = tensor([-1])]; + tensor norm1_2_weight_to_fp16 = const()[name = tensor("norm1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33650944)))]; + tensor norm1_2_bias_to_fp16 = const()[name = tensor("norm1_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33653056)))]; + tensor var_698_to_fp16 = const()[name = tensor("op_698_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_23_cast_fp16 = layer_norm(axes = input_23_axes_0, beta = norm1_2_bias_to_fp16, epsilon = var_698_to_fp16, gamma = norm1_2_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor linear1_1_weight_to_fp16 = const()[name = tensor("linear1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33655168)))]; + tensor linear_7_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear1_1_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_25_mode_0 = const()[name = tensor("input_25_mode_0"), val = tensor("EXACT")]; + tensor input_25_cast_fp16 = gelu(mode = input_25_mode_0, x = linear_7_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor linear1_2_weight_to_fp16 = const()[name = tensor("linear1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42043840)))]; + tensor linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear1_2_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor input_27_cast_fp16 = add(x = input_21_cast_fp16, y = linear_8_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor input_29_axes_0 = const()[name = tensor("input_29_axes_0"), val = tensor([-1])]; + tensor norm2_1_weight_to_fp16 = const()[name = tensor("norm2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50432512)))]; + tensor norm2_1_bias_to_fp16 = const()[name = tensor("norm2_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50434624)))]; + tensor var_715_to_fp16 = const()[name = tensor("op_715_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_29_cast_fp16 = layer_norm(axes = input_29_axes_0, beta = norm2_1_bias_to_fp16, epsilon = var_715_to_fp16, gamma = norm2_1_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor attn2_in_proj_weight_to_fp16 = const()[name = tensor("attn2_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50436736)))]; + tensor linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn2_in_proj_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor var_728 = const()[name = tensor("op_728"), val = tensor([1, 3, 16, 64])]; + tensor qkv_5_cast_fp16 = reshape(shape = var_728, x = linear_9_cast_fp16)[name = tensor("qkv_5_cast_fp16")]; + tensor x_9_begin_0 = const()[name = tensor("x_9_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_9_end_0 = const()[name = tensor("x_9_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_9_end_mask_0 = const()[name = tensor("x_9_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_9_squeeze_mask_0 = const()[name = tensor("x_9_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_9_cast_fp16 = slice_by_index(begin = x_9_begin_0, end = x_9_end_0, end_mask = x_9_end_mask_0, squeeze_mask = x_9_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("x_9_cast_fp16")]; + tensor x_11_begin_0 = const()[name = tensor("x_11_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_11_end_0 = const()[name = tensor("x_11_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_11_end_mask_0 = const()[name = tensor("x_11_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_11_squeeze_mask_0 = const()[name = tensor("x_11_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_11_cast_fp16 = slice_by_index(begin = x_11_begin_0, end = x_11_end_0, end_mask = x_11_end_mask_0, squeeze_mask = x_11_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("x_11_cast_fp16")]; + tensor v_5_begin_0 = const()[name = tensor("v_5_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_5_end_0 = const()[name = tensor("v_5_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_5_end_mask_0 = const()[name = tensor("v_5_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_5_squeeze_mask_0 = const()[name = tensor("v_5_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_5_cast_fp16 = slice_by_index(begin = v_5_begin_0, end = v_5_end_0, end_mask = v_5_end_mask_0, squeeze_mask = v_5_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("v_5_cast_fp16")]; + tensor var_757 = const()[name = tensor("op_757"), val = tensor([1, 1, 1])]; + tensor position2_to_fp16_dtype_0 = const()[name = tensor("position2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position2_to_fp16 = cast(dtype = position2_to_fp16_dtype_0, x = position2)[name = tensor("cast_561")]; + tensor var_758_cast_fp16 = reshape(shape = var_757, x = position2_to_fp16)[name = tensor("op_758_cast_fp16")]; + tensor angles_9_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_758_cast_fp16)[name = tensor("angles_9_cast_fp16")]; + tensor rotr_9_cast_fp16 = cos(x = angles_9_cast_fp16)[name = tensor("rotr_9_cast_fp16")]; + tensor roti_9_cast_fp16 = sin(x = angles_9_cast_fp16)[name = tensor("roti_9_cast_fp16")]; + tensor var_766 = const()[name = tensor("op_766"), val = tensor([1, 16, 32, 2])]; + tensor pairs_9_cast_fp16 = reshape(shape = var_766, x = x_9_cast_fp16)[name = tensor("pairs_9_cast_fp16")]; + tensor xr_9_begin_0 = const()[name = tensor("xr_9_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_9_end_0 = const()[name = tensor("xr_9_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_9_end_mask_0 = const()[name = tensor("xr_9_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_9_squeeze_mask_0 = const()[name = tensor("xr_9_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_9_cast_fp16 = slice_by_index(begin = xr_9_begin_0, end = xr_9_end_0, end_mask = xr_9_end_mask_0, squeeze_mask = xr_9_squeeze_mask_0, x = pairs_9_cast_fp16)[name = tensor("xr_9_cast_fp16")]; + tensor xi_9_begin_0 = const()[name = tensor("xi_9_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_9_end_0 = const()[name = tensor("xi_9_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_9_end_mask_0 = const()[name = tensor("xi_9_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_9_squeeze_mask_0 = const()[name = tensor("xi_9_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_9_cast_fp16 = slice_by_index(begin = xi_9_begin_0, end = xi_9_end_0, end_mask = xi_9_end_mask_0, squeeze_mask = xi_9_squeeze_mask_0, x = pairs_9_cast_fp16)[name = tensor("xi_9_cast_fp16")]; + tensor var_774_cast_fp16 = mul(x = xr_9_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_774_cast_fp16")]; + tensor var_775_cast_fp16 = mul(x = xi_9_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_775_cast_fp16")]; + tensor xor__9_cast_fp16 = sub(x = var_774_cast_fp16, y = var_775_cast_fp16)[name = tensor("xor__9_cast_fp16")]; + tensor var_778_cast_fp16 = mul(x = xr_9_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_778_cast_fp16")]; + tensor var_779_cast_fp16 = mul(x = xi_9_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_779_cast_fp16")]; + tensor xoi_9_cast_fp16 = add(x = var_778_cast_fp16, y = var_779_cast_fp16)[name = tensor("xoi_9_cast_fp16")]; + tensor out_9_axis_0 = const()[name = tensor("out_9_axis_0"), val = tensor(-1)]; + tensor out_9_cast_fp16 = stack(axis = out_9_axis_0, values = (xor__9_cast_fp16, xoi_9_cast_fp16))[name = tensor("out_9_cast_fp16")]; + tensor var_802 = const()[name = tensor("op_802"), val = tensor([1, 16, 32, 2])]; + tensor pairs_11_cast_fp16 = reshape(shape = var_802, x = x_11_cast_fp16)[name = tensor("pairs_11_cast_fp16")]; + tensor xr_11_begin_0 = const()[name = tensor("xr_11_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_11_end_0 = const()[name = tensor("xr_11_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_11_end_mask_0 = const()[name = tensor("xr_11_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_11_squeeze_mask_0 = const()[name = tensor("xr_11_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_11_cast_fp16 = slice_by_index(begin = xr_11_begin_0, end = xr_11_end_0, end_mask = xr_11_end_mask_0, squeeze_mask = xr_11_squeeze_mask_0, x = pairs_11_cast_fp16)[name = tensor("xr_11_cast_fp16")]; + tensor xi_11_begin_0 = const()[name = tensor("xi_11_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_11_end_0 = const()[name = tensor("xi_11_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_11_end_mask_0 = const()[name = tensor("xi_11_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_11_squeeze_mask_0 = const()[name = tensor("xi_11_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_11_cast_fp16 = slice_by_index(begin = xi_11_begin_0, end = xi_11_end_0, end_mask = xi_11_end_mask_0, squeeze_mask = xi_11_squeeze_mask_0, x = pairs_11_cast_fp16)[name = tensor("xi_11_cast_fp16")]; + tensor var_810_cast_fp16 = mul(x = xr_11_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_810_cast_fp16")]; + tensor var_811_cast_fp16 = mul(x = xi_11_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_811_cast_fp16")]; + tensor xor__11_cast_fp16 = sub(x = var_810_cast_fp16, y = var_811_cast_fp16)[name = tensor("xor__11_cast_fp16")]; + tensor var_814_cast_fp16 = mul(x = xr_11_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_814_cast_fp16")]; + tensor var_815_cast_fp16 = mul(x = xi_11_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_815_cast_fp16")]; + tensor xoi_11_cast_fp16 = add(x = var_814_cast_fp16, y = var_815_cast_fp16)[name = tensor("xoi_11_cast_fp16")]; + tensor out_11_axis_0 = const()[name = tensor("out_11_axis_0"), val = tensor(-1)]; + tensor out_11_cast_fp16 = stack(axis = out_11_axis_0, values = (xor__11_cast_fp16, xoi_11_cast_fp16))[name = tensor("out_11_cast_fp16")]; + tensor _inversed_827_y_0_to_fp16 = const()[name = tensor("_inversed_827_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_827_cast_fp16 = mul(x = position2_to_fp16, y = _inversed_827_y_0_to_fp16)[name = tensor("_inversed_827_cast_fp16")]; + tensor var_828_cast_fp16 = floor(x = _inversed_827_cast_fp16)[name = tensor("op_828_cast_fp16")]; + tensor var_829_to_fp16 = const()[name = tensor("op_829_to_fp16"), val = tensor(0x1p+9)]; + tensor var_830_cast_fp16 = mul(x = var_828_cast_fp16, y = var_829_to_fp16)[name = tensor("op_830_cast_fp16")]; + tensor write_idx_5_cast_fp16 = sub(x = position2_to_fp16, y = var_830_cast_fp16)[name = tensor("write_idx_5_cast_fp16")]; + tensor var_833_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_5_cast_fp16)[name = tensor("op_833_cast_fp16")]; + tensor var_843 = const()[name = tensor("op_843"), val = tensor([1, 512, 1, 1])]; + tensor onehot_9_to_fp16_dtype_0 = const()[name = tensor("onehot_9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_833_cast_fp16_to_fp16 = cast(dtype = onehot_9_to_fp16_dtype_0, x = var_833_cast_fp16)[name = tensor("cast_560")]; + tensor onehot_11_cast_fp16 = reshape(shape = var_843, x = var_833_cast_fp16_to_fp16)[name = tensor("onehot_11_cast_fp16")]; + tensor var_845_to_fp16 = const()[name = tensor("op_845_to_fp16"), val = tensor(0x1p+0)]; + tensor var_847_cast_fp16 = sub(x = var_845_to_fp16, y = onehot_11_cast_fp16)[name = tensor("op_847_cast_fp16")]; + tensor k_cache2_to_fp16_dtype_0 = const()[name = tensor("k_cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache2_to_fp16 = cast(dtype = k_cache2_to_fp16_dtype_0, x = k_cache2)[name = tensor("cast_559")]; + tensor var_848_cast_fp16 = mul(x = k_cache2_to_fp16, y = var_847_cast_fp16)[name = tensor("op_848_cast_fp16")]; + tensor var_853 = const()[name = tensor("op_853"), val = tensor([1, 1, 16, 64])]; + tensor var_854_cast_fp16 = reshape(shape = var_853, x = out_11_cast_fp16)[name = tensor("op_854_cast_fp16")]; + tensor var_855_cast_fp16 = mul(x = var_854_cast_fp16, y = onehot_11_cast_fp16)[name = tensor("op_855_cast_fp16")]; + tensor new_k_cache_5_cast_fp16 = add(x = var_848_cast_fp16, y = var_855_cast_fp16)[name = tensor("new_k_cache_5_cast_fp16")]; + tensor new_k_cache_5_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_5_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache2_to_fp16_dtype_0 = const()[name = tensor("v_cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache2_to_fp16 = cast(dtype = v_cache2_to_fp16_dtype_0, x = v_cache2)[name = tensor("cast_557")]; + tensor var_861_cast_fp16 = mul(x = v_cache2_to_fp16, y = var_847_cast_fp16)[name = tensor("op_861_cast_fp16")]; + tensor var_866 = const()[name = tensor("op_866"), val = tensor([1, 1, 16, 64])]; + tensor var_867_cast_fp16 = reshape(shape = var_866, x = v_5_cast_fp16)[name = tensor("op_867_cast_fp16")]; + tensor var_868_cast_fp16 = mul(x = var_867_cast_fp16, y = onehot_11_cast_fp16)[name = tensor("op_868_cast_fp16")]; + tensor new_v_cache_5_cast_fp16 = add(x = var_861_cast_fp16, y = var_868_cast_fp16)[name = tensor("new_v_cache_5_cast_fp16")]; + tensor new_v_cache_5_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_5_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_2_perm_0 = const()[name = tensor("transpose_2_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_881 = const()[name = tensor("op_881"), val = tensor([0, 2, 1, 3])]; + tensor var_887 = const()[name = tensor("op_887"), val = tensor([1, 16, 1, 64])]; + tensor q4_5_cast_fp16 = reshape(shape = var_887, x = out_9_cast_fp16)[name = tensor("q4_5_cast_fp16")]; + tensor var_892_transpose_x_0 = const()[name = tensor("op_892_transpose_x_0"), val = tensor(false)]; + tensor var_892_transpose_y_0 = const()[name = tensor("op_892_transpose_y_0"), val = tensor(false)]; + tensor transpose_2_cast_fp16 = transpose(perm = transpose_2_perm_0, x = new_k_cache_5_cast_fp16)[name = tensor("transpose_89")]; + tensor var_892_cast_fp16 = matmul(transpose_x = var_892_transpose_x_0, transpose_y = var_892_transpose_y_0, x = q4_5_cast_fp16, y = transpose_2_cast_fp16)[name = tensor("op_892_cast_fp16")]; + tensor var_893_to_fp16 = const()[name = tensor("op_893_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_9_cast_fp16 = mul(x = var_892_cast_fp16, y = var_893_to_fp16)[name = tensor("scores_9_cast_fp16")]; + tensor var_897 = const()[name = tensor("op_897"), val = tensor([1, 1])]; + tensor var_898_cast_fp16 = reshape(shape = var_897, x = position2_to_fp16)[name = tensor("op_898_cast_fp16")]; + tensor var_899_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_898_cast_fp16)[name = tensor("op_899_cast_fp16")]; + tensor mask_5_to_fp16_dtype_0 = const()[name = tensor("mask_5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_906_to_fp16 = const()[name = tensor("op_906_to_fp16"), val = tensor(0x1p+0)]; + tensor var_899_cast_fp16_to_fp16 = cast(dtype = mask_5_to_fp16_dtype_0, x = var_899_cast_fp16)[name = tensor("cast_555")]; + tensor var_907_cast_fp16 = sub(x = var_899_cast_fp16_to_fp16, y = var_906_to_fp16)[name = tensor("op_907_cast_fp16")]; + tensor var_912 = const()[name = tensor("op_912"), val = tensor([1, 1, 1, 512])]; + tensor var_913_cast_fp16 = reshape(shape = var_912, x = var_907_cast_fp16)[name = tensor("op_913_cast_fp16")]; + tensor var_914_to_fp16 = const()[name = tensor("op_914_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_915_cast_fp16 = mul(x = var_913_cast_fp16, y = var_914_to_fp16)[name = tensor("op_915_cast_fp16")]; + tensor scores_11_cast_fp16 = add(x = scores_9_cast_fp16, y = var_915_cast_fp16)[name = tensor("scores_11_cast_fp16")]; + tensor var_918 = const()[name = tensor("op_918"), val = tensor(-1)]; + tensor weights_5_cast_fp16 = softmax(axis = var_918, x = scores_11_cast_fp16)[name = tensor("weights_5_cast_fp16")]; + tensor attn_5_transpose_x_0 = const()[name = tensor("attn_5_transpose_x_0"), val = tensor(false)]; + tensor attn_5_transpose_y_0 = const()[name = tensor("attn_5_transpose_y_0"), val = tensor(false)]; + tensor values_5_cast_fp16 = transpose(perm = var_881, x = new_v_cache_5_cast_fp16)[name = tensor("transpose_88")]; + tensor attn_5_cast_fp16 = matmul(transpose_x = attn_5_transpose_x_0, transpose_y = attn_5_transpose_y_0, x = weights_5_cast_fp16, y = values_5_cast_fp16)[name = tensor("attn_5_cast_fp16")]; + tensor var_926 = const()[name = tensor("op_926"), val = tensor([0, 2, 1, 3])]; + tensor var_931 = const()[name = tensor("op_931"), val = tensor([1, 1, 1024])]; + tensor var_927_cast_fp16 = transpose(perm = var_926, x = attn_5_cast_fp16)[name = tensor("transpose_87")]; + tensor input_31_cast_fp16 = reshape(shape = var_931, x = var_927_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor attn2_out_proj_weight_to_fp16 = const()[name = tensor("attn2_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(56728256)))]; + tensor linear_10_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn2_out_proj_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor var_937_to_fp16 = const()[name = tensor("op_937_to_fp16"), val = tensor(0x1p+0)]; + tensor var_938_cast_fp16 = add(x = position2_to_fp16, y = var_937_to_fp16)[name = tensor("op_938_cast_fp16")]; + tensor var_938_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_938_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_33_cast_fp16 = add(x = input_27_cast_fp16, y = linear_10_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor input_35_axes_0 = const()[name = tensor("input_35_axes_0"), val = tensor([-1])]; + tensor norm2_2_weight_to_fp16 = const()[name = tensor("norm2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58825472)))]; + tensor norm2_2_bias_to_fp16 = const()[name = tensor("norm2_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58827584)))]; + tensor var_942_to_fp16 = const()[name = tensor("op_942_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_35_cast_fp16 = layer_norm(axes = input_35_axes_0, beta = norm2_2_bias_to_fp16, epsilon = var_942_to_fp16, gamma = norm2_2_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor linear2_1_weight_to_fp16 = const()[name = tensor("linear2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58829696)))]; + tensor linear_11_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear2_1_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor input_37_mode_0 = const()[name = tensor("input_37_mode_0"), val = tensor("EXACT")]; + tensor input_37_cast_fp16 = gelu(mode = input_37_mode_0, x = linear_11_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor linear2_2_weight_to_fp16 = const()[name = tensor("linear2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67218368)))]; + tensor linear_12_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear2_2_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor input_39_cast_fp16 = add(x = input_33_cast_fp16, y = linear_12_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor input_41_axes_0 = const()[name = tensor("input_41_axes_0"), val = tensor([-1])]; + tensor norm3_1_weight_to_fp16 = const()[name = tensor("norm3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75607040)))]; + tensor norm3_1_bias_to_fp16 = const()[name = tensor("norm3_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75609152)))]; + tensor var_959_to_fp16 = const()[name = tensor("op_959_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_41_cast_fp16 = layer_norm(axes = input_41_axes_0, beta = norm3_1_bias_to_fp16, epsilon = var_959_to_fp16, gamma = norm3_1_weight_to_fp16, x = input_39_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor attn3_in_proj_weight_to_fp16 = const()[name = tensor("attn3_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75611264)))]; + tensor linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn3_in_proj_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor var_972 = const()[name = tensor("op_972"), val = tensor([1, 3, 16, 64])]; + tensor qkv_7_cast_fp16 = reshape(shape = var_972, x = linear_13_cast_fp16)[name = tensor("qkv_7_cast_fp16")]; + tensor x_13_begin_0 = const()[name = tensor("x_13_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_13_end_0 = const()[name = tensor("x_13_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_13_end_mask_0 = const()[name = tensor("x_13_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_13_squeeze_mask_0 = const()[name = tensor("x_13_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_13_cast_fp16 = slice_by_index(begin = x_13_begin_0, end = x_13_end_0, end_mask = x_13_end_mask_0, squeeze_mask = x_13_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("x_13_cast_fp16")]; + tensor x_15_begin_0 = const()[name = tensor("x_15_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_15_end_0 = const()[name = tensor("x_15_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_15_end_mask_0 = const()[name = tensor("x_15_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_15_squeeze_mask_0 = const()[name = tensor("x_15_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_15_cast_fp16 = slice_by_index(begin = x_15_begin_0, end = x_15_end_0, end_mask = x_15_end_mask_0, squeeze_mask = x_15_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("x_15_cast_fp16")]; + tensor v_7_begin_0 = const()[name = tensor("v_7_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_7_end_0 = const()[name = tensor("v_7_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_7_end_mask_0 = const()[name = tensor("v_7_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_7_squeeze_mask_0 = const()[name = tensor("v_7_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_7_cast_fp16 = slice_by_index(begin = v_7_begin_0, end = v_7_end_0, end_mask = v_7_end_mask_0, squeeze_mask = v_7_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("v_7_cast_fp16")]; + tensor var_1001 = const()[name = tensor("op_1001"), val = tensor([1, 1, 1])]; + tensor position3_to_fp16_dtype_0 = const()[name = tensor("position3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position3_to_fp16 = cast(dtype = position3_to_fp16_dtype_0, x = position3)[name = tensor("cast_553")]; + tensor var_1002_cast_fp16 = reshape(shape = var_1001, x = position3_to_fp16)[name = tensor("op_1002_cast_fp16")]; + tensor angles_13_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_1002_cast_fp16)[name = tensor("angles_13_cast_fp16")]; + tensor rotr_13_cast_fp16 = cos(x = angles_13_cast_fp16)[name = tensor("rotr_13_cast_fp16")]; + tensor roti_13_cast_fp16 = sin(x = angles_13_cast_fp16)[name = tensor("roti_13_cast_fp16")]; + tensor var_1010 = const()[name = tensor("op_1010"), val = tensor([1, 16, 32, 2])]; + tensor pairs_13_cast_fp16 = reshape(shape = var_1010, x = x_13_cast_fp16)[name = tensor("pairs_13_cast_fp16")]; + tensor xr_13_begin_0 = const()[name = tensor("xr_13_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_13_end_0 = const()[name = tensor("xr_13_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_13_end_mask_0 = const()[name = tensor("xr_13_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_13_squeeze_mask_0 = const()[name = tensor("xr_13_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_13_cast_fp16 = slice_by_index(begin = xr_13_begin_0, end = xr_13_end_0, end_mask = xr_13_end_mask_0, squeeze_mask = xr_13_squeeze_mask_0, x = pairs_13_cast_fp16)[name = tensor("xr_13_cast_fp16")]; + tensor xi_13_begin_0 = const()[name = tensor("xi_13_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_13_end_0 = const()[name = tensor("xi_13_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_13_end_mask_0 = const()[name = tensor("xi_13_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_13_squeeze_mask_0 = const()[name = tensor("xi_13_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_13_cast_fp16 = slice_by_index(begin = xi_13_begin_0, end = xi_13_end_0, end_mask = xi_13_end_mask_0, squeeze_mask = xi_13_squeeze_mask_0, x = pairs_13_cast_fp16)[name = tensor("xi_13_cast_fp16")]; + tensor var_1018_cast_fp16 = mul(x = xr_13_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_1018_cast_fp16")]; + tensor var_1019_cast_fp16 = mul(x = xi_13_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_1019_cast_fp16")]; + tensor xor__13_cast_fp16 = sub(x = var_1018_cast_fp16, y = var_1019_cast_fp16)[name = tensor("xor__13_cast_fp16")]; + tensor var_1022_cast_fp16 = mul(x = xr_13_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_1022_cast_fp16")]; + tensor var_1023_cast_fp16 = mul(x = xi_13_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_1023_cast_fp16")]; + tensor xoi_13_cast_fp16 = add(x = var_1022_cast_fp16, y = var_1023_cast_fp16)[name = tensor("xoi_13_cast_fp16")]; + tensor out_13_axis_0 = const()[name = tensor("out_13_axis_0"), val = tensor(-1)]; + tensor out_13_cast_fp16 = stack(axis = out_13_axis_0, values = (xor__13_cast_fp16, xoi_13_cast_fp16))[name = tensor("out_13_cast_fp16")]; + tensor var_1046 = const()[name = tensor("op_1046"), val = tensor([1, 16, 32, 2])]; + tensor pairs_15_cast_fp16 = reshape(shape = var_1046, x = x_15_cast_fp16)[name = tensor("pairs_15_cast_fp16")]; + tensor xr_15_begin_0 = const()[name = tensor("xr_15_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_15_end_0 = const()[name = tensor("xr_15_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_15_end_mask_0 = const()[name = tensor("xr_15_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_15_squeeze_mask_0 = const()[name = tensor("xr_15_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_15_cast_fp16 = slice_by_index(begin = xr_15_begin_0, end = xr_15_end_0, end_mask = xr_15_end_mask_0, squeeze_mask = xr_15_squeeze_mask_0, x = pairs_15_cast_fp16)[name = tensor("xr_15_cast_fp16")]; + tensor xi_15_begin_0 = const()[name = tensor("xi_15_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_15_end_0 = const()[name = tensor("xi_15_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_15_end_mask_0 = const()[name = tensor("xi_15_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_15_squeeze_mask_0 = const()[name = tensor("xi_15_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_15_cast_fp16 = slice_by_index(begin = xi_15_begin_0, end = xi_15_end_0, end_mask = xi_15_end_mask_0, squeeze_mask = xi_15_squeeze_mask_0, x = pairs_15_cast_fp16)[name = tensor("xi_15_cast_fp16")]; + tensor var_1054_cast_fp16 = mul(x = xr_15_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_1054_cast_fp16")]; + tensor var_1055_cast_fp16 = mul(x = xi_15_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_1055_cast_fp16")]; + tensor xor__15_cast_fp16 = sub(x = var_1054_cast_fp16, y = var_1055_cast_fp16)[name = tensor("xor__15_cast_fp16")]; + tensor var_1058_cast_fp16 = mul(x = xr_15_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_1058_cast_fp16")]; + tensor var_1059_cast_fp16 = mul(x = xi_15_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_1059_cast_fp16")]; + tensor xoi_15_cast_fp16 = add(x = var_1058_cast_fp16, y = var_1059_cast_fp16)[name = tensor("xoi_15_cast_fp16")]; + tensor out_15_axis_0 = const()[name = tensor("out_15_axis_0"), val = tensor(-1)]; + tensor out_15_cast_fp16 = stack(axis = out_15_axis_0, values = (xor__15_cast_fp16, xoi_15_cast_fp16))[name = tensor("out_15_cast_fp16")]; + tensor _inversed_1071_y_0_to_fp16 = const()[name = tensor("_inversed_1071_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1071_cast_fp16 = mul(x = position3_to_fp16, y = _inversed_1071_y_0_to_fp16)[name = tensor("_inversed_1071_cast_fp16")]; + tensor var_1072_cast_fp16 = floor(x = _inversed_1071_cast_fp16)[name = tensor("op_1072_cast_fp16")]; + tensor var_1073_to_fp16 = const()[name = tensor("op_1073_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1074_cast_fp16 = mul(x = var_1072_cast_fp16, y = var_1073_to_fp16)[name = tensor("op_1074_cast_fp16")]; + tensor write_idx_7_cast_fp16 = sub(x = position3_to_fp16, y = var_1074_cast_fp16)[name = tensor("write_idx_7_cast_fp16")]; + tensor var_1077_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_7_cast_fp16)[name = tensor("op_1077_cast_fp16")]; + tensor var_1087 = const()[name = tensor("op_1087"), val = tensor([1, 512, 1, 1])]; + tensor onehot_13_to_fp16_dtype_0 = const()[name = tensor("onehot_13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_1077_cast_fp16_to_fp16 = cast(dtype = onehot_13_to_fp16_dtype_0, x = var_1077_cast_fp16)[name = tensor("cast_552")]; + tensor onehot_15_cast_fp16 = reshape(shape = var_1087, x = var_1077_cast_fp16_to_fp16)[name = tensor("onehot_15_cast_fp16")]; + tensor var_1089_to_fp16 = const()[name = tensor("op_1089_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1091_cast_fp16 = sub(x = var_1089_to_fp16, y = onehot_15_cast_fp16)[name = tensor("op_1091_cast_fp16")]; + tensor k_cache3_to_fp16_dtype_0 = const()[name = tensor("k_cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache3_to_fp16 = cast(dtype = k_cache3_to_fp16_dtype_0, x = k_cache3)[name = tensor("cast_551")]; + tensor var_1092_cast_fp16 = mul(x = k_cache3_to_fp16, y = var_1091_cast_fp16)[name = tensor("op_1092_cast_fp16")]; + tensor var_1097 = const()[name = tensor("op_1097"), val = tensor([1, 1, 16, 64])]; + tensor var_1098_cast_fp16 = reshape(shape = var_1097, x = out_15_cast_fp16)[name = tensor("op_1098_cast_fp16")]; + tensor var_1099_cast_fp16 = mul(x = var_1098_cast_fp16, y = onehot_15_cast_fp16)[name = tensor("op_1099_cast_fp16")]; + tensor new_k_cache_7_cast_fp16 = add(x = var_1092_cast_fp16, y = var_1099_cast_fp16)[name = tensor("new_k_cache_7_cast_fp16")]; + tensor new_k_cache_7_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_7_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache3_to_fp16_dtype_0 = const()[name = tensor("v_cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache3_to_fp16 = cast(dtype = v_cache3_to_fp16_dtype_0, x = v_cache3)[name = tensor("cast_549")]; + tensor var_1105_cast_fp16 = mul(x = v_cache3_to_fp16, y = var_1091_cast_fp16)[name = tensor("op_1105_cast_fp16")]; + tensor var_1110 = const()[name = tensor("op_1110"), val = tensor([1, 1, 16, 64])]; + tensor var_1111_cast_fp16 = reshape(shape = var_1110, x = v_7_cast_fp16)[name = tensor("op_1111_cast_fp16")]; + tensor var_1112_cast_fp16 = mul(x = var_1111_cast_fp16, y = onehot_15_cast_fp16)[name = tensor("op_1112_cast_fp16")]; + tensor new_v_cache_7_cast_fp16 = add(x = var_1105_cast_fp16, y = var_1112_cast_fp16)[name = tensor("new_v_cache_7_cast_fp16")]; + tensor new_v_cache_7_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_7_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_3_perm_0 = const()[name = tensor("transpose_3_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_1125 = const()[name = tensor("op_1125"), val = tensor([0, 2, 1, 3])]; + tensor var_1131 = const()[name = tensor("op_1131"), val = tensor([1, 16, 1, 64])]; + tensor q4_7_cast_fp16 = reshape(shape = var_1131, x = out_13_cast_fp16)[name = tensor("q4_7_cast_fp16")]; + tensor var_1136_transpose_x_0 = const()[name = tensor("op_1136_transpose_x_0"), val = tensor(false)]; + tensor var_1136_transpose_y_0 = const()[name = tensor("op_1136_transpose_y_0"), val = tensor(false)]; + tensor transpose_3_cast_fp16 = transpose(perm = transpose_3_perm_0, x = new_k_cache_7_cast_fp16)[name = tensor("transpose_86")]; + tensor var_1136_cast_fp16 = matmul(transpose_x = var_1136_transpose_x_0, transpose_y = var_1136_transpose_y_0, x = q4_7_cast_fp16, y = transpose_3_cast_fp16)[name = tensor("op_1136_cast_fp16")]; + tensor var_1137_to_fp16 = const()[name = tensor("op_1137_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_13_cast_fp16 = mul(x = var_1136_cast_fp16, y = var_1137_to_fp16)[name = tensor("scores_13_cast_fp16")]; + tensor var_1141 = const()[name = tensor("op_1141"), val = tensor([1, 1])]; + tensor var_1142_cast_fp16 = reshape(shape = var_1141, x = position3_to_fp16)[name = tensor("op_1142_cast_fp16")]; + tensor var_1143_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_1142_cast_fp16)[name = tensor("op_1143_cast_fp16")]; + tensor mask_7_to_fp16_dtype_0 = const()[name = tensor("mask_7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_1150_to_fp16 = const()[name = tensor("op_1150_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1143_cast_fp16_to_fp16 = cast(dtype = mask_7_to_fp16_dtype_0, x = var_1143_cast_fp16)[name = tensor("cast_547")]; + tensor var_1151_cast_fp16 = sub(x = var_1143_cast_fp16_to_fp16, y = var_1150_to_fp16)[name = tensor("op_1151_cast_fp16")]; + tensor var_1156 = const()[name = tensor("op_1156"), val = tensor([1, 1, 1, 512])]; + tensor var_1157_cast_fp16 = reshape(shape = var_1156, x = var_1151_cast_fp16)[name = tensor("op_1157_cast_fp16")]; + tensor var_1158_to_fp16 = const()[name = tensor("op_1158_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_1159_cast_fp16 = mul(x = var_1157_cast_fp16, y = var_1158_to_fp16)[name = tensor("op_1159_cast_fp16")]; + tensor scores_15_cast_fp16 = add(x = scores_13_cast_fp16, y = var_1159_cast_fp16)[name = tensor("scores_15_cast_fp16")]; + tensor var_1162 = const()[name = tensor("op_1162"), val = tensor(-1)]; + tensor weights_7_cast_fp16 = softmax(axis = var_1162, x = scores_15_cast_fp16)[name = tensor("weights_7_cast_fp16")]; + tensor attn_7_transpose_x_0 = const()[name = tensor("attn_7_transpose_x_0"), val = tensor(false)]; + tensor attn_7_transpose_y_0 = const()[name = tensor("attn_7_transpose_y_0"), val = tensor(false)]; + tensor values_7_cast_fp16 = transpose(perm = var_1125, x = new_v_cache_7_cast_fp16)[name = tensor("transpose_85")]; + tensor attn_7_cast_fp16 = matmul(transpose_x = attn_7_transpose_x_0, transpose_y = attn_7_transpose_y_0, x = weights_7_cast_fp16, y = values_7_cast_fp16)[name = tensor("attn_7_cast_fp16")]; + tensor var_1170 = const()[name = tensor("op_1170"), val = tensor([0, 2, 1, 3])]; + tensor var_1175 = const()[name = tensor("op_1175"), val = tensor([1, 1, 1024])]; + tensor var_1171_cast_fp16 = transpose(perm = var_1170, x = attn_7_cast_fp16)[name = tensor("transpose_84")]; + tensor input_43_cast_fp16 = reshape(shape = var_1175, x = var_1171_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor attn3_out_proj_weight_to_fp16 = const()[name = tensor("attn3_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(81902784)))]; + tensor linear_14_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn3_out_proj_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor var_1181_to_fp16 = const()[name = tensor("op_1181_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1182_cast_fp16 = add(x = position3_to_fp16, y = var_1181_to_fp16)[name = tensor("op_1182_cast_fp16")]; + tensor var_1182_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1182_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_45_cast_fp16 = add(x = input_39_cast_fp16, y = linear_14_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor input_47_axes_0 = const()[name = tensor("input_47_axes_0"), val = tensor([-1])]; + tensor norm3_2_weight_to_fp16 = const()[name = tensor("norm3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84000000)))]; + tensor norm3_2_bias_to_fp16 = const()[name = tensor("norm3_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84002112)))]; + tensor var_1186_to_fp16 = const()[name = tensor("op_1186_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_47_cast_fp16 = layer_norm(axes = input_47_axes_0, beta = norm3_2_bias_to_fp16, epsilon = var_1186_to_fp16, gamma = norm3_2_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor linear3_1_weight_to_fp16 = const()[name = tensor("linear3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84004224)))]; + tensor linear_15_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear3_1_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor input_49_mode_0 = const()[name = tensor("input_49_mode_0"), val = tensor("EXACT")]; + tensor input_49_cast_fp16 = gelu(mode = input_49_mode_0, x = linear_15_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor linear3_2_weight_to_fp16 = const()[name = tensor("linear3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(92392896)))]; + tensor linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear3_2_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor input_51_cast_fp16 = add(x = input_45_cast_fp16, y = linear_16_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor input_53_axes_0 = const()[name = tensor("input_53_axes_0"), val = tensor([-1])]; + tensor norm4_1_weight_to_fp16 = const()[name = tensor("norm4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100781568)))]; + tensor norm4_1_bias_to_fp16 = const()[name = tensor("norm4_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100783680)))]; + tensor var_1203_to_fp16 = const()[name = tensor("op_1203_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_53_cast_fp16 = layer_norm(axes = input_53_axes_0, beta = norm4_1_bias_to_fp16, epsilon = var_1203_to_fp16, gamma = norm4_1_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor attn4_in_proj_weight_to_fp16 = const()[name = tensor("attn4_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100785792)))]; + tensor linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn4_in_proj_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor var_1216 = const()[name = tensor("op_1216"), val = tensor([1, 3, 16, 64])]; + tensor qkv_9_cast_fp16 = reshape(shape = var_1216, x = linear_17_cast_fp16)[name = tensor("qkv_9_cast_fp16")]; + tensor x_17_begin_0 = const()[name = tensor("x_17_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_17_end_0 = const()[name = tensor("x_17_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_17_end_mask_0 = const()[name = tensor("x_17_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_17_squeeze_mask_0 = const()[name = tensor("x_17_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_17_cast_fp16 = slice_by_index(begin = x_17_begin_0, end = x_17_end_0, end_mask = x_17_end_mask_0, squeeze_mask = x_17_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("x_17_cast_fp16")]; + tensor x_19_begin_0 = const()[name = tensor("x_19_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_19_end_0 = const()[name = tensor("x_19_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_19_end_mask_0 = const()[name = tensor("x_19_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_19_squeeze_mask_0 = const()[name = tensor("x_19_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_19_cast_fp16 = slice_by_index(begin = x_19_begin_0, end = x_19_end_0, end_mask = x_19_end_mask_0, squeeze_mask = x_19_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("x_19_cast_fp16")]; + tensor v_9_begin_0 = const()[name = tensor("v_9_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_9_end_0 = const()[name = tensor("v_9_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_9_end_mask_0 = const()[name = tensor("v_9_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_9_squeeze_mask_0 = const()[name = tensor("v_9_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_9_cast_fp16 = slice_by_index(begin = v_9_begin_0, end = v_9_end_0, end_mask = v_9_end_mask_0, squeeze_mask = v_9_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("v_9_cast_fp16")]; + tensor var_1245 = const()[name = tensor("op_1245"), val = tensor([1, 1, 1])]; + tensor position4_to_fp16_dtype_0 = const()[name = tensor("position4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position4_to_fp16 = cast(dtype = position4_to_fp16_dtype_0, x = position4)[name = tensor("cast_545")]; + tensor var_1246_cast_fp16 = reshape(shape = var_1245, x = position4_to_fp16)[name = tensor("op_1246_cast_fp16")]; + tensor angles_17_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_1246_cast_fp16)[name = tensor("angles_17_cast_fp16")]; + tensor rotr_17_cast_fp16 = cos(x = angles_17_cast_fp16)[name = tensor("rotr_17_cast_fp16")]; + tensor roti_17_cast_fp16 = sin(x = angles_17_cast_fp16)[name = tensor("roti_17_cast_fp16")]; + tensor var_1254 = const()[name = tensor("op_1254"), val = tensor([1, 16, 32, 2])]; + tensor pairs_17_cast_fp16 = reshape(shape = var_1254, x = x_17_cast_fp16)[name = tensor("pairs_17_cast_fp16")]; + tensor xr_17_begin_0 = const()[name = tensor("xr_17_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_17_end_0 = const()[name = tensor("xr_17_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_17_end_mask_0 = const()[name = tensor("xr_17_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_17_squeeze_mask_0 = const()[name = tensor("xr_17_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_17_cast_fp16 = slice_by_index(begin = xr_17_begin_0, end = xr_17_end_0, end_mask = xr_17_end_mask_0, squeeze_mask = xr_17_squeeze_mask_0, x = pairs_17_cast_fp16)[name = tensor("xr_17_cast_fp16")]; + tensor xi_17_begin_0 = const()[name = tensor("xi_17_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_17_end_0 = const()[name = tensor("xi_17_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_17_end_mask_0 = const()[name = tensor("xi_17_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_17_squeeze_mask_0 = const()[name = tensor("xi_17_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_17_cast_fp16 = slice_by_index(begin = xi_17_begin_0, end = xi_17_end_0, end_mask = xi_17_end_mask_0, squeeze_mask = xi_17_squeeze_mask_0, x = pairs_17_cast_fp16)[name = tensor("xi_17_cast_fp16")]; + tensor var_1262_cast_fp16 = mul(x = xr_17_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_1262_cast_fp16")]; + tensor var_1263_cast_fp16 = mul(x = xi_17_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_1263_cast_fp16")]; + tensor xor__17_cast_fp16 = sub(x = var_1262_cast_fp16, y = var_1263_cast_fp16)[name = tensor("xor__17_cast_fp16")]; + tensor var_1266_cast_fp16 = mul(x = xr_17_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_1266_cast_fp16")]; + tensor var_1267_cast_fp16 = mul(x = xi_17_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_1267_cast_fp16")]; + tensor xoi_17_cast_fp16 = add(x = var_1266_cast_fp16, y = var_1267_cast_fp16)[name = tensor("xoi_17_cast_fp16")]; + tensor out_17_axis_0 = const()[name = tensor("out_17_axis_0"), val = tensor(-1)]; + tensor out_17_cast_fp16 = stack(axis = out_17_axis_0, values = (xor__17_cast_fp16, xoi_17_cast_fp16))[name = tensor("out_17_cast_fp16")]; + tensor var_1290 = const()[name = tensor("op_1290"), val = tensor([1, 16, 32, 2])]; + tensor pairs_19_cast_fp16 = reshape(shape = var_1290, x = x_19_cast_fp16)[name = tensor("pairs_19_cast_fp16")]; + tensor xr_19_begin_0 = const()[name = tensor("xr_19_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_19_end_0 = const()[name = tensor("xr_19_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_19_end_mask_0 = const()[name = tensor("xr_19_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_19_squeeze_mask_0 = const()[name = tensor("xr_19_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_19_cast_fp16 = slice_by_index(begin = xr_19_begin_0, end = xr_19_end_0, end_mask = xr_19_end_mask_0, squeeze_mask = xr_19_squeeze_mask_0, x = pairs_19_cast_fp16)[name = tensor("xr_19_cast_fp16")]; + tensor xi_19_begin_0 = const()[name = tensor("xi_19_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_19_end_0 = const()[name = tensor("xi_19_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_19_end_mask_0 = const()[name = tensor("xi_19_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_19_squeeze_mask_0 = const()[name = tensor("xi_19_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_19_cast_fp16 = slice_by_index(begin = xi_19_begin_0, end = xi_19_end_0, end_mask = xi_19_end_mask_0, squeeze_mask = xi_19_squeeze_mask_0, x = pairs_19_cast_fp16)[name = tensor("xi_19_cast_fp16")]; + tensor var_1298_cast_fp16 = mul(x = xr_19_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_1298_cast_fp16")]; + tensor var_1299_cast_fp16 = mul(x = xi_19_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_1299_cast_fp16")]; + tensor xor__19_cast_fp16 = sub(x = var_1298_cast_fp16, y = var_1299_cast_fp16)[name = tensor("xor__19_cast_fp16")]; + tensor var_1302_cast_fp16 = mul(x = xr_19_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_1302_cast_fp16")]; + tensor var_1303_cast_fp16 = mul(x = xi_19_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_1303_cast_fp16")]; + tensor xoi_19_cast_fp16 = add(x = var_1302_cast_fp16, y = var_1303_cast_fp16)[name = tensor("xoi_19_cast_fp16")]; + tensor out_19_axis_0 = const()[name = tensor("out_19_axis_0"), val = tensor(-1)]; + tensor out_19_cast_fp16 = stack(axis = out_19_axis_0, values = (xor__19_cast_fp16, xoi_19_cast_fp16))[name = tensor("out_19_cast_fp16")]; + tensor _inversed_1315_y_0_to_fp16 = const()[name = tensor("_inversed_1315_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1315_cast_fp16 = mul(x = position4_to_fp16, y = _inversed_1315_y_0_to_fp16)[name = tensor("_inversed_1315_cast_fp16")]; + tensor var_1316_cast_fp16 = floor(x = _inversed_1315_cast_fp16)[name = tensor("op_1316_cast_fp16")]; + tensor var_1317_to_fp16 = const()[name = tensor("op_1317_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1318_cast_fp16 = mul(x = var_1316_cast_fp16, y = var_1317_to_fp16)[name = tensor("op_1318_cast_fp16")]; + tensor write_idx_9_cast_fp16 = sub(x = position4_to_fp16, y = var_1318_cast_fp16)[name = tensor("write_idx_9_cast_fp16")]; + tensor var_1321_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_9_cast_fp16)[name = tensor("op_1321_cast_fp16")]; + tensor var_1331 = const()[name = tensor("op_1331"), val = tensor([1, 512, 1, 1])]; + tensor onehot_17_to_fp16_dtype_0 = const()[name = tensor("onehot_17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_1321_cast_fp16_to_fp16 = cast(dtype = onehot_17_to_fp16_dtype_0, x = var_1321_cast_fp16)[name = tensor("cast_544")]; + tensor onehot_19_cast_fp16 = reshape(shape = var_1331, x = var_1321_cast_fp16_to_fp16)[name = tensor("onehot_19_cast_fp16")]; + tensor var_1333_to_fp16 = const()[name = tensor("op_1333_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1335_cast_fp16 = sub(x = var_1333_to_fp16, y = onehot_19_cast_fp16)[name = tensor("op_1335_cast_fp16")]; + tensor k_cache4_to_fp16_dtype_0 = const()[name = tensor("k_cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache4_to_fp16 = cast(dtype = k_cache4_to_fp16_dtype_0, x = k_cache4)[name = tensor("cast_543")]; + tensor var_1336_cast_fp16 = mul(x = k_cache4_to_fp16, y = var_1335_cast_fp16)[name = tensor("op_1336_cast_fp16")]; + tensor var_1341 = const()[name = tensor("op_1341"), val = tensor([1, 1, 16, 64])]; + tensor var_1342_cast_fp16 = reshape(shape = var_1341, x = out_19_cast_fp16)[name = tensor("op_1342_cast_fp16")]; + tensor var_1343_cast_fp16 = mul(x = var_1342_cast_fp16, y = onehot_19_cast_fp16)[name = tensor("op_1343_cast_fp16")]; + tensor new_k_cache_9_cast_fp16 = add(x = var_1336_cast_fp16, y = var_1343_cast_fp16)[name = tensor("new_k_cache_9_cast_fp16")]; + tensor new_k_cache_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache4_to_fp16_dtype_0 = const()[name = tensor("v_cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache4_to_fp16 = cast(dtype = v_cache4_to_fp16_dtype_0, x = v_cache4)[name = tensor("cast_541")]; + tensor var_1349_cast_fp16 = mul(x = v_cache4_to_fp16, y = var_1335_cast_fp16)[name = tensor("op_1349_cast_fp16")]; + tensor var_1354 = const()[name = tensor("op_1354"), val = tensor([1, 1, 16, 64])]; + tensor var_1355_cast_fp16 = reshape(shape = var_1354, x = v_9_cast_fp16)[name = tensor("op_1355_cast_fp16")]; + tensor var_1356_cast_fp16 = mul(x = var_1355_cast_fp16, y = onehot_19_cast_fp16)[name = tensor("op_1356_cast_fp16")]; + tensor new_v_cache_9_cast_fp16 = add(x = var_1349_cast_fp16, y = var_1356_cast_fp16)[name = tensor("new_v_cache_9_cast_fp16")]; + tensor new_v_cache_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_4_perm_0 = const()[name = tensor("transpose_4_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_1369 = const()[name = tensor("op_1369"), val = tensor([0, 2, 1, 3])]; + tensor var_1375 = const()[name = tensor("op_1375"), val = tensor([1, 16, 1, 64])]; + tensor q4_9_cast_fp16 = reshape(shape = var_1375, x = out_17_cast_fp16)[name = tensor("q4_9_cast_fp16")]; + tensor var_1380_transpose_x_0 = const()[name = tensor("op_1380_transpose_x_0"), val = tensor(false)]; + tensor var_1380_transpose_y_0 = const()[name = tensor("op_1380_transpose_y_0"), val = tensor(false)]; + tensor transpose_4_cast_fp16 = transpose(perm = transpose_4_perm_0, x = new_k_cache_9_cast_fp16)[name = tensor("transpose_83")]; + tensor var_1380_cast_fp16 = matmul(transpose_x = var_1380_transpose_x_0, transpose_y = var_1380_transpose_y_0, x = q4_9_cast_fp16, y = transpose_4_cast_fp16)[name = tensor("op_1380_cast_fp16")]; + tensor var_1381_to_fp16 = const()[name = tensor("op_1381_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_17_cast_fp16 = mul(x = var_1380_cast_fp16, y = var_1381_to_fp16)[name = tensor("scores_17_cast_fp16")]; + tensor var_1385 = const()[name = tensor("op_1385"), val = tensor([1, 1])]; + tensor var_1386_cast_fp16 = reshape(shape = var_1385, x = position4_to_fp16)[name = tensor("op_1386_cast_fp16")]; + tensor var_1387_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_1386_cast_fp16)[name = tensor("op_1387_cast_fp16")]; + tensor mask_9_to_fp16_dtype_0 = const()[name = tensor("mask_9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_1394_to_fp16 = const()[name = tensor("op_1394_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1387_cast_fp16_to_fp16 = cast(dtype = mask_9_to_fp16_dtype_0, x = var_1387_cast_fp16)[name = tensor("cast_539")]; + tensor var_1395_cast_fp16 = sub(x = var_1387_cast_fp16_to_fp16, y = var_1394_to_fp16)[name = tensor("op_1395_cast_fp16")]; + tensor var_1400 = const()[name = tensor("op_1400"), val = tensor([1, 1, 1, 512])]; + tensor var_1401_cast_fp16 = reshape(shape = var_1400, x = var_1395_cast_fp16)[name = tensor("op_1401_cast_fp16")]; + tensor var_1402_to_fp16 = const()[name = tensor("op_1402_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_1403_cast_fp16 = mul(x = var_1401_cast_fp16, y = var_1402_to_fp16)[name = tensor("op_1403_cast_fp16")]; + tensor scores_19_cast_fp16 = add(x = scores_17_cast_fp16, y = var_1403_cast_fp16)[name = tensor("scores_19_cast_fp16")]; + tensor var_1406 = const()[name = tensor("op_1406"), val = tensor(-1)]; + tensor weights_9_cast_fp16 = softmax(axis = var_1406, x = scores_19_cast_fp16)[name = tensor("weights_9_cast_fp16")]; + tensor attn_9_transpose_x_0 = const()[name = tensor("attn_9_transpose_x_0"), val = tensor(false)]; + tensor attn_9_transpose_y_0 = const()[name = tensor("attn_9_transpose_y_0"), val = tensor(false)]; + tensor values_9_cast_fp16 = transpose(perm = var_1369, x = new_v_cache_9_cast_fp16)[name = tensor("transpose_82")]; + tensor attn_9_cast_fp16 = matmul(transpose_x = attn_9_transpose_x_0, transpose_y = attn_9_transpose_y_0, x = weights_9_cast_fp16, y = values_9_cast_fp16)[name = tensor("attn_9_cast_fp16")]; + tensor var_1414 = const()[name = tensor("op_1414"), val = tensor([0, 2, 1, 3])]; + tensor var_1419 = const()[name = tensor("op_1419"), val = tensor([1, 1, 1024])]; + tensor var_1415_cast_fp16 = transpose(perm = var_1414, x = attn_9_cast_fp16)[name = tensor("transpose_81")]; + tensor input_55_cast_fp16 = reshape(shape = var_1419, x = var_1415_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor attn4_out_proj_weight_to_fp16 = const()[name = tensor("attn4_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(107077312)))]; + tensor linear_18_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn4_out_proj_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor var_1425_to_fp16 = const()[name = tensor("op_1425_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1426_cast_fp16 = add(x = position4_to_fp16, y = var_1425_to_fp16)[name = tensor("op_1426_cast_fp16")]; + tensor var_1426_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1426_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_57_cast_fp16 = add(x = input_51_cast_fp16, y = linear_18_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor input_59_axes_0 = const()[name = tensor("input_59_axes_0"), val = tensor([-1])]; + tensor norm4_2_weight_to_fp16 = const()[name = tensor("norm4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109174528)))]; + tensor norm4_2_bias_to_fp16 = const()[name = tensor("norm4_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109176640)))]; + tensor var_1430_to_fp16 = const()[name = tensor("op_1430_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_59_cast_fp16 = layer_norm(axes = input_59_axes_0, beta = norm4_2_bias_to_fp16, epsilon = var_1430_to_fp16, gamma = norm4_2_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor linear4_1_weight_to_fp16 = const()[name = tensor("linear4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109178752)))]; + tensor linear_19_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear4_1_weight_to_fp16, x = input_59_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_61_mode_0 = const()[name = tensor("input_61_mode_0"), val = tensor("EXACT")]; + tensor input_61_cast_fp16 = gelu(mode = input_61_mode_0, x = linear_19_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor linear4_2_weight_to_fp16 = const()[name = tensor("linear4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(117567424)))]; + tensor linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear4_2_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor input_63_cast_fp16 = add(x = input_57_cast_fp16, y = linear_20_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor input_65_axes_0 = const()[name = tensor("input_65_axes_0"), val = tensor([-1])]; + tensor norm5_1_weight_to_fp16 = const()[name = tensor("norm5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125956096)))]; + tensor norm5_1_bias_to_fp16 = const()[name = tensor("norm5_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125958208)))]; + tensor var_1447_to_fp16 = const()[name = tensor("op_1447_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_65_cast_fp16 = layer_norm(axes = input_65_axes_0, beta = norm5_1_bias_to_fp16, epsilon = var_1447_to_fp16, gamma = norm5_1_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("input_65_cast_fp16")]; + tensor attn5_in_proj_weight_to_fp16 = const()[name = tensor("attn5_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125960320)))]; + tensor linear_21_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn5_in_proj_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("linear_21_cast_fp16")]; + tensor var_1460 = const()[name = tensor("op_1460"), val = tensor([1, 3, 16, 64])]; + tensor qkv_11_cast_fp16 = reshape(shape = var_1460, x = linear_21_cast_fp16)[name = tensor("qkv_11_cast_fp16")]; + tensor x_21_begin_0 = const()[name = tensor("x_21_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_21_end_0 = const()[name = tensor("x_21_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_21_end_mask_0 = const()[name = tensor("x_21_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_21_squeeze_mask_0 = const()[name = tensor("x_21_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_21_cast_fp16 = slice_by_index(begin = x_21_begin_0, end = x_21_end_0, end_mask = x_21_end_mask_0, squeeze_mask = x_21_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("x_21_cast_fp16")]; + tensor x_23_begin_0 = const()[name = tensor("x_23_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_23_end_0 = const()[name = tensor("x_23_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_23_end_mask_0 = const()[name = tensor("x_23_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_23_squeeze_mask_0 = const()[name = tensor("x_23_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_23_cast_fp16 = slice_by_index(begin = x_23_begin_0, end = x_23_end_0, end_mask = x_23_end_mask_0, squeeze_mask = x_23_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("x_23_cast_fp16")]; + tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_11_cast_fp16 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("v_11_cast_fp16")]; + tensor var_1489 = const()[name = tensor("op_1489"), val = tensor([1, 1, 1])]; + tensor position5_to_fp16_dtype_0 = const()[name = tensor("position5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position5_to_fp16 = cast(dtype = position5_to_fp16_dtype_0, x = position5)[name = tensor("cast_537")]; + tensor var_1490_cast_fp16 = reshape(shape = var_1489, x = position5_to_fp16)[name = tensor("op_1490_cast_fp16")]; + tensor angles_21_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_1490_cast_fp16)[name = tensor("angles_21_cast_fp16")]; + tensor rotr_21_cast_fp16 = cos(x = angles_21_cast_fp16)[name = tensor("rotr_21_cast_fp16")]; + tensor roti_21_cast_fp16 = sin(x = angles_21_cast_fp16)[name = tensor("roti_21_cast_fp16")]; + tensor var_1498 = const()[name = tensor("op_1498"), val = tensor([1, 16, 32, 2])]; + tensor pairs_21_cast_fp16 = reshape(shape = var_1498, x = x_21_cast_fp16)[name = tensor("pairs_21_cast_fp16")]; + tensor xr_21_begin_0 = const()[name = tensor("xr_21_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_21_end_0 = const()[name = tensor("xr_21_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_21_end_mask_0 = const()[name = tensor("xr_21_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_21_squeeze_mask_0 = const()[name = tensor("xr_21_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_21_cast_fp16 = slice_by_index(begin = xr_21_begin_0, end = xr_21_end_0, end_mask = xr_21_end_mask_0, squeeze_mask = xr_21_squeeze_mask_0, x = pairs_21_cast_fp16)[name = tensor("xr_21_cast_fp16")]; + tensor xi_21_begin_0 = const()[name = tensor("xi_21_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_21_end_0 = const()[name = tensor("xi_21_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_21_end_mask_0 = const()[name = tensor("xi_21_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_21_squeeze_mask_0 = const()[name = tensor("xi_21_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_21_cast_fp16 = slice_by_index(begin = xi_21_begin_0, end = xi_21_end_0, end_mask = xi_21_end_mask_0, squeeze_mask = xi_21_squeeze_mask_0, x = pairs_21_cast_fp16)[name = tensor("xi_21_cast_fp16")]; + tensor var_1506_cast_fp16 = mul(x = xr_21_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_1506_cast_fp16")]; + tensor var_1507_cast_fp16 = mul(x = xi_21_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_1507_cast_fp16")]; + tensor xor__21_cast_fp16 = sub(x = var_1506_cast_fp16, y = var_1507_cast_fp16)[name = tensor("xor__21_cast_fp16")]; + tensor var_1510_cast_fp16 = mul(x = xr_21_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_1510_cast_fp16")]; + tensor var_1511_cast_fp16 = mul(x = xi_21_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_1511_cast_fp16")]; + tensor xoi_21_cast_fp16 = add(x = var_1510_cast_fp16, y = var_1511_cast_fp16)[name = tensor("xoi_21_cast_fp16")]; + tensor out_21_axis_0 = const()[name = tensor("out_21_axis_0"), val = tensor(-1)]; + tensor out_21_cast_fp16 = stack(axis = out_21_axis_0, values = (xor__21_cast_fp16, xoi_21_cast_fp16))[name = tensor("out_21_cast_fp16")]; + tensor var_1534 = const()[name = tensor("op_1534"), val = tensor([1, 16, 32, 2])]; + tensor pairs_23_cast_fp16 = reshape(shape = var_1534, x = x_23_cast_fp16)[name = tensor("pairs_23_cast_fp16")]; + tensor xr_23_begin_0 = const()[name = tensor("xr_23_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_23_end_0 = const()[name = tensor("xr_23_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_23_end_mask_0 = const()[name = tensor("xr_23_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_23_squeeze_mask_0 = const()[name = tensor("xr_23_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_23_cast_fp16 = slice_by_index(begin = xr_23_begin_0, end = xr_23_end_0, end_mask = xr_23_end_mask_0, squeeze_mask = xr_23_squeeze_mask_0, x = pairs_23_cast_fp16)[name = tensor("xr_23_cast_fp16")]; + tensor xi_23_begin_0 = const()[name = tensor("xi_23_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_23_end_0 = const()[name = tensor("xi_23_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_23_end_mask_0 = const()[name = tensor("xi_23_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_23_squeeze_mask_0 = const()[name = tensor("xi_23_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_23_cast_fp16 = slice_by_index(begin = xi_23_begin_0, end = xi_23_end_0, end_mask = xi_23_end_mask_0, squeeze_mask = xi_23_squeeze_mask_0, x = pairs_23_cast_fp16)[name = tensor("xi_23_cast_fp16")]; + tensor var_1542_cast_fp16 = mul(x = xr_23_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_1542_cast_fp16")]; + tensor var_1543_cast_fp16 = mul(x = xi_23_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_1543_cast_fp16")]; + tensor xor__23_cast_fp16 = sub(x = var_1542_cast_fp16, y = var_1543_cast_fp16)[name = tensor("xor__23_cast_fp16")]; + tensor var_1546_cast_fp16 = mul(x = xr_23_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_1546_cast_fp16")]; + tensor var_1547_cast_fp16 = mul(x = xi_23_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_1547_cast_fp16")]; + tensor xoi_23_cast_fp16 = add(x = var_1546_cast_fp16, y = var_1547_cast_fp16)[name = tensor("xoi_23_cast_fp16")]; + tensor out_23_axis_0 = const()[name = tensor("out_23_axis_0"), val = tensor(-1)]; + tensor out_23_cast_fp16 = stack(axis = out_23_axis_0, values = (xor__23_cast_fp16, xoi_23_cast_fp16))[name = tensor("out_23_cast_fp16")]; + tensor _inversed_1559_y_0_to_fp16 = const()[name = tensor("_inversed_1559_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1559_cast_fp16 = mul(x = position5_to_fp16, y = _inversed_1559_y_0_to_fp16)[name = tensor("_inversed_1559_cast_fp16")]; + tensor var_1560_cast_fp16 = floor(x = _inversed_1559_cast_fp16)[name = tensor("op_1560_cast_fp16")]; + tensor var_1561_to_fp16 = const()[name = tensor("op_1561_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1562_cast_fp16 = mul(x = var_1560_cast_fp16, y = var_1561_to_fp16)[name = tensor("op_1562_cast_fp16")]; + tensor write_idx_11_cast_fp16 = sub(x = position5_to_fp16, y = var_1562_cast_fp16)[name = tensor("write_idx_11_cast_fp16")]; + tensor var_1565_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_11_cast_fp16)[name = tensor("op_1565_cast_fp16")]; + tensor var_1575 = const()[name = tensor("op_1575"), val = tensor([1, 512, 1, 1])]; + tensor onehot_21_to_fp16_dtype_0 = const()[name = tensor("onehot_21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_1565_cast_fp16_to_fp16 = cast(dtype = onehot_21_to_fp16_dtype_0, x = var_1565_cast_fp16)[name = tensor("cast_536")]; + tensor onehot_23_cast_fp16 = reshape(shape = var_1575, x = var_1565_cast_fp16_to_fp16)[name = tensor("onehot_23_cast_fp16")]; + tensor var_1577_to_fp16 = const()[name = tensor("op_1577_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1579_cast_fp16 = sub(x = var_1577_to_fp16, y = onehot_23_cast_fp16)[name = tensor("op_1579_cast_fp16")]; + tensor k_cache5_to_fp16_dtype_0 = const()[name = tensor("k_cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache5_to_fp16 = cast(dtype = k_cache5_to_fp16_dtype_0, x = k_cache5)[name = tensor("cast_535")]; + tensor var_1580_cast_fp16 = mul(x = k_cache5_to_fp16, y = var_1579_cast_fp16)[name = tensor("op_1580_cast_fp16")]; + tensor var_1585 = const()[name = tensor("op_1585"), val = tensor([1, 1, 16, 64])]; + tensor var_1586_cast_fp16 = reshape(shape = var_1585, x = out_23_cast_fp16)[name = tensor("op_1586_cast_fp16")]; + tensor var_1587_cast_fp16 = mul(x = var_1586_cast_fp16, y = onehot_23_cast_fp16)[name = tensor("op_1587_cast_fp16")]; + tensor new_k_cache_11_cast_fp16 = add(x = var_1580_cast_fp16, y = var_1587_cast_fp16)[name = tensor("new_k_cache_11_cast_fp16")]; + tensor new_k_cache_11_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_11_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache5_to_fp16_dtype_0 = const()[name = tensor("v_cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache5_to_fp16 = cast(dtype = v_cache5_to_fp16_dtype_0, x = v_cache5)[name = tensor("cast_533")]; + tensor var_1593_cast_fp16 = mul(x = v_cache5_to_fp16, y = var_1579_cast_fp16)[name = tensor("op_1593_cast_fp16")]; + tensor var_1598 = const()[name = tensor("op_1598"), val = tensor([1, 1, 16, 64])]; + tensor var_1599_cast_fp16 = reshape(shape = var_1598, x = v_11_cast_fp16)[name = tensor("op_1599_cast_fp16")]; + tensor var_1600_cast_fp16 = mul(x = var_1599_cast_fp16, y = onehot_23_cast_fp16)[name = tensor("op_1600_cast_fp16")]; + tensor new_v_cache_11_cast_fp16 = add(x = var_1593_cast_fp16, y = var_1600_cast_fp16)[name = tensor("new_v_cache_11_cast_fp16")]; + tensor new_v_cache_11_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_11_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_5_perm_0 = const()[name = tensor("transpose_5_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_1613 = const()[name = tensor("op_1613"), val = tensor([0, 2, 1, 3])]; + tensor var_1619 = const()[name = tensor("op_1619"), val = tensor([1, 16, 1, 64])]; + tensor q4_11_cast_fp16 = reshape(shape = var_1619, x = out_21_cast_fp16)[name = tensor("q4_11_cast_fp16")]; + tensor var_1624_transpose_x_0 = const()[name = tensor("op_1624_transpose_x_0"), val = tensor(false)]; + tensor var_1624_transpose_y_0 = const()[name = tensor("op_1624_transpose_y_0"), val = tensor(false)]; + tensor transpose_5_cast_fp16 = transpose(perm = transpose_5_perm_0, x = new_k_cache_11_cast_fp16)[name = tensor("transpose_80")]; + tensor var_1624_cast_fp16 = matmul(transpose_x = var_1624_transpose_x_0, transpose_y = var_1624_transpose_y_0, x = q4_11_cast_fp16, y = transpose_5_cast_fp16)[name = tensor("op_1624_cast_fp16")]; + tensor var_1625_to_fp16 = const()[name = tensor("op_1625_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_21_cast_fp16 = mul(x = var_1624_cast_fp16, y = var_1625_to_fp16)[name = tensor("scores_21_cast_fp16")]; + tensor var_1629 = const()[name = tensor("op_1629"), val = tensor([1, 1])]; + tensor var_1630_cast_fp16 = reshape(shape = var_1629, x = position5_to_fp16)[name = tensor("op_1630_cast_fp16")]; + tensor var_1631_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_1630_cast_fp16)[name = tensor("op_1631_cast_fp16")]; + tensor mask_11_to_fp16_dtype_0 = const()[name = tensor("mask_11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_1638_to_fp16 = const()[name = tensor("op_1638_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1631_cast_fp16_to_fp16 = cast(dtype = mask_11_to_fp16_dtype_0, x = var_1631_cast_fp16)[name = tensor("cast_531")]; + tensor var_1639_cast_fp16 = sub(x = var_1631_cast_fp16_to_fp16, y = var_1638_to_fp16)[name = tensor("op_1639_cast_fp16")]; + tensor var_1644 = const()[name = tensor("op_1644"), val = tensor([1, 1, 1, 512])]; + tensor var_1645_cast_fp16 = reshape(shape = var_1644, x = var_1639_cast_fp16)[name = tensor("op_1645_cast_fp16")]; + tensor var_1646_to_fp16 = const()[name = tensor("op_1646_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_1647_cast_fp16 = mul(x = var_1645_cast_fp16, y = var_1646_to_fp16)[name = tensor("op_1647_cast_fp16")]; + tensor scores_23_cast_fp16 = add(x = scores_21_cast_fp16, y = var_1647_cast_fp16)[name = tensor("scores_23_cast_fp16")]; + tensor var_1650 = const()[name = tensor("op_1650"), val = tensor(-1)]; + tensor weights_11_cast_fp16 = softmax(axis = var_1650, x = scores_23_cast_fp16)[name = tensor("weights_11_cast_fp16")]; + tensor attn_11_transpose_x_0 = const()[name = tensor("attn_11_transpose_x_0"), val = tensor(false)]; + tensor attn_11_transpose_y_0 = const()[name = tensor("attn_11_transpose_y_0"), val = tensor(false)]; + tensor values_11_cast_fp16 = transpose(perm = var_1613, x = new_v_cache_11_cast_fp16)[name = tensor("transpose_79")]; + tensor attn_11_cast_fp16 = matmul(transpose_x = attn_11_transpose_x_0, transpose_y = attn_11_transpose_y_0, x = weights_11_cast_fp16, y = values_11_cast_fp16)[name = tensor("attn_11_cast_fp16")]; + tensor var_1658 = const()[name = tensor("op_1658"), val = tensor([0, 2, 1, 3])]; + tensor var_1663 = const()[name = tensor("op_1663"), val = tensor([1, 1, 1024])]; + tensor var_1659_cast_fp16 = transpose(perm = var_1658, x = attn_11_cast_fp16)[name = tensor("transpose_78")]; + tensor input_67_cast_fp16 = reshape(shape = var_1663, x = var_1659_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor attn5_out_proj_weight_to_fp16 = const()[name = tensor("attn5_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132251840)))]; + tensor linear_22_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn5_out_proj_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("linear_22_cast_fp16")]; + tensor var_1669_to_fp16 = const()[name = tensor("op_1669_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1670_cast_fp16 = add(x = position5_to_fp16, y = var_1669_to_fp16)[name = tensor("op_1670_cast_fp16")]; + tensor var_1670_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1670_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_69_cast_fp16 = add(x = input_63_cast_fp16, y = linear_22_cast_fp16)[name = tensor("input_69_cast_fp16")]; + tensor input_71_axes_0 = const()[name = tensor("input_71_axes_0"), val = tensor([-1])]; + tensor norm5_2_weight_to_fp16 = const()[name = tensor("norm5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134349056)))]; + tensor norm5_2_bias_to_fp16 = const()[name = tensor("norm5_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134351168)))]; + tensor var_1674_to_fp16 = const()[name = tensor("op_1674_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_71_cast_fp16 = layer_norm(axes = input_71_axes_0, beta = norm5_2_bias_to_fp16, epsilon = var_1674_to_fp16, gamma = norm5_2_weight_to_fp16, x = input_69_cast_fp16)[name = tensor("input_71_cast_fp16")]; + tensor linear5_1_weight_to_fp16 = const()[name = tensor("linear5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134353280)))]; + tensor linear_23_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear5_1_weight_to_fp16, x = input_71_cast_fp16)[name = tensor("linear_23_cast_fp16")]; + tensor input_73_mode_0 = const()[name = tensor("input_73_mode_0"), val = tensor("EXACT")]; + tensor input_73_cast_fp16 = gelu(mode = input_73_mode_0, x = linear_23_cast_fp16)[name = tensor("input_73_cast_fp16")]; + tensor linear5_2_weight_to_fp16 = const()[name = tensor("linear5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142741952)))]; + tensor linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear5_2_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("linear_24_cast_fp16")]; + tensor input_75_cast_fp16 = add(x = input_69_cast_fp16, y = linear_24_cast_fp16)[name = tensor("input_75_cast_fp16")]; + tensor input_77_axes_0 = const()[name = tensor("input_77_axes_0"), val = tensor([-1])]; + tensor norm6_1_weight_to_fp16 = const()[name = tensor("norm6_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151130624)))]; + tensor norm6_1_bias_to_fp16 = const()[name = tensor("norm6_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151132736)))]; + tensor var_1691_to_fp16 = const()[name = tensor("op_1691_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_77_cast_fp16 = layer_norm(axes = input_77_axes_0, beta = norm6_1_bias_to_fp16, epsilon = var_1691_to_fp16, gamma = norm6_1_weight_to_fp16, x = input_75_cast_fp16)[name = tensor("input_77_cast_fp16")]; + tensor attn6_in_proj_weight_to_fp16 = const()[name = tensor("attn6_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151134848)))]; + tensor linear_25_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn6_in_proj_weight_to_fp16, x = input_77_cast_fp16)[name = tensor("linear_25_cast_fp16")]; + tensor var_1704 = const()[name = tensor("op_1704"), val = tensor([1, 3, 16, 64])]; + tensor qkv_13_cast_fp16 = reshape(shape = var_1704, x = linear_25_cast_fp16)[name = tensor("qkv_13_cast_fp16")]; + tensor x_25_begin_0 = const()[name = tensor("x_25_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_25_end_0 = const()[name = tensor("x_25_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_25_end_mask_0 = const()[name = tensor("x_25_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_25_squeeze_mask_0 = const()[name = tensor("x_25_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_25_cast_fp16 = slice_by_index(begin = x_25_begin_0, end = x_25_end_0, end_mask = x_25_end_mask_0, squeeze_mask = x_25_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("x_25_cast_fp16")]; + tensor x_27_begin_0 = const()[name = tensor("x_27_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_27_end_0 = const()[name = tensor("x_27_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_27_end_mask_0 = const()[name = tensor("x_27_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_27_squeeze_mask_0 = const()[name = tensor("x_27_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_27_cast_fp16 = slice_by_index(begin = x_27_begin_0, end = x_27_end_0, end_mask = x_27_end_mask_0, squeeze_mask = x_27_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("x_27_cast_fp16")]; + tensor v_13_begin_0 = const()[name = tensor("v_13_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_13_end_0 = const()[name = tensor("v_13_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_13_end_mask_0 = const()[name = tensor("v_13_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_13_squeeze_mask_0 = const()[name = tensor("v_13_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_13_cast_fp16 = slice_by_index(begin = v_13_begin_0, end = v_13_end_0, end_mask = v_13_end_mask_0, squeeze_mask = v_13_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("v_13_cast_fp16")]; + tensor var_1733 = const()[name = tensor("op_1733"), val = tensor([1, 1, 1])]; + tensor position6_to_fp16_dtype_0 = const()[name = tensor("position6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position6_to_fp16 = cast(dtype = position6_to_fp16_dtype_0, x = position6)[name = tensor("cast_529")]; + tensor var_1734_cast_fp16 = reshape(shape = var_1733, x = position6_to_fp16)[name = tensor("op_1734_cast_fp16")]; + tensor angles_25_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_1734_cast_fp16)[name = tensor("angles_25_cast_fp16")]; + tensor rotr_25_cast_fp16 = cos(x = angles_25_cast_fp16)[name = tensor("rotr_25_cast_fp16")]; + tensor roti_25_cast_fp16 = sin(x = angles_25_cast_fp16)[name = tensor("roti_25_cast_fp16")]; + tensor var_1742 = const()[name = tensor("op_1742"), val = tensor([1, 16, 32, 2])]; + tensor pairs_25_cast_fp16 = reshape(shape = var_1742, x = x_25_cast_fp16)[name = tensor("pairs_25_cast_fp16")]; + tensor xr_25_begin_0 = const()[name = tensor("xr_25_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_25_end_0 = const()[name = tensor("xr_25_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_25_end_mask_0 = const()[name = tensor("xr_25_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_25_squeeze_mask_0 = const()[name = tensor("xr_25_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_25_cast_fp16 = slice_by_index(begin = xr_25_begin_0, end = xr_25_end_0, end_mask = xr_25_end_mask_0, squeeze_mask = xr_25_squeeze_mask_0, x = pairs_25_cast_fp16)[name = tensor("xr_25_cast_fp16")]; + tensor xi_25_begin_0 = const()[name = tensor("xi_25_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_25_end_0 = const()[name = tensor("xi_25_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_25_end_mask_0 = const()[name = tensor("xi_25_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_25_squeeze_mask_0 = const()[name = tensor("xi_25_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_25_cast_fp16 = slice_by_index(begin = xi_25_begin_0, end = xi_25_end_0, end_mask = xi_25_end_mask_0, squeeze_mask = xi_25_squeeze_mask_0, x = pairs_25_cast_fp16)[name = tensor("xi_25_cast_fp16")]; + tensor var_1750_cast_fp16 = mul(x = xr_25_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_1750_cast_fp16")]; + tensor var_1751_cast_fp16 = mul(x = xi_25_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_1751_cast_fp16")]; + tensor xor__25_cast_fp16 = sub(x = var_1750_cast_fp16, y = var_1751_cast_fp16)[name = tensor("xor__25_cast_fp16")]; + tensor var_1754_cast_fp16 = mul(x = xr_25_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_1754_cast_fp16")]; + tensor var_1755_cast_fp16 = mul(x = xi_25_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_1755_cast_fp16")]; + tensor xoi_25_cast_fp16 = add(x = var_1754_cast_fp16, y = var_1755_cast_fp16)[name = tensor("xoi_25_cast_fp16")]; + tensor out_25_axis_0 = const()[name = tensor("out_25_axis_0"), val = tensor(-1)]; + tensor out_25_cast_fp16 = stack(axis = out_25_axis_0, values = (xor__25_cast_fp16, xoi_25_cast_fp16))[name = tensor("out_25_cast_fp16")]; + tensor var_1778 = const()[name = tensor("op_1778"), val = tensor([1, 16, 32, 2])]; + tensor pairs_27_cast_fp16 = reshape(shape = var_1778, x = x_27_cast_fp16)[name = tensor("pairs_27_cast_fp16")]; + tensor xr_27_begin_0 = const()[name = tensor("xr_27_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_27_end_0 = const()[name = tensor("xr_27_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_27_end_mask_0 = const()[name = tensor("xr_27_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_27_squeeze_mask_0 = const()[name = tensor("xr_27_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_27_cast_fp16 = slice_by_index(begin = xr_27_begin_0, end = xr_27_end_0, end_mask = xr_27_end_mask_0, squeeze_mask = xr_27_squeeze_mask_0, x = pairs_27_cast_fp16)[name = tensor("xr_27_cast_fp16")]; + tensor xi_27_begin_0 = const()[name = tensor("xi_27_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_27_end_0 = const()[name = tensor("xi_27_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_27_end_mask_0 = const()[name = tensor("xi_27_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_27_squeeze_mask_0 = const()[name = tensor("xi_27_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_27_cast_fp16 = slice_by_index(begin = xi_27_begin_0, end = xi_27_end_0, end_mask = xi_27_end_mask_0, squeeze_mask = xi_27_squeeze_mask_0, x = pairs_27_cast_fp16)[name = tensor("xi_27_cast_fp16")]; + tensor var_1786_cast_fp16 = mul(x = xr_27_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_1786_cast_fp16")]; + tensor var_1787_cast_fp16 = mul(x = xi_27_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_1787_cast_fp16")]; + tensor xor__27_cast_fp16 = sub(x = var_1786_cast_fp16, y = var_1787_cast_fp16)[name = tensor("xor__27_cast_fp16")]; + tensor var_1790_cast_fp16 = mul(x = xr_27_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_1790_cast_fp16")]; + tensor var_1791_cast_fp16 = mul(x = xi_27_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_1791_cast_fp16")]; + tensor xoi_27_cast_fp16 = add(x = var_1790_cast_fp16, y = var_1791_cast_fp16)[name = tensor("xoi_27_cast_fp16")]; + tensor out_27_axis_0 = const()[name = tensor("out_27_axis_0"), val = tensor(-1)]; + tensor out_27_cast_fp16 = stack(axis = out_27_axis_0, values = (xor__27_cast_fp16, xoi_27_cast_fp16))[name = tensor("out_27_cast_fp16")]; + tensor _inversed_1803_y_0_to_fp16 = const()[name = tensor("_inversed_1803_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1803_cast_fp16 = mul(x = position6_to_fp16, y = _inversed_1803_y_0_to_fp16)[name = tensor("_inversed_1803_cast_fp16")]; + tensor var_1804_cast_fp16 = floor(x = _inversed_1803_cast_fp16)[name = tensor("op_1804_cast_fp16")]; + tensor var_1805_to_fp16 = const()[name = tensor("op_1805_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1806_cast_fp16 = mul(x = var_1804_cast_fp16, y = var_1805_to_fp16)[name = tensor("op_1806_cast_fp16")]; + tensor write_idx_13_cast_fp16 = sub(x = position6_to_fp16, y = var_1806_cast_fp16)[name = tensor("write_idx_13_cast_fp16")]; + tensor var_1809_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_13_cast_fp16)[name = tensor("op_1809_cast_fp16")]; + tensor var_1819 = const()[name = tensor("op_1819"), val = tensor([1, 512, 1, 1])]; + tensor onehot_25_to_fp16_dtype_0 = const()[name = tensor("onehot_25_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_1809_cast_fp16_to_fp16 = cast(dtype = onehot_25_to_fp16_dtype_0, x = var_1809_cast_fp16)[name = tensor("cast_528")]; + tensor onehot_27_cast_fp16 = reshape(shape = var_1819, x = var_1809_cast_fp16_to_fp16)[name = tensor("onehot_27_cast_fp16")]; + tensor var_1821_to_fp16 = const()[name = tensor("op_1821_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1823_cast_fp16 = sub(x = var_1821_to_fp16, y = onehot_27_cast_fp16)[name = tensor("op_1823_cast_fp16")]; + tensor k_cache6_to_fp16_dtype_0 = const()[name = tensor("k_cache6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache6_to_fp16 = cast(dtype = k_cache6_to_fp16_dtype_0, x = k_cache6)[name = tensor("cast_527")]; + tensor var_1824_cast_fp16 = mul(x = k_cache6_to_fp16, y = var_1823_cast_fp16)[name = tensor("op_1824_cast_fp16")]; + tensor var_1829 = const()[name = tensor("op_1829"), val = tensor([1, 1, 16, 64])]; + tensor var_1830_cast_fp16 = reshape(shape = var_1829, x = out_27_cast_fp16)[name = tensor("op_1830_cast_fp16")]; + tensor var_1831_cast_fp16 = mul(x = var_1830_cast_fp16, y = onehot_27_cast_fp16)[name = tensor("op_1831_cast_fp16")]; + tensor new_k_cache_13_cast_fp16 = add(x = var_1824_cast_fp16, y = var_1831_cast_fp16)[name = tensor("new_k_cache_13_cast_fp16")]; + tensor new_k_cache_13_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_13_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache6_to_fp16_dtype_0 = const()[name = tensor("v_cache6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache6_to_fp16 = cast(dtype = v_cache6_to_fp16_dtype_0, x = v_cache6)[name = tensor("cast_525")]; + tensor var_1837_cast_fp16 = mul(x = v_cache6_to_fp16, y = var_1823_cast_fp16)[name = tensor("op_1837_cast_fp16")]; + tensor var_1842 = const()[name = tensor("op_1842"), val = tensor([1, 1, 16, 64])]; + tensor var_1843_cast_fp16 = reshape(shape = var_1842, x = v_13_cast_fp16)[name = tensor("op_1843_cast_fp16")]; + tensor var_1844_cast_fp16 = mul(x = var_1843_cast_fp16, y = onehot_27_cast_fp16)[name = tensor("op_1844_cast_fp16")]; + tensor new_v_cache_13_cast_fp16 = add(x = var_1837_cast_fp16, y = var_1844_cast_fp16)[name = tensor("new_v_cache_13_cast_fp16")]; + tensor new_v_cache_13_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_13_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_6_perm_0 = const()[name = tensor("transpose_6_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_1857 = const()[name = tensor("op_1857"), val = tensor([0, 2, 1, 3])]; + tensor var_1863 = const()[name = tensor("op_1863"), val = tensor([1, 16, 1, 64])]; + tensor q4_13_cast_fp16 = reshape(shape = var_1863, x = out_25_cast_fp16)[name = tensor("q4_13_cast_fp16")]; + tensor var_1868_transpose_x_0 = const()[name = tensor("op_1868_transpose_x_0"), val = tensor(false)]; + tensor var_1868_transpose_y_0 = const()[name = tensor("op_1868_transpose_y_0"), val = tensor(false)]; + tensor transpose_6_cast_fp16 = transpose(perm = transpose_6_perm_0, x = new_k_cache_13_cast_fp16)[name = tensor("transpose_77")]; + tensor var_1868_cast_fp16 = matmul(transpose_x = var_1868_transpose_x_0, transpose_y = var_1868_transpose_y_0, x = q4_13_cast_fp16, y = transpose_6_cast_fp16)[name = tensor("op_1868_cast_fp16")]; + tensor var_1869_to_fp16 = const()[name = tensor("op_1869_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_25_cast_fp16 = mul(x = var_1868_cast_fp16, y = var_1869_to_fp16)[name = tensor("scores_25_cast_fp16")]; + tensor var_1873 = const()[name = tensor("op_1873"), val = tensor([1, 1])]; + tensor var_1874_cast_fp16 = reshape(shape = var_1873, x = position6_to_fp16)[name = tensor("op_1874_cast_fp16")]; + tensor var_1875_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_1874_cast_fp16)[name = tensor("op_1875_cast_fp16")]; + tensor mask_13_to_fp16_dtype_0 = const()[name = tensor("mask_13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_1882_to_fp16 = const()[name = tensor("op_1882_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1875_cast_fp16_to_fp16 = cast(dtype = mask_13_to_fp16_dtype_0, x = var_1875_cast_fp16)[name = tensor("cast_523")]; + tensor var_1883_cast_fp16 = sub(x = var_1875_cast_fp16_to_fp16, y = var_1882_to_fp16)[name = tensor("op_1883_cast_fp16")]; + tensor var_1888 = const()[name = tensor("op_1888"), val = tensor([1, 1, 1, 512])]; + tensor var_1889_cast_fp16 = reshape(shape = var_1888, x = var_1883_cast_fp16)[name = tensor("op_1889_cast_fp16")]; + tensor var_1890_to_fp16 = const()[name = tensor("op_1890_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_1891_cast_fp16 = mul(x = var_1889_cast_fp16, y = var_1890_to_fp16)[name = tensor("op_1891_cast_fp16")]; + tensor scores_27_cast_fp16 = add(x = scores_25_cast_fp16, y = var_1891_cast_fp16)[name = tensor("scores_27_cast_fp16")]; + tensor var_1894 = const()[name = tensor("op_1894"), val = tensor(-1)]; + tensor weights_13_cast_fp16 = softmax(axis = var_1894, x = scores_27_cast_fp16)[name = tensor("weights_13_cast_fp16")]; + tensor attn_13_transpose_x_0 = const()[name = tensor("attn_13_transpose_x_0"), val = tensor(false)]; + tensor attn_13_transpose_y_0 = const()[name = tensor("attn_13_transpose_y_0"), val = tensor(false)]; + tensor values_13_cast_fp16 = transpose(perm = var_1857, x = new_v_cache_13_cast_fp16)[name = tensor("transpose_76")]; + tensor attn_13_cast_fp16 = matmul(transpose_x = attn_13_transpose_x_0, transpose_y = attn_13_transpose_y_0, x = weights_13_cast_fp16, y = values_13_cast_fp16)[name = tensor("attn_13_cast_fp16")]; + tensor var_1902 = const()[name = tensor("op_1902"), val = tensor([0, 2, 1, 3])]; + tensor var_1907 = const()[name = tensor("op_1907"), val = tensor([1, 1, 1024])]; + tensor var_1903_cast_fp16 = transpose(perm = var_1902, x = attn_13_cast_fp16)[name = tensor("transpose_75")]; + tensor input_79_cast_fp16 = reshape(shape = var_1907, x = var_1903_cast_fp16)[name = tensor("input_79_cast_fp16")]; + tensor attn6_out_proj_weight_to_fp16 = const()[name = tensor("attn6_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(157426368)))]; + tensor linear_26_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn6_out_proj_weight_to_fp16, x = input_79_cast_fp16)[name = tensor("linear_26_cast_fp16")]; + tensor var_1913_to_fp16 = const()[name = tensor("op_1913_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1914_cast_fp16 = add(x = position6_to_fp16, y = var_1913_to_fp16)[name = tensor("op_1914_cast_fp16")]; + tensor var_1914_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1914_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_81_cast_fp16 = add(x = input_75_cast_fp16, y = linear_26_cast_fp16)[name = tensor("input_81_cast_fp16")]; + tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([-1])]; + tensor norm6_2_weight_to_fp16 = const()[name = tensor("norm6_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159523584)))]; + tensor norm6_2_bias_to_fp16 = const()[name = tensor("norm6_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159525696)))]; + tensor var_1918_to_fp16 = const()[name = tensor("op_1918_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_83_cast_fp16 = layer_norm(axes = input_83_axes_0, beta = norm6_2_bias_to_fp16, epsilon = var_1918_to_fp16, gamma = norm6_2_weight_to_fp16, x = input_81_cast_fp16)[name = tensor("input_83_cast_fp16")]; + tensor linear6_1_weight_to_fp16 = const()[name = tensor("linear6_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159527808)))]; + tensor linear_27_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear6_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_27_cast_fp16")]; + tensor input_85_mode_0 = const()[name = tensor("input_85_mode_0"), val = tensor("EXACT")]; + tensor input_85_cast_fp16 = gelu(mode = input_85_mode_0, x = linear_27_cast_fp16)[name = tensor("input_85_cast_fp16")]; + tensor linear6_2_weight_to_fp16 = const()[name = tensor("linear6_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(167916480)))]; + tensor linear_28_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear6_2_weight_to_fp16, x = input_85_cast_fp16)[name = tensor("linear_28_cast_fp16")]; + tensor input_87_cast_fp16 = add(x = input_81_cast_fp16, y = linear_28_cast_fp16)[name = tensor("input_87_cast_fp16")]; + tensor input_89_axes_0 = const()[name = tensor("input_89_axes_0"), val = tensor([-1])]; + tensor norm7_1_weight_to_fp16 = const()[name = tensor("norm7_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176305152)))]; + tensor norm7_1_bias_to_fp16 = const()[name = tensor("norm7_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176307264)))]; + tensor var_1935_to_fp16 = const()[name = tensor("op_1935_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_89_cast_fp16 = layer_norm(axes = input_89_axes_0, beta = norm7_1_bias_to_fp16, epsilon = var_1935_to_fp16, gamma = norm7_1_weight_to_fp16, x = input_87_cast_fp16)[name = tensor("input_89_cast_fp16")]; + tensor attn7_in_proj_weight_to_fp16 = const()[name = tensor("attn7_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176309376)))]; + tensor linear_29_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn7_in_proj_weight_to_fp16, x = input_89_cast_fp16)[name = tensor("linear_29_cast_fp16")]; + tensor var_1948 = const()[name = tensor("op_1948"), val = tensor([1, 3, 16, 64])]; + tensor qkv_15_cast_fp16 = reshape(shape = var_1948, x = linear_29_cast_fp16)[name = tensor("qkv_15_cast_fp16")]; + tensor x_29_begin_0 = const()[name = tensor("x_29_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_29_end_0 = const()[name = tensor("x_29_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_29_end_mask_0 = const()[name = tensor("x_29_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_29_squeeze_mask_0 = const()[name = tensor("x_29_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_29_cast_fp16 = slice_by_index(begin = x_29_begin_0, end = x_29_end_0, end_mask = x_29_end_mask_0, squeeze_mask = x_29_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("x_29_cast_fp16")]; + tensor x_31_begin_0 = const()[name = tensor("x_31_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_31_end_0 = const()[name = tensor("x_31_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_31_end_mask_0 = const()[name = tensor("x_31_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_31_squeeze_mask_0 = const()[name = tensor("x_31_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_31_cast_fp16 = slice_by_index(begin = x_31_begin_0, end = x_31_end_0, end_mask = x_31_end_mask_0, squeeze_mask = x_31_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("x_31_cast_fp16")]; + tensor v_15_begin_0 = const()[name = tensor("v_15_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_15_end_0 = const()[name = tensor("v_15_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_15_end_mask_0 = const()[name = tensor("v_15_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_15_squeeze_mask_0 = const()[name = tensor("v_15_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_15_cast_fp16 = slice_by_index(begin = v_15_begin_0, end = v_15_end_0, end_mask = v_15_end_mask_0, squeeze_mask = v_15_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("v_15_cast_fp16")]; + tensor var_1977 = const()[name = tensor("op_1977"), val = tensor([1, 1, 1])]; + tensor position7_to_fp16_dtype_0 = const()[name = tensor("position7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position7_to_fp16 = cast(dtype = position7_to_fp16_dtype_0, x = position7)[name = tensor("cast_521")]; + tensor var_1978_cast_fp16 = reshape(shape = var_1977, x = position7_to_fp16)[name = tensor("op_1978_cast_fp16")]; + tensor angles_29_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_1978_cast_fp16)[name = tensor("angles_29_cast_fp16")]; + tensor rotr_29_cast_fp16 = cos(x = angles_29_cast_fp16)[name = tensor("rotr_29_cast_fp16")]; + tensor roti_29_cast_fp16 = sin(x = angles_29_cast_fp16)[name = tensor("roti_29_cast_fp16")]; + tensor var_1986 = const()[name = tensor("op_1986"), val = tensor([1, 16, 32, 2])]; + tensor pairs_29_cast_fp16 = reshape(shape = var_1986, x = x_29_cast_fp16)[name = tensor("pairs_29_cast_fp16")]; + tensor xr_29_begin_0 = const()[name = tensor("xr_29_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_29_end_0 = const()[name = tensor("xr_29_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_29_end_mask_0 = const()[name = tensor("xr_29_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_29_squeeze_mask_0 = const()[name = tensor("xr_29_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_29_cast_fp16 = slice_by_index(begin = xr_29_begin_0, end = xr_29_end_0, end_mask = xr_29_end_mask_0, squeeze_mask = xr_29_squeeze_mask_0, x = pairs_29_cast_fp16)[name = tensor("xr_29_cast_fp16")]; + tensor xi_29_begin_0 = const()[name = tensor("xi_29_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_29_end_0 = const()[name = tensor("xi_29_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_29_end_mask_0 = const()[name = tensor("xi_29_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_29_squeeze_mask_0 = const()[name = tensor("xi_29_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_29_cast_fp16 = slice_by_index(begin = xi_29_begin_0, end = xi_29_end_0, end_mask = xi_29_end_mask_0, squeeze_mask = xi_29_squeeze_mask_0, x = pairs_29_cast_fp16)[name = tensor("xi_29_cast_fp16")]; + tensor var_1994_cast_fp16 = mul(x = xr_29_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_1994_cast_fp16")]; + tensor var_1995_cast_fp16 = mul(x = xi_29_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_1995_cast_fp16")]; + tensor xor__29_cast_fp16 = sub(x = var_1994_cast_fp16, y = var_1995_cast_fp16)[name = tensor("xor__29_cast_fp16")]; + tensor var_1998_cast_fp16 = mul(x = xr_29_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_1998_cast_fp16")]; + tensor var_1999_cast_fp16 = mul(x = xi_29_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_1999_cast_fp16")]; + tensor xoi_29_cast_fp16 = add(x = var_1998_cast_fp16, y = var_1999_cast_fp16)[name = tensor("xoi_29_cast_fp16")]; + tensor out_29_axis_0 = const()[name = tensor("out_29_axis_0"), val = tensor(-1)]; + tensor out_29_cast_fp16 = stack(axis = out_29_axis_0, values = (xor__29_cast_fp16, xoi_29_cast_fp16))[name = tensor("out_29_cast_fp16")]; + tensor var_2022 = const()[name = tensor("op_2022"), val = tensor([1, 16, 32, 2])]; + tensor pairs_31_cast_fp16 = reshape(shape = var_2022, x = x_31_cast_fp16)[name = tensor("pairs_31_cast_fp16")]; + tensor xr_31_begin_0 = const()[name = tensor("xr_31_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_31_end_0 = const()[name = tensor("xr_31_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_31_end_mask_0 = const()[name = tensor("xr_31_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_31_squeeze_mask_0 = const()[name = tensor("xr_31_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_31_cast_fp16 = slice_by_index(begin = xr_31_begin_0, end = xr_31_end_0, end_mask = xr_31_end_mask_0, squeeze_mask = xr_31_squeeze_mask_0, x = pairs_31_cast_fp16)[name = tensor("xr_31_cast_fp16")]; + tensor xi_31_begin_0 = const()[name = tensor("xi_31_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_31_end_0 = const()[name = tensor("xi_31_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_31_end_mask_0 = const()[name = tensor("xi_31_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_31_squeeze_mask_0 = const()[name = tensor("xi_31_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_31_cast_fp16 = slice_by_index(begin = xi_31_begin_0, end = xi_31_end_0, end_mask = xi_31_end_mask_0, squeeze_mask = xi_31_squeeze_mask_0, x = pairs_31_cast_fp16)[name = tensor("xi_31_cast_fp16")]; + tensor var_2030_cast_fp16 = mul(x = xr_31_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_2030_cast_fp16")]; + tensor var_2031_cast_fp16 = mul(x = xi_31_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_2031_cast_fp16")]; + tensor xor__31_cast_fp16 = sub(x = var_2030_cast_fp16, y = var_2031_cast_fp16)[name = tensor("xor__31_cast_fp16")]; + tensor var_2034_cast_fp16 = mul(x = xr_31_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_2034_cast_fp16")]; + tensor var_2035_cast_fp16 = mul(x = xi_31_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_2035_cast_fp16")]; + tensor xoi_31_cast_fp16 = add(x = var_2034_cast_fp16, y = var_2035_cast_fp16)[name = tensor("xoi_31_cast_fp16")]; + tensor out_31_axis_0 = const()[name = tensor("out_31_axis_0"), val = tensor(-1)]; + tensor out_31_cast_fp16 = stack(axis = out_31_axis_0, values = (xor__31_cast_fp16, xoi_31_cast_fp16))[name = tensor("out_31_cast_fp16")]; + tensor _inversed_2047_y_0_to_fp16 = const()[name = tensor("_inversed_2047_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2047_cast_fp16 = mul(x = position7_to_fp16, y = _inversed_2047_y_0_to_fp16)[name = tensor("_inversed_2047_cast_fp16")]; + tensor var_2048_cast_fp16 = floor(x = _inversed_2047_cast_fp16)[name = tensor("op_2048_cast_fp16")]; + tensor var_2049_to_fp16 = const()[name = tensor("op_2049_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2050_cast_fp16 = mul(x = var_2048_cast_fp16, y = var_2049_to_fp16)[name = tensor("op_2050_cast_fp16")]; + tensor write_idx_15_cast_fp16 = sub(x = position7_to_fp16, y = var_2050_cast_fp16)[name = tensor("write_idx_15_cast_fp16")]; + tensor var_2053_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_15_cast_fp16)[name = tensor("op_2053_cast_fp16")]; + tensor var_2063 = const()[name = tensor("op_2063"), val = tensor([1, 512, 1, 1])]; + tensor onehot_29_to_fp16_dtype_0 = const()[name = tensor("onehot_29_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_2053_cast_fp16_to_fp16 = cast(dtype = onehot_29_to_fp16_dtype_0, x = var_2053_cast_fp16)[name = tensor("cast_520")]; + tensor onehot_31_cast_fp16 = reshape(shape = var_2063, x = var_2053_cast_fp16_to_fp16)[name = tensor("onehot_31_cast_fp16")]; + tensor var_2065_to_fp16 = const()[name = tensor("op_2065_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2067_cast_fp16 = sub(x = var_2065_to_fp16, y = onehot_31_cast_fp16)[name = tensor("op_2067_cast_fp16")]; + tensor k_cache7_to_fp16_dtype_0 = const()[name = tensor("k_cache7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache7_to_fp16 = cast(dtype = k_cache7_to_fp16_dtype_0, x = k_cache7)[name = tensor("cast_519")]; + tensor var_2068_cast_fp16 = mul(x = k_cache7_to_fp16, y = var_2067_cast_fp16)[name = tensor("op_2068_cast_fp16")]; + tensor var_2073 = const()[name = tensor("op_2073"), val = tensor([1, 1, 16, 64])]; + tensor var_2074_cast_fp16 = reshape(shape = var_2073, x = out_31_cast_fp16)[name = tensor("op_2074_cast_fp16")]; + tensor var_2075_cast_fp16 = mul(x = var_2074_cast_fp16, y = onehot_31_cast_fp16)[name = tensor("op_2075_cast_fp16")]; + tensor new_k_cache_15_cast_fp16 = add(x = var_2068_cast_fp16, y = var_2075_cast_fp16)[name = tensor("new_k_cache_15_cast_fp16")]; + tensor new_k_cache_15_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_15_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache7_to_fp16_dtype_0 = const()[name = tensor("v_cache7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache7_to_fp16 = cast(dtype = v_cache7_to_fp16_dtype_0, x = v_cache7)[name = tensor("cast_517")]; + tensor var_2081_cast_fp16 = mul(x = v_cache7_to_fp16, y = var_2067_cast_fp16)[name = tensor("op_2081_cast_fp16")]; + tensor var_2086 = const()[name = tensor("op_2086"), val = tensor([1, 1, 16, 64])]; + tensor var_2087_cast_fp16 = reshape(shape = var_2086, x = v_15_cast_fp16)[name = tensor("op_2087_cast_fp16")]; + tensor var_2088_cast_fp16 = mul(x = var_2087_cast_fp16, y = onehot_31_cast_fp16)[name = tensor("op_2088_cast_fp16")]; + tensor new_v_cache_15_cast_fp16 = add(x = var_2081_cast_fp16, y = var_2088_cast_fp16)[name = tensor("new_v_cache_15_cast_fp16")]; + tensor new_v_cache_15_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_15_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_7_perm_0 = const()[name = tensor("transpose_7_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_2101 = const()[name = tensor("op_2101"), val = tensor([0, 2, 1, 3])]; + tensor var_2107 = const()[name = tensor("op_2107"), val = tensor([1, 16, 1, 64])]; + tensor q4_15_cast_fp16 = reshape(shape = var_2107, x = out_29_cast_fp16)[name = tensor("q4_15_cast_fp16")]; + tensor var_2112_transpose_x_0 = const()[name = tensor("op_2112_transpose_x_0"), val = tensor(false)]; + tensor var_2112_transpose_y_0 = const()[name = tensor("op_2112_transpose_y_0"), val = tensor(false)]; + tensor transpose_7_cast_fp16 = transpose(perm = transpose_7_perm_0, x = new_k_cache_15_cast_fp16)[name = tensor("transpose_74")]; + tensor var_2112_cast_fp16 = matmul(transpose_x = var_2112_transpose_x_0, transpose_y = var_2112_transpose_y_0, x = q4_15_cast_fp16, y = transpose_7_cast_fp16)[name = tensor("op_2112_cast_fp16")]; + tensor var_2113_to_fp16 = const()[name = tensor("op_2113_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_29_cast_fp16 = mul(x = var_2112_cast_fp16, y = var_2113_to_fp16)[name = tensor("scores_29_cast_fp16")]; + tensor var_2117 = const()[name = tensor("op_2117"), val = tensor([1, 1])]; + tensor var_2118_cast_fp16 = reshape(shape = var_2117, x = position7_to_fp16)[name = tensor("op_2118_cast_fp16")]; + tensor var_2119_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_2118_cast_fp16)[name = tensor("op_2119_cast_fp16")]; + tensor mask_15_to_fp16_dtype_0 = const()[name = tensor("mask_15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_2126_to_fp16 = const()[name = tensor("op_2126_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2119_cast_fp16_to_fp16 = cast(dtype = mask_15_to_fp16_dtype_0, x = var_2119_cast_fp16)[name = tensor("cast_515")]; + tensor var_2127_cast_fp16 = sub(x = var_2119_cast_fp16_to_fp16, y = var_2126_to_fp16)[name = tensor("op_2127_cast_fp16")]; + tensor var_2132 = const()[name = tensor("op_2132"), val = tensor([1, 1, 1, 512])]; + tensor var_2133_cast_fp16 = reshape(shape = var_2132, x = var_2127_cast_fp16)[name = tensor("op_2133_cast_fp16")]; + tensor var_2134_to_fp16 = const()[name = tensor("op_2134_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_2135_cast_fp16 = mul(x = var_2133_cast_fp16, y = var_2134_to_fp16)[name = tensor("op_2135_cast_fp16")]; + tensor scores_31_cast_fp16 = add(x = scores_29_cast_fp16, y = var_2135_cast_fp16)[name = tensor("scores_31_cast_fp16")]; + tensor var_2138 = const()[name = tensor("op_2138"), val = tensor(-1)]; + tensor weights_15_cast_fp16 = softmax(axis = var_2138, x = scores_31_cast_fp16)[name = tensor("weights_15_cast_fp16")]; + tensor attn_15_transpose_x_0 = const()[name = tensor("attn_15_transpose_x_0"), val = tensor(false)]; + tensor attn_15_transpose_y_0 = const()[name = tensor("attn_15_transpose_y_0"), val = tensor(false)]; + tensor values_15_cast_fp16 = transpose(perm = var_2101, x = new_v_cache_15_cast_fp16)[name = tensor("transpose_73")]; + tensor attn_15_cast_fp16 = matmul(transpose_x = attn_15_transpose_x_0, transpose_y = attn_15_transpose_y_0, x = weights_15_cast_fp16, y = values_15_cast_fp16)[name = tensor("attn_15_cast_fp16")]; + tensor var_2146 = const()[name = tensor("op_2146"), val = tensor([0, 2, 1, 3])]; + tensor var_2151 = const()[name = tensor("op_2151"), val = tensor([1, 1, 1024])]; + tensor var_2147_cast_fp16 = transpose(perm = var_2146, x = attn_15_cast_fp16)[name = tensor("transpose_72")]; + tensor input_91_cast_fp16 = reshape(shape = var_2151, x = var_2147_cast_fp16)[name = tensor("input_91_cast_fp16")]; + tensor attn7_out_proj_weight_to_fp16 = const()[name = tensor("attn7_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(182600896)))]; + tensor linear_30_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn7_out_proj_weight_to_fp16, x = input_91_cast_fp16)[name = tensor("linear_30_cast_fp16")]; + tensor var_2157_to_fp16 = const()[name = tensor("op_2157_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2158_cast_fp16 = add(x = position7_to_fp16, y = var_2157_to_fp16)[name = tensor("op_2158_cast_fp16")]; + tensor var_2158_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2158_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_93_cast_fp16 = add(x = input_87_cast_fp16, y = linear_30_cast_fp16)[name = tensor("input_93_cast_fp16")]; + tensor input_95_axes_0 = const()[name = tensor("input_95_axes_0"), val = tensor([-1])]; + tensor norm7_2_weight_to_fp16 = const()[name = tensor("norm7_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184698112)))]; + tensor norm7_2_bias_to_fp16 = const()[name = tensor("norm7_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184700224)))]; + tensor var_2162_to_fp16 = const()[name = tensor("op_2162_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_95_cast_fp16 = layer_norm(axes = input_95_axes_0, beta = norm7_2_bias_to_fp16, epsilon = var_2162_to_fp16, gamma = norm7_2_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("input_95_cast_fp16")]; + tensor linear7_1_weight_to_fp16 = const()[name = tensor("linear7_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184702336)))]; + tensor linear_31_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear7_1_weight_to_fp16, x = input_95_cast_fp16)[name = tensor("linear_31_cast_fp16")]; + tensor input_97_mode_0 = const()[name = tensor("input_97_mode_0"), val = tensor("EXACT")]; + tensor input_97_cast_fp16 = gelu(mode = input_97_mode_0, x = linear_31_cast_fp16)[name = tensor("input_97_cast_fp16")]; + tensor linear7_2_weight_to_fp16 = const()[name = tensor("linear7_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(193091008)))]; + tensor linear_32_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear7_2_weight_to_fp16, x = input_97_cast_fp16)[name = tensor("linear_32_cast_fp16")]; + tensor input_99_cast_fp16 = add(x = input_93_cast_fp16, y = linear_32_cast_fp16)[name = tensor("input_99_cast_fp16")]; + tensor input_101_axes_0 = const()[name = tensor("input_101_axes_0"), val = tensor([-1])]; + tensor norm8_1_weight_to_fp16 = const()[name = tensor("norm8_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201479680)))]; + tensor norm8_1_bias_to_fp16 = const()[name = tensor("norm8_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201481792)))]; + tensor var_2179_to_fp16 = const()[name = tensor("op_2179_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_101_cast_fp16 = layer_norm(axes = input_101_axes_0, beta = norm8_1_bias_to_fp16, epsilon = var_2179_to_fp16, gamma = norm8_1_weight_to_fp16, x = input_99_cast_fp16)[name = tensor("input_101_cast_fp16")]; + tensor attn8_in_proj_weight_to_fp16 = const()[name = tensor("attn8_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201483904)))]; + tensor linear_33_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn8_in_proj_weight_to_fp16, x = input_101_cast_fp16)[name = tensor("linear_33_cast_fp16")]; + tensor var_2192 = const()[name = tensor("op_2192"), val = tensor([1, 3, 16, 64])]; + tensor qkv_17_cast_fp16 = reshape(shape = var_2192, x = linear_33_cast_fp16)[name = tensor("qkv_17_cast_fp16")]; + tensor x_33_begin_0 = const()[name = tensor("x_33_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_33_end_0 = const()[name = tensor("x_33_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_33_end_mask_0 = const()[name = tensor("x_33_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_33_squeeze_mask_0 = const()[name = tensor("x_33_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_33_cast_fp16 = slice_by_index(begin = x_33_begin_0, end = x_33_end_0, end_mask = x_33_end_mask_0, squeeze_mask = x_33_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("x_33_cast_fp16")]; + tensor x_35_begin_0 = const()[name = tensor("x_35_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_35_end_0 = const()[name = tensor("x_35_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_35_end_mask_0 = const()[name = tensor("x_35_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_35_squeeze_mask_0 = const()[name = tensor("x_35_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_35_cast_fp16 = slice_by_index(begin = x_35_begin_0, end = x_35_end_0, end_mask = x_35_end_mask_0, squeeze_mask = x_35_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("x_35_cast_fp16")]; + tensor v_17_begin_0 = const()[name = tensor("v_17_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_17_end_0 = const()[name = tensor("v_17_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_17_end_mask_0 = const()[name = tensor("v_17_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_17_squeeze_mask_0 = const()[name = tensor("v_17_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_17_cast_fp16 = slice_by_index(begin = v_17_begin_0, end = v_17_end_0, end_mask = v_17_end_mask_0, squeeze_mask = v_17_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("v_17_cast_fp16")]; + tensor var_2221 = const()[name = tensor("op_2221"), val = tensor([1, 1, 1])]; + tensor position8_to_fp16_dtype_0 = const()[name = tensor("position8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position8_to_fp16 = cast(dtype = position8_to_fp16_dtype_0, x = position8)[name = tensor("cast_513")]; + tensor var_2222_cast_fp16 = reshape(shape = var_2221, x = position8_to_fp16)[name = tensor("op_2222_cast_fp16")]; + tensor angles_33_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_2222_cast_fp16)[name = tensor("angles_33_cast_fp16")]; + tensor rotr_33_cast_fp16 = cos(x = angles_33_cast_fp16)[name = tensor("rotr_33_cast_fp16")]; + tensor roti_33_cast_fp16 = sin(x = angles_33_cast_fp16)[name = tensor("roti_33_cast_fp16")]; + tensor var_2230 = const()[name = tensor("op_2230"), val = tensor([1, 16, 32, 2])]; + tensor pairs_33_cast_fp16 = reshape(shape = var_2230, x = x_33_cast_fp16)[name = tensor("pairs_33_cast_fp16")]; + tensor xr_33_begin_0 = const()[name = tensor("xr_33_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_33_end_0 = const()[name = tensor("xr_33_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_33_end_mask_0 = const()[name = tensor("xr_33_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_33_squeeze_mask_0 = const()[name = tensor("xr_33_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_33_cast_fp16 = slice_by_index(begin = xr_33_begin_0, end = xr_33_end_0, end_mask = xr_33_end_mask_0, squeeze_mask = xr_33_squeeze_mask_0, x = pairs_33_cast_fp16)[name = tensor("xr_33_cast_fp16")]; + tensor xi_33_begin_0 = const()[name = tensor("xi_33_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_33_end_0 = const()[name = tensor("xi_33_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_33_end_mask_0 = const()[name = tensor("xi_33_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_33_squeeze_mask_0 = const()[name = tensor("xi_33_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_33_cast_fp16 = slice_by_index(begin = xi_33_begin_0, end = xi_33_end_0, end_mask = xi_33_end_mask_0, squeeze_mask = xi_33_squeeze_mask_0, x = pairs_33_cast_fp16)[name = tensor("xi_33_cast_fp16")]; + tensor var_2238_cast_fp16 = mul(x = xr_33_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_2238_cast_fp16")]; + tensor var_2239_cast_fp16 = mul(x = xi_33_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_2239_cast_fp16")]; + tensor xor__33_cast_fp16 = sub(x = var_2238_cast_fp16, y = var_2239_cast_fp16)[name = tensor("xor__33_cast_fp16")]; + tensor var_2242_cast_fp16 = mul(x = xr_33_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_2242_cast_fp16")]; + tensor var_2243_cast_fp16 = mul(x = xi_33_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_2243_cast_fp16")]; + tensor xoi_33_cast_fp16 = add(x = var_2242_cast_fp16, y = var_2243_cast_fp16)[name = tensor("xoi_33_cast_fp16")]; + tensor out_33_axis_0 = const()[name = tensor("out_33_axis_0"), val = tensor(-1)]; + tensor out_33_cast_fp16 = stack(axis = out_33_axis_0, values = (xor__33_cast_fp16, xoi_33_cast_fp16))[name = tensor("out_33_cast_fp16")]; + tensor var_2266 = const()[name = tensor("op_2266"), val = tensor([1, 16, 32, 2])]; + tensor pairs_35_cast_fp16 = reshape(shape = var_2266, x = x_35_cast_fp16)[name = tensor("pairs_35_cast_fp16")]; + tensor xr_35_begin_0 = const()[name = tensor("xr_35_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_35_end_0 = const()[name = tensor("xr_35_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_35_end_mask_0 = const()[name = tensor("xr_35_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_35_squeeze_mask_0 = const()[name = tensor("xr_35_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_35_cast_fp16 = slice_by_index(begin = xr_35_begin_0, end = xr_35_end_0, end_mask = xr_35_end_mask_0, squeeze_mask = xr_35_squeeze_mask_0, x = pairs_35_cast_fp16)[name = tensor("xr_35_cast_fp16")]; + tensor xi_35_begin_0 = const()[name = tensor("xi_35_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_35_end_0 = const()[name = tensor("xi_35_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_35_end_mask_0 = const()[name = tensor("xi_35_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_35_squeeze_mask_0 = const()[name = tensor("xi_35_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_35_cast_fp16 = slice_by_index(begin = xi_35_begin_0, end = xi_35_end_0, end_mask = xi_35_end_mask_0, squeeze_mask = xi_35_squeeze_mask_0, x = pairs_35_cast_fp16)[name = tensor("xi_35_cast_fp16")]; + tensor var_2274_cast_fp16 = mul(x = xr_35_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_2274_cast_fp16")]; + tensor var_2275_cast_fp16 = mul(x = xi_35_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_2275_cast_fp16")]; + tensor xor__35_cast_fp16 = sub(x = var_2274_cast_fp16, y = var_2275_cast_fp16)[name = tensor("xor__35_cast_fp16")]; + tensor var_2278_cast_fp16 = mul(x = xr_35_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_2278_cast_fp16")]; + tensor var_2279_cast_fp16 = mul(x = xi_35_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_2279_cast_fp16")]; + tensor xoi_35_cast_fp16 = add(x = var_2278_cast_fp16, y = var_2279_cast_fp16)[name = tensor("xoi_35_cast_fp16")]; + tensor out_35_axis_0 = const()[name = tensor("out_35_axis_0"), val = tensor(-1)]; + tensor out_35_cast_fp16 = stack(axis = out_35_axis_0, values = (xor__35_cast_fp16, xoi_35_cast_fp16))[name = tensor("out_35_cast_fp16")]; + tensor _inversed_2291_y_0_to_fp16 = const()[name = tensor("_inversed_2291_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2291_cast_fp16 = mul(x = position8_to_fp16, y = _inversed_2291_y_0_to_fp16)[name = tensor("_inversed_2291_cast_fp16")]; + tensor var_2292_cast_fp16 = floor(x = _inversed_2291_cast_fp16)[name = tensor("op_2292_cast_fp16")]; + tensor var_2293_to_fp16 = const()[name = tensor("op_2293_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2294_cast_fp16 = mul(x = var_2292_cast_fp16, y = var_2293_to_fp16)[name = tensor("op_2294_cast_fp16")]; + tensor write_idx_17_cast_fp16 = sub(x = position8_to_fp16, y = var_2294_cast_fp16)[name = tensor("write_idx_17_cast_fp16")]; + tensor var_2297_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_17_cast_fp16)[name = tensor("op_2297_cast_fp16")]; + tensor var_2307 = const()[name = tensor("op_2307"), val = tensor([1, 512, 1, 1])]; + tensor onehot_33_to_fp16_dtype_0 = const()[name = tensor("onehot_33_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_2297_cast_fp16_to_fp16 = cast(dtype = onehot_33_to_fp16_dtype_0, x = var_2297_cast_fp16)[name = tensor("cast_512")]; + tensor onehot_35_cast_fp16 = reshape(shape = var_2307, x = var_2297_cast_fp16_to_fp16)[name = tensor("onehot_35_cast_fp16")]; + tensor var_2309_to_fp16 = const()[name = tensor("op_2309_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2311_cast_fp16 = sub(x = var_2309_to_fp16, y = onehot_35_cast_fp16)[name = tensor("op_2311_cast_fp16")]; + tensor k_cache8_to_fp16_dtype_0 = const()[name = tensor("k_cache8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache8_to_fp16 = cast(dtype = k_cache8_to_fp16_dtype_0, x = k_cache8)[name = tensor("cast_511")]; + tensor var_2312_cast_fp16 = mul(x = k_cache8_to_fp16, y = var_2311_cast_fp16)[name = tensor("op_2312_cast_fp16")]; + tensor var_2317 = const()[name = tensor("op_2317"), val = tensor([1, 1, 16, 64])]; + tensor var_2318_cast_fp16 = reshape(shape = var_2317, x = out_35_cast_fp16)[name = tensor("op_2318_cast_fp16")]; + tensor var_2319_cast_fp16 = mul(x = var_2318_cast_fp16, y = onehot_35_cast_fp16)[name = tensor("op_2319_cast_fp16")]; + tensor new_k_cache_17_cast_fp16 = add(x = var_2312_cast_fp16, y = var_2319_cast_fp16)[name = tensor("new_k_cache_17_cast_fp16")]; + tensor new_k_cache_17_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_17_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache8_to_fp16_dtype_0 = const()[name = tensor("v_cache8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache8_to_fp16 = cast(dtype = v_cache8_to_fp16_dtype_0, x = v_cache8)[name = tensor("cast_509")]; + tensor var_2325_cast_fp16 = mul(x = v_cache8_to_fp16, y = var_2311_cast_fp16)[name = tensor("op_2325_cast_fp16")]; + tensor var_2330 = const()[name = tensor("op_2330"), val = tensor([1, 1, 16, 64])]; + tensor var_2331_cast_fp16 = reshape(shape = var_2330, x = v_17_cast_fp16)[name = tensor("op_2331_cast_fp16")]; + tensor var_2332_cast_fp16 = mul(x = var_2331_cast_fp16, y = onehot_35_cast_fp16)[name = tensor("op_2332_cast_fp16")]; + tensor new_v_cache_17_cast_fp16 = add(x = var_2325_cast_fp16, y = var_2332_cast_fp16)[name = tensor("new_v_cache_17_cast_fp16")]; + tensor new_v_cache_17_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_17_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_8_perm_0 = const()[name = tensor("transpose_8_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_2345 = const()[name = tensor("op_2345"), val = tensor([0, 2, 1, 3])]; + tensor var_2351 = const()[name = tensor("op_2351"), val = tensor([1, 16, 1, 64])]; + tensor q4_17_cast_fp16 = reshape(shape = var_2351, x = out_33_cast_fp16)[name = tensor("q4_17_cast_fp16")]; + tensor var_2356_transpose_x_0 = const()[name = tensor("op_2356_transpose_x_0"), val = tensor(false)]; + tensor var_2356_transpose_y_0 = const()[name = tensor("op_2356_transpose_y_0"), val = tensor(false)]; + tensor transpose_8_cast_fp16 = transpose(perm = transpose_8_perm_0, x = new_k_cache_17_cast_fp16)[name = tensor("transpose_71")]; + tensor var_2356_cast_fp16 = matmul(transpose_x = var_2356_transpose_x_0, transpose_y = var_2356_transpose_y_0, x = q4_17_cast_fp16, y = transpose_8_cast_fp16)[name = tensor("op_2356_cast_fp16")]; + tensor var_2357_to_fp16 = const()[name = tensor("op_2357_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_33_cast_fp16 = mul(x = var_2356_cast_fp16, y = var_2357_to_fp16)[name = tensor("scores_33_cast_fp16")]; + tensor var_2361 = const()[name = tensor("op_2361"), val = tensor([1, 1])]; + tensor var_2362_cast_fp16 = reshape(shape = var_2361, x = position8_to_fp16)[name = tensor("op_2362_cast_fp16")]; + tensor var_2363_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_2362_cast_fp16)[name = tensor("op_2363_cast_fp16")]; + tensor mask_17_to_fp16_dtype_0 = const()[name = tensor("mask_17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_2370_to_fp16 = const()[name = tensor("op_2370_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2363_cast_fp16_to_fp16 = cast(dtype = mask_17_to_fp16_dtype_0, x = var_2363_cast_fp16)[name = tensor("cast_507")]; + tensor var_2371_cast_fp16 = sub(x = var_2363_cast_fp16_to_fp16, y = var_2370_to_fp16)[name = tensor("op_2371_cast_fp16")]; + tensor var_2376 = const()[name = tensor("op_2376"), val = tensor([1, 1, 1, 512])]; + tensor var_2377_cast_fp16 = reshape(shape = var_2376, x = var_2371_cast_fp16)[name = tensor("op_2377_cast_fp16")]; + tensor var_2378_to_fp16 = const()[name = tensor("op_2378_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_2379_cast_fp16 = mul(x = var_2377_cast_fp16, y = var_2378_to_fp16)[name = tensor("op_2379_cast_fp16")]; + tensor scores_35_cast_fp16 = add(x = scores_33_cast_fp16, y = var_2379_cast_fp16)[name = tensor("scores_35_cast_fp16")]; + tensor var_2382 = const()[name = tensor("op_2382"), val = tensor(-1)]; + tensor weights_17_cast_fp16 = softmax(axis = var_2382, x = scores_35_cast_fp16)[name = tensor("weights_17_cast_fp16")]; + tensor attn_17_transpose_x_0 = const()[name = tensor("attn_17_transpose_x_0"), val = tensor(false)]; + tensor attn_17_transpose_y_0 = const()[name = tensor("attn_17_transpose_y_0"), val = tensor(false)]; + tensor values_17_cast_fp16 = transpose(perm = var_2345, x = new_v_cache_17_cast_fp16)[name = tensor("transpose_70")]; + tensor attn_17_cast_fp16 = matmul(transpose_x = attn_17_transpose_x_0, transpose_y = attn_17_transpose_y_0, x = weights_17_cast_fp16, y = values_17_cast_fp16)[name = tensor("attn_17_cast_fp16")]; + tensor var_2390 = const()[name = tensor("op_2390"), val = tensor([0, 2, 1, 3])]; + tensor var_2395 = const()[name = tensor("op_2395"), val = tensor([1, 1, 1024])]; + tensor var_2391_cast_fp16 = transpose(perm = var_2390, x = attn_17_cast_fp16)[name = tensor("transpose_69")]; + tensor input_103_cast_fp16 = reshape(shape = var_2395, x = var_2391_cast_fp16)[name = tensor("input_103_cast_fp16")]; + tensor attn8_out_proj_weight_to_fp16 = const()[name = tensor("attn8_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(207775424)))]; + tensor linear_34_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn8_out_proj_weight_to_fp16, x = input_103_cast_fp16)[name = tensor("linear_34_cast_fp16")]; + tensor var_2401_to_fp16 = const()[name = tensor("op_2401_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2402_cast_fp16 = add(x = position8_to_fp16, y = var_2401_to_fp16)[name = tensor("op_2402_cast_fp16")]; + tensor var_2402_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2402_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_105_cast_fp16 = add(x = input_99_cast_fp16, y = linear_34_cast_fp16)[name = tensor("input_105_cast_fp16")]; + tensor input_107_axes_0 = const()[name = tensor("input_107_axes_0"), val = tensor([-1])]; + tensor norm8_2_weight_to_fp16 = const()[name = tensor("norm8_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209872640)))]; + tensor norm8_2_bias_to_fp16 = const()[name = tensor("norm8_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209874752)))]; + tensor var_2406_to_fp16 = const()[name = tensor("op_2406_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_107_cast_fp16 = layer_norm(axes = input_107_axes_0, beta = norm8_2_bias_to_fp16, epsilon = var_2406_to_fp16, gamma = norm8_2_weight_to_fp16, x = input_105_cast_fp16)[name = tensor("input_107_cast_fp16")]; + tensor linear8_1_weight_to_fp16 = const()[name = tensor("linear8_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209876864)))]; + tensor linear_35_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear8_1_weight_to_fp16, x = input_107_cast_fp16)[name = tensor("linear_35_cast_fp16")]; + tensor input_109_mode_0 = const()[name = tensor("input_109_mode_0"), val = tensor("EXACT")]; + tensor input_109_cast_fp16 = gelu(mode = input_109_mode_0, x = linear_35_cast_fp16)[name = tensor("input_109_cast_fp16")]; + tensor linear8_2_weight_to_fp16 = const()[name = tensor("linear8_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(218265536)))]; + tensor linear_36_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear8_2_weight_to_fp16, x = input_109_cast_fp16)[name = tensor("linear_36_cast_fp16")]; + tensor input_111_cast_fp16 = add(x = input_105_cast_fp16, y = linear_36_cast_fp16)[name = tensor("input_111_cast_fp16")]; + tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([-1])]; + tensor norm9_1_weight_to_fp16 = const()[name = tensor("norm9_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226654208)))]; + tensor norm9_1_bias_to_fp16 = const()[name = tensor("norm9_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226656320)))]; + tensor var_2423_to_fp16 = const()[name = tensor("op_2423_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_113_cast_fp16 = layer_norm(axes = input_113_axes_0, beta = norm9_1_bias_to_fp16, epsilon = var_2423_to_fp16, gamma = norm9_1_weight_to_fp16, x = input_111_cast_fp16)[name = tensor("input_113_cast_fp16")]; + tensor attn9_in_proj_weight_to_fp16 = const()[name = tensor("attn9_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226658432)))]; + tensor linear_37_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn9_in_proj_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("linear_37_cast_fp16")]; + tensor var_2436 = const()[name = tensor("op_2436"), val = tensor([1, 3, 16, 64])]; + tensor qkv_19_cast_fp16 = reshape(shape = var_2436, x = linear_37_cast_fp16)[name = tensor("qkv_19_cast_fp16")]; + tensor x_37_begin_0 = const()[name = tensor("x_37_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_37_end_0 = const()[name = tensor("x_37_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_37_end_mask_0 = const()[name = tensor("x_37_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_37_squeeze_mask_0 = const()[name = tensor("x_37_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_37_cast_fp16 = slice_by_index(begin = x_37_begin_0, end = x_37_end_0, end_mask = x_37_end_mask_0, squeeze_mask = x_37_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("x_37_cast_fp16")]; + tensor x_39_begin_0 = const()[name = tensor("x_39_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_39_end_0 = const()[name = tensor("x_39_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_39_end_mask_0 = const()[name = tensor("x_39_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_39_squeeze_mask_0 = const()[name = tensor("x_39_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_39_cast_fp16 = slice_by_index(begin = x_39_begin_0, end = x_39_end_0, end_mask = x_39_end_mask_0, squeeze_mask = x_39_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("x_39_cast_fp16")]; + tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_19_cast_fp16 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("v_19_cast_fp16")]; + tensor var_2465 = const()[name = tensor("op_2465"), val = tensor([1, 1, 1])]; + tensor position9_to_fp16_dtype_0 = const()[name = tensor("position9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position9_to_fp16 = cast(dtype = position9_to_fp16_dtype_0, x = position9)[name = tensor("cast_505")]; + tensor var_2466_cast_fp16 = reshape(shape = var_2465, x = position9_to_fp16)[name = tensor("op_2466_cast_fp16")]; + tensor angles_37_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_2466_cast_fp16)[name = tensor("angles_37_cast_fp16")]; + tensor rotr_37_cast_fp16 = cos(x = angles_37_cast_fp16)[name = tensor("rotr_37_cast_fp16")]; + tensor roti_37_cast_fp16 = sin(x = angles_37_cast_fp16)[name = tensor("roti_37_cast_fp16")]; + tensor var_2474 = const()[name = tensor("op_2474"), val = tensor([1, 16, 32, 2])]; + tensor pairs_37_cast_fp16 = reshape(shape = var_2474, x = x_37_cast_fp16)[name = tensor("pairs_37_cast_fp16")]; + tensor xr_37_begin_0 = const()[name = tensor("xr_37_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_37_end_0 = const()[name = tensor("xr_37_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_37_end_mask_0 = const()[name = tensor("xr_37_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_37_squeeze_mask_0 = const()[name = tensor("xr_37_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_37_cast_fp16 = slice_by_index(begin = xr_37_begin_0, end = xr_37_end_0, end_mask = xr_37_end_mask_0, squeeze_mask = xr_37_squeeze_mask_0, x = pairs_37_cast_fp16)[name = tensor("xr_37_cast_fp16")]; + tensor xi_37_begin_0 = const()[name = tensor("xi_37_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_37_end_0 = const()[name = tensor("xi_37_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_37_end_mask_0 = const()[name = tensor("xi_37_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_37_squeeze_mask_0 = const()[name = tensor("xi_37_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_37_cast_fp16 = slice_by_index(begin = xi_37_begin_0, end = xi_37_end_0, end_mask = xi_37_end_mask_0, squeeze_mask = xi_37_squeeze_mask_0, x = pairs_37_cast_fp16)[name = tensor("xi_37_cast_fp16")]; + tensor var_2482_cast_fp16 = mul(x = xr_37_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_2482_cast_fp16")]; + tensor var_2483_cast_fp16 = mul(x = xi_37_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_2483_cast_fp16")]; + tensor xor__37_cast_fp16 = sub(x = var_2482_cast_fp16, y = var_2483_cast_fp16)[name = tensor("xor__37_cast_fp16")]; + tensor var_2486_cast_fp16 = mul(x = xr_37_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_2486_cast_fp16")]; + tensor var_2487_cast_fp16 = mul(x = xi_37_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_2487_cast_fp16")]; + tensor xoi_37_cast_fp16 = add(x = var_2486_cast_fp16, y = var_2487_cast_fp16)[name = tensor("xoi_37_cast_fp16")]; + tensor out_37_axis_0 = const()[name = tensor("out_37_axis_0"), val = tensor(-1)]; + tensor out_37_cast_fp16 = stack(axis = out_37_axis_0, values = (xor__37_cast_fp16, xoi_37_cast_fp16))[name = tensor("out_37_cast_fp16")]; + tensor var_2510 = const()[name = tensor("op_2510"), val = tensor([1, 16, 32, 2])]; + tensor pairs_39_cast_fp16 = reshape(shape = var_2510, x = x_39_cast_fp16)[name = tensor("pairs_39_cast_fp16")]; + tensor xr_39_begin_0 = const()[name = tensor("xr_39_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_39_end_0 = const()[name = tensor("xr_39_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_39_end_mask_0 = const()[name = tensor("xr_39_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_39_squeeze_mask_0 = const()[name = tensor("xr_39_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_39_cast_fp16 = slice_by_index(begin = xr_39_begin_0, end = xr_39_end_0, end_mask = xr_39_end_mask_0, squeeze_mask = xr_39_squeeze_mask_0, x = pairs_39_cast_fp16)[name = tensor("xr_39_cast_fp16")]; + tensor xi_39_begin_0 = const()[name = tensor("xi_39_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_39_end_0 = const()[name = tensor("xi_39_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_39_end_mask_0 = const()[name = tensor("xi_39_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_39_squeeze_mask_0 = const()[name = tensor("xi_39_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_39_cast_fp16 = slice_by_index(begin = xi_39_begin_0, end = xi_39_end_0, end_mask = xi_39_end_mask_0, squeeze_mask = xi_39_squeeze_mask_0, x = pairs_39_cast_fp16)[name = tensor("xi_39_cast_fp16")]; + tensor var_2518_cast_fp16 = mul(x = xr_39_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_2518_cast_fp16")]; + tensor var_2519_cast_fp16 = mul(x = xi_39_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_2519_cast_fp16")]; + tensor xor__39_cast_fp16 = sub(x = var_2518_cast_fp16, y = var_2519_cast_fp16)[name = tensor("xor__39_cast_fp16")]; + tensor var_2522_cast_fp16 = mul(x = xr_39_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_2522_cast_fp16")]; + tensor var_2523_cast_fp16 = mul(x = xi_39_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_2523_cast_fp16")]; + tensor xoi_39_cast_fp16 = add(x = var_2522_cast_fp16, y = var_2523_cast_fp16)[name = tensor("xoi_39_cast_fp16")]; + tensor out_39_axis_0 = const()[name = tensor("out_39_axis_0"), val = tensor(-1)]; + tensor out_39_cast_fp16 = stack(axis = out_39_axis_0, values = (xor__39_cast_fp16, xoi_39_cast_fp16))[name = tensor("out_39_cast_fp16")]; + tensor _inversed_2535_y_0_to_fp16 = const()[name = tensor("_inversed_2535_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2535_cast_fp16 = mul(x = position9_to_fp16, y = _inversed_2535_y_0_to_fp16)[name = tensor("_inversed_2535_cast_fp16")]; + tensor var_2536_cast_fp16 = floor(x = _inversed_2535_cast_fp16)[name = tensor("op_2536_cast_fp16")]; + tensor var_2537_to_fp16 = const()[name = tensor("op_2537_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2538_cast_fp16 = mul(x = var_2536_cast_fp16, y = var_2537_to_fp16)[name = tensor("op_2538_cast_fp16")]; + tensor write_idx_19_cast_fp16 = sub(x = position9_to_fp16, y = var_2538_cast_fp16)[name = tensor("write_idx_19_cast_fp16")]; + tensor var_2541_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_19_cast_fp16)[name = tensor("op_2541_cast_fp16")]; + tensor var_2551 = const()[name = tensor("op_2551"), val = tensor([1, 512, 1, 1])]; + tensor onehot_37_to_fp16_dtype_0 = const()[name = tensor("onehot_37_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_2541_cast_fp16_to_fp16 = cast(dtype = onehot_37_to_fp16_dtype_0, x = var_2541_cast_fp16)[name = tensor("cast_504")]; + tensor onehot_39_cast_fp16 = reshape(shape = var_2551, x = var_2541_cast_fp16_to_fp16)[name = tensor("onehot_39_cast_fp16")]; + tensor var_2553_to_fp16 = const()[name = tensor("op_2553_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2555_cast_fp16 = sub(x = var_2553_to_fp16, y = onehot_39_cast_fp16)[name = tensor("op_2555_cast_fp16")]; + tensor k_cache9_to_fp16_dtype_0 = const()[name = tensor("k_cache9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache9_to_fp16 = cast(dtype = k_cache9_to_fp16_dtype_0, x = k_cache9)[name = tensor("cast_503")]; + tensor var_2556_cast_fp16 = mul(x = k_cache9_to_fp16, y = var_2555_cast_fp16)[name = tensor("op_2556_cast_fp16")]; + tensor var_2561 = const()[name = tensor("op_2561"), val = tensor([1, 1, 16, 64])]; + tensor var_2562_cast_fp16 = reshape(shape = var_2561, x = out_39_cast_fp16)[name = tensor("op_2562_cast_fp16")]; + tensor var_2563_cast_fp16 = mul(x = var_2562_cast_fp16, y = onehot_39_cast_fp16)[name = tensor("op_2563_cast_fp16")]; + tensor new_k_cache_19_cast_fp16 = add(x = var_2556_cast_fp16, y = var_2563_cast_fp16)[name = tensor("new_k_cache_19_cast_fp16")]; + tensor new_k_cache_19_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_19_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache9_to_fp16_dtype_0 = const()[name = tensor("v_cache9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache9_to_fp16 = cast(dtype = v_cache9_to_fp16_dtype_0, x = v_cache9)[name = tensor("cast_501")]; + tensor var_2569_cast_fp16 = mul(x = v_cache9_to_fp16, y = var_2555_cast_fp16)[name = tensor("op_2569_cast_fp16")]; + tensor var_2574 = const()[name = tensor("op_2574"), val = tensor([1, 1, 16, 64])]; + tensor var_2575_cast_fp16 = reshape(shape = var_2574, x = v_19_cast_fp16)[name = tensor("op_2575_cast_fp16")]; + tensor var_2576_cast_fp16 = mul(x = var_2575_cast_fp16, y = onehot_39_cast_fp16)[name = tensor("op_2576_cast_fp16")]; + tensor new_v_cache_19_cast_fp16 = add(x = var_2569_cast_fp16, y = var_2576_cast_fp16)[name = tensor("new_v_cache_19_cast_fp16")]; + tensor new_v_cache_19_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_19_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_9_perm_0 = const()[name = tensor("transpose_9_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_2589 = const()[name = tensor("op_2589"), val = tensor([0, 2, 1, 3])]; + tensor var_2595 = const()[name = tensor("op_2595"), val = tensor([1, 16, 1, 64])]; + tensor q4_19_cast_fp16 = reshape(shape = var_2595, x = out_37_cast_fp16)[name = tensor("q4_19_cast_fp16")]; + tensor var_2600_transpose_x_0 = const()[name = tensor("op_2600_transpose_x_0"), val = tensor(false)]; + tensor var_2600_transpose_y_0 = const()[name = tensor("op_2600_transpose_y_0"), val = tensor(false)]; + tensor transpose_9_cast_fp16 = transpose(perm = transpose_9_perm_0, x = new_k_cache_19_cast_fp16)[name = tensor("transpose_68")]; + tensor var_2600_cast_fp16 = matmul(transpose_x = var_2600_transpose_x_0, transpose_y = var_2600_transpose_y_0, x = q4_19_cast_fp16, y = transpose_9_cast_fp16)[name = tensor("op_2600_cast_fp16")]; + tensor var_2601_to_fp16 = const()[name = tensor("op_2601_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_37_cast_fp16 = mul(x = var_2600_cast_fp16, y = var_2601_to_fp16)[name = tensor("scores_37_cast_fp16")]; + tensor var_2605 = const()[name = tensor("op_2605"), val = tensor([1, 1])]; + tensor var_2606_cast_fp16 = reshape(shape = var_2605, x = position9_to_fp16)[name = tensor("op_2606_cast_fp16")]; + tensor var_2607_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_2606_cast_fp16)[name = tensor("op_2607_cast_fp16")]; + tensor mask_19_to_fp16_dtype_0 = const()[name = tensor("mask_19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_2614_to_fp16 = const()[name = tensor("op_2614_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2607_cast_fp16_to_fp16 = cast(dtype = mask_19_to_fp16_dtype_0, x = var_2607_cast_fp16)[name = tensor("cast_499")]; + tensor var_2615_cast_fp16 = sub(x = var_2607_cast_fp16_to_fp16, y = var_2614_to_fp16)[name = tensor("op_2615_cast_fp16")]; + tensor var_2620 = const()[name = tensor("op_2620"), val = tensor([1, 1, 1, 512])]; + tensor var_2621_cast_fp16 = reshape(shape = var_2620, x = var_2615_cast_fp16)[name = tensor("op_2621_cast_fp16")]; + tensor var_2622_to_fp16 = const()[name = tensor("op_2622_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_2623_cast_fp16 = mul(x = var_2621_cast_fp16, y = var_2622_to_fp16)[name = tensor("op_2623_cast_fp16")]; + tensor scores_39_cast_fp16 = add(x = scores_37_cast_fp16, y = var_2623_cast_fp16)[name = tensor("scores_39_cast_fp16")]; + tensor var_2626 = const()[name = tensor("op_2626"), val = tensor(-1)]; + tensor weights_19_cast_fp16 = softmax(axis = var_2626, x = scores_39_cast_fp16)[name = tensor("weights_19_cast_fp16")]; + tensor attn_19_transpose_x_0 = const()[name = tensor("attn_19_transpose_x_0"), val = tensor(false)]; + tensor attn_19_transpose_y_0 = const()[name = tensor("attn_19_transpose_y_0"), val = tensor(false)]; + tensor values_19_cast_fp16 = transpose(perm = var_2589, x = new_v_cache_19_cast_fp16)[name = tensor("transpose_67")]; + tensor attn_19_cast_fp16 = matmul(transpose_x = attn_19_transpose_x_0, transpose_y = attn_19_transpose_y_0, x = weights_19_cast_fp16, y = values_19_cast_fp16)[name = tensor("attn_19_cast_fp16")]; + tensor var_2634 = const()[name = tensor("op_2634"), val = tensor([0, 2, 1, 3])]; + tensor var_2639 = const()[name = tensor("op_2639"), val = tensor([1, 1, 1024])]; + tensor var_2635_cast_fp16 = transpose(perm = var_2634, x = attn_19_cast_fp16)[name = tensor("transpose_66")]; + tensor input_115_cast_fp16 = reshape(shape = var_2639, x = var_2635_cast_fp16)[name = tensor("input_115_cast_fp16")]; + tensor attn9_out_proj_weight_to_fp16 = const()[name = tensor("attn9_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(232949952)))]; + tensor linear_38_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn9_out_proj_weight_to_fp16, x = input_115_cast_fp16)[name = tensor("linear_38_cast_fp16")]; + tensor var_2645_to_fp16 = const()[name = tensor("op_2645_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2646_cast_fp16 = add(x = position9_to_fp16, y = var_2645_to_fp16)[name = tensor("op_2646_cast_fp16")]; + tensor var_2646_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2646_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_117_cast_fp16 = add(x = input_111_cast_fp16, y = linear_38_cast_fp16)[name = tensor("input_117_cast_fp16")]; + tensor input_119_axes_0 = const()[name = tensor("input_119_axes_0"), val = tensor([-1])]; + tensor norm9_2_weight_to_fp16 = const()[name = tensor("norm9_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(235047168)))]; + tensor norm9_2_bias_to_fp16 = const()[name = tensor("norm9_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(235049280)))]; + tensor var_2650_to_fp16 = const()[name = tensor("op_2650_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_119_cast_fp16 = layer_norm(axes = input_119_axes_0, beta = norm9_2_bias_to_fp16, epsilon = var_2650_to_fp16, gamma = norm9_2_weight_to_fp16, x = input_117_cast_fp16)[name = tensor("input_119_cast_fp16")]; + tensor linear9_1_weight_to_fp16 = const()[name = tensor("linear9_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(235051392)))]; + tensor linear_39_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear9_1_weight_to_fp16, x = input_119_cast_fp16)[name = tensor("linear_39_cast_fp16")]; + tensor input_121_mode_0 = const()[name = tensor("input_121_mode_0"), val = tensor("EXACT")]; + tensor input_121_cast_fp16 = gelu(mode = input_121_mode_0, x = linear_39_cast_fp16)[name = tensor("input_121_cast_fp16")]; + tensor linear9_2_weight_to_fp16 = const()[name = tensor("linear9_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(243440064)))]; + tensor linear_40_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear9_2_weight_to_fp16, x = input_121_cast_fp16)[name = tensor("linear_40_cast_fp16")]; + tensor input_123_cast_fp16 = add(x = input_117_cast_fp16, y = linear_40_cast_fp16)[name = tensor("input_123_cast_fp16")]; + tensor input_125_axes_0 = const()[name = tensor("input_125_axes_0"), val = tensor([-1])]; + tensor norm10_1_weight_to_fp16 = const()[name = tensor("norm10_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(251828736)))]; + tensor norm10_1_bias_to_fp16 = const()[name = tensor("norm10_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(251830848)))]; + tensor var_2667_to_fp16 = const()[name = tensor("op_2667_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_125_cast_fp16 = layer_norm(axes = input_125_axes_0, beta = norm10_1_bias_to_fp16, epsilon = var_2667_to_fp16, gamma = norm10_1_weight_to_fp16, x = input_123_cast_fp16)[name = tensor("input_125_cast_fp16")]; + tensor attn10_in_proj_weight_to_fp16 = const()[name = tensor("attn10_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(251832960)))]; + tensor linear_41_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn10_in_proj_weight_to_fp16, x = input_125_cast_fp16)[name = tensor("linear_41_cast_fp16")]; + tensor var_2680 = const()[name = tensor("op_2680"), val = tensor([1, 3, 16, 64])]; + tensor qkv_21_cast_fp16 = reshape(shape = var_2680, x = linear_41_cast_fp16)[name = tensor("qkv_21_cast_fp16")]; + tensor x_41_begin_0 = const()[name = tensor("x_41_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_41_end_0 = const()[name = tensor("x_41_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_41_end_mask_0 = const()[name = tensor("x_41_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_41_squeeze_mask_0 = const()[name = tensor("x_41_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_41_cast_fp16 = slice_by_index(begin = x_41_begin_0, end = x_41_end_0, end_mask = x_41_end_mask_0, squeeze_mask = x_41_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("x_41_cast_fp16")]; + tensor x_43_begin_0 = const()[name = tensor("x_43_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_43_end_0 = const()[name = tensor("x_43_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_43_end_mask_0 = const()[name = tensor("x_43_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_43_squeeze_mask_0 = const()[name = tensor("x_43_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_43_cast_fp16 = slice_by_index(begin = x_43_begin_0, end = x_43_end_0, end_mask = x_43_end_mask_0, squeeze_mask = x_43_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("x_43_cast_fp16")]; + tensor v_21_begin_0 = const()[name = tensor("v_21_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_21_end_0 = const()[name = tensor("v_21_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_21_end_mask_0 = const()[name = tensor("v_21_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_21_squeeze_mask_0 = const()[name = tensor("v_21_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_21_cast_fp16 = slice_by_index(begin = v_21_begin_0, end = v_21_end_0, end_mask = v_21_end_mask_0, squeeze_mask = v_21_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("v_21_cast_fp16")]; + tensor var_2709 = const()[name = tensor("op_2709"), val = tensor([1, 1, 1])]; + tensor position10_to_fp16_dtype_0 = const()[name = tensor("position10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position10_to_fp16 = cast(dtype = position10_to_fp16_dtype_0, x = position10)[name = tensor("cast_497")]; + tensor var_2710_cast_fp16 = reshape(shape = var_2709, x = position10_to_fp16)[name = tensor("op_2710_cast_fp16")]; + tensor angles_41_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_2710_cast_fp16)[name = tensor("angles_41_cast_fp16")]; + tensor rotr_41_cast_fp16 = cos(x = angles_41_cast_fp16)[name = tensor("rotr_41_cast_fp16")]; + tensor roti_41_cast_fp16 = sin(x = angles_41_cast_fp16)[name = tensor("roti_41_cast_fp16")]; + tensor var_2718 = const()[name = tensor("op_2718"), val = tensor([1, 16, 32, 2])]; + tensor pairs_41_cast_fp16 = reshape(shape = var_2718, x = x_41_cast_fp16)[name = tensor("pairs_41_cast_fp16")]; + tensor xr_41_begin_0 = const()[name = tensor("xr_41_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_41_end_0 = const()[name = tensor("xr_41_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_41_end_mask_0 = const()[name = tensor("xr_41_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_41_squeeze_mask_0 = const()[name = tensor("xr_41_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_41_cast_fp16 = slice_by_index(begin = xr_41_begin_0, end = xr_41_end_0, end_mask = xr_41_end_mask_0, squeeze_mask = xr_41_squeeze_mask_0, x = pairs_41_cast_fp16)[name = tensor("xr_41_cast_fp16")]; + tensor xi_41_begin_0 = const()[name = tensor("xi_41_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_41_end_0 = const()[name = tensor("xi_41_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_41_end_mask_0 = const()[name = tensor("xi_41_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_41_squeeze_mask_0 = const()[name = tensor("xi_41_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_41_cast_fp16 = slice_by_index(begin = xi_41_begin_0, end = xi_41_end_0, end_mask = xi_41_end_mask_0, squeeze_mask = xi_41_squeeze_mask_0, x = pairs_41_cast_fp16)[name = tensor("xi_41_cast_fp16")]; + tensor var_2726_cast_fp16 = mul(x = xr_41_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_2726_cast_fp16")]; + tensor var_2727_cast_fp16 = mul(x = xi_41_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_2727_cast_fp16")]; + tensor xor__41_cast_fp16 = sub(x = var_2726_cast_fp16, y = var_2727_cast_fp16)[name = tensor("xor__41_cast_fp16")]; + tensor var_2730_cast_fp16 = mul(x = xr_41_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_2730_cast_fp16")]; + tensor var_2731_cast_fp16 = mul(x = xi_41_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_2731_cast_fp16")]; + tensor xoi_41_cast_fp16 = add(x = var_2730_cast_fp16, y = var_2731_cast_fp16)[name = tensor("xoi_41_cast_fp16")]; + tensor out_41_axis_0 = const()[name = tensor("out_41_axis_0"), val = tensor(-1)]; + tensor out_41_cast_fp16 = stack(axis = out_41_axis_0, values = (xor__41_cast_fp16, xoi_41_cast_fp16))[name = tensor("out_41_cast_fp16")]; + tensor var_2754 = const()[name = tensor("op_2754"), val = tensor([1, 16, 32, 2])]; + tensor pairs_43_cast_fp16 = reshape(shape = var_2754, x = x_43_cast_fp16)[name = tensor("pairs_43_cast_fp16")]; + tensor xr_43_begin_0 = const()[name = tensor("xr_43_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_43_end_0 = const()[name = tensor("xr_43_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_43_end_mask_0 = const()[name = tensor("xr_43_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_43_squeeze_mask_0 = const()[name = tensor("xr_43_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_43_cast_fp16 = slice_by_index(begin = xr_43_begin_0, end = xr_43_end_0, end_mask = xr_43_end_mask_0, squeeze_mask = xr_43_squeeze_mask_0, x = pairs_43_cast_fp16)[name = tensor("xr_43_cast_fp16")]; + tensor xi_43_begin_0 = const()[name = tensor("xi_43_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_43_end_0 = const()[name = tensor("xi_43_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_43_end_mask_0 = const()[name = tensor("xi_43_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_43_squeeze_mask_0 = const()[name = tensor("xi_43_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_43_cast_fp16 = slice_by_index(begin = xi_43_begin_0, end = xi_43_end_0, end_mask = xi_43_end_mask_0, squeeze_mask = xi_43_squeeze_mask_0, x = pairs_43_cast_fp16)[name = tensor("xi_43_cast_fp16")]; + tensor var_2762_cast_fp16 = mul(x = xr_43_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_2762_cast_fp16")]; + tensor var_2763_cast_fp16 = mul(x = xi_43_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_2763_cast_fp16")]; + tensor xor__43_cast_fp16 = sub(x = var_2762_cast_fp16, y = var_2763_cast_fp16)[name = tensor("xor__43_cast_fp16")]; + tensor var_2766_cast_fp16 = mul(x = xr_43_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_2766_cast_fp16")]; + tensor var_2767_cast_fp16 = mul(x = xi_43_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_2767_cast_fp16")]; + tensor xoi_43_cast_fp16 = add(x = var_2766_cast_fp16, y = var_2767_cast_fp16)[name = tensor("xoi_43_cast_fp16")]; + tensor out_43_axis_0 = const()[name = tensor("out_43_axis_0"), val = tensor(-1)]; + tensor out_43_cast_fp16 = stack(axis = out_43_axis_0, values = (xor__43_cast_fp16, xoi_43_cast_fp16))[name = tensor("out_43_cast_fp16")]; + tensor _inversed_2779_y_0_to_fp16 = const()[name = tensor("_inversed_2779_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2779_cast_fp16 = mul(x = position10_to_fp16, y = _inversed_2779_y_0_to_fp16)[name = tensor("_inversed_2779_cast_fp16")]; + tensor var_2780_cast_fp16 = floor(x = _inversed_2779_cast_fp16)[name = tensor("op_2780_cast_fp16")]; + tensor var_2781_to_fp16 = const()[name = tensor("op_2781_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2782_cast_fp16 = mul(x = var_2780_cast_fp16, y = var_2781_to_fp16)[name = tensor("op_2782_cast_fp16")]; + tensor write_idx_21_cast_fp16 = sub(x = position10_to_fp16, y = var_2782_cast_fp16)[name = tensor("write_idx_21_cast_fp16")]; + tensor var_2785_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_21_cast_fp16)[name = tensor("op_2785_cast_fp16")]; + tensor var_2795 = const()[name = tensor("op_2795"), val = tensor([1, 512, 1, 1])]; + tensor onehot_41_to_fp16_dtype_0 = const()[name = tensor("onehot_41_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_2785_cast_fp16_to_fp16 = cast(dtype = onehot_41_to_fp16_dtype_0, x = var_2785_cast_fp16)[name = tensor("cast_496")]; + tensor onehot_43_cast_fp16 = reshape(shape = var_2795, x = var_2785_cast_fp16_to_fp16)[name = tensor("onehot_43_cast_fp16")]; + tensor var_2797_to_fp16 = const()[name = tensor("op_2797_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2799_cast_fp16 = sub(x = var_2797_to_fp16, y = onehot_43_cast_fp16)[name = tensor("op_2799_cast_fp16")]; + tensor k_cache10_to_fp16_dtype_0 = const()[name = tensor("k_cache10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache10_to_fp16 = cast(dtype = k_cache10_to_fp16_dtype_0, x = k_cache10)[name = tensor("cast_495")]; + tensor var_2800_cast_fp16 = mul(x = k_cache10_to_fp16, y = var_2799_cast_fp16)[name = tensor("op_2800_cast_fp16")]; + tensor var_2805 = const()[name = tensor("op_2805"), val = tensor([1, 1, 16, 64])]; + tensor var_2806_cast_fp16 = reshape(shape = var_2805, x = out_43_cast_fp16)[name = tensor("op_2806_cast_fp16")]; + tensor var_2807_cast_fp16 = mul(x = var_2806_cast_fp16, y = onehot_43_cast_fp16)[name = tensor("op_2807_cast_fp16")]; + tensor new_k_cache_21_cast_fp16 = add(x = var_2800_cast_fp16, y = var_2807_cast_fp16)[name = tensor("new_k_cache_21_cast_fp16")]; + tensor new_k_cache_21_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_21_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache10_to_fp16_dtype_0 = const()[name = tensor("v_cache10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache10_to_fp16 = cast(dtype = v_cache10_to_fp16_dtype_0, x = v_cache10)[name = tensor("cast_493")]; + tensor var_2813_cast_fp16 = mul(x = v_cache10_to_fp16, y = var_2799_cast_fp16)[name = tensor("op_2813_cast_fp16")]; + tensor var_2818 = const()[name = tensor("op_2818"), val = tensor([1, 1, 16, 64])]; + tensor var_2819_cast_fp16 = reshape(shape = var_2818, x = v_21_cast_fp16)[name = tensor("op_2819_cast_fp16")]; + tensor var_2820_cast_fp16 = mul(x = var_2819_cast_fp16, y = onehot_43_cast_fp16)[name = tensor("op_2820_cast_fp16")]; + tensor new_v_cache_21_cast_fp16 = add(x = var_2813_cast_fp16, y = var_2820_cast_fp16)[name = tensor("new_v_cache_21_cast_fp16")]; + tensor new_v_cache_21_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_21_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_10_perm_0 = const()[name = tensor("transpose_10_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_2833 = const()[name = tensor("op_2833"), val = tensor([0, 2, 1, 3])]; + tensor var_2839 = const()[name = tensor("op_2839"), val = tensor([1, 16, 1, 64])]; + tensor q4_21_cast_fp16 = reshape(shape = var_2839, x = out_41_cast_fp16)[name = tensor("q4_21_cast_fp16")]; + tensor var_2844_transpose_x_0 = const()[name = tensor("op_2844_transpose_x_0"), val = tensor(false)]; + tensor var_2844_transpose_y_0 = const()[name = tensor("op_2844_transpose_y_0"), val = tensor(false)]; + tensor transpose_10_cast_fp16 = transpose(perm = transpose_10_perm_0, x = new_k_cache_21_cast_fp16)[name = tensor("transpose_65")]; + tensor var_2844_cast_fp16 = matmul(transpose_x = var_2844_transpose_x_0, transpose_y = var_2844_transpose_y_0, x = q4_21_cast_fp16, y = transpose_10_cast_fp16)[name = tensor("op_2844_cast_fp16")]; + tensor var_2845_to_fp16 = const()[name = tensor("op_2845_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_41_cast_fp16 = mul(x = var_2844_cast_fp16, y = var_2845_to_fp16)[name = tensor("scores_41_cast_fp16")]; + tensor var_2849 = const()[name = tensor("op_2849"), val = tensor([1, 1])]; + tensor var_2850_cast_fp16 = reshape(shape = var_2849, x = position10_to_fp16)[name = tensor("op_2850_cast_fp16")]; + tensor var_2851_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_2850_cast_fp16)[name = tensor("op_2851_cast_fp16")]; + tensor mask_21_to_fp16_dtype_0 = const()[name = tensor("mask_21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_2858_to_fp16 = const()[name = tensor("op_2858_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2851_cast_fp16_to_fp16 = cast(dtype = mask_21_to_fp16_dtype_0, x = var_2851_cast_fp16)[name = tensor("cast_491")]; + tensor var_2859_cast_fp16 = sub(x = var_2851_cast_fp16_to_fp16, y = var_2858_to_fp16)[name = tensor("op_2859_cast_fp16")]; + tensor var_2864 = const()[name = tensor("op_2864"), val = tensor([1, 1, 1, 512])]; + tensor var_2865_cast_fp16 = reshape(shape = var_2864, x = var_2859_cast_fp16)[name = tensor("op_2865_cast_fp16")]; + tensor var_2866_to_fp16 = const()[name = tensor("op_2866_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_2867_cast_fp16 = mul(x = var_2865_cast_fp16, y = var_2866_to_fp16)[name = tensor("op_2867_cast_fp16")]; + tensor scores_43_cast_fp16 = add(x = scores_41_cast_fp16, y = var_2867_cast_fp16)[name = tensor("scores_43_cast_fp16")]; + tensor var_2870 = const()[name = tensor("op_2870"), val = tensor(-1)]; + tensor weights_21_cast_fp16 = softmax(axis = var_2870, x = scores_43_cast_fp16)[name = tensor("weights_21_cast_fp16")]; + tensor attn_21_transpose_x_0 = const()[name = tensor("attn_21_transpose_x_0"), val = tensor(false)]; + tensor attn_21_transpose_y_0 = const()[name = tensor("attn_21_transpose_y_0"), val = tensor(false)]; + tensor values_21_cast_fp16 = transpose(perm = var_2833, x = new_v_cache_21_cast_fp16)[name = tensor("transpose_64")]; + tensor attn_21_cast_fp16 = matmul(transpose_x = attn_21_transpose_x_0, transpose_y = attn_21_transpose_y_0, x = weights_21_cast_fp16, y = values_21_cast_fp16)[name = tensor("attn_21_cast_fp16")]; + tensor var_2878 = const()[name = tensor("op_2878"), val = tensor([0, 2, 1, 3])]; + tensor var_2883 = const()[name = tensor("op_2883"), val = tensor([1, 1, 1024])]; + tensor var_2879_cast_fp16 = transpose(perm = var_2878, x = attn_21_cast_fp16)[name = tensor("transpose_63")]; + tensor input_127_cast_fp16 = reshape(shape = var_2883, x = var_2879_cast_fp16)[name = tensor("input_127_cast_fp16")]; + tensor attn10_out_proj_weight_to_fp16 = const()[name = tensor("attn10_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(258124480)))]; + tensor linear_42_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn10_out_proj_weight_to_fp16, x = input_127_cast_fp16)[name = tensor("linear_42_cast_fp16")]; + tensor var_2889_to_fp16 = const()[name = tensor("op_2889_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2890_cast_fp16 = add(x = position10_to_fp16, y = var_2889_to_fp16)[name = tensor("op_2890_cast_fp16")]; + tensor var_2890_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2890_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_129_cast_fp16 = add(x = input_123_cast_fp16, y = linear_42_cast_fp16)[name = tensor("input_129_cast_fp16")]; + tensor input_131_axes_0 = const()[name = tensor("input_131_axes_0"), val = tensor([-1])]; + tensor norm10_2_weight_to_fp16 = const()[name = tensor("norm10_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260221696)))]; + tensor norm10_2_bias_to_fp16 = const()[name = tensor("norm10_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260223808)))]; + tensor var_2894_to_fp16 = const()[name = tensor("op_2894_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_131_cast_fp16 = layer_norm(axes = input_131_axes_0, beta = norm10_2_bias_to_fp16, epsilon = var_2894_to_fp16, gamma = norm10_2_weight_to_fp16, x = input_129_cast_fp16)[name = tensor("input_131_cast_fp16")]; + tensor linear10_1_weight_to_fp16 = const()[name = tensor("linear10_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260225920)))]; + tensor linear_43_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear10_1_weight_to_fp16, x = input_131_cast_fp16)[name = tensor("linear_43_cast_fp16")]; + tensor input_133_mode_0 = const()[name = tensor("input_133_mode_0"), val = tensor("EXACT")]; + tensor input_133_cast_fp16 = gelu(mode = input_133_mode_0, x = linear_43_cast_fp16)[name = tensor("input_133_cast_fp16")]; + tensor linear10_2_weight_to_fp16 = const()[name = tensor("linear10_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(268614592)))]; + tensor linear_44_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear10_2_weight_to_fp16, x = input_133_cast_fp16)[name = tensor("linear_44_cast_fp16")]; + tensor input_135_cast_fp16 = add(x = input_129_cast_fp16, y = linear_44_cast_fp16)[name = tensor("input_135_cast_fp16")]; + tensor input_137_axes_0 = const()[name = tensor("input_137_axes_0"), val = tensor([-1])]; + tensor norm11_1_weight_to_fp16 = const()[name = tensor("norm11_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277003264)))]; + tensor norm11_1_bias_to_fp16 = const()[name = tensor("norm11_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277005376)))]; + tensor var_2911_to_fp16 = const()[name = tensor("op_2911_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_137_cast_fp16 = layer_norm(axes = input_137_axes_0, beta = norm11_1_bias_to_fp16, epsilon = var_2911_to_fp16, gamma = norm11_1_weight_to_fp16, x = input_135_cast_fp16)[name = tensor("input_137_cast_fp16")]; + tensor attn11_in_proj_weight_to_fp16 = const()[name = tensor("attn11_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277007488)))]; + tensor linear_45_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn11_in_proj_weight_to_fp16, x = input_137_cast_fp16)[name = tensor("linear_45_cast_fp16")]; + tensor var_2924 = const()[name = tensor("op_2924"), val = tensor([1, 3, 16, 64])]; + tensor qkv_23_cast_fp16 = reshape(shape = var_2924, x = linear_45_cast_fp16)[name = tensor("qkv_23_cast_fp16")]; + tensor x_45_begin_0 = const()[name = tensor("x_45_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_45_end_0 = const()[name = tensor("x_45_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_45_end_mask_0 = const()[name = tensor("x_45_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_45_squeeze_mask_0 = const()[name = tensor("x_45_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_45_cast_fp16 = slice_by_index(begin = x_45_begin_0, end = x_45_end_0, end_mask = x_45_end_mask_0, squeeze_mask = x_45_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("x_45_cast_fp16")]; + tensor x_47_begin_0 = const()[name = tensor("x_47_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_47_end_0 = const()[name = tensor("x_47_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_47_end_mask_0 = const()[name = tensor("x_47_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_47_squeeze_mask_0 = const()[name = tensor("x_47_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_47_cast_fp16 = slice_by_index(begin = x_47_begin_0, end = x_47_end_0, end_mask = x_47_end_mask_0, squeeze_mask = x_47_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("x_47_cast_fp16")]; + tensor v_23_begin_0 = const()[name = tensor("v_23_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_23_end_0 = const()[name = tensor("v_23_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_23_end_mask_0 = const()[name = tensor("v_23_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_23_squeeze_mask_0 = const()[name = tensor("v_23_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_23_cast_fp16 = slice_by_index(begin = v_23_begin_0, end = v_23_end_0, end_mask = v_23_end_mask_0, squeeze_mask = v_23_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("v_23_cast_fp16")]; + tensor var_2953 = const()[name = tensor("op_2953"), val = tensor([1, 1, 1])]; + tensor position11_to_fp16_dtype_0 = const()[name = tensor("position11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position11_to_fp16 = cast(dtype = position11_to_fp16_dtype_0, x = position11)[name = tensor("cast_489")]; + tensor var_2954_cast_fp16 = reshape(shape = var_2953, x = position11_to_fp16)[name = tensor("op_2954_cast_fp16")]; + tensor angles_45_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_2954_cast_fp16)[name = tensor("angles_45_cast_fp16")]; + tensor rotr_45_cast_fp16 = cos(x = angles_45_cast_fp16)[name = tensor("rotr_45_cast_fp16")]; + tensor roti_45_cast_fp16 = sin(x = angles_45_cast_fp16)[name = tensor("roti_45_cast_fp16")]; + tensor var_2962 = const()[name = tensor("op_2962"), val = tensor([1, 16, 32, 2])]; + tensor pairs_45_cast_fp16 = reshape(shape = var_2962, x = x_45_cast_fp16)[name = tensor("pairs_45_cast_fp16")]; + tensor xr_45_begin_0 = const()[name = tensor("xr_45_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_45_end_0 = const()[name = tensor("xr_45_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_45_end_mask_0 = const()[name = tensor("xr_45_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_45_squeeze_mask_0 = const()[name = tensor("xr_45_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_45_cast_fp16 = slice_by_index(begin = xr_45_begin_0, end = xr_45_end_0, end_mask = xr_45_end_mask_0, squeeze_mask = xr_45_squeeze_mask_0, x = pairs_45_cast_fp16)[name = tensor("xr_45_cast_fp16")]; + tensor xi_45_begin_0 = const()[name = tensor("xi_45_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_45_end_0 = const()[name = tensor("xi_45_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_45_end_mask_0 = const()[name = tensor("xi_45_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_45_squeeze_mask_0 = const()[name = tensor("xi_45_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_45_cast_fp16 = slice_by_index(begin = xi_45_begin_0, end = xi_45_end_0, end_mask = xi_45_end_mask_0, squeeze_mask = xi_45_squeeze_mask_0, x = pairs_45_cast_fp16)[name = tensor("xi_45_cast_fp16")]; + tensor var_2970_cast_fp16 = mul(x = xr_45_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_2970_cast_fp16")]; + tensor var_2971_cast_fp16 = mul(x = xi_45_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_2971_cast_fp16")]; + tensor xor__45_cast_fp16 = sub(x = var_2970_cast_fp16, y = var_2971_cast_fp16)[name = tensor("xor__45_cast_fp16")]; + tensor var_2974_cast_fp16 = mul(x = xr_45_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_2974_cast_fp16")]; + tensor var_2975_cast_fp16 = mul(x = xi_45_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_2975_cast_fp16")]; + tensor xoi_45_cast_fp16 = add(x = var_2974_cast_fp16, y = var_2975_cast_fp16)[name = tensor("xoi_45_cast_fp16")]; + tensor out_45_axis_0 = const()[name = tensor("out_45_axis_0"), val = tensor(-1)]; + tensor out_45_cast_fp16 = stack(axis = out_45_axis_0, values = (xor__45_cast_fp16, xoi_45_cast_fp16))[name = tensor("out_45_cast_fp16")]; + tensor var_2998 = const()[name = tensor("op_2998"), val = tensor([1, 16, 32, 2])]; + tensor pairs_47_cast_fp16 = reshape(shape = var_2998, x = x_47_cast_fp16)[name = tensor("pairs_47_cast_fp16")]; + tensor xr_47_begin_0 = const()[name = tensor("xr_47_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_47_end_0 = const()[name = tensor("xr_47_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_47_end_mask_0 = const()[name = tensor("xr_47_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_47_squeeze_mask_0 = const()[name = tensor("xr_47_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_47_cast_fp16 = slice_by_index(begin = xr_47_begin_0, end = xr_47_end_0, end_mask = xr_47_end_mask_0, squeeze_mask = xr_47_squeeze_mask_0, x = pairs_47_cast_fp16)[name = tensor("xr_47_cast_fp16")]; + tensor xi_47_begin_0 = const()[name = tensor("xi_47_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_47_end_0 = const()[name = tensor("xi_47_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_47_end_mask_0 = const()[name = tensor("xi_47_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_47_squeeze_mask_0 = const()[name = tensor("xi_47_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_47_cast_fp16 = slice_by_index(begin = xi_47_begin_0, end = xi_47_end_0, end_mask = xi_47_end_mask_0, squeeze_mask = xi_47_squeeze_mask_0, x = pairs_47_cast_fp16)[name = tensor("xi_47_cast_fp16")]; + tensor var_3006_cast_fp16 = mul(x = xr_47_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_3006_cast_fp16")]; + tensor var_3007_cast_fp16 = mul(x = xi_47_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_3007_cast_fp16")]; + tensor xor__47_cast_fp16 = sub(x = var_3006_cast_fp16, y = var_3007_cast_fp16)[name = tensor("xor__47_cast_fp16")]; + tensor var_3010_cast_fp16 = mul(x = xr_47_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_3010_cast_fp16")]; + tensor var_3011_cast_fp16 = mul(x = xi_47_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_3011_cast_fp16")]; + tensor xoi_47_cast_fp16 = add(x = var_3010_cast_fp16, y = var_3011_cast_fp16)[name = tensor("xoi_47_cast_fp16")]; + tensor out_47_axis_0 = const()[name = tensor("out_47_axis_0"), val = tensor(-1)]; + tensor out_47_cast_fp16 = stack(axis = out_47_axis_0, values = (xor__47_cast_fp16, xoi_47_cast_fp16))[name = tensor("out_47_cast_fp16")]; + tensor _inversed_3023_y_0_to_fp16 = const()[name = tensor("_inversed_3023_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3023_cast_fp16 = mul(x = position11_to_fp16, y = _inversed_3023_y_0_to_fp16)[name = tensor("_inversed_3023_cast_fp16")]; + tensor var_3024_cast_fp16 = floor(x = _inversed_3023_cast_fp16)[name = tensor("op_3024_cast_fp16")]; + tensor var_3025_to_fp16 = const()[name = tensor("op_3025_to_fp16"), val = tensor(0x1p+9)]; + tensor var_3026_cast_fp16 = mul(x = var_3024_cast_fp16, y = var_3025_to_fp16)[name = tensor("op_3026_cast_fp16")]; + tensor write_idx_23_cast_fp16 = sub(x = position11_to_fp16, y = var_3026_cast_fp16)[name = tensor("write_idx_23_cast_fp16")]; + tensor var_3029_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_23_cast_fp16)[name = tensor("op_3029_cast_fp16")]; + tensor var_3039 = const()[name = tensor("op_3039"), val = tensor([1, 512, 1, 1])]; + tensor onehot_45_to_fp16_dtype_0 = const()[name = tensor("onehot_45_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_3029_cast_fp16_to_fp16 = cast(dtype = onehot_45_to_fp16_dtype_0, x = var_3029_cast_fp16)[name = tensor("cast_488")]; + tensor onehot_47_cast_fp16 = reshape(shape = var_3039, x = var_3029_cast_fp16_to_fp16)[name = tensor("onehot_47_cast_fp16")]; + tensor var_3041_to_fp16 = const()[name = tensor("op_3041_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3043_cast_fp16 = sub(x = var_3041_to_fp16, y = onehot_47_cast_fp16)[name = tensor("op_3043_cast_fp16")]; + tensor k_cache11_to_fp16_dtype_0 = const()[name = tensor("k_cache11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache11_to_fp16 = cast(dtype = k_cache11_to_fp16_dtype_0, x = k_cache11)[name = tensor("cast_487")]; + tensor var_3044_cast_fp16 = mul(x = k_cache11_to_fp16, y = var_3043_cast_fp16)[name = tensor("op_3044_cast_fp16")]; + tensor var_3049 = const()[name = tensor("op_3049"), val = tensor([1, 1, 16, 64])]; + tensor var_3050_cast_fp16 = reshape(shape = var_3049, x = out_47_cast_fp16)[name = tensor("op_3050_cast_fp16")]; + tensor var_3051_cast_fp16 = mul(x = var_3050_cast_fp16, y = onehot_47_cast_fp16)[name = tensor("op_3051_cast_fp16")]; + tensor new_k_cache_23_cast_fp16 = add(x = var_3044_cast_fp16, y = var_3051_cast_fp16)[name = tensor("new_k_cache_23_cast_fp16")]; + tensor new_k_cache_23_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_23_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache11_to_fp16_dtype_0 = const()[name = tensor("v_cache11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache11_to_fp16 = cast(dtype = v_cache11_to_fp16_dtype_0, x = v_cache11)[name = tensor("cast_485")]; + tensor var_3057_cast_fp16 = mul(x = v_cache11_to_fp16, y = var_3043_cast_fp16)[name = tensor("op_3057_cast_fp16")]; + tensor var_3062 = const()[name = tensor("op_3062"), val = tensor([1, 1, 16, 64])]; + tensor var_3063_cast_fp16 = reshape(shape = var_3062, x = v_23_cast_fp16)[name = tensor("op_3063_cast_fp16")]; + tensor var_3064_cast_fp16 = mul(x = var_3063_cast_fp16, y = onehot_47_cast_fp16)[name = tensor("op_3064_cast_fp16")]; + tensor new_v_cache_23_cast_fp16 = add(x = var_3057_cast_fp16, y = var_3064_cast_fp16)[name = tensor("new_v_cache_23_cast_fp16")]; + tensor new_v_cache_23_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_23_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_11_perm_0 = const()[name = tensor("transpose_11_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_3077 = const()[name = tensor("op_3077"), val = tensor([0, 2, 1, 3])]; + tensor var_3083 = const()[name = tensor("op_3083"), val = tensor([1, 16, 1, 64])]; + tensor q4_23_cast_fp16 = reshape(shape = var_3083, x = out_45_cast_fp16)[name = tensor("q4_23_cast_fp16")]; + tensor var_3088_transpose_x_0 = const()[name = tensor("op_3088_transpose_x_0"), val = tensor(false)]; + tensor var_3088_transpose_y_0 = const()[name = tensor("op_3088_transpose_y_0"), val = tensor(false)]; + tensor transpose_11_cast_fp16 = transpose(perm = transpose_11_perm_0, x = new_k_cache_23_cast_fp16)[name = tensor("transpose_62")]; + tensor var_3088_cast_fp16 = matmul(transpose_x = var_3088_transpose_x_0, transpose_y = var_3088_transpose_y_0, x = q4_23_cast_fp16, y = transpose_11_cast_fp16)[name = tensor("op_3088_cast_fp16")]; + tensor var_3089_to_fp16 = const()[name = tensor("op_3089_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_45_cast_fp16 = mul(x = var_3088_cast_fp16, y = var_3089_to_fp16)[name = tensor("scores_45_cast_fp16")]; + tensor var_3093 = const()[name = tensor("op_3093"), val = tensor([1, 1])]; + tensor var_3094_cast_fp16 = reshape(shape = var_3093, x = position11_to_fp16)[name = tensor("op_3094_cast_fp16")]; + tensor var_3095_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_3094_cast_fp16)[name = tensor("op_3095_cast_fp16")]; + tensor mask_23_to_fp16_dtype_0 = const()[name = tensor("mask_23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_3102_to_fp16 = const()[name = tensor("op_3102_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3095_cast_fp16_to_fp16 = cast(dtype = mask_23_to_fp16_dtype_0, x = var_3095_cast_fp16)[name = tensor("cast_483")]; + tensor var_3103_cast_fp16 = sub(x = var_3095_cast_fp16_to_fp16, y = var_3102_to_fp16)[name = tensor("op_3103_cast_fp16")]; + tensor var_3108 = const()[name = tensor("op_3108"), val = tensor([1, 1, 1, 512])]; + tensor var_3109_cast_fp16 = reshape(shape = var_3108, x = var_3103_cast_fp16)[name = tensor("op_3109_cast_fp16")]; + tensor var_3110_to_fp16 = const()[name = tensor("op_3110_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_3111_cast_fp16 = mul(x = var_3109_cast_fp16, y = var_3110_to_fp16)[name = tensor("op_3111_cast_fp16")]; + tensor scores_47_cast_fp16 = add(x = scores_45_cast_fp16, y = var_3111_cast_fp16)[name = tensor("scores_47_cast_fp16")]; + tensor var_3114 = const()[name = tensor("op_3114"), val = tensor(-1)]; + tensor weights_23_cast_fp16 = softmax(axis = var_3114, x = scores_47_cast_fp16)[name = tensor("weights_23_cast_fp16")]; + tensor attn_23_transpose_x_0 = const()[name = tensor("attn_23_transpose_x_0"), val = tensor(false)]; + tensor attn_23_transpose_y_0 = const()[name = tensor("attn_23_transpose_y_0"), val = tensor(false)]; + tensor values_23_cast_fp16 = transpose(perm = var_3077, x = new_v_cache_23_cast_fp16)[name = tensor("transpose_61")]; + tensor attn_23_cast_fp16 = matmul(transpose_x = attn_23_transpose_x_0, transpose_y = attn_23_transpose_y_0, x = weights_23_cast_fp16, y = values_23_cast_fp16)[name = tensor("attn_23_cast_fp16")]; + tensor var_3122 = const()[name = tensor("op_3122"), val = tensor([0, 2, 1, 3])]; + tensor var_3127 = const()[name = tensor("op_3127"), val = tensor([1, 1, 1024])]; + tensor var_3123_cast_fp16 = transpose(perm = var_3122, x = attn_23_cast_fp16)[name = tensor("transpose_60")]; + tensor input_139_cast_fp16 = reshape(shape = var_3127, x = var_3123_cast_fp16)[name = tensor("input_139_cast_fp16")]; + tensor attn11_out_proj_weight_to_fp16 = const()[name = tensor("attn11_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(283299008)))]; + tensor linear_46_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn11_out_proj_weight_to_fp16, x = input_139_cast_fp16)[name = tensor("linear_46_cast_fp16")]; + tensor var_3133_to_fp16 = const()[name = tensor("op_3133_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3134_cast_fp16 = add(x = position11_to_fp16, y = var_3133_to_fp16)[name = tensor("op_3134_cast_fp16")]; + tensor var_3134_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3134_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_141_cast_fp16 = add(x = input_135_cast_fp16, y = linear_46_cast_fp16)[name = tensor("input_141_cast_fp16")]; + tensor input_143_axes_0 = const()[name = tensor("input_143_axes_0"), val = tensor([-1])]; + tensor norm11_2_weight_to_fp16 = const()[name = tensor("norm11_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285396224)))]; + tensor norm11_2_bias_to_fp16 = const()[name = tensor("norm11_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285398336)))]; + tensor var_3138_to_fp16 = const()[name = tensor("op_3138_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_143_cast_fp16 = layer_norm(axes = input_143_axes_0, beta = norm11_2_bias_to_fp16, epsilon = var_3138_to_fp16, gamma = norm11_2_weight_to_fp16, x = input_141_cast_fp16)[name = tensor("input_143_cast_fp16")]; + tensor linear11_1_weight_to_fp16 = const()[name = tensor("linear11_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285400448)))]; + tensor linear_47_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear11_1_weight_to_fp16, x = input_143_cast_fp16)[name = tensor("linear_47_cast_fp16")]; + tensor input_145_mode_0 = const()[name = tensor("input_145_mode_0"), val = tensor("EXACT")]; + tensor input_145_cast_fp16 = gelu(mode = input_145_mode_0, x = linear_47_cast_fp16)[name = tensor("input_145_cast_fp16")]; + tensor linear11_2_weight_to_fp16 = const()[name = tensor("linear11_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293789120)))]; + tensor linear_48_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear11_2_weight_to_fp16, x = input_145_cast_fp16)[name = tensor("linear_48_cast_fp16")]; + tensor input_147_cast_fp16 = add(x = input_141_cast_fp16, y = linear_48_cast_fp16)[name = tensor("input_147_cast_fp16")]; + tensor input_149_axes_0 = const()[name = tensor("input_149_axes_0"), val = tensor([-1])]; + tensor norm12_1_weight_to_fp16 = const()[name = tensor("norm12_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302177792)))]; + tensor norm12_1_bias_to_fp16 = const()[name = tensor("norm12_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302179904)))]; + tensor var_3155_to_fp16 = const()[name = tensor("op_3155_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_149_cast_fp16 = layer_norm(axes = input_149_axes_0, beta = norm12_1_bias_to_fp16, epsilon = var_3155_to_fp16, gamma = norm12_1_weight_to_fp16, x = input_147_cast_fp16)[name = tensor("input_149_cast_fp16")]; + tensor attn12_in_proj_weight_to_fp16 = const()[name = tensor("attn12_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302182016)))]; + tensor linear_49_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn12_in_proj_weight_to_fp16, x = input_149_cast_fp16)[name = tensor("linear_49_cast_fp16")]; + tensor var_3168 = const()[name = tensor("op_3168"), val = tensor([1, 3, 16, 64])]; + tensor qkv_25_cast_fp16 = reshape(shape = var_3168, x = linear_49_cast_fp16)[name = tensor("qkv_25_cast_fp16")]; + tensor x_49_begin_0 = const()[name = tensor("x_49_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_49_end_0 = const()[name = tensor("x_49_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_49_end_mask_0 = const()[name = tensor("x_49_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_49_squeeze_mask_0 = const()[name = tensor("x_49_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_49_cast_fp16 = slice_by_index(begin = x_49_begin_0, end = x_49_end_0, end_mask = x_49_end_mask_0, squeeze_mask = x_49_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("x_49_cast_fp16")]; + tensor x_51_begin_0 = const()[name = tensor("x_51_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_51_end_0 = const()[name = tensor("x_51_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_51_end_mask_0 = const()[name = tensor("x_51_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_51_squeeze_mask_0 = const()[name = tensor("x_51_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_51_cast_fp16 = slice_by_index(begin = x_51_begin_0, end = x_51_end_0, end_mask = x_51_end_mask_0, squeeze_mask = x_51_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("x_51_cast_fp16")]; + tensor v_25_begin_0 = const()[name = tensor("v_25_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_25_end_0 = const()[name = tensor("v_25_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_25_end_mask_0 = const()[name = tensor("v_25_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_25_squeeze_mask_0 = const()[name = tensor("v_25_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_25_cast_fp16 = slice_by_index(begin = v_25_begin_0, end = v_25_end_0, end_mask = v_25_end_mask_0, squeeze_mask = v_25_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("v_25_cast_fp16")]; + tensor var_3197 = const()[name = tensor("op_3197"), val = tensor([1, 1, 1])]; + tensor position12_to_fp16_dtype_0 = const()[name = tensor("position12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position12_to_fp16 = cast(dtype = position12_to_fp16_dtype_0, x = position12)[name = tensor("cast_481")]; + tensor var_3198_cast_fp16 = reshape(shape = var_3197, x = position12_to_fp16)[name = tensor("op_3198_cast_fp16")]; + tensor angles_49_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_3198_cast_fp16)[name = tensor("angles_49_cast_fp16")]; + tensor rotr_49_cast_fp16 = cos(x = angles_49_cast_fp16)[name = tensor("rotr_49_cast_fp16")]; + tensor roti_49_cast_fp16 = sin(x = angles_49_cast_fp16)[name = tensor("roti_49_cast_fp16")]; + tensor var_3206 = const()[name = tensor("op_3206"), val = tensor([1, 16, 32, 2])]; + tensor pairs_49_cast_fp16 = reshape(shape = var_3206, x = x_49_cast_fp16)[name = tensor("pairs_49_cast_fp16")]; + tensor xr_49_begin_0 = const()[name = tensor("xr_49_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_49_end_0 = const()[name = tensor("xr_49_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_49_end_mask_0 = const()[name = tensor("xr_49_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_49_squeeze_mask_0 = const()[name = tensor("xr_49_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_49_cast_fp16 = slice_by_index(begin = xr_49_begin_0, end = xr_49_end_0, end_mask = xr_49_end_mask_0, squeeze_mask = xr_49_squeeze_mask_0, x = pairs_49_cast_fp16)[name = tensor("xr_49_cast_fp16")]; + tensor xi_49_begin_0 = const()[name = tensor("xi_49_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_49_end_0 = const()[name = tensor("xi_49_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_49_end_mask_0 = const()[name = tensor("xi_49_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_49_squeeze_mask_0 = const()[name = tensor("xi_49_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_49_cast_fp16 = slice_by_index(begin = xi_49_begin_0, end = xi_49_end_0, end_mask = xi_49_end_mask_0, squeeze_mask = xi_49_squeeze_mask_0, x = pairs_49_cast_fp16)[name = tensor("xi_49_cast_fp16")]; + tensor var_3214_cast_fp16 = mul(x = xr_49_cast_fp16, y = rotr_49_cast_fp16)[name = tensor("op_3214_cast_fp16")]; + tensor var_3215_cast_fp16 = mul(x = xi_49_cast_fp16, y = roti_49_cast_fp16)[name = tensor("op_3215_cast_fp16")]; + tensor xor__49_cast_fp16 = sub(x = var_3214_cast_fp16, y = var_3215_cast_fp16)[name = tensor("xor__49_cast_fp16")]; + tensor var_3218_cast_fp16 = mul(x = xr_49_cast_fp16, y = roti_49_cast_fp16)[name = tensor("op_3218_cast_fp16")]; + tensor var_3219_cast_fp16 = mul(x = xi_49_cast_fp16, y = rotr_49_cast_fp16)[name = tensor("op_3219_cast_fp16")]; + tensor xoi_49_cast_fp16 = add(x = var_3218_cast_fp16, y = var_3219_cast_fp16)[name = tensor("xoi_49_cast_fp16")]; + tensor out_49_axis_0 = const()[name = tensor("out_49_axis_0"), val = tensor(-1)]; + tensor out_49_cast_fp16 = stack(axis = out_49_axis_0, values = (xor__49_cast_fp16, xoi_49_cast_fp16))[name = tensor("out_49_cast_fp16")]; + tensor var_3242 = const()[name = tensor("op_3242"), val = tensor([1, 16, 32, 2])]; + tensor pairs_51_cast_fp16 = reshape(shape = var_3242, x = x_51_cast_fp16)[name = tensor("pairs_51_cast_fp16")]; + tensor xr_51_begin_0 = const()[name = tensor("xr_51_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_51_end_0 = const()[name = tensor("xr_51_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_51_end_mask_0 = const()[name = tensor("xr_51_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_51_squeeze_mask_0 = const()[name = tensor("xr_51_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_51_cast_fp16 = slice_by_index(begin = xr_51_begin_0, end = xr_51_end_0, end_mask = xr_51_end_mask_0, squeeze_mask = xr_51_squeeze_mask_0, x = pairs_51_cast_fp16)[name = tensor("xr_51_cast_fp16")]; + tensor xi_51_begin_0 = const()[name = tensor("xi_51_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_51_end_0 = const()[name = tensor("xi_51_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_51_end_mask_0 = const()[name = tensor("xi_51_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_51_squeeze_mask_0 = const()[name = tensor("xi_51_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_51_cast_fp16 = slice_by_index(begin = xi_51_begin_0, end = xi_51_end_0, end_mask = xi_51_end_mask_0, squeeze_mask = xi_51_squeeze_mask_0, x = pairs_51_cast_fp16)[name = tensor("xi_51_cast_fp16")]; + tensor var_3250_cast_fp16 = mul(x = xr_51_cast_fp16, y = rotr_49_cast_fp16)[name = tensor("op_3250_cast_fp16")]; + tensor var_3251_cast_fp16 = mul(x = xi_51_cast_fp16, y = roti_49_cast_fp16)[name = tensor("op_3251_cast_fp16")]; + tensor xor__51_cast_fp16 = sub(x = var_3250_cast_fp16, y = var_3251_cast_fp16)[name = tensor("xor__51_cast_fp16")]; + tensor var_3254_cast_fp16 = mul(x = xr_51_cast_fp16, y = roti_49_cast_fp16)[name = tensor("op_3254_cast_fp16")]; + tensor var_3255_cast_fp16 = mul(x = xi_51_cast_fp16, y = rotr_49_cast_fp16)[name = tensor("op_3255_cast_fp16")]; + tensor xoi_51_cast_fp16 = add(x = var_3254_cast_fp16, y = var_3255_cast_fp16)[name = tensor("xoi_51_cast_fp16")]; + tensor out_51_axis_0 = const()[name = tensor("out_51_axis_0"), val = tensor(-1)]; + tensor out_51_cast_fp16 = stack(axis = out_51_axis_0, values = (xor__51_cast_fp16, xoi_51_cast_fp16))[name = tensor("out_51_cast_fp16")]; + tensor _inversed_3267_y_0_to_fp16 = const()[name = tensor("_inversed_3267_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3267_cast_fp16 = mul(x = position12_to_fp16, y = _inversed_3267_y_0_to_fp16)[name = tensor("_inversed_3267_cast_fp16")]; + tensor var_3268_cast_fp16 = floor(x = _inversed_3267_cast_fp16)[name = tensor("op_3268_cast_fp16")]; + tensor var_3269_to_fp16 = const()[name = tensor("op_3269_to_fp16"), val = tensor(0x1p+9)]; + tensor var_3270_cast_fp16 = mul(x = var_3268_cast_fp16, y = var_3269_to_fp16)[name = tensor("op_3270_cast_fp16")]; + tensor write_idx_25_cast_fp16 = sub(x = position12_to_fp16, y = var_3270_cast_fp16)[name = tensor("write_idx_25_cast_fp16")]; + tensor var_3273_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_25_cast_fp16)[name = tensor("op_3273_cast_fp16")]; + tensor var_3283 = const()[name = tensor("op_3283"), val = tensor([1, 512, 1, 1])]; + tensor onehot_49_to_fp16_dtype_0 = const()[name = tensor("onehot_49_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_3273_cast_fp16_to_fp16 = cast(dtype = onehot_49_to_fp16_dtype_0, x = var_3273_cast_fp16)[name = tensor("cast_480")]; + tensor onehot_51_cast_fp16 = reshape(shape = var_3283, x = var_3273_cast_fp16_to_fp16)[name = tensor("onehot_51_cast_fp16")]; + tensor var_3285_to_fp16 = const()[name = tensor("op_3285_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3287_cast_fp16 = sub(x = var_3285_to_fp16, y = onehot_51_cast_fp16)[name = tensor("op_3287_cast_fp16")]; + tensor k_cache12_to_fp16_dtype_0 = const()[name = tensor("k_cache12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache12_to_fp16 = cast(dtype = k_cache12_to_fp16_dtype_0, x = k_cache12)[name = tensor("cast_479")]; + tensor var_3288_cast_fp16 = mul(x = k_cache12_to_fp16, y = var_3287_cast_fp16)[name = tensor("op_3288_cast_fp16")]; + tensor var_3293 = const()[name = tensor("op_3293"), val = tensor([1, 1, 16, 64])]; + tensor var_3294_cast_fp16 = reshape(shape = var_3293, x = out_51_cast_fp16)[name = tensor("op_3294_cast_fp16")]; + tensor var_3295_cast_fp16 = mul(x = var_3294_cast_fp16, y = onehot_51_cast_fp16)[name = tensor("op_3295_cast_fp16")]; + tensor new_k_cache_25_cast_fp16 = add(x = var_3288_cast_fp16, y = var_3295_cast_fp16)[name = tensor("new_k_cache_25_cast_fp16")]; + tensor new_k_cache_25_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_25_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache12_to_fp16_dtype_0 = const()[name = tensor("v_cache12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache12_to_fp16 = cast(dtype = v_cache12_to_fp16_dtype_0, x = v_cache12)[name = tensor("cast_477")]; + tensor var_3301_cast_fp16 = mul(x = v_cache12_to_fp16, y = var_3287_cast_fp16)[name = tensor("op_3301_cast_fp16")]; + tensor var_3306 = const()[name = tensor("op_3306"), val = tensor([1, 1, 16, 64])]; + tensor var_3307_cast_fp16 = reshape(shape = var_3306, x = v_25_cast_fp16)[name = tensor("op_3307_cast_fp16")]; + tensor var_3308_cast_fp16 = mul(x = var_3307_cast_fp16, y = onehot_51_cast_fp16)[name = tensor("op_3308_cast_fp16")]; + tensor new_v_cache_25_cast_fp16 = add(x = var_3301_cast_fp16, y = var_3308_cast_fp16)[name = tensor("new_v_cache_25_cast_fp16")]; + tensor new_v_cache_25_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_25_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_12_perm_0 = const()[name = tensor("transpose_12_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_3321 = const()[name = tensor("op_3321"), val = tensor([0, 2, 1, 3])]; + tensor var_3327 = const()[name = tensor("op_3327"), val = tensor([1, 16, 1, 64])]; + tensor q4_25_cast_fp16 = reshape(shape = var_3327, x = out_49_cast_fp16)[name = tensor("q4_25_cast_fp16")]; + tensor var_3332_transpose_x_0 = const()[name = tensor("op_3332_transpose_x_0"), val = tensor(false)]; + tensor var_3332_transpose_y_0 = const()[name = tensor("op_3332_transpose_y_0"), val = tensor(false)]; + tensor transpose_12_cast_fp16 = transpose(perm = transpose_12_perm_0, x = new_k_cache_25_cast_fp16)[name = tensor("transpose_59")]; + tensor var_3332_cast_fp16 = matmul(transpose_x = var_3332_transpose_x_0, transpose_y = var_3332_transpose_y_0, x = q4_25_cast_fp16, y = transpose_12_cast_fp16)[name = tensor("op_3332_cast_fp16")]; + tensor var_3333_to_fp16 = const()[name = tensor("op_3333_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_49_cast_fp16 = mul(x = var_3332_cast_fp16, y = var_3333_to_fp16)[name = tensor("scores_49_cast_fp16")]; + tensor var_3337 = const()[name = tensor("op_3337"), val = tensor([1, 1])]; + tensor var_3338_cast_fp16 = reshape(shape = var_3337, x = position12_to_fp16)[name = tensor("op_3338_cast_fp16")]; + tensor var_3339_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_3338_cast_fp16)[name = tensor("op_3339_cast_fp16")]; + tensor mask_25_to_fp16_dtype_0 = const()[name = tensor("mask_25_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_3346_to_fp16 = const()[name = tensor("op_3346_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3339_cast_fp16_to_fp16 = cast(dtype = mask_25_to_fp16_dtype_0, x = var_3339_cast_fp16)[name = tensor("cast_475")]; + tensor var_3347_cast_fp16 = sub(x = var_3339_cast_fp16_to_fp16, y = var_3346_to_fp16)[name = tensor("op_3347_cast_fp16")]; + tensor var_3352 = const()[name = tensor("op_3352"), val = tensor([1, 1, 1, 512])]; + tensor var_3353_cast_fp16 = reshape(shape = var_3352, x = var_3347_cast_fp16)[name = tensor("op_3353_cast_fp16")]; + tensor var_3354_to_fp16 = const()[name = tensor("op_3354_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_3355_cast_fp16 = mul(x = var_3353_cast_fp16, y = var_3354_to_fp16)[name = tensor("op_3355_cast_fp16")]; + tensor scores_51_cast_fp16 = add(x = scores_49_cast_fp16, y = var_3355_cast_fp16)[name = tensor("scores_51_cast_fp16")]; + tensor var_3358 = const()[name = tensor("op_3358"), val = tensor(-1)]; + tensor weights_25_cast_fp16 = softmax(axis = var_3358, x = scores_51_cast_fp16)[name = tensor("weights_25_cast_fp16")]; + tensor attn_25_transpose_x_0 = const()[name = tensor("attn_25_transpose_x_0"), val = tensor(false)]; + tensor attn_25_transpose_y_0 = const()[name = tensor("attn_25_transpose_y_0"), val = tensor(false)]; + tensor values_25_cast_fp16 = transpose(perm = var_3321, x = new_v_cache_25_cast_fp16)[name = tensor("transpose_58")]; + tensor attn_25_cast_fp16 = matmul(transpose_x = attn_25_transpose_x_0, transpose_y = attn_25_transpose_y_0, x = weights_25_cast_fp16, y = values_25_cast_fp16)[name = tensor("attn_25_cast_fp16")]; + tensor var_3366 = const()[name = tensor("op_3366"), val = tensor([0, 2, 1, 3])]; + tensor var_3371 = const()[name = tensor("op_3371"), val = tensor([1, 1, 1024])]; + tensor var_3367_cast_fp16 = transpose(perm = var_3366, x = attn_25_cast_fp16)[name = tensor("transpose_57")]; + tensor input_151_cast_fp16 = reshape(shape = var_3371, x = var_3367_cast_fp16)[name = tensor("input_151_cast_fp16")]; + tensor attn12_out_proj_weight_to_fp16 = const()[name = tensor("attn12_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(308473536)))]; + tensor linear_50_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn12_out_proj_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_50_cast_fp16")]; + tensor var_3377_to_fp16 = const()[name = tensor("op_3377_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3378_cast_fp16 = add(x = position12_to_fp16, y = var_3377_to_fp16)[name = tensor("op_3378_cast_fp16")]; + tensor var_3378_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3378_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_153_cast_fp16 = add(x = input_147_cast_fp16, y = linear_50_cast_fp16)[name = tensor("input_153_cast_fp16")]; + tensor input_155_axes_0 = const()[name = tensor("input_155_axes_0"), val = tensor([-1])]; + tensor norm12_2_weight_to_fp16 = const()[name = tensor("norm12_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(310570752)))]; + tensor norm12_2_bias_to_fp16 = const()[name = tensor("norm12_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(310572864)))]; + tensor var_3382_to_fp16 = const()[name = tensor("op_3382_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_155_cast_fp16 = layer_norm(axes = input_155_axes_0, beta = norm12_2_bias_to_fp16, epsilon = var_3382_to_fp16, gamma = norm12_2_weight_to_fp16, x = input_153_cast_fp16)[name = tensor("input_155_cast_fp16")]; + tensor linear12_1_weight_to_fp16 = const()[name = tensor("linear12_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(310574976)))]; + tensor linear_51_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear12_1_weight_to_fp16, x = input_155_cast_fp16)[name = tensor("linear_51_cast_fp16")]; + tensor input_157_mode_0 = const()[name = tensor("input_157_mode_0"), val = tensor("EXACT")]; + tensor input_157_cast_fp16 = gelu(mode = input_157_mode_0, x = linear_51_cast_fp16)[name = tensor("input_157_cast_fp16")]; + tensor linear12_2_weight_to_fp16 = const()[name = tensor("linear12_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(318963648)))]; + tensor linear_52_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear12_2_weight_to_fp16, x = input_157_cast_fp16)[name = tensor("linear_52_cast_fp16")]; + tensor input_159_cast_fp16 = add(x = input_153_cast_fp16, y = linear_52_cast_fp16)[name = tensor("input_159_cast_fp16")]; + tensor input_161_axes_0 = const()[name = tensor("input_161_axes_0"), val = tensor([-1])]; + tensor norm13_1_weight_to_fp16 = const()[name = tensor("norm13_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(327352320)))]; + tensor norm13_1_bias_to_fp16 = const()[name = tensor("norm13_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(327354432)))]; + tensor var_3399_to_fp16 = const()[name = tensor("op_3399_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_161_cast_fp16 = layer_norm(axes = input_161_axes_0, beta = norm13_1_bias_to_fp16, epsilon = var_3399_to_fp16, gamma = norm13_1_weight_to_fp16, x = input_159_cast_fp16)[name = tensor("input_161_cast_fp16")]; + tensor attn13_in_proj_weight_to_fp16 = const()[name = tensor("attn13_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(327356544)))]; + tensor linear_53_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn13_in_proj_weight_to_fp16, x = input_161_cast_fp16)[name = tensor("linear_53_cast_fp16")]; + tensor var_3412 = const()[name = tensor("op_3412"), val = tensor([1, 3, 16, 64])]; + tensor qkv_27_cast_fp16 = reshape(shape = var_3412, x = linear_53_cast_fp16)[name = tensor("qkv_27_cast_fp16")]; + tensor x_53_begin_0 = const()[name = tensor("x_53_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_53_end_0 = const()[name = tensor("x_53_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_53_end_mask_0 = const()[name = tensor("x_53_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_53_squeeze_mask_0 = const()[name = tensor("x_53_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_53_cast_fp16 = slice_by_index(begin = x_53_begin_0, end = x_53_end_0, end_mask = x_53_end_mask_0, squeeze_mask = x_53_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("x_53_cast_fp16")]; + tensor x_55_begin_0 = const()[name = tensor("x_55_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_55_end_0 = const()[name = tensor("x_55_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_55_end_mask_0 = const()[name = tensor("x_55_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_55_squeeze_mask_0 = const()[name = tensor("x_55_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_55_cast_fp16 = slice_by_index(begin = x_55_begin_0, end = x_55_end_0, end_mask = x_55_end_mask_0, squeeze_mask = x_55_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("x_55_cast_fp16")]; + tensor v_27_begin_0 = const()[name = tensor("v_27_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_27_end_0 = const()[name = tensor("v_27_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_27_end_mask_0 = const()[name = tensor("v_27_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_27_squeeze_mask_0 = const()[name = tensor("v_27_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_27_cast_fp16 = slice_by_index(begin = v_27_begin_0, end = v_27_end_0, end_mask = v_27_end_mask_0, squeeze_mask = v_27_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("v_27_cast_fp16")]; + tensor var_3441 = const()[name = tensor("op_3441"), val = tensor([1, 1, 1])]; + tensor position13_to_fp16_dtype_0 = const()[name = tensor("position13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position13_to_fp16 = cast(dtype = position13_to_fp16_dtype_0, x = position13)[name = tensor("cast_473")]; + tensor var_3442_cast_fp16 = reshape(shape = var_3441, x = position13_to_fp16)[name = tensor("op_3442_cast_fp16")]; + tensor angles_53_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_3442_cast_fp16)[name = tensor("angles_53_cast_fp16")]; + tensor rotr_53_cast_fp16 = cos(x = angles_53_cast_fp16)[name = tensor("rotr_53_cast_fp16")]; + tensor roti_53_cast_fp16 = sin(x = angles_53_cast_fp16)[name = tensor("roti_53_cast_fp16")]; + tensor var_3450 = const()[name = tensor("op_3450"), val = tensor([1, 16, 32, 2])]; + tensor pairs_53_cast_fp16 = reshape(shape = var_3450, x = x_53_cast_fp16)[name = tensor("pairs_53_cast_fp16")]; + tensor xr_53_begin_0 = const()[name = tensor("xr_53_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_53_end_0 = const()[name = tensor("xr_53_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_53_end_mask_0 = const()[name = tensor("xr_53_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_53_squeeze_mask_0 = const()[name = tensor("xr_53_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_53_cast_fp16 = slice_by_index(begin = xr_53_begin_0, end = xr_53_end_0, end_mask = xr_53_end_mask_0, squeeze_mask = xr_53_squeeze_mask_0, x = pairs_53_cast_fp16)[name = tensor("xr_53_cast_fp16")]; + tensor xi_53_begin_0 = const()[name = tensor("xi_53_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_53_end_0 = const()[name = tensor("xi_53_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_53_end_mask_0 = const()[name = tensor("xi_53_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_53_squeeze_mask_0 = const()[name = tensor("xi_53_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_53_cast_fp16 = slice_by_index(begin = xi_53_begin_0, end = xi_53_end_0, end_mask = xi_53_end_mask_0, squeeze_mask = xi_53_squeeze_mask_0, x = pairs_53_cast_fp16)[name = tensor("xi_53_cast_fp16")]; + tensor var_3458_cast_fp16 = mul(x = xr_53_cast_fp16, y = rotr_53_cast_fp16)[name = tensor("op_3458_cast_fp16")]; + tensor var_3459_cast_fp16 = mul(x = xi_53_cast_fp16, y = roti_53_cast_fp16)[name = tensor("op_3459_cast_fp16")]; + tensor xor__53_cast_fp16 = sub(x = var_3458_cast_fp16, y = var_3459_cast_fp16)[name = tensor("xor__53_cast_fp16")]; + tensor var_3462_cast_fp16 = mul(x = xr_53_cast_fp16, y = roti_53_cast_fp16)[name = tensor("op_3462_cast_fp16")]; + tensor var_3463_cast_fp16 = mul(x = xi_53_cast_fp16, y = rotr_53_cast_fp16)[name = tensor("op_3463_cast_fp16")]; + tensor xoi_53_cast_fp16 = add(x = var_3462_cast_fp16, y = var_3463_cast_fp16)[name = tensor("xoi_53_cast_fp16")]; + tensor out_53_axis_0 = const()[name = tensor("out_53_axis_0"), val = tensor(-1)]; + tensor out_53_cast_fp16 = stack(axis = out_53_axis_0, values = (xor__53_cast_fp16, xoi_53_cast_fp16))[name = tensor("out_53_cast_fp16")]; + tensor var_3486 = const()[name = tensor("op_3486"), val = tensor([1, 16, 32, 2])]; + tensor pairs_55_cast_fp16 = reshape(shape = var_3486, x = x_55_cast_fp16)[name = tensor("pairs_55_cast_fp16")]; + tensor xr_55_begin_0 = const()[name = tensor("xr_55_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_55_end_0 = const()[name = tensor("xr_55_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_55_end_mask_0 = const()[name = tensor("xr_55_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_55_squeeze_mask_0 = const()[name = tensor("xr_55_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_55_cast_fp16 = slice_by_index(begin = xr_55_begin_0, end = xr_55_end_0, end_mask = xr_55_end_mask_0, squeeze_mask = xr_55_squeeze_mask_0, x = pairs_55_cast_fp16)[name = tensor("xr_55_cast_fp16")]; + tensor xi_55_begin_0 = const()[name = tensor("xi_55_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_55_end_0 = const()[name = tensor("xi_55_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_55_end_mask_0 = const()[name = tensor("xi_55_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_55_squeeze_mask_0 = const()[name = tensor("xi_55_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_55_cast_fp16 = slice_by_index(begin = xi_55_begin_0, end = xi_55_end_0, end_mask = xi_55_end_mask_0, squeeze_mask = xi_55_squeeze_mask_0, x = pairs_55_cast_fp16)[name = tensor("xi_55_cast_fp16")]; + tensor var_3494_cast_fp16 = mul(x = xr_55_cast_fp16, y = rotr_53_cast_fp16)[name = tensor("op_3494_cast_fp16")]; + tensor var_3495_cast_fp16 = mul(x = xi_55_cast_fp16, y = roti_53_cast_fp16)[name = tensor("op_3495_cast_fp16")]; + tensor xor__55_cast_fp16 = sub(x = var_3494_cast_fp16, y = var_3495_cast_fp16)[name = tensor("xor__55_cast_fp16")]; + tensor var_3498_cast_fp16 = mul(x = xr_55_cast_fp16, y = roti_53_cast_fp16)[name = tensor("op_3498_cast_fp16")]; + tensor var_3499_cast_fp16 = mul(x = xi_55_cast_fp16, y = rotr_53_cast_fp16)[name = tensor("op_3499_cast_fp16")]; + tensor xoi_55_cast_fp16 = add(x = var_3498_cast_fp16, y = var_3499_cast_fp16)[name = tensor("xoi_55_cast_fp16")]; + tensor out_55_axis_0 = const()[name = tensor("out_55_axis_0"), val = tensor(-1)]; + tensor out_55_cast_fp16 = stack(axis = out_55_axis_0, values = (xor__55_cast_fp16, xoi_55_cast_fp16))[name = tensor("out_55_cast_fp16")]; + tensor _inversed_3511_y_0_to_fp16 = const()[name = tensor("_inversed_3511_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3511_cast_fp16 = mul(x = position13_to_fp16, y = _inversed_3511_y_0_to_fp16)[name = tensor("_inversed_3511_cast_fp16")]; + tensor var_3512_cast_fp16 = floor(x = _inversed_3511_cast_fp16)[name = tensor("op_3512_cast_fp16")]; + tensor var_3513_to_fp16 = const()[name = tensor("op_3513_to_fp16"), val = tensor(0x1p+9)]; + tensor var_3514_cast_fp16 = mul(x = var_3512_cast_fp16, y = var_3513_to_fp16)[name = tensor("op_3514_cast_fp16")]; + tensor write_idx_27_cast_fp16 = sub(x = position13_to_fp16, y = var_3514_cast_fp16)[name = tensor("write_idx_27_cast_fp16")]; + tensor var_3517_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_27_cast_fp16)[name = tensor("op_3517_cast_fp16")]; + tensor var_3527 = const()[name = tensor("op_3527"), val = tensor([1, 512, 1, 1])]; + tensor onehot_53_to_fp16_dtype_0 = const()[name = tensor("onehot_53_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_3517_cast_fp16_to_fp16 = cast(dtype = onehot_53_to_fp16_dtype_0, x = var_3517_cast_fp16)[name = tensor("cast_472")]; + tensor onehot_55_cast_fp16 = reshape(shape = var_3527, x = var_3517_cast_fp16_to_fp16)[name = tensor("onehot_55_cast_fp16")]; + tensor var_3529_to_fp16 = const()[name = tensor("op_3529_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3531_cast_fp16 = sub(x = var_3529_to_fp16, y = onehot_55_cast_fp16)[name = tensor("op_3531_cast_fp16")]; + tensor k_cache13_to_fp16_dtype_0 = const()[name = tensor("k_cache13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache13_to_fp16 = cast(dtype = k_cache13_to_fp16_dtype_0, x = k_cache13)[name = tensor("cast_471")]; + tensor var_3532_cast_fp16 = mul(x = k_cache13_to_fp16, y = var_3531_cast_fp16)[name = tensor("op_3532_cast_fp16")]; + tensor var_3537 = const()[name = tensor("op_3537"), val = tensor([1, 1, 16, 64])]; + tensor var_3538_cast_fp16 = reshape(shape = var_3537, x = out_55_cast_fp16)[name = tensor("op_3538_cast_fp16")]; + tensor var_3539_cast_fp16 = mul(x = var_3538_cast_fp16, y = onehot_55_cast_fp16)[name = tensor("op_3539_cast_fp16")]; + tensor new_k_cache_27_cast_fp16 = add(x = var_3532_cast_fp16, y = var_3539_cast_fp16)[name = tensor("new_k_cache_27_cast_fp16")]; + tensor new_k_cache_27_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_27_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache13_to_fp16_dtype_0 = const()[name = tensor("v_cache13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache13_to_fp16 = cast(dtype = v_cache13_to_fp16_dtype_0, x = v_cache13)[name = tensor("cast_469")]; + tensor var_3545_cast_fp16 = mul(x = v_cache13_to_fp16, y = var_3531_cast_fp16)[name = tensor("op_3545_cast_fp16")]; + tensor var_3550 = const()[name = tensor("op_3550"), val = tensor([1, 1, 16, 64])]; + tensor var_3551_cast_fp16 = reshape(shape = var_3550, x = v_27_cast_fp16)[name = tensor("op_3551_cast_fp16")]; + tensor var_3552_cast_fp16 = mul(x = var_3551_cast_fp16, y = onehot_55_cast_fp16)[name = tensor("op_3552_cast_fp16")]; + tensor new_v_cache_27_cast_fp16 = add(x = var_3545_cast_fp16, y = var_3552_cast_fp16)[name = tensor("new_v_cache_27_cast_fp16")]; + tensor new_v_cache_27_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_27_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_13_perm_0 = const()[name = tensor("transpose_13_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_3565 = const()[name = tensor("op_3565"), val = tensor([0, 2, 1, 3])]; + tensor var_3571 = const()[name = tensor("op_3571"), val = tensor([1, 16, 1, 64])]; + tensor q4_27_cast_fp16 = reshape(shape = var_3571, x = out_53_cast_fp16)[name = tensor("q4_27_cast_fp16")]; + tensor var_3576_transpose_x_0 = const()[name = tensor("op_3576_transpose_x_0"), val = tensor(false)]; + tensor var_3576_transpose_y_0 = const()[name = tensor("op_3576_transpose_y_0"), val = tensor(false)]; + tensor transpose_13_cast_fp16 = transpose(perm = transpose_13_perm_0, x = new_k_cache_27_cast_fp16)[name = tensor("transpose_56")]; + tensor var_3576_cast_fp16 = matmul(transpose_x = var_3576_transpose_x_0, transpose_y = var_3576_transpose_y_0, x = q4_27_cast_fp16, y = transpose_13_cast_fp16)[name = tensor("op_3576_cast_fp16")]; + tensor var_3577_to_fp16 = const()[name = tensor("op_3577_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_53_cast_fp16 = mul(x = var_3576_cast_fp16, y = var_3577_to_fp16)[name = tensor("scores_53_cast_fp16")]; + tensor var_3581 = const()[name = tensor("op_3581"), val = tensor([1, 1])]; + tensor var_3582_cast_fp16 = reshape(shape = var_3581, x = position13_to_fp16)[name = tensor("op_3582_cast_fp16")]; + tensor var_3583_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_3582_cast_fp16)[name = tensor("op_3583_cast_fp16")]; + tensor mask_27_to_fp16_dtype_0 = const()[name = tensor("mask_27_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_3590_to_fp16 = const()[name = tensor("op_3590_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3583_cast_fp16_to_fp16 = cast(dtype = mask_27_to_fp16_dtype_0, x = var_3583_cast_fp16)[name = tensor("cast_467")]; + tensor var_3591_cast_fp16 = sub(x = var_3583_cast_fp16_to_fp16, y = var_3590_to_fp16)[name = tensor("op_3591_cast_fp16")]; + tensor var_3596 = const()[name = tensor("op_3596"), val = tensor([1, 1, 1, 512])]; + tensor var_3597_cast_fp16 = reshape(shape = var_3596, x = var_3591_cast_fp16)[name = tensor("op_3597_cast_fp16")]; + tensor var_3598_to_fp16 = const()[name = tensor("op_3598_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_3599_cast_fp16 = mul(x = var_3597_cast_fp16, y = var_3598_to_fp16)[name = tensor("op_3599_cast_fp16")]; + tensor scores_55_cast_fp16 = add(x = scores_53_cast_fp16, y = var_3599_cast_fp16)[name = tensor("scores_55_cast_fp16")]; + tensor var_3602 = const()[name = tensor("op_3602"), val = tensor(-1)]; + tensor weights_27_cast_fp16 = softmax(axis = var_3602, x = scores_55_cast_fp16)[name = tensor("weights_27_cast_fp16")]; + tensor attn_27_transpose_x_0 = const()[name = tensor("attn_27_transpose_x_0"), val = tensor(false)]; + tensor attn_27_transpose_y_0 = const()[name = tensor("attn_27_transpose_y_0"), val = tensor(false)]; + tensor values_27_cast_fp16 = transpose(perm = var_3565, x = new_v_cache_27_cast_fp16)[name = tensor("transpose_55")]; + tensor attn_27_cast_fp16 = matmul(transpose_x = attn_27_transpose_x_0, transpose_y = attn_27_transpose_y_0, x = weights_27_cast_fp16, y = values_27_cast_fp16)[name = tensor("attn_27_cast_fp16")]; + tensor var_3610 = const()[name = tensor("op_3610"), val = tensor([0, 2, 1, 3])]; + tensor var_3615 = const()[name = tensor("op_3615"), val = tensor([1, 1, 1024])]; + tensor var_3611_cast_fp16 = transpose(perm = var_3610, x = attn_27_cast_fp16)[name = tensor("transpose_54")]; + tensor input_163_cast_fp16 = reshape(shape = var_3615, x = var_3611_cast_fp16)[name = tensor("input_163_cast_fp16")]; + tensor attn13_out_proj_weight_to_fp16 = const()[name = tensor("attn13_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(333648064)))]; + tensor linear_54_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn13_out_proj_weight_to_fp16, x = input_163_cast_fp16)[name = tensor("linear_54_cast_fp16")]; + tensor var_3621_to_fp16 = const()[name = tensor("op_3621_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3622_cast_fp16 = add(x = position13_to_fp16, y = var_3621_to_fp16)[name = tensor("op_3622_cast_fp16")]; + tensor var_3622_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3622_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_165_cast_fp16 = add(x = input_159_cast_fp16, y = linear_54_cast_fp16)[name = tensor("input_165_cast_fp16")]; + tensor input_167_axes_0 = const()[name = tensor("input_167_axes_0"), val = tensor([-1])]; + tensor norm13_2_weight_to_fp16 = const()[name = tensor("norm13_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335745280)))]; + tensor norm13_2_bias_to_fp16 = const()[name = tensor("norm13_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335747392)))]; + tensor var_3626_to_fp16 = const()[name = tensor("op_3626_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_167_cast_fp16 = layer_norm(axes = input_167_axes_0, beta = norm13_2_bias_to_fp16, epsilon = var_3626_to_fp16, gamma = norm13_2_weight_to_fp16, x = input_165_cast_fp16)[name = tensor("input_167_cast_fp16")]; + tensor linear13_1_weight_to_fp16 = const()[name = tensor("linear13_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335749504)))]; + tensor linear_55_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear13_1_weight_to_fp16, x = input_167_cast_fp16)[name = tensor("linear_55_cast_fp16")]; + tensor input_169_mode_0 = const()[name = tensor("input_169_mode_0"), val = tensor("EXACT")]; + tensor input_169_cast_fp16 = gelu(mode = input_169_mode_0, x = linear_55_cast_fp16)[name = tensor("input_169_cast_fp16")]; + tensor linear13_2_weight_to_fp16 = const()[name = tensor("linear13_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(344138176)))]; + tensor linear_56_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear13_2_weight_to_fp16, x = input_169_cast_fp16)[name = tensor("linear_56_cast_fp16")]; + tensor input_171_cast_fp16 = add(x = input_165_cast_fp16, y = linear_56_cast_fp16)[name = tensor("input_171_cast_fp16")]; + tensor input_173_axes_0 = const()[name = tensor("input_173_axes_0"), val = tensor([-1])]; + tensor norm14_1_weight_to_fp16 = const()[name = tensor("norm14_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(352526848)))]; + tensor norm14_1_bias_to_fp16 = const()[name = tensor("norm14_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(352528960)))]; + tensor var_3643_to_fp16 = const()[name = tensor("op_3643_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_173_cast_fp16 = layer_norm(axes = input_173_axes_0, beta = norm14_1_bias_to_fp16, epsilon = var_3643_to_fp16, gamma = norm14_1_weight_to_fp16, x = input_171_cast_fp16)[name = tensor("input_173_cast_fp16")]; + tensor attn14_in_proj_weight_to_fp16 = const()[name = tensor("attn14_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(352531072)))]; + tensor linear_57_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn14_in_proj_weight_to_fp16, x = input_173_cast_fp16)[name = tensor("linear_57_cast_fp16")]; + tensor var_3656 = const()[name = tensor("op_3656"), val = tensor([1, 3, 16, 64])]; + tensor qkv_29_cast_fp16 = reshape(shape = var_3656, x = linear_57_cast_fp16)[name = tensor("qkv_29_cast_fp16")]; + tensor x_57_begin_0 = const()[name = tensor("x_57_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_57_end_0 = const()[name = tensor("x_57_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_57_end_mask_0 = const()[name = tensor("x_57_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_57_squeeze_mask_0 = const()[name = tensor("x_57_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_57_cast_fp16 = slice_by_index(begin = x_57_begin_0, end = x_57_end_0, end_mask = x_57_end_mask_0, squeeze_mask = x_57_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("x_57_cast_fp16")]; + tensor x_59_begin_0 = const()[name = tensor("x_59_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_59_end_0 = const()[name = tensor("x_59_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_59_end_mask_0 = const()[name = tensor("x_59_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_59_squeeze_mask_0 = const()[name = tensor("x_59_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_59_cast_fp16 = slice_by_index(begin = x_59_begin_0, end = x_59_end_0, end_mask = x_59_end_mask_0, squeeze_mask = x_59_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("x_59_cast_fp16")]; + tensor v_29_begin_0 = const()[name = tensor("v_29_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_29_end_0 = const()[name = tensor("v_29_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_29_end_mask_0 = const()[name = tensor("v_29_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_29_squeeze_mask_0 = const()[name = tensor("v_29_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_29_cast_fp16 = slice_by_index(begin = v_29_begin_0, end = v_29_end_0, end_mask = v_29_end_mask_0, squeeze_mask = v_29_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("v_29_cast_fp16")]; + tensor var_3685 = const()[name = tensor("op_3685"), val = tensor([1, 1, 1])]; + tensor position14_to_fp16_dtype_0 = const()[name = tensor("position14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position14_to_fp16 = cast(dtype = position14_to_fp16_dtype_0, x = position14)[name = tensor("cast_465")]; + tensor var_3686_cast_fp16 = reshape(shape = var_3685, x = position14_to_fp16)[name = tensor("op_3686_cast_fp16")]; + tensor angles_57_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_3686_cast_fp16)[name = tensor("angles_57_cast_fp16")]; + tensor rotr_57_cast_fp16 = cos(x = angles_57_cast_fp16)[name = tensor("rotr_57_cast_fp16")]; + tensor roti_57_cast_fp16 = sin(x = angles_57_cast_fp16)[name = tensor("roti_57_cast_fp16")]; + tensor var_3694 = const()[name = tensor("op_3694"), val = tensor([1, 16, 32, 2])]; + tensor pairs_57_cast_fp16 = reshape(shape = var_3694, x = x_57_cast_fp16)[name = tensor("pairs_57_cast_fp16")]; + tensor xr_57_begin_0 = const()[name = tensor("xr_57_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_57_end_0 = const()[name = tensor("xr_57_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_57_end_mask_0 = const()[name = tensor("xr_57_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_57_squeeze_mask_0 = const()[name = tensor("xr_57_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_57_cast_fp16 = slice_by_index(begin = xr_57_begin_0, end = xr_57_end_0, end_mask = xr_57_end_mask_0, squeeze_mask = xr_57_squeeze_mask_0, x = pairs_57_cast_fp16)[name = tensor("xr_57_cast_fp16")]; + tensor xi_57_begin_0 = const()[name = tensor("xi_57_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_57_end_0 = const()[name = tensor("xi_57_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_57_end_mask_0 = const()[name = tensor("xi_57_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_57_squeeze_mask_0 = const()[name = tensor("xi_57_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_57_cast_fp16 = slice_by_index(begin = xi_57_begin_0, end = xi_57_end_0, end_mask = xi_57_end_mask_0, squeeze_mask = xi_57_squeeze_mask_0, x = pairs_57_cast_fp16)[name = tensor("xi_57_cast_fp16")]; + tensor var_3702_cast_fp16 = mul(x = xr_57_cast_fp16, y = rotr_57_cast_fp16)[name = tensor("op_3702_cast_fp16")]; + tensor var_3703_cast_fp16 = mul(x = xi_57_cast_fp16, y = roti_57_cast_fp16)[name = tensor("op_3703_cast_fp16")]; + tensor xor__57_cast_fp16 = sub(x = var_3702_cast_fp16, y = var_3703_cast_fp16)[name = tensor("xor__57_cast_fp16")]; + tensor var_3706_cast_fp16 = mul(x = xr_57_cast_fp16, y = roti_57_cast_fp16)[name = tensor("op_3706_cast_fp16")]; + tensor var_3707_cast_fp16 = mul(x = xi_57_cast_fp16, y = rotr_57_cast_fp16)[name = tensor("op_3707_cast_fp16")]; + tensor xoi_57_cast_fp16 = add(x = var_3706_cast_fp16, y = var_3707_cast_fp16)[name = tensor("xoi_57_cast_fp16")]; + tensor out_57_axis_0 = const()[name = tensor("out_57_axis_0"), val = tensor(-1)]; + tensor out_57_cast_fp16 = stack(axis = out_57_axis_0, values = (xor__57_cast_fp16, xoi_57_cast_fp16))[name = tensor("out_57_cast_fp16")]; + tensor var_3730 = const()[name = tensor("op_3730"), val = tensor([1, 16, 32, 2])]; + tensor pairs_59_cast_fp16 = reshape(shape = var_3730, x = x_59_cast_fp16)[name = tensor("pairs_59_cast_fp16")]; + tensor xr_59_begin_0 = const()[name = tensor("xr_59_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_59_end_0 = const()[name = tensor("xr_59_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_59_end_mask_0 = const()[name = tensor("xr_59_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_59_squeeze_mask_0 = const()[name = tensor("xr_59_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_59_cast_fp16 = slice_by_index(begin = xr_59_begin_0, end = xr_59_end_0, end_mask = xr_59_end_mask_0, squeeze_mask = xr_59_squeeze_mask_0, x = pairs_59_cast_fp16)[name = tensor("xr_59_cast_fp16")]; + tensor xi_59_begin_0 = const()[name = tensor("xi_59_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_59_end_0 = const()[name = tensor("xi_59_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_59_end_mask_0 = const()[name = tensor("xi_59_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_59_squeeze_mask_0 = const()[name = tensor("xi_59_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_59_cast_fp16 = slice_by_index(begin = xi_59_begin_0, end = xi_59_end_0, end_mask = xi_59_end_mask_0, squeeze_mask = xi_59_squeeze_mask_0, x = pairs_59_cast_fp16)[name = tensor("xi_59_cast_fp16")]; + tensor var_3738_cast_fp16 = mul(x = xr_59_cast_fp16, y = rotr_57_cast_fp16)[name = tensor("op_3738_cast_fp16")]; + tensor var_3739_cast_fp16 = mul(x = xi_59_cast_fp16, y = roti_57_cast_fp16)[name = tensor("op_3739_cast_fp16")]; + tensor xor__59_cast_fp16 = sub(x = var_3738_cast_fp16, y = var_3739_cast_fp16)[name = tensor("xor__59_cast_fp16")]; + tensor var_3742_cast_fp16 = mul(x = xr_59_cast_fp16, y = roti_57_cast_fp16)[name = tensor("op_3742_cast_fp16")]; + tensor var_3743_cast_fp16 = mul(x = xi_59_cast_fp16, y = rotr_57_cast_fp16)[name = tensor("op_3743_cast_fp16")]; + tensor xoi_59_cast_fp16 = add(x = var_3742_cast_fp16, y = var_3743_cast_fp16)[name = tensor("xoi_59_cast_fp16")]; + tensor out_59_axis_0 = const()[name = tensor("out_59_axis_0"), val = tensor(-1)]; + tensor out_59_cast_fp16 = stack(axis = out_59_axis_0, values = (xor__59_cast_fp16, xoi_59_cast_fp16))[name = tensor("out_59_cast_fp16")]; + tensor _inversed_3755_y_0_to_fp16 = const()[name = tensor("_inversed_3755_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3755_cast_fp16 = mul(x = position14_to_fp16, y = _inversed_3755_y_0_to_fp16)[name = tensor("_inversed_3755_cast_fp16")]; + tensor var_3756_cast_fp16 = floor(x = _inversed_3755_cast_fp16)[name = tensor("op_3756_cast_fp16")]; + tensor var_3757_to_fp16 = const()[name = tensor("op_3757_to_fp16"), val = tensor(0x1p+9)]; + tensor var_3758_cast_fp16 = mul(x = var_3756_cast_fp16, y = var_3757_to_fp16)[name = tensor("op_3758_cast_fp16")]; + tensor write_idx_29_cast_fp16 = sub(x = position14_to_fp16, y = var_3758_cast_fp16)[name = tensor("write_idx_29_cast_fp16")]; + tensor var_3761_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_29_cast_fp16)[name = tensor("op_3761_cast_fp16")]; + tensor var_3771 = const()[name = tensor("op_3771"), val = tensor([1, 512, 1, 1])]; + tensor onehot_57_to_fp16_dtype_0 = const()[name = tensor("onehot_57_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_3761_cast_fp16_to_fp16 = cast(dtype = onehot_57_to_fp16_dtype_0, x = var_3761_cast_fp16)[name = tensor("cast_464")]; + tensor onehot_59_cast_fp16 = reshape(shape = var_3771, x = var_3761_cast_fp16_to_fp16)[name = tensor("onehot_59_cast_fp16")]; + tensor var_3773_to_fp16 = const()[name = tensor("op_3773_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3775_cast_fp16 = sub(x = var_3773_to_fp16, y = onehot_59_cast_fp16)[name = tensor("op_3775_cast_fp16")]; + tensor k_cache14_to_fp16_dtype_0 = const()[name = tensor("k_cache14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache14_to_fp16 = cast(dtype = k_cache14_to_fp16_dtype_0, x = k_cache14)[name = tensor("cast_463")]; + tensor var_3776_cast_fp16 = mul(x = k_cache14_to_fp16, y = var_3775_cast_fp16)[name = tensor("op_3776_cast_fp16")]; + tensor var_3781 = const()[name = tensor("op_3781"), val = tensor([1, 1, 16, 64])]; + tensor var_3782_cast_fp16 = reshape(shape = var_3781, x = out_59_cast_fp16)[name = tensor("op_3782_cast_fp16")]; + tensor var_3783_cast_fp16 = mul(x = var_3782_cast_fp16, y = onehot_59_cast_fp16)[name = tensor("op_3783_cast_fp16")]; + tensor new_k_cache_29_cast_fp16 = add(x = var_3776_cast_fp16, y = var_3783_cast_fp16)[name = tensor("new_k_cache_29_cast_fp16")]; + tensor new_k_cache_29_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_29_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache14_to_fp16_dtype_0 = const()[name = tensor("v_cache14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache14_to_fp16 = cast(dtype = v_cache14_to_fp16_dtype_0, x = v_cache14)[name = tensor("cast_461")]; + tensor var_3789_cast_fp16 = mul(x = v_cache14_to_fp16, y = var_3775_cast_fp16)[name = tensor("op_3789_cast_fp16")]; + tensor var_3794 = const()[name = tensor("op_3794"), val = tensor([1, 1, 16, 64])]; + tensor var_3795_cast_fp16 = reshape(shape = var_3794, x = v_29_cast_fp16)[name = tensor("op_3795_cast_fp16")]; + tensor var_3796_cast_fp16 = mul(x = var_3795_cast_fp16, y = onehot_59_cast_fp16)[name = tensor("op_3796_cast_fp16")]; + tensor new_v_cache_29_cast_fp16 = add(x = var_3789_cast_fp16, y = var_3796_cast_fp16)[name = tensor("new_v_cache_29_cast_fp16")]; + tensor new_v_cache_29_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_29_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_14_perm_0 = const()[name = tensor("transpose_14_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_3809 = const()[name = tensor("op_3809"), val = tensor([0, 2, 1, 3])]; + tensor var_3815 = const()[name = tensor("op_3815"), val = tensor([1, 16, 1, 64])]; + tensor q4_29_cast_fp16 = reshape(shape = var_3815, x = out_57_cast_fp16)[name = tensor("q4_29_cast_fp16")]; + tensor var_3820_transpose_x_0 = const()[name = tensor("op_3820_transpose_x_0"), val = tensor(false)]; + tensor var_3820_transpose_y_0 = const()[name = tensor("op_3820_transpose_y_0"), val = tensor(false)]; + tensor transpose_14_cast_fp16 = transpose(perm = transpose_14_perm_0, x = new_k_cache_29_cast_fp16)[name = tensor("transpose_53")]; + tensor var_3820_cast_fp16 = matmul(transpose_x = var_3820_transpose_x_0, transpose_y = var_3820_transpose_y_0, x = q4_29_cast_fp16, y = transpose_14_cast_fp16)[name = tensor("op_3820_cast_fp16")]; + tensor var_3821_to_fp16 = const()[name = tensor("op_3821_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_57_cast_fp16 = mul(x = var_3820_cast_fp16, y = var_3821_to_fp16)[name = tensor("scores_57_cast_fp16")]; + tensor var_3825 = const()[name = tensor("op_3825"), val = tensor([1, 1])]; + tensor var_3826_cast_fp16 = reshape(shape = var_3825, x = position14_to_fp16)[name = tensor("op_3826_cast_fp16")]; + tensor var_3827_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_3826_cast_fp16)[name = tensor("op_3827_cast_fp16")]; + tensor mask_29_to_fp16_dtype_0 = const()[name = tensor("mask_29_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_3834_to_fp16 = const()[name = tensor("op_3834_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3827_cast_fp16_to_fp16 = cast(dtype = mask_29_to_fp16_dtype_0, x = var_3827_cast_fp16)[name = tensor("cast_459")]; + tensor var_3835_cast_fp16 = sub(x = var_3827_cast_fp16_to_fp16, y = var_3834_to_fp16)[name = tensor("op_3835_cast_fp16")]; + tensor var_3840 = const()[name = tensor("op_3840"), val = tensor([1, 1, 1, 512])]; + tensor var_3841_cast_fp16 = reshape(shape = var_3840, x = var_3835_cast_fp16)[name = tensor("op_3841_cast_fp16")]; + tensor var_3842_to_fp16 = const()[name = tensor("op_3842_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_3843_cast_fp16 = mul(x = var_3841_cast_fp16, y = var_3842_to_fp16)[name = tensor("op_3843_cast_fp16")]; + tensor scores_59_cast_fp16 = add(x = scores_57_cast_fp16, y = var_3843_cast_fp16)[name = tensor("scores_59_cast_fp16")]; + tensor var_3846 = const()[name = tensor("op_3846"), val = tensor(-1)]; + tensor weights_29_cast_fp16 = softmax(axis = var_3846, x = scores_59_cast_fp16)[name = tensor("weights_29_cast_fp16")]; + tensor attn_29_transpose_x_0 = const()[name = tensor("attn_29_transpose_x_0"), val = tensor(false)]; + tensor attn_29_transpose_y_0 = const()[name = tensor("attn_29_transpose_y_0"), val = tensor(false)]; + tensor values_29_cast_fp16 = transpose(perm = var_3809, x = new_v_cache_29_cast_fp16)[name = tensor("transpose_52")]; + tensor attn_29_cast_fp16 = matmul(transpose_x = attn_29_transpose_x_0, transpose_y = attn_29_transpose_y_0, x = weights_29_cast_fp16, y = values_29_cast_fp16)[name = tensor("attn_29_cast_fp16")]; + tensor var_3854 = const()[name = tensor("op_3854"), val = tensor([0, 2, 1, 3])]; + tensor var_3859 = const()[name = tensor("op_3859"), val = tensor([1, 1, 1024])]; + tensor var_3855_cast_fp16 = transpose(perm = var_3854, x = attn_29_cast_fp16)[name = tensor("transpose_51")]; + tensor input_175_cast_fp16 = reshape(shape = var_3859, x = var_3855_cast_fp16)[name = tensor("input_175_cast_fp16")]; + tensor attn14_out_proj_weight_to_fp16 = const()[name = tensor("attn14_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(358822592)))]; + tensor linear_58_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn14_out_proj_weight_to_fp16, x = input_175_cast_fp16)[name = tensor("linear_58_cast_fp16")]; + tensor var_3865_to_fp16 = const()[name = tensor("op_3865_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3866_cast_fp16 = add(x = position14_to_fp16, y = var_3865_to_fp16)[name = tensor("op_3866_cast_fp16")]; + tensor var_3866_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3866_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_177_cast_fp16 = add(x = input_171_cast_fp16, y = linear_58_cast_fp16)[name = tensor("input_177_cast_fp16")]; + tensor input_179_axes_0 = const()[name = tensor("input_179_axes_0"), val = tensor([-1])]; + tensor norm14_2_weight_to_fp16 = const()[name = tensor("norm14_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(360919808)))]; + tensor norm14_2_bias_to_fp16 = const()[name = tensor("norm14_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(360921920)))]; + tensor var_3870_to_fp16 = const()[name = tensor("op_3870_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_179_cast_fp16 = layer_norm(axes = input_179_axes_0, beta = norm14_2_bias_to_fp16, epsilon = var_3870_to_fp16, gamma = norm14_2_weight_to_fp16, x = input_177_cast_fp16)[name = tensor("input_179_cast_fp16")]; + tensor linear14_1_weight_to_fp16 = const()[name = tensor("linear14_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(360924032)))]; + tensor linear_59_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear14_1_weight_to_fp16, x = input_179_cast_fp16)[name = tensor("linear_59_cast_fp16")]; + tensor input_181_mode_0 = const()[name = tensor("input_181_mode_0"), val = tensor("EXACT")]; + tensor input_181_cast_fp16 = gelu(mode = input_181_mode_0, x = linear_59_cast_fp16)[name = tensor("input_181_cast_fp16")]; + tensor linear14_2_weight_to_fp16 = const()[name = tensor("linear14_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(369312704)))]; + tensor linear_60_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear14_2_weight_to_fp16, x = input_181_cast_fp16)[name = tensor("linear_60_cast_fp16")]; + tensor input_183_cast_fp16 = add(x = input_177_cast_fp16, y = linear_60_cast_fp16)[name = tensor("input_183_cast_fp16")]; + tensor input_185_axes_0 = const()[name = tensor("input_185_axes_0"), val = tensor([-1])]; + tensor norm15_1_weight_to_fp16 = const()[name = tensor("norm15_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(377701376)))]; + tensor norm15_1_bias_to_fp16 = const()[name = tensor("norm15_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(377703488)))]; + tensor var_3887_to_fp16 = const()[name = tensor("op_3887_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_185_cast_fp16 = layer_norm(axes = input_185_axes_0, beta = norm15_1_bias_to_fp16, epsilon = var_3887_to_fp16, gamma = norm15_1_weight_to_fp16, x = input_183_cast_fp16)[name = tensor("input_185_cast_fp16")]; + tensor attn15_in_proj_weight_to_fp16 = const()[name = tensor("attn15_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(377705600)))]; + tensor linear_61_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn15_in_proj_weight_to_fp16, x = input_185_cast_fp16)[name = tensor("linear_61_cast_fp16")]; + tensor var_3900 = const()[name = tensor("op_3900"), val = tensor([1, 3, 16, 64])]; + tensor qkv_31_cast_fp16 = reshape(shape = var_3900, x = linear_61_cast_fp16)[name = tensor("qkv_31_cast_fp16")]; + tensor x_61_begin_0 = const()[name = tensor("x_61_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_61_end_0 = const()[name = tensor("x_61_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_61_end_mask_0 = const()[name = tensor("x_61_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_61_squeeze_mask_0 = const()[name = tensor("x_61_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_61_cast_fp16 = slice_by_index(begin = x_61_begin_0, end = x_61_end_0, end_mask = x_61_end_mask_0, squeeze_mask = x_61_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("x_61_cast_fp16")]; + tensor x_63_begin_0 = const()[name = tensor("x_63_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_63_end_0 = const()[name = tensor("x_63_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_63_end_mask_0 = const()[name = tensor("x_63_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_63_squeeze_mask_0 = const()[name = tensor("x_63_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_63_cast_fp16 = slice_by_index(begin = x_63_begin_0, end = x_63_end_0, end_mask = x_63_end_mask_0, squeeze_mask = x_63_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("x_63_cast_fp16")]; + tensor v_31_begin_0 = const()[name = tensor("v_31_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_31_end_0 = const()[name = tensor("v_31_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_31_end_mask_0 = const()[name = tensor("v_31_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_31_squeeze_mask_0 = const()[name = tensor("v_31_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_31_cast_fp16 = slice_by_index(begin = v_31_begin_0, end = v_31_end_0, end_mask = v_31_end_mask_0, squeeze_mask = v_31_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("v_31_cast_fp16")]; + tensor var_3929 = const()[name = tensor("op_3929"), val = tensor([1, 1, 1])]; + tensor position15_to_fp16_dtype_0 = const()[name = tensor("position15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position15_to_fp16 = cast(dtype = position15_to_fp16_dtype_0, x = position15)[name = tensor("cast_457")]; + tensor var_3930_cast_fp16 = reshape(shape = var_3929, x = position15_to_fp16)[name = tensor("op_3930_cast_fp16")]; + tensor angles_61_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_3930_cast_fp16)[name = tensor("angles_61_cast_fp16")]; + tensor rotr_61_cast_fp16 = cos(x = angles_61_cast_fp16)[name = tensor("rotr_61_cast_fp16")]; + tensor roti_61_cast_fp16 = sin(x = angles_61_cast_fp16)[name = tensor("roti_61_cast_fp16")]; + tensor var_3938 = const()[name = tensor("op_3938"), val = tensor([1, 16, 32, 2])]; + tensor pairs_61_cast_fp16 = reshape(shape = var_3938, x = x_61_cast_fp16)[name = tensor("pairs_61_cast_fp16")]; + tensor xr_61_begin_0 = const()[name = tensor("xr_61_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_61_end_0 = const()[name = tensor("xr_61_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_61_end_mask_0 = const()[name = tensor("xr_61_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_61_squeeze_mask_0 = const()[name = tensor("xr_61_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_61_cast_fp16 = slice_by_index(begin = xr_61_begin_0, end = xr_61_end_0, end_mask = xr_61_end_mask_0, squeeze_mask = xr_61_squeeze_mask_0, x = pairs_61_cast_fp16)[name = tensor("xr_61_cast_fp16")]; + tensor xi_61_begin_0 = const()[name = tensor("xi_61_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_61_end_0 = const()[name = tensor("xi_61_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_61_end_mask_0 = const()[name = tensor("xi_61_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_61_squeeze_mask_0 = const()[name = tensor("xi_61_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_61_cast_fp16 = slice_by_index(begin = xi_61_begin_0, end = xi_61_end_0, end_mask = xi_61_end_mask_0, squeeze_mask = xi_61_squeeze_mask_0, x = pairs_61_cast_fp16)[name = tensor("xi_61_cast_fp16")]; + tensor var_3946_cast_fp16 = mul(x = xr_61_cast_fp16, y = rotr_61_cast_fp16)[name = tensor("op_3946_cast_fp16")]; + tensor var_3947_cast_fp16 = mul(x = xi_61_cast_fp16, y = roti_61_cast_fp16)[name = tensor("op_3947_cast_fp16")]; + tensor xor__61_cast_fp16 = sub(x = var_3946_cast_fp16, y = var_3947_cast_fp16)[name = tensor("xor__61_cast_fp16")]; + tensor var_3950_cast_fp16 = mul(x = xr_61_cast_fp16, y = roti_61_cast_fp16)[name = tensor("op_3950_cast_fp16")]; + tensor var_3951_cast_fp16 = mul(x = xi_61_cast_fp16, y = rotr_61_cast_fp16)[name = tensor("op_3951_cast_fp16")]; + tensor xoi_61_cast_fp16 = add(x = var_3950_cast_fp16, y = var_3951_cast_fp16)[name = tensor("xoi_61_cast_fp16")]; + tensor out_61_axis_0 = const()[name = tensor("out_61_axis_0"), val = tensor(-1)]; + tensor out_61_cast_fp16 = stack(axis = out_61_axis_0, values = (xor__61_cast_fp16, xoi_61_cast_fp16))[name = tensor("out_61_cast_fp16")]; + tensor var_3974 = const()[name = tensor("op_3974"), val = tensor([1, 16, 32, 2])]; + tensor pairs_63_cast_fp16 = reshape(shape = var_3974, x = x_63_cast_fp16)[name = tensor("pairs_63_cast_fp16")]; + tensor xr_63_begin_0 = const()[name = tensor("xr_63_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_63_end_0 = const()[name = tensor("xr_63_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_63_end_mask_0 = const()[name = tensor("xr_63_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_63_squeeze_mask_0 = const()[name = tensor("xr_63_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_63_cast_fp16 = slice_by_index(begin = xr_63_begin_0, end = xr_63_end_0, end_mask = xr_63_end_mask_0, squeeze_mask = xr_63_squeeze_mask_0, x = pairs_63_cast_fp16)[name = tensor("xr_63_cast_fp16")]; + tensor xi_63_begin_0 = const()[name = tensor("xi_63_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_63_end_0 = const()[name = tensor("xi_63_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_63_end_mask_0 = const()[name = tensor("xi_63_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_63_squeeze_mask_0 = const()[name = tensor("xi_63_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_63_cast_fp16 = slice_by_index(begin = xi_63_begin_0, end = xi_63_end_0, end_mask = xi_63_end_mask_0, squeeze_mask = xi_63_squeeze_mask_0, x = pairs_63_cast_fp16)[name = tensor("xi_63_cast_fp16")]; + tensor var_3982_cast_fp16 = mul(x = xr_63_cast_fp16, y = rotr_61_cast_fp16)[name = tensor("op_3982_cast_fp16")]; + tensor var_3983_cast_fp16 = mul(x = xi_63_cast_fp16, y = roti_61_cast_fp16)[name = tensor("op_3983_cast_fp16")]; + tensor xor__63_cast_fp16 = sub(x = var_3982_cast_fp16, y = var_3983_cast_fp16)[name = tensor("xor__63_cast_fp16")]; + tensor var_3986_cast_fp16 = mul(x = xr_63_cast_fp16, y = roti_61_cast_fp16)[name = tensor("op_3986_cast_fp16")]; + tensor var_3987_cast_fp16 = mul(x = xi_63_cast_fp16, y = rotr_61_cast_fp16)[name = tensor("op_3987_cast_fp16")]; + tensor xoi_63_cast_fp16 = add(x = var_3986_cast_fp16, y = var_3987_cast_fp16)[name = tensor("xoi_63_cast_fp16")]; + tensor out_63_axis_0 = const()[name = tensor("out_63_axis_0"), val = tensor(-1)]; + tensor out_63_cast_fp16 = stack(axis = out_63_axis_0, values = (xor__63_cast_fp16, xoi_63_cast_fp16))[name = tensor("out_63_cast_fp16")]; + tensor _inversed_3999_y_0_to_fp16 = const()[name = tensor("_inversed_3999_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3999_cast_fp16 = mul(x = position15_to_fp16, y = _inversed_3999_y_0_to_fp16)[name = tensor("_inversed_3999_cast_fp16")]; + tensor var_4000_cast_fp16 = floor(x = _inversed_3999_cast_fp16)[name = tensor("op_4000_cast_fp16")]; + tensor var_4001_to_fp16 = const()[name = tensor("op_4001_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4002_cast_fp16 = mul(x = var_4000_cast_fp16, y = var_4001_to_fp16)[name = tensor("op_4002_cast_fp16")]; + tensor write_idx_31_cast_fp16 = sub(x = position15_to_fp16, y = var_4002_cast_fp16)[name = tensor("write_idx_31_cast_fp16")]; + tensor var_4005_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_31_cast_fp16)[name = tensor("op_4005_cast_fp16")]; + tensor var_4015 = const()[name = tensor("op_4015"), val = tensor([1, 512, 1, 1])]; + tensor onehot_61_to_fp16_dtype_0 = const()[name = tensor("onehot_61_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_4005_cast_fp16_to_fp16 = cast(dtype = onehot_61_to_fp16_dtype_0, x = var_4005_cast_fp16)[name = tensor("cast_456")]; + tensor onehot_63_cast_fp16 = reshape(shape = var_4015, x = var_4005_cast_fp16_to_fp16)[name = tensor("onehot_63_cast_fp16")]; + tensor var_4017_to_fp16 = const()[name = tensor("op_4017_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4019_cast_fp16 = sub(x = var_4017_to_fp16, y = onehot_63_cast_fp16)[name = tensor("op_4019_cast_fp16")]; + tensor k_cache15_to_fp16_dtype_0 = const()[name = tensor("k_cache15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache15_to_fp16 = cast(dtype = k_cache15_to_fp16_dtype_0, x = k_cache15)[name = tensor("cast_455")]; + tensor var_4020_cast_fp16 = mul(x = k_cache15_to_fp16, y = var_4019_cast_fp16)[name = tensor("op_4020_cast_fp16")]; + tensor var_4025 = const()[name = tensor("op_4025"), val = tensor([1, 1, 16, 64])]; + tensor var_4026_cast_fp16 = reshape(shape = var_4025, x = out_63_cast_fp16)[name = tensor("op_4026_cast_fp16")]; + tensor var_4027_cast_fp16 = mul(x = var_4026_cast_fp16, y = onehot_63_cast_fp16)[name = tensor("op_4027_cast_fp16")]; + tensor new_k_cache_31_cast_fp16 = add(x = var_4020_cast_fp16, y = var_4027_cast_fp16)[name = tensor("new_k_cache_31_cast_fp16")]; + tensor new_k_cache_31_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_31_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache15_to_fp16_dtype_0 = const()[name = tensor("v_cache15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache15_to_fp16 = cast(dtype = v_cache15_to_fp16_dtype_0, x = v_cache15)[name = tensor("cast_453")]; + tensor var_4033_cast_fp16 = mul(x = v_cache15_to_fp16, y = var_4019_cast_fp16)[name = tensor("op_4033_cast_fp16")]; + tensor var_4038 = const()[name = tensor("op_4038"), val = tensor([1, 1, 16, 64])]; + tensor var_4039_cast_fp16 = reshape(shape = var_4038, x = v_31_cast_fp16)[name = tensor("op_4039_cast_fp16")]; + tensor var_4040_cast_fp16 = mul(x = var_4039_cast_fp16, y = onehot_63_cast_fp16)[name = tensor("op_4040_cast_fp16")]; + tensor new_v_cache_31_cast_fp16 = add(x = var_4033_cast_fp16, y = var_4040_cast_fp16)[name = tensor("new_v_cache_31_cast_fp16")]; + tensor new_v_cache_31_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_31_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_15_perm_0 = const()[name = tensor("transpose_15_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_4053 = const()[name = tensor("op_4053"), val = tensor([0, 2, 1, 3])]; + tensor var_4059 = const()[name = tensor("op_4059"), val = tensor([1, 16, 1, 64])]; + tensor q4_31_cast_fp16 = reshape(shape = var_4059, x = out_61_cast_fp16)[name = tensor("q4_31_cast_fp16")]; + tensor var_4064_transpose_x_0 = const()[name = tensor("op_4064_transpose_x_0"), val = tensor(false)]; + tensor var_4064_transpose_y_0 = const()[name = tensor("op_4064_transpose_y_0"), val = tensor(false)]; + tensor transpose_15_cast_fp16 = transpose(perm = transpose_15_perm_0, x = new_k_cache_31_cast_fp16)[name = tensor("transpose_50")]; + tensor var_4064_cast_fp16 = matmul(transpose_x = var_4064_transpose_x_0, transpose_y = var_4064_transpose_y_0, x = q4_31_cast_fp16, y = transpose_15_cast_fp16)[name = tensor("op_4064_cast_fp16")]; + tensor var_4065_to_fp16 = const()[name = tensor("op_4065_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_61_cast_fp16 = mul(x = var_4064_cast_fp16, y = var_4065_to_fp16)[name = tensor("scores_61_cast_fp16")]; + tensor var_4069 = const()[name = tensor("op_4069"), val = tensor([1, 1])]; + tensor var_4070_cast_fp16 = reshape(shape = var_4069, x = position15_to_fp16)[name = tensor("op_4070_cast_fp16")]; + tensor var_4071_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_4070_cast_fp16)[name = tensor("op_4071_cast_fp16")]; + tensor mask_31_to_fp16_dtype_0 = const()[name = tensor("mask_31_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_4078_to_fp16 = const()[name = tensor("op_4078_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4071_cast_fp16_to_fp16 = cast(dtype = mask_31_to_fp16_dtype_0, x = var_4071_cast_fp16)[name = tensor("cast_451")]; + tensor var_4079_cast_fp16 = sub(x = var_4071_cast_fp16_to_fp16, y = var_4078_to_fp16)[name = tensor("op_4079_cast_fp16")]; + tensor var_4084 = const()[name = tensor("op_4084"), val = tensor([1, 1, 1, 512])]; + tensor var_4085_cast_fp16 = reshape(shape = var_4084, x = var_4079_cast_fp16)[name = tensor("op_4085_cast_fp16")]; + tensor var_4086_to_fp16 = const()[name = tensor("op_4086_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_4087_cast_fp16 = mul(x = var_4085_cast_fp16, y = var_4086_to_fp16)[name = tensor("op_4087_cast_fp16")]; + tensor scores_63_cast_fp16 = add(x = scores_61_cast_fp16, y = var_4087_cast_fp16)[name = tensor("scores_63_cast_fp16")]; + tensor var_4090 = const()[name = tensor("op_4090"), val = tensor(-1)]; + tensor weights_31_cast_fp16 = softmax(axis = var_4090, x = scores_63_cast_fp16)[name = tensor("weights_31_cast_fp16")]; + tensor attn_31_transpose_x_0 = const()[name = tensor("attn_31_transpose_x_0"), val = tensor(false)]; + tensor attn_31_transpose_y_0 = const()[name = tensor("attn_31_transpose_y_0"), val = tensor(false)]; + tensor values_31_cast_fp16 = transpose(perm = var_4053, x = new_v_cache_31_cast_fp16)[name = tensor("transpose_49")]; + tensor attn_31_cast_fp16 = matmul(transpose_x = attn_31_transpose_x_0, transpose_y = attn_31_transpose_y_0, x = weights_31_cast_fp16, y = values_31_cast_fp16)[name = tensor("attn_31_cast_fp16")]; + tensor var_4098 = const()[name = tensor("op_4098"), val = tensor([0, 2, 1, 3])]; + tensor var_4103 = const()[name = tensor("op_4103"), val = tensor([1, 1, 1024])]; + tensor var_4099_cast_fp16 = transpose(perm = var_4098, x = attn_31_cast_fp16)[name = tensor("transpose_48")]; + tensor input_187_cast_fp16 = reshape(shape = var_4103, x = var_4099_cast_fp16)[name = tensor("input_187_cast_fp16")]; + tensor attn15_out_proj_weight_to_fp16 = const()[name = tensor("attn15_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(383997120)))]; + tensor linear_62_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn15_out_proj_weight_to_fp16, x = input_187_cast_fp16)[name = tensor("linear_62_cast_fp16")]; + tensor var_4109_to_fp16 = const()[name = tensor("op_4109_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4110_cast_fp16 = add(x = position15_to_fp16, y = var_4109_to_fp16)[name = tensor("op_4110_cast_fp16")]; + tensor var_4110_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4110_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_189_cast_fp16 = add(x = input_183_cast_fp16, y = linear_62_cast_fp16)[name = tensor("input_189_cast_fp16")]; + tensor input_191_axes_0 = const()[name = tensor("input_191_axes_0"), val = tensor([-1])]; + tensor norm15_2_weight_to_fp16 = const()[name = tensor("norm15_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(386094336)))]; + tensor norm15_2_bias_to_fp16 = const()[name = tensor("norm15_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(386096448)))]; + tensor var_4114_to_fp16 = const()[name = tensor("op_4114_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_191_cast_fp16 = layer_norm(axes = input_191_axes_0, beta = norm15_2_bias_to_fp16, epsilon = var_4114_to_fp16, gamma = norm15_2_weight_to_fp16, x = input_189_cast_fp16)[name = tensor("input_191_cast_fp16")]; + tensor linear15_1_weight_to_fp16 = const()[name = tensor("linear15_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(386098560)))]; + tensor linear_63_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear15_1_weight_to_fp16, x = input_191_cast_fp16)[name = tensor("linear_63_cast_fp16")]; + tensor input_193_mode_0 = const()[name = tensor("input_193_mode_0"), val = tensor("EXACT")]; + tensor input_193_cast_fp16 = gelu(mode = input_193_mode_0, x = linear_63_cast_fp16)[name = tensor("input_193_cast_fp16")]; + tensor linear15_2_weight_to_fp16 = const()[name = tensor("linear15_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(394487232)))]; + tensor linear_64_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear15_2_weight_to_fp16, x = input_193_cast_fp16)[name = tensor("linear_64_cast_fp16")]; + tensor input_195_cast_fp16 = add(x = input_189_cast_fp16, y = linear_64_cast_fp16)[name = tensor("input_195_cast_fp16")]; + tensor input_197_axes_0 = const()[name = tensor("input_197_axes_0"), val = tensor([-1])]; + tensor norm16_1_weight_to_fp16 = const()[name = tensor("norm16_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(402875904)))]; + tensor norm16_1_bias_to_fp16 = const()[name = tensor("norm16_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(402878016)))]; + tensor var_4131_to_fp16 = const()[name = tensor("op_4131_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_197_cast_fp16 = layer_norm(axes = input_197_axes_0, beta = norm16_1_bias_to_fp16, epsilon = var_4131_to_fp16, gamma = norm16_1_weight_to_fp16, x = input_195_cast_fp16)[name = tensor("input_197_cast_fp16")]; + tensor attn16_in_proj_weight_to_fp16 = const()[name = tensor("attn16_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(402880128)))]; + tensor linear_65_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn16_in_proj_weight_to_fp16, x = input_197_cast_fp16)[name = tensor("linear_65_cast_fp16")]; + tensor var_4144 = const()[name = tensor("op_4144"), val = tensor([1, 3, 16, 64])]; + tensor qkv_33_cast_fp16 = reshape(shape = var_4144, x = linear_65_cast_fp16)[name = tensor("qkv_33_cast_fp16")]; + tensor x_65_begin_0 = const()[name = tensor("x_65_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_65_end_0 = const()[name = tensor("x_65_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_65_end_mask_0 = const()[name = tensor("x_65_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_65_squeeze_mask_0 = const()[name = tensor("x_65_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_65_cast_fp16 = slice_by_index(begin = x_65_begin_0, end = x_65_end_0, end_mask = x_65_end_mask_0, squeeze_mask = x_65_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("x_65_cast_fp16")]; + tensor x_67_begin_0 = const()[name = tensor("x_67_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_67_end_0 = const()[name = tensor("x_67_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_67_end_mask_0 = const()[name = tensor("x_67_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_67_squeeze_mask_0 = const()[name = tensor("x_67_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_67_cast_fp16 = slice_by_index(begin = x_67_begin_0, end = x_67_end_0, end_mask = x_67_end_mask_0, squeeze_mask = x_67_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("x_67_cast_fp16")]; + tensor v_33_begin_0 = const()[name = tensor("v_33_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_33_end_0 = const()[name = tensor("v_33_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_33_end_mask_0 = const()[name = tensor("v_33_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_33_squeeze_mask_0 = const()[name = tensor("v_33_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_33_cast_fp16 = slice_by_index(begin = v_33_begin_0, end = v_33_end_0, end_mask = v_33_end_mask_0, squeeze_mask = v_33_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("v_33_cast_fp16")]; + tensor var_4173 = const()[name = tensor("op_4173"), val = tensor([1, 1, 1])]; + tensor position16_to_fp16_dtype_0 = const()[name = tensor("position16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position16_to_fp16 = cast(dtype = position16_to_fp16_dtype_0, x = position16)[name = tensor("cast_449")]; + tensor var_4174_cast_fp16 = reshape(shape = var_4173, x = position16_to_fp16)[name = tensor("op_4174_cast_fp16")]; + tensor angles_65_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_4174_cast_fp16)[name = tensor("angles_65_cast_fp16")]; + tensor rotr_65_cast_fp16 = cos(x = angles_65_cast_fp16)[name = tensor("rotr_65_cast_fp16")]; + tensor roti_65_cast_fp16 = sin(x = angles_65_cast_fp16)[name = tensor("roti_65_cast_fp16")]; + tensor var_4182 = const()[name = tensor("op_4182"), val = tensor([1, 16, 32, 2])]; + tensor pairs_65_cast_fp16 = reshape(shape = var_4182, x = x_65_cast_fp16)[name = tensor("pairs_65_cast_fp16")]; + tensor xr_65_begin_0 = const()[name = tensor("xr_65_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_65_end_0 = const()[name = tensor("xr_65_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_65_end_mask_0 = const()[name = tensor("xr_65_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_65_squeeze_mask_0 = const()[name = tensor("xr_65_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_65_cast_fp16 = slice_by_index(begin = xr_65_begin_0, end = xr_65_end_0, end_mask = xr_65_end_mask_0, squeeze_mask = xr_65_squeeze_mask_0, x = pairs_65_cast_fp16)[name = tensor("xr_65_cast_fp16")]; + tensor xi_65_begin_0 = const()[name = tensor("xi_65_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_65_end_0 = const()[name = tensor("xi_65_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_65_end_mask_0 = const()[name = tensor("xi_65_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_65_squeeze_mask_0 = const()[name = tensor("xi_65_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_65_cast_fp16 = slice_by_index(begin = xi_65_begin_0, end = xi_65_end_0, end_mask = xi_65_end_mask_0, squeeze_mask = xi_65_squeeze_mask_0, x = pairs_65_cast_fp16)[name = tensor("xi_65_cast_fp16")]; + tensor var_4190_cast_fp16 = mul(x = xr_65_cast_fp16, y = rotr_65_cast_fp16)[name = tensor("op_4190_cast_fp16")]; + tensor var_4191_cast_fp16 = mul(x = xi_65_cast_fp16, y = roti_65_cast_fp16)[name = tensor("op_4191_cast_fp16")]; + tensor xor__65_cast_fp16 = sub(x = var_4190_cast_fp16, y = var_4191_cast_fp16)[name = tensor("xor__65_cast_fp16")]; + tensor var_4194_cast_fp16 = mul(x = xr_65_cast_fp16, y = roti_65_cast_fp16)[name = tensor("op_4194_cast_fp16")]; + tensor var_4195_cast_fp16 = mul(x = xi_65_cast_fp16, y = rotr_65_cast_fp16)[name = tensor("op_4195_cast_fp16")]; + tensor xoi_65_cast_fp16 = add(x = var_4194_cast_fp16, y = var_4195_cast_fp16)[name = tensor("xoi_65_cast_fp16")]; + tensor out_65_axis_0 = const()[name = tensor("out_65_axis_0"), val = tensor(-1)]; + tensor out_65_cast_fp16 = stack(axis = out_65_axis_0, values = (xor__65_cast_fp16, xoi_65_cast_fp16))[name = tensor("out_65_cast_fp16")]; + tensor var_4218 = const()[name = tensor("op_4218"), val = tensor([1, 16, 32, 2])]; + tensor pairs_67_cast_fp16 = reshape(shape = var_4218, x = x_67_cast_fp16)[name = tensor("pairs_67_cast_fp16")]; + tensor xr_67_begin_0 = const()[name = tensor("xr_67_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_67_end_0 = const()[name = tensor("xr_67_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_67_end_mask_0 = const()[name = tensor("xr_67_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_67_squeeze_mask_0 = const()[name = tensor("xr_67_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_67_cast_fp16 = slice_by_index(begin = xr_67_begin_0, end = xr_67_end_0, end_mask = xr_67_end_mask_0, squeeze_mask = xr_67_squeeze_mask_0, x = pairs_67_cast_fp16)[name = tensor("xr_67_cast_fp16")]; + tensor xi_67_begin_0 = const()[name = tensor("xi_67_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_67_end_0 = const()[name = tensor("xi_67_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_67_end_mask_0 = const()[name = tensor("xi_67_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_67_squeeze_mask_0 = const()[name = tensor("xi_67_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_67_cast_fp16 = slice_by_index(begin = xi_67_begin_0, end = xi_67_end_0, end_mask = xi_67_end_mask_0, squeeze_mask = xi_67_squeeze_mask_0, x = pairs_67_cast_fp16)[name = tensor("xi_67_cast_fp16")]; + tensor var_4226_cast_fp16 = mul(x = xr_67_cast_fp16, y = rotr_65_cast_fp16)[name = tensor("op_4226_cast_fp16")]; + tensor var_4227_cast_fp16 = mul(x = xi_67_cast_fp16, y = roti_65_cast_fp16)[name = tensor("op_4227_cast_fp16")]; + tensor xor__67_cast_fp16 = sub(x = var_4226_cast_fp16, y = var_4227_cast_fp16)[name = tensor("xor__67_cast_fp16")]; + tensor var_4230_cast_fp16 = mul(x = xr_67_cast_fp16, y = roti_65_cast_fp16)[name = tensor("op_4230_cast_fp16")]; + tensor var_4231_cast_fp16 = mul(x = xi_67_cast_fp16, y = rotr_65_cast_fp16)[name = tensor("op_4231_cast_fp16")]; + tensor xoi_67_cast_fp16 = add(x = var_4230_cast_fp16, y = var_4231_cast_fp16)[name = tensor("xoi_67_cast_fp16")]; + tensor out_67_axis_0 = const()[name = tensor("out_67_axis_0"), val = tensor(-1)]; + tensor out_67_cast_fp16 = stack(axis = out_67_axis_0, values = (xor__67_cast_fp16, xoi_67_cast_fp16))[name = tensor("out_67_cast_fp16")]; + tensor _inversed_4243_y_0_to_fp16 = const()[name = tensor("_inversed_4243_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_4243_cast_fp16 = mul(x = position16_to_fp16, y = _inversed_4243_y_0_to_fp16)[name = tensor("_inversed_4243_cast_fp16")]; + tensor var_4244_cast_fp16 = floor(x = _inversed_4243_cast_fp16)[name = tensor("op_4244_cast_fp16")]; + tensor var_4245_to_fp16 = const()[name = tensor("op_4245_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4246_cast_fp16 = mul(x = var_4244_cast_fp16, y = var_4245_to_fp16)[name = tensor("op_4246_cast_fp16")]; + tensor write_idx_33_cast_fp16 = sub(x = position16_to_fp16, y = var_4246_cast_fp16)[name = tensor("write_idx_33_cast_fp16")]; + tensor var_4249_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_33_cast_fp16)[name = tensor("op_4249_cast_fp16")]; + tensor var_4259 = const()[name = tensor("op_4259"), val = tensor([1, 512, 1, 1])]; + tensor onehot_65_to_fp16_dtype_0 = const()[name = tensor("onehot_65_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_4249_cast_fp16_to_fp16 = cast(dtype = onehot_65_to_fp16_dtype_0, x = var_4249_cast_fp16)[name = tensor("cast_448")]; + tensor onehot_67_cast_fp16 = reshape(shape = var_4259, x = var_4249_cast_fp16_to_fp16)[name = tensor("onehot_67_cast_fp16")]; + tensor var_4261_to_fp16 = const()[name = tensor("op_4261_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4263_cast_fp16 = sub(x = var_4261_to_fp16, y = onehot_67_cast_fp16)[name = tensor("op_4263_cast_fp16")]; + tensor k_cache16_to_fp16_dtype_0 = const()[name = tensor("k_cache16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache16_to_fp16 = cast(dtype = k_cache16_to_fp16_dtype_0, x = k_cache16)[name = tensor("cast_447")]; + tensor var_4264_cast_fp16 = mul(x = k_cache16_to_fp16, y = var_4263_cast_fp16)[name = tensor("op_4264_cast_fp16")]; + tensor var_4269 = const()[name = tensor("op_4269"), val = tensor([1, 1, 16, 64])]; + tensor var_4270_cast_fp16 = reshape(shape = var_4269, x = out_67_cast_fp16)[name = tensor("op_4270_cast_fp16")]; + tensor var_4271_cast_fp16 = mul(x = var_4270_cast_fp16, y = onehot_67_cast_fp16)[name = tensor("op_4271_cast_fp16")]; + tensor new_k_cache_33_cast_fp16 = add(x = var_4264_cast_fp16, y = var_4271_cast_fp16)[name = tensor("new_k_cache_33_cast_fp16")]; + tensor new_k_cache_33_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_33_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache16_to_fp16_dtype_0 = const()[name = tensor("v_cache16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache16_to_fp16 = cast(dtype = v_cache16_to_fp16_dtype_0, x = v_cache16)[name = tensor("cast_445")]; + tensor var_4277_cast_fp16 = mul(x = v_cache16_to_fp16, y = var_4263_cast_fp16)[name = tensor("op_4277_cast_fp16")]; + tensor var_4282 = const()[name = tensor("op_4282"), val = tensor([1, 1, 16, 64])]; + tensor var_4283_cast_fp16 = reshape(shape = var_4282, x = v_33_cast_fp16)[name = tensor("op_4283_cast_fp16")]; + tensor var_4284_cast_fp16 = mul(x = var_4283_cast_fp16, y = onehot_67_cast_fp16)[name = tensor("op_4284_cast_fp16")]; + tensor new_v_cache_33_cast_fp16 = add(x = var_4277_cast_fp16, y = var_4284_cast_fp16)[name = tensor("new_v_cache_33_cast_fp16")]; + tensor new_v_cache_33_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_33_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_16_perm_0 = const()[name = tensor("transpose_16_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_4297 = const()[name = tensor("op_4297"), val = tensor([0, 2, 1, 3])]; + tensor var_4303 = const()[name = tensor("op_4303"), val = tensor([1, 16, 1, 64])]; + tensor q4_33_cast_fp16 = reshape(shape = var_4303, x = out_65_cast_fp16)[name = tensor("q4_33_cast_fp16")]; + tensor var_4308_transpose_x_0 = const()[name = tensor("op_4308_transpose_x_0"), val = tensor(false)]; + tensor var_4308_transpose_y_0 = const()[name = tensor("op_4308_transpose_y_0"), val = tensor(false)]; + tensor transpose_16_cast_fp16 = transpose(perm = transpose_16_perm_0, x = new_k_cache_33_cast_fp16)[name = tensor("transpose_47")]; + tensor var_4308_cast_fp16 = matmul(transpose_x = var_4308_transpose_x_0, transpose_y = var_4308_transpose_y_0, x = q4_33_cast_fp16, y = transpose_16_cast_fp16)[name = tensor("op_4308_cast_fp16")]; + tensor var_4309_to_fp16 = const()[name = tensor("op_4309_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_65_cast_fp16 = mul(x = var_4308_cast_fp16, y = var_4309_to_fp16)[name = tensor("scores_65_cast_fp16")]; + tensor var_4313 = const()[name = tensor("op_4313"), val = tensor([1, 1])]; + tensor var_4314_cast_fp16 = reshape(shape = var_4313, x = position16_to_fp16)[name = tensor("op_4314_cast_fp16")]; + tensor var_4315_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_4314_cast_fp16)[name = tensor("op_4315_cast_fp16")]; + tensor mask_33_to_fp16_dtype_0 = const()[name = tensor("mask_33_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_4322_to_fp16 = const()[name = tensor("op_4322_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4315_cast_fp16_to_fp16 = cast(dtype = mask_33_to_fp16_dtype_0, x = var_4315_cast_fp16)[name = tensor("cast_443")]; + tensor var_4323_cast_fp16 = sub(x = var_4315_cast_fp16_to_fp16, y = var_4322_to_fp16)[name = tensor("op_4323_cast_fp16")]; + tensor var_4328 = const()[name = tensor("op_4328"), val = tensor([1, 1, 1, 512])]; + tensor var_4329_cast_fp16 = reshape(shape = var_4328, x = var_4323_cast_fp16)[name = tensor("op_4329_cast_fp16")]; + tensor var_4330_to_fp16 = const()[name = tensor("op_4330_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_4331_cast_fp16 = mul(x = var_4329_cast_fp16, y = var_4330_to_fp16)[name = tensor("op_4331_cast_fp16")]; + tensor scores_67_cast_fp16 = add(x = scores_65_cast_fp16, y = var_4331_cast_fp16)[name = tensor("scores_67_cast_fp16")]; + tensor var_4334 = const()[name = tensor("op_4334"), val = tensor(-1)]; + tensor weights_33_cast_fp16 = softmax(axis = var_4334, x = scores_67_cast_fp16)[name = tensor("weights_33_cast_fp16")]; + tensor attn_33_transpose_x_0 = const()[name = tensor("attn_33_transpose_x_0"), val = tensor(false)]; + tensor attn_33_transpose_y_0 = const()[name = tensor("attn_33_transpose_y_0"), val = tensor(false)]; + tensor values_33_cast_fp16 = transpose(perm = var_4297, x = new_v_cache_33_cast_fp16)[name = tensor("transpose_46")]; + tensor attn_33_cast_fp16 = matmul(transpose_x = attn_33_transpose_x_0, transpose_y = attn_33_transpose_y_0, x = weights_33_cast_fp16, y = values_33_cast_fp16)[name = tensor("attn_33_cast_fp16")]; + tensor var_4342 = const()[name = tensor("op_4342"), val = tensor([0, 2, 1, 3])]; + tensor var_4347 = const()[name = tensor("op_4347"), val = tensor([1, 1, 1024])]; + tensor var_4343_cast_fp16 = transpose(perm = var_4342, x = attn_33_cast_fp16)[name = tensor("transpose_45")]; + tensor input_199_cast_fp16 = reshape(shape = var_4347, x = var_4343_cast_fp16)[name = tensor("input_199_cast_fp16")]; + tensor attn16_out_proj_weight_to_fp16 = const()[name = tensor("attn16_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(409171648)))]; + tensor linear_66_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn16_out_proj_weight_to_fp16, x = input_199_cast_fp16)[name = tensor("linear_66_cast_fp16")]; + tensor var_4353_to_fp16 = const()[name = tensor("op_4353_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4354_cast_fp16 = add(x = position16_to_fp16, y = var_4353_to_fp16)[name = tensor("op_4354_cast_fp16")]; + tensor var_4354_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4354_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_201_cast_fp16 = add(x = input_195_cast_fp16, y = linear_66_cast_fp16)[name = tensor("input_201_cast_fp16")]; + tensor input_203_axes_0 = const()[name = tensor("input_203_axes_0"), val = tensor([-1])]; + tensor norm16_2_weight_to_fp16 = const()[name = tensor("norm16_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411268864)))]; + tensor norm16_2_bias_to_fp16 = const()[name = tensor("norm16_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411270976)))]; + tensor var_4358_to_fp16 = const()[name = tensor("op_4358_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_203_cast_fp16 = layer_norm(axes = input_203_axes_0, beta = norm16_2_bias_to_fp16, epsilon = var_4358_to_fp16, gamma = norm16_2_weight_to_fp16, x = input_201_cast_fp16)[name = tensor("input_203_cast_fp16")]; + tensor linear16_1_weight_to_fp16 = const()[name = tensor("linear16_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411273088)))]; + tensor linear_67_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear16_1_weight_to_fp16, x = input_203_cast_fp16)[name = tensor("linear_67_cast_fp16")]; + tensor input_205_mode_0 = const()[name = tensor("input_205_mode_0"), val = tensor("EXACT")]; + tensor input_205_cast_fp16 = gelu(mode = input_205_mode_0, x = linear_67_cast_fp16)[name = tensor("input_205_cast_fp16")]; + tensor linear16_2_weight_to_fp16 = const()[name = tensor("linear16_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(419661760)))]; + tensor linear_68_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear16_2_weight_to_fp16, x = input_205_cast_fp16)[name = tensor("linear_68_cast_fp16")]; + tensor input_207_cast_fp16 = add(x = input_201_cast_fp16, y = linear_68_cast_fp16)[name = tensor("input_207_cast_fp16")]; + tensor input_209_axes_0 = const()[name = tensor("input_209_axes_0"), val = tensor([-1])]; + tensor norm17_1_weight_to_fp16 = const()[name = tensor("norm17_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(428050432)))]; + tensor norm17_1_bias_to_fp16 = const()[name = tensor("norm17_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(428052544)))]; + tensor var_4375_to_fp16 = const()[name = tensor("op_4375_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_209_cast_fp16 = layer_norm(axes = input_209_axes_0, beta = norm17_1_bias_to_fp16, epsilon = var_4375_to_fp16, gamma = norm17_1_weight_to_fp16, x = input_207_cast_fp16)[name = tensor("input_209_cast_fp16")]; + tensor attn17_in_proj_weight_to_fp16 = const()[name = tensor("attn17_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(428054656)))]; + tensor linear_69_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn17_in_proj_weight_to_fp16, x = input_209_cast_fp16)[name = tensor("linear_69_cast_fp16")]; + tensor var_4388 = const()[name = tensor("op_4388"), val = tensor([1, 3, 16, 64])]; + tensor qkv_35_cast_fp16 = reshape(shape = var_4388, x = linear_69_cast_fp16)[name = tensor("qkv_35_cast_fp16")]; + tensor x_69_begin_0 = const()[name = tensor("x_69_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_69_end_0 = const()[name = tensor("x_69_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_69_end_mask_0 = const()[name = tensor("x_69_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_69_squeeze_mask_0 = const()[name = tensor("x_69_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_69_cast_fp16 = slice_by_index(begin = x_69_begin_0, end = x_69_end_0, end_mask = x_69_end_mask_0, squeeze_mask = x_69_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("x_69_cast_fp16")]; + tensor x_71_begin_0 = const()[name = tensor("x_71_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_71_end_0 = const()[name = tensor("x_71_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_71_end_mask_0 = const()[name = tensor("x_71_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_71_squeeze_mask_0 = const()[name = tensor("x_71_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_71_cast_fp16 = slice_by_index(begin = x_71_begin_0, end = x_71_end_0, end_mask = x_71_end_mask_0, squeeze_mask = x_71_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("x_71_cast_fp16")]; + tensor v_35_begin_0 = const()[name = tensor("v_35_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_35_end_0 = const()[name = tensor("v_35_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_35_end_mask_0 = const()[name = tensor("v_35_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_35_squeeze_mask_0 = const()[name = tensor("v_35_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_35_cast_fp16 = slice_by_index(begin = v_35_begin_0, end = v_35_end_0, end_mask = v_35_end_mask_0, squeeze_mask = v_35_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("v_35_cast_fp16")]; + tensor var_4417 = const()[name = tensor("op_4417"), val = tensor([1, 1, 1])]; + tensor position17_to_fp16_dtype_0 = const()[name = tensor("position17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position17_to_fp16 = cast(dtype = position17_to_fp16_dtype_0, x = position17)[name = tensor("cast_441")]; + tensor var_4418_cast_fp16 = reshape(shape = var_4417, x = position17_to_fp16)[name = tensor("op_4418_cast_fp16")]; + tensor angles_69_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_4418_cast_fp16)[name = tensor("angles_69_cast_fp16")]; + tensor rotr_69_cast_fp16 = cos(x = angles_69_cast_fp16)[name = tensor("rotr_69_cast_fp16")]; + tensor roti_69_cast_fp16 = sin(x = angles_69_cast_fp16)[name = tensor("roti_69_cast_fp16")]; + tensor var_4426 = const()[name = tensor("op_4426"), val = tensor([1, 16, 32, 2])]; + tensor pairs_69_cast_fp16 = reshape(shape = var_4426, x = x_69_cast_fp16)[name = tensor("pairs_69_cast_fp16")]; + tensor xr_69_begin_0 = const()[name = tensor("xr_69_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_69_end_0 = const()[name = tensor("xr_69_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_69_end_mask_0 = const()[name = tensor("xr_69_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_69_squeeze_mask_0 = const()[name = tensor("xr_69_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_69_cast_fp16 = slice_by_index(begin = xr_69_begin_0, end = xr_69_end_0, end_mask = xr_69_end_mask_0, squeeze_mask = xr_69_squeeze_mask_0, x = pairs_69_cast_fp16)[name = tensor("xr_69_cast_fp16")]; + tensor xi_69_begin_0 = const()[name = tensor("xi_69_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_69_end_0 = const()[name = tensor("xi_69_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_69_end_mask_0 = const()[name = tensor("xi_69_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_69_squeeze_mask_0 = const()[name = tensor("xi_69_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_69_cast_fp16 = slice_by_index(begin = xi_69_begin_0, end = xi_69_end_0, end_mask = xi_69_end_mask_0, squeeze_mask = xi_69_squeeze_mask_0, x = pairs_69_cast_fp16)[name = tensor("xi_69_cast_fp16")]; + tensor var_4434_cast_fp16 = mul(x = xr_69_cast_fp16, y = rotr_69_cast_fp16)[name = tensor("op_4434_cast_fp16")]; + tensor var_4435_cast_fp16 = mul(x = xi_69_cast_fp16, y = roti_69_cast_fp16)[name = tensor("op_4435_cast_fp16")]; + tensor xor__69_cast_fp16 = sub(x = var_4434_cast_fp16, y = var_4435_cast_fp16)[name = tensor("xor__69_cast_fp16")]; + tensor var_4438_cast_fp16 = mul(x = xr_69_cast_fp16, y = roti_69_cast_fp16)[name = tensor("op_4438_cast_fp16")]; + tensor var_4439_cast_fp16 = mul(x = xi_69_cast_fp16, y = rotr_69_cast_fp16)[name = tensor("op_4439_cast_fp16")]; + tensor xoi_69_cast_fp16 = add(x = var_4438_cast_fp16, y = var_4439_cast_fp16)[name = tensor("xoi_69_cast_fp16")]; + tensor out_69_axis_0 = const()[name = tensor("out_69_axis_0"), val = tensor(-1)]; + tensor out_69_cast_fp16 = stack(axis = out_69_axis_0, values = (xor__69_cast_fp16, xoi_69_cast_fp16))[name = tensor("out_69_cast_fp16")]; + tensor var_4462 = const()[name = tensor("op_4462"), val = tensor([1, 16, 32, 2])]; + tensor pairs_71_cast_fp16 = reshape(shape = var_4462, x = x_71_cast_fp16)[name = tensor("pairs_71_cast_fp16")]; + tensor xr_71_begin_0 = const()[name = tensor("xr_71_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_71_end_0 = const()[name = tensor("xr_71_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_71_end_mask_0 = const()[name = tensor("xr_71_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_71_squeeze_mask_0 = const()[name = tensor("xr_71_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_71_cast_fp16 = slice_by_index(begin = xr_71_begin_0, end = xr_71_end_0, end_mask = xr_71_end_mask_0, squeeze_mask = xr_71_squeeze_mask_0, x = pairs_71_cast_fp16)[name = tensor("xr_71_cast_fp16")]; + tensor xi_71_begin_0 = const()[name = tensor("xi_71_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_71_end_0 = const()[name = tensor("xi_71_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_71_end_mask_0 = const()[name = tensor("xi_71_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_71_squeeze_mask_0 = const()[name = tensor("xi_71_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_71_cast_fp16 = slice_by_index(begin = xi_71_begin_0, end = xi_71_end_0, end_mask = xi_71_end_mask_0, squeeze_mask = xi_71_squeeze_mask_0, x = pairs_71_cast_fp16)[name = tensor("xi_71_cast_fp16")]; + tensor var_4470_cast_fp16 = mul(x = xr_71_cast_fp16, y = rotr_69_cast_fp16)[name = tensor("op_4470_cast_fp16")]; + tensor var_4471_cast_fp16 = mul(x = xi_71_cast_fp16, y = roti_69_cast_fp16)[name = tensor("op_4471_cast_fp16")]; + tensor xor__71_cast_fp16 = sub(x = var_4470_cast_fp16, y = var_4471_cast_fp16)[name = tensor("xor__71_cast_fp16")]; + tensor var_4474_cast_fp16 = mul(x = xr_71_cast_fp16, y = roti_69_cast_fp16)[name = tensor("op_4474_cast_fp16")]; + tensor var_4475_cast_fp16 = mul(x = xi_71_cast_fp16, y = rotr_69_cast_fp16)[name = tensor("op_4475_cast_fp16")]; + tensor xoi_71_cast_fp16 = add(x = var_4474_cast_fp16, y = var_4475_cast_fp16)[name = tensor("xoi_71_cast_fp16")]; + tensor out_71_axis_0 = const()[name = tensor("out_71_axis_0"), val = tensor(-1)]; + tensor out_71_cast_fp16 = stack(axis = out_71_axis_0, values = (xor__71_cast_fp16, xoi_71_cast_fp16))[name = tensor("out_71_cast_fp16")]; + tensor _inversed_4487_y_0_to_fp16 = const()[name = tensor("_inversed_4487_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_4487_cast_fp16 = mul(x = position17_to_fp16, y = _inversed_4487_y_0_to_fp16)[name = tensor("_inversed_4487_cast_fp16")]; + tensor var_4488_cast_fp16 = floor(x = _inversed_4487_cast_fp16)[name = tensor("op_4488_cast_fp16")]; + tensor var_4489_to_fp16 = const()[name = tensor("op_4489_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4490_cast_fp16 = mul(x = var_4488_cast_fp16, y = var_4489_to_fp16)[name = tensor("op_4490_cast_fp16")]; + tensor write_idx_35_cast_fp16 = sub(x = position17_to_fp16, y = var_4490_cast_fp16)[name = tensor("write_idx_35_cast_fp16")]; + tensor var_4493_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_35_cast_fp16)[name = tensor("op_4493_cast_fp16")]; + tensor var_4503 = const()[name = tensor("op_4503"), val = tensor([1, 512, 1, 1])]; + tensor onehot_69_to_fp16_dtype_0 = const()[name = tensor("onehot_69_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_4493_cast_fp16_to_fp16 = cast(dtype = onehot_69_to_fp16_dtype_0, x = var_4493_cast_fp16)[name = tensor("cast_440")]; + tensor onehot_71_cast_fp16 = reshape(shape = var_4503, x = var_4493_cast_fp16_to_fp16)[name = tensor("onehot_71_cast_fp16")]; + tensor var_4505_to_fp16 = const()[name = tensor("op_4505_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4507_cast_fp16 = sub(x = var_4505_to_fp16, y = onehot_71_cast_fp16)[name = tensor("op_4507_cast_fp16")]; + tensor k_cache17_to_fp16_dtype_0 = const()[name = tensor("k_cache17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache17_to_fp16 = cast(dtype = k_cache17_to_fp16_dtype_0, x = k_cache17)[name = tensor("cast_439")]; + tensor var_4508_cast_fp16 = mul(x = k_cache17_to_fp16, y = var_4507_cast_fp16)[name = tensor("op_4508_cast_fp16")]; + tensor var_4513 = const()[name = tensor("op_4513"), val = tensor([1, 1, 16, 64])]; + tensor var_4514_cast_fp16 = reshape(shape = var_4513, x = out_71_cast_fp16)[name = tensor("op_4514_cast_fp16")]; + tensor var_4515_cast_fp16 = mul(x = var_4514_cast_fp16, y = onehot_71_cast_fp16)[name = tensor("op_4515_cast_fp16")]; + tensor new_k_cache_35_cast_fp16 = add(x = var_4508_cast_fp16, y = var_4515_cast_fp16)[name = tensor("new_k_cache_35_cast_fp16")]; + tensor new_k_cache_35_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_35_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache17_to_fp16_dtype_0 = const()[name = tensor("v_cache17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache17_to_fp16 = cast(dtype = v_cache17_to_fp16_dtype_0, x = v_cache17)[name = tensor("cast_437")]; + tensor var_4521_cast_fp16 = mul(x = v_cache17_to_fp16, y = var_4507_cast_fp16)[name = tensor("op_4521_cast_fp16")]; + tensor var_4526 = const()[name = tensor("op_4526"), val = tensor([1, 1, 16, 64])]; + tensor var_4527_cast_fp16 = reshape(shape = var_4526, x = v_35_cast_fp16)[name = tensor("op_4527_cast_fp16")]; + tensor var_4528_cast_fp16 = mul(x = var_4527_cast_fp16, y = onehot_71_cast_fp16)[name = tensor("op_4528_cast_fp16")]; + tensor new_v_cache_35_cast_fp16 = add(x = var_4521_cast_fp16, y = var_4528_cast_fp16)[name = tensor("new_v_cache_35_cast_fp16")]; + tensor new_v_cache_35_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_35_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_17_perm_0 = const()[name = tensor("transpose_17_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_4541 = const()[name = tensor("op_4541"), val = tensor([0, 2, 1, 3])]; + tensor var_4547 = const()[name = tensor("op_4547"), val = tensor([1, 16, 1, 64])]; + tensor q4_35_cast_fp16 = reshape(shape = var_4547, x = out_69_cast_fp16)[name = tensor("q4_35_cast_fp16")]; + tensor var_4552_transpose_x_0 = const()[name = tensor("op_4552_transpose_x_0"), val = tensor(false)]; + tensor var_4552_transpose_y_0 = const()[name = tensor("op_4552_transpose_y_0"), val = tensor(false)]; + tensor transpose_17_cast_fp16 = transpose(perm = transpose_17_perm_0, x = new_k_cache_35_cast_fp16)[name = tensor("transpose_44")]; + tensor var_4552_cast_fp16 = matmul(transpose_x = var_4552_transpose_x_0, transpose_y = var_4552_transpose_y_0, x = q4_35_cast_fp16, y = transpose_17_cast_fp16)[name = tensor("op_4552_cast_fp16")]; + tensor var_4553_to_fp16 = const()[name = tensor("op_4553_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_69_cast_fp16 = mul(x = var_4552_cast_fp16, y = var_4553_to_fp16)[name = tensor("scores_69_cast_fp16")]; + tensor var_4557 = const()[name = tensor("op_4557"), val = tensor([1, 1])]; + tensor var_4558_cast_fp16 = reshape(shape = var_4557, x = position17_to_fp16)[name = tensor("op_4558_cast_fp16")]; + tensor var_4559_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_4558_cast_fp16)[name = tensor("op_4559_cast_fp16")]; + tensor mask_35_to_fp16_dtype_0 = const()[name = tensor("mask_35_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_4566_to_fp16 = const()[name = tensor("op_4566_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4559_cast_fp16_to_fp16 = cast(dtype = mask_35_to_fp16_dtype_0, x = var_4559_cast_fp16)[name = tensor("cast_435")]; + tensor var_4567_cast_fp16 = sub(x = var_4559_cast_fp16_to_fp16, y = var_4566_to_fp16)[name = tensor("op_4567_cast_fp16")]; + tensor var_4572 = const()[name = tensor("op_4572"), val = tensor([1, 1, 1, 512])]; + tensor var_4573_cast_fp16 = reshape(shape = var_4572, x = var_4567_cast_fp16)[name = tensor("op_4573_cast_fp16")]; + tensor var_4574_to_fp16 = const()[name = tensor("op_4574_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_4575_cast_fp16 = mul(x = var_4573_cast_fp16, y = var_4574_to_fp16)[name = tensor("op_4575_cast_fp16")]; + tensor scores_71_cast_fp16 = add(x = scores_69_cast_fp16, y = var_4575_cast_fp16)[name = tensor("scores_71_cast_fp16")]; + tensor var_4578 = const()[name = tensor("op_4578"), val = tensor(-1)]; + tensor weights_35_cast_fp16 = softmax(axis = var_4578, x = scores_71_cast_fp16)[name = tensor("weights_35_cast_fp16")]; + tensor attn_35_transpose_x_0 = const()[name = tensor("attn_35_transpose_x_0"), val = tensor(false)]; + tensor attn_35_transpose_y_0 = const()[name = tensor("attn_35_transpose_y_0"), val = tensor(false)]; + tensor values_35_cast_fp16 = transpose(perm = var_4541, x = new_v_cache_35_cast_fp16)[name = tensor("transpose_43")]; + tensor attn_35_cast_fp16 = matmul(transpose_x = attn_35_transpose_x_0, transpose_y = attn_35_transpose_y_0, x = weights_35_cast_fp16, y = values_35_cast_fp16)[name = tensor("attn_35_cast_fp16")]; + tensor var_4586 = const()[name = tensor("op_4586"), val = tensor([0, 2, 1, 3])]; + tensor var_4591 = const()[name = tensor("op_4591"), val = tensor([1, 1, 1024])]; + tensor var_4587_cast_fp16 = transpose(perm = var_4586, x = attn_35_cast_fp16)[name = tensor("transpose_42")]; + tensor input_211_cast_fp16 = reshape(shape = var_4591, x = var_4587_cast_fp16)[name = tensor("input_211_cast_fp16")]; + tensor attn17_out_proj_weight_to_fp16 = const()[name = tensor("attn17_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(434346176)))]; + tensor linear_70_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn17_out_proj_weight_to_fp16, x = input_211_cast_fp16)[name = tensor("linear_70_cast_fp16")]; + tensor var_4597_to_fp16 = const()[name = tensor("op_4597_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4598_cast_fp16 = add(x = position17_to_fp16, y = var_4597_to_fp16)[name = tensor("op_4598_cast_fp16")]; + tensor var_4598_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4598_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_213_cast_fp16 = add(x = input_207_cast_fp16, y = linear_70_cast_fp16)[name = tensor("input_213_cast_fp16")]; + tensor input_215_axes_0 = const()[name = tensor("input_215_axes_0"), val = tensor([-1])]; + tensor norm17_2_weight_to_fp16 = const()[name = tensor("norm17_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(436443392)))]; + tensor norm17_2_bias_to_fp16 = const()[name = tensor("norm17_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(436445504)))]; + tensor var_4602_to_fp16 = const()[name = tensor("op_4602_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_215_cast_fp16 = layer_norm(axes = input_215_axes_0, beta = norm17_2_bias_to_fp16, epsilon = var_4602_to_fp16, gamma = norm17_2_weight_to_fp16, x = input_213_cast_fp16)[name = tensor("input_215_cast_fp16")]; + tensor linear17_1_weight_to_fp16 = const()[name = tensor("linear17_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(436447616)))]; + tensor linear_71_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear17_1_weight_to_fp16, x = input_215_cast_fp16)[name = tensor("linear_71_cast_fp16")]; + tensor input_217_mode_0 = const()[name = tensor("input_217_mode_0"), val = tensor("EXACT")]; + tensor input_217_cast_fp16 = gelu(mode = input_217_mode_0, x = linear_71_cast_fp16)[name = tensor("input_217_cast_fp16")]; + tensor linear17_2_weight_to_fp16 = const()[name = tensor("linear17_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(444836288)))]; + tensor linear_72_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear17_2_weight_to_fp16, x = input_217_cast_fp16)[name = tensor("linear_72_cast_fp16")]; + tensor input_219_cast_fp16 = add(x = input_213_cast_fp16, y = linear_72_cast_fp16)[name = tensor("input_219_cast_fp16")]; + tensor input_221_axes_0 = const()[name = tensor("input_221_axes_0"), val = tensor([-1])]; + tensor norm18_1_weight_to_fp16 = const()[name = tensor("norm18_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(453224960)))]; + tensor norm18_1_bias_to_fp16 = const()[name = tensor("norm18_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(453227072)))]; + tensor var_4619_to_fp16 = const()[name = tensor("op_4619_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_221_cast_fp16 = layer_norm(axes = input_221_axes_0, beta = norm18_1_bias_to_fp16, epsilon = var_4619_to_fp16, gamma = norm18_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("input_221_cast_fp16")]; + tensor attn18_in_proj_weight_to_fp16 = const()[name = tensor("attn18_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(453229184)))]; + tensor linear_73_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn18_in_proj_weight_to_fp16, x = input_221_cast_fp16)[name = tensor("linear_73_cast_fp16")]; + tensor var_4632 = const()[name = tensor("op_4632"), val = tensor([1, 3, 16, 64])]; + tensor qkv_37_cast_fp16 = reshape(shape = var_4632, x = linear_73_cast_fp16)[name = tensor("qkv_37_cast_fp16")]; + tensor x_73_begin_0 = const()[name = tensor("x_73_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_73_end_0 = const()[name = tensor("x_73_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_73_end_mask_0 = const()[name = tensor("x_73_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_73_squeeze_mask_0 = const()[name = tensor("x_73_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_73_cast_fp16 = slice_by_index(begin = x_73_begin_0, end = x_73_end_0, end_mask = x_73_end_mask_0, squeeze_mask = x_73_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("x_73_cast_fp16")]; + tensor x_75_begin_0 = const()[name = tensor("x_75_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_75_end_0 = const()[name = tensor("x_75_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_75_end_mask_0 = const()[name = tensor("x_75_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_75_squeeze_mask_0 = const()[name = tensor("x_75_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_75_cast_fp16 = slice_by_index(begin = x_75_begin_0, end = x_75_end_0, end_mask = x_75_end_mask_0, squeeze_mask = x_75_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("x_75_cast_fp16")]; + tensor v_37_begin_0 = const()[name = tensor("v_37_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_37_end_0 = const()[name = tensor("v_37_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_37_end_mask_0 = const()[name = tensor("v_37_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_37_squeeze_mask_0 = const()[name = tensor("v_37_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_37_cast_fp16 = slice_by_index(begin = v_37_begin_0, end = v_37_end_0, end_mask = v_37_end_mask_0, squeeze_mask = v_37_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("v_37_cast_fp16")]; + tensor var_4661 = const()[name = tensor("op_4661"), val = tensor([1, 1, 1])]; + tensor position18_to_fp16_dtype_0 = const()[name = tensor("position18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position18_to_fp16 = cast(dtype = position18_to_fp16_dtype_0, x = position18)[name = tensor("cast_433")]; + tensor var_4662_cast_fp16 = reshape(shape = var_4661, x = position18_to_fp16)[name = tensor("op_4662_cast_fp16")]; + tensor angles_73_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_4662_cast_fp16)[name = tensor("angles_73_cast_fp16")]; + tensor rotr_73_cast_fp16 = cos(x = angles_73_cast_fp16)[name = tensor("rotr_73_cast_fp16")]; + tensor roti_73_cast_fp16 = sin(x = angles_73_cast_fp16)[name = tensor("roti_73_cast_fp16")]; + tensor var_4670 = const()[name = tensor("op_4670"), val = tensor([1, 16, 32, 2])]; + tensor pairs_73_cast_fp16 = reshape(shape = var_4670, x = x_73_cast_fp16)[name = tensor("pairs_73_cast_fp16")]; + tensor xr_73_begin_0 = const()[name = tensor("xr_73_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_73_end_0 = const()[name = tensor("xr_73_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_73_end_mask_0 = const()[name = tensor("xr_73_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_73_squeeze_mask_0 = const()[name = tensor("xr_73_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_73_cast_fp16 = slice_by_index(begin = xr_73_begin_0, end = xr_73_end_0, end_mask = xr_73_end_mask_0, squeeze_mask = xr_73_squeeze_mask_0, x = pairs_73_cast_fp16)[name = tensor("xr_73_cast_fp16")]; + tensor xi_73_begin_0 = const()[name = tensor("xi_73_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_73_end_0 = const()[name = tensor("xi_73_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_73_end_mask_0 = const()[name = tensor("xi_73_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_73_squeeze_mask_0 = const()[name = tensor("xi_73_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_73_cast_fp16 = slice_by_index(begin = xi_73_begin_0, end = xi_73_end_0, end_mask = xi_73_end_mask_0, squeeze_mask = xi_73_squeeze_mask_0, x = pairs_73_cast_fp16)[name = tensor("xi_73_cast_fp16")]; + tensor var_4678_cast_fp16 = mul(x = xr_73_cast_fp16, y = rotr_73_cast_fp16)[name = tensor("op_4678_cast_fp16")]; + tensor var_4679_cast_fp16 = mul(x = xi_73_cast_fp16, y = roti_73_cast_fp16)[name = tensor("op_4679_cast_fp16")]; + tensor xor__73_cast_fp16 = sub(x = var_4678_cast_fp16, y = var_4679_cast_fp16)[name = tensor("xor__73_cast_fp16")]; + tensor var_4682_cast_fp16 = mul(x = xr_73_cast_fp16, y = roti_73_cast_fp16)[name = tensor("op_4682_cast_fp16")]; + tensor var_4683_cast_fp16 = mul(x = xi_73_cast_fp16, y = rotr_73_cast_fp16)[name = tensor("op_4683_cast_fp16")]; + tensor xoi_73_cast_fp16 = add(x = var_4682_cast_fp16, y = var_4683_cast_fp16)[name = tensor("xoi_73_cast_fp16")]; + tensor out_73_axis_0 = const()[name = tensor("out_73_axis_0"), val = tensor(-1)]; + tensor out_73_cast_fp16 = stack(axis = out_73_axis_0, values = (xor__73_cast_fp16, xoi_73_cast_fp16))[name = tensor("out_73_cast_fp16")]; + tensor var_4706 = const()[name = tensor("op_4706"), val = tensor([1, 16, 32, 2])]; + tensor pairs_75_cast_fp16 = reshape(shape = var_4706, x = x_75_cast_fp16)[name = tensor("pairs_75_cast_fp16")]; + tensor xr_75_begin_0 = const()[name = tensor("xr_75_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_75_end_0 = const()[name = tensor("xr_75_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_75_end_mask_0 = const()[name = tensor("xr_75_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_75_squeeze_mask_0 = const()[name = tensor("xr_75_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_75_cast_fp16 = slice_by_index(begin = xr_75_begin_0, end = xr_75_end_0, end_mask = xr_75_end_mask_0, squeeze_mask = xr_75_squeeze_mask_0, x = pairs_75_cast_fp16)[name = tensor("xr_75_cast_fp16")]; + tensor xi_75_begin_0 = const()[name = tensor("xi_75_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_75_end_0 = const()[name = tensor("xi_75_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_75_end_mask_0 = const()[name = tensor("xi_75_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_75_squeeze_mask_0 = const()[name = tensor("xi_75_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_75_cast_fp16 = slice_by_index(begin = xi_75_begin_0, end = xi_75_end_0, end_mask = xi_75_end_mask_0, squeeze_mask = xi_75_squeeze_mask_0, x = pairs_75_cast_fp16)[name = tensor("xi_75_cast_fp16")]; + tensor var_4714_cast_fp16 = mul(x = xr_75_cast_fp16, y = rotr_73_cast_fp16)[name = tensor("op_4714_cast_fp16")]; + tensor var_4715_cast_fp16 = mul(x = xi_75_cast_fp16, y = roti_73_cast_fp16)[name = tensor("op_4715_cast_fp16")]; + tensor xor__75_cast_fp16 = sub(x = var_4714_cast_fp16, y = var_4715_cast_fp16)[name = tensor("xor__75_cast_fp16")]; + tensor var_4718_cast_fp16 = mul(x = xr_75_cast_fp16, y = roti_73_cast_fp16)[name = tensor("op_4718_cast_fp16")]; + tensor var_4719_cast_fp16 = mul(x = xi_75_cast_fp16, y = rotr_73_cast_fp16)[name = tensor("op_4719_cast_fp16")]; + tensor xoi_75_cast_fp16 = add(x = var_4718_cast_fp16, y = var_4719_cast_fp16)[name = tensor("xoi_75_cast_fp16")]; + tensor out_75_axis_0 = const()[name = tensor("out_75_axis_0"), val = tensor(-1)]; + tensor out_75_cast_fp16 = stack(axis = out_75_axis_0, values = (xor__75_cast_fp16, xoi_75_cast_fp16))[name = tensor("out_75_cast_fp16")]; + tensor _inversed_4731_y_0_to_fp16 = const()[name = tensor("_inversed_4731_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_4731_cast_fp16 = mul(x = position18_to_fp16, y = _inversed_4731_y_0_to_fp16)[name = tensor("_inversed_4731_cast_fp16")]; + tensor var_4732_cast_fp16 = floor(x = _inversed_4731_cast_fp16)[name = tensor("op_4732_cast_fp16")]; + tensor var_4733_to_fp16 = const()[name = tensor("op_4733_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4734_cast_fp16 = mul(x = var_4732_cast_fp16, y = var_4733_to_fp16)[name = tensor("op_4734_cast_fp16")]; + tensor write_idx_37_cast_fp16 = sub(x = position18_to_fp16, y = var_4734_cast_fp16)[name = tensor("write_idx_37_cast_fp16")]; + tensor var_4737_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_37_cast_fp16)[name = tensor("op_4737_cast_fp16")]; + tensor var_4747 = const()[name = tensor("op_4747"), val = tensor([1, 512, 1, 1])]; + tensor onehot_73_to_fp16_dtype_0 = const()[name = tensor("onehot_73_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_4737_cast_fp16_to_fp16 = cast(dtype = onehot_73_to_fp16_dtype_0, x = var_4737_cast_fp16)[name = tensor("cast_432")]; + tensor onehot_75_cast_fp16 = reshape(shape = var_4747, x = var_4737_cast_fp16_to_fp16)[name = tensor("onehot_75_cast_fp16")]; + tensor var_4749_to_fp16 = const()[name = tensor("op_4749_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4751_cast_fp16 = sub(x = var_4749_to_fp16, y = onehot_75_cast_fp16)[name = tensor("op_4751_cast_fp16")]; + tensor k_cache18_to_fp16_dtype_0 = const()[name = tensor("k_cache18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache18_to_fp16 = cast(dtype = k_cache18_to_fp16_dtype_0, x = k_cache18)[name = tensor("cast_431")]; + tensor var_4752_cast_fp16 = mul(x = k_cache18_to_fp16, y = var_4751_cast_fp16)[name = tensor("op_4752_cast_fp16")]; + tensor var_4757 = const()[name = tensor("op_4757"), val = tensor([1, 1, 16, 64])]; + tensor var_4758_cast_fp16 = reshape(shape = var_4757, x = out_75_cast_fp16)[name = tensor("op_4758_cast_fp16")]; + tensor var_4759_cast_fp16 = mul(x = var_4758_cast_fp16, y = onehot_75_cast_fp16)[name = tensor("op_4759_cast_fp16")]; + tensor new_k_cache_37_cast_fp16 = add(x = var_4752_cast_fp16, y = var_4759_cast_fp16)[name = tensor("new_k_cache_37_cast_fp16")]; + tensor new_k_cache_37_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_37_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache18_to_fp16_dtype_0 = const()[name = tensor("v_cache18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache18_to_fp16 = cast(dtype = v_cache18_to_fp16_dtype_0, x = v_cache18)[name = tensor("cast_429")]; + tensor var_4765_cast_fp16 = mul(x = v_cache18_to_fp16, y = var_4751_cast_fp16)[name = tensor("op_4765_cast_fp16")]; + tensor var_4770 = const()[name = tensor("op_4770"), val = tensor([1, 1, 16, 64])]; + tensor var_4771_cast_fp16 = reshape(shape = var_4770, x = v_37_cast_fp16)[name = tensor("op_4771_cast_fp16")]; + tensor var_4772_cast_fp16 = mul(x = var_4771_cast_fp16, y = onehot_75_cast_fp16)[name = tensor("op_4772_cast_fp16")]; + tensor new_v_cache_37_cast_fp16 = add(x = var_4765_cast_fp16, y = var_4772_cast_fp16)[name = tensor("new_v_cache_37_cast_fp16")]; + tensor new_v_cache_37_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_37_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_18_perm_0 = const()[name = tensor("transpose_18_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_4785 = const()[name = tensor("op_4785"), val = tensor([0, 2, 1, 3])]; + tensor var_4791 = const()[name = tensor("op_4791"), val = tensor([1, 16, 1, 64])]; + tensor q4_37_cast_fp16 = reshape(shape = var_4791, x = out_73_cast_fp16)[name = tensor("q4_37_cast_fp16")]; + tensor var_4796_transpose_x_0 = const()[name = tensor("op_4796_transpose_x_0"), val = tensor(false)]; + tensor var_4796_transpose_y_0 = const()[name = tensor("op_4796_transpose_y_0"), val = tensor(false)]; + tensor transpose_18_cast_fp16 = transpose(perm = transpose_18_perm_0, x = new_k_cache_37_cast_fp16)[name = tensor("transpose_41")]; + tensor var_4796_cast_fp16 = matmul(transpose_x = var_4796_transpose_x_0, transpose_y = var_4796_transpose_y_0, x = q4_37_cast_fp16, y = transpose_18_cast_fp16)[name = tensor("op_4796_cast_fp16")]; + tensor var_4797_to_fp16 = const()[name = tensor("op_4797_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_73_cast_fp16 = mul(x = var_4796_cast_fp16, y = var_4797_to_fp16)[name = tensor("scores_73_cast_fp16")]; + tensor var_4801 = const()[name = tensor("op_4801"), val = tensor([1, 1])]; + tensor var_4802_cast_fp16 = reshape(shape = var_4801, x = position18_to_fp16)[name = tensor("op_4802_cast_fp16")]; + tensor var_4803_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_4802_cast_fp16)[name = tensor("op_4803_cast_fp16")]; + tensor mask_37_to_fp16_dtype_0 = const()[name = tensor("mask_37_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_4810_to_fp16 = const()[name = tensor("op_4810_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4803_cast_fp16_to_fp16 = cast(dtype = mask_37_to_fp16_dtype_0, x = var_4803_cast_fp16)[name = tensor("cast_427")]; + tensor var_4811_cast_fp16 = sub(x = var_4803_cast_fp16_to_fp16, y = var_4810_to_fp16)[name = tensor("op_4811_cast_fp16")]; + tensor var_4816 = const()[name = tensor("op_4816"), val = tensor([1, 1, 1, 512])]; + tensor var_4817_cast_fp16 = reshape(shape = var_4816, x = var_4811_cast_fp16)[name = tensor("op_4817_cast_fp16")]; + tensor var_4818_to_fp16 = const()[name = tensor("op_4818_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_4819_cast_fp16 = mul(x = var_4817_cast_fp16, y = var_4818_to_fp16)[name = tensor("op_4819_cast_fp16")]; + tensor scores_75_cast_fp16 = add(x = scores_73_cast_fp16, y = var_4819_cast_fp16)[name = tensor("scores_75_cast_fp16")]; + tensor var_4822 = const()[name = tensor("op_4822"), val = tensor(-1)]; + tensor weights_37_cast_fp16 = softmax(axis = var_4822, x = scores_75_cast_fp16)[name = tensor("weights_37_cast_fp16")]; + tensor attn_37_transpose_x_0 = const()[name = tensor("attn_37_transpose_x_0"), val = tensor(false)]; + tensor attn_37_transpose_y_0 = const()[name = tensor("attn_37_transpose_y_0"), val = tensor(false)]; + tensor values_37_cast_fp16 = transpose(perm = var_4785, x = new_v_cache_37_cast_fp16)[name = tensor("transpose_40")]; + tensor attn_37_cast_fp16 = matmul(transpose_x = attn_37_transpose_x_0, transpose_y = attn_37_transpose_y_0, x = weights_37_cast_fp16, y = values_37_cast_fp16)[name = tensor("attn_37_cast_fp16")]; + tensor var_4830 = const()[name = tensor("op_4830"), val = tensor([0, 2, 1, 3])]; + tensor var_4835 = const()[name = tensor("op_4835"), val = tensor([1, 1, 1024])]; + tensor var_4831_cast_fp16 = transpose(perm = var_4830, x = attn_37_cast_fp16)[name = tensor("transpose_39")]; + tensor input_223_cast_fp16 = reshape(shape = var_4835, x = var_4831_cast_fp16)[name = tensor("input_223_cast_fp16")]; + tensor attn18_out_proj_weight_to_fp16 = const()[name = tensor("attn18_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(459520704)))]; + tensor linear_74_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn18_out_proj_weight_to_fp16, x = input_223_cast_fp16)[name = tensor("linear_74_cast_fp16")]; + tensor var_4841_to_fp16 = const()[name = tensor("op_4841_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4842_cast_fp16 = add(x = position18_to_fp16, y = var_4841_to_fp16)[name = tensor("op_4842_cast_fp16")]; + tensor var_4842_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4842_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_225_cast_fp16 = add(x = input_219_cast_fp16, y = linear_74_cast_fp16)[name = tensor("input_225_cast_fp16")]; + tensor input_227_axes_0 = const()[name = tensor("input_227_axes_0"), val = tensor([-1])]; + tensor norm18_2_weight_to_fp16 = const()[name = tensor("norm18_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(461617920)))]; + tensor norm18_2_bias_to_fp16 = const()[name = tensor("norm18_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(461620032)))]; + tensor var_4846_to_fp16 = const()[name = tensor("op_4846_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_227_cast_fp16 = layer_norm(axes = input_227_axes_0, beta = norm18_2_bias_to_fp16, epsilon = var_4846_to_fp16, gamma = norm18_2_weight_to_fp16, x = input_225_cast_fp16)[name = tensor("input_227_cast_fp16")]; + tensor linear18_1_weight_to_fp16 = const()[name = tensor("linear18_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(461622144)))]; + tensor linear_75_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear18_1_weight_to_fp16, x = input_227_cast_fp16)[name = tensor("linear_75_cast_fp16")]; + tensor input_229_mode_0 = const()[name = tensor("input_229_mode_0"), val = tensor("EXACT")]; + tensor input_229_cast_fp16 = gelu(mode = input_229_mode_0, x = linear_75_cast_fp16)[name = tensor("input_229_cast_fp16")]; + tensor linear18_2_weight_to_fp16 = const()[name = tensor("linear18_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(470010816)))]; + tensor linear_76_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear18_2_weight_to_fp16, x = input_229_cast_fp16)[name = tensor("linear_76_cast_fp16")]; + tensor input_231_cast_fp16 = add(x = input_225_cast_fp16, y = linear_76_cast_fp16)[name = tensor("input_231_cast_fp16")]; + tensor input_233_axes_0 = const()[name = tensor("input_233_axes_0"), val = tensor([-1])]; + tensor norm19_1_weight_to_fp16 = const()[name = tensor("norm19_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(478399488)))]; + tensor norm19_1_bias_to_fp16 = const()[name = tensor("norm19_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(478401600)))]; + tensor var_4863_to_fp16 = const()[name = tensor("op_4863_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_233_cast_fp16 = layer_norm(axes = input_233_axes_0, beta = norm19_1_bias_to_fp16, epsilon = var_4863_to_fp16, gamma = norm19_1_weight_to_fp16, x = input_231_cast_fp16)[name = tensor("input_233_cast_fp16")]; + tensor attn19_in_proj_weight_to_fp16 = const()[name = tensor("attn19_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(478403712)))]; + tensor linear_77_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn19_in_proj_weight_to_fp16, x = input_233_cast_fp16)[name = tensor("linear_77_cast_fp16")]; + tensor var_4876 = const()[name = tensor("op_4876"), val = tensor([1, 3, 16, 64])]; + tensor qkv_39_cast_fp16 = reshape(shape = var_4876, x = linear_77_cast_fp16)[name = tensor("qkv_39_cast_fp16")]; + tensor x_77_begin_0 = const()[name = tensor("x_77_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_77_end_0 = const()[name = tensor("x_77_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_77_end_mask_0 = const()[name = tensor("x_77_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_77_squeeze_mask_0 = const()[name = tensor("x_77_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_77_cast_fp16 = slice_by_index(begin = x_77_begin_0, end = x_77_end_0, end_mask = x_77_end_mask_0, squeeze_mask = x_77_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("x_77_cast_fp16")]; + tensor x_79_begin_0 = const()[name = tensor("x_79_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_79_end_0 = const()[name = tensor("x_79_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_79_end_mask_0 = const()[name = tensor("x_79_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_79_squeeze_mask_0 = const()[name = tensor("x_79_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_79_cast_fp16 = slice_by_index(begin = x_79_begin_0, end = x_79_end_0, end_mask = x_79_end_mask_0, squeeze_mask = x_79_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("x_79_cast_fp16")]; + tensor v_39_begin_0 = const()[name = tensor("v_39_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_39_end_0 = const()[name = tensor("v_39_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_39_end_mask_0 = const()[name = tensor("v_39_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_39_squeeze_mask_0 = const()[name = tensor("v_39_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_39_cast_fp16 = slice_by_index(begin = v_39_begin_0, end = v_39_end_0, end_mask = v_39_end_mask_0, squeeze_mask = v_39_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("v_39_cast_fp16")]; + tensor var_4905 = const()[name = tensor("op_4905"), val = tensor([1, 1, 1])]; + tensor position19_to_fp16_dtype_0 = const()[name = tensor("position19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position19_to_fp16 = cast(dtype = position19_to_fp16_dtype_0, x = position19)[name = tensor("cast_425")]; + tensor var_4906_cast_fp16 = reshape(shape = var_4905, x = position19_to_fp16)[name = tensor("op_4906_cast_fp16")]; + tensor angles_77_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_4906_cast_fp16)[name = tensor("angles_77_cast_fp16")]; + tensor rotr_77_cast_fp16 = cos(x = angles_77_cast_fp16)[name = tensor("rotr_77_cast_fp16")]; + tensor roti_77_cast_fp16 = sin(x = angles_77_cast_fp16)[name = tensor("roti_77_cast_fp16")]; + tensor var_4914 = const()[name = tensor("op_4914"), val = tensor([1, 16, 32, 2])]; + tensor pairs_77_cast_fp16 = reshape(shape = var_4914, x = x_77_cast_fp16)[name = tensor("pairs_77_cast_fp16")]; + tensor xr_77_begin_0 = const()[name = tensor("xr_77_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_77_end_0 = const()[name = tensor("xr_77_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_77_end_mask_0 = const()[name = tensor("xr_77_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_77_squeeze_mask_0 = const()[name = tensor("xr_77_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_77_cast_fp16 = slice_by_index(begin = xr_77_begin_0, end = xr_77_end_0, end_mask = xr_77_end_mask_0, squeeze_mask = xr_77_squeeze_mask_0, x = pairs_77_cast_fp16)[name = tensor("xr_77_cast_fp16")]; + tensor xi_77_begin_0 = const()[name = tensor("xi_77_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_77_end_0 = const()[name = tensor("xi_77_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_77_end_mask_0 = const()[name = tensor("xi_77_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_77_squeeze_mask_0 = const()[name = tensor("xi_77_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_77_cast_fp16 = slice_by_index(begin = xi_77_begin_0, end = xi_77_end_0, end_mask = xi_77_end_mask_0, squeeze_mask = xi_77_squeeze_mask_0, x = pairs_77_cast_fp16)[name = tensor("xi_77_cast_fp16")]; + tensor var_4922_cast_fp16 = mul(x = xr_77_cast_fp16, y = rotr_77_cast_fp16)[name = tensor("op_4922_cast_fp16")]; + tensor var_4923_cast_fp16 = mul(x = xi_77_cast_fp16, y = roti_77_cast_fp16)[name = tensor("op_4923_cast_fp16")]; + tensor xor__77_cast_fp16 = sub(x = var_4922_cast_fp16, y = var_4923_cast_fp16)[name = tensor("xor__77_cast_fp16")]; + tensor var_4926_cast_fp16 = mul(x = xr_77_cast_fp16, y = roti_77_cast_fp16)[name = tensor("op_4926_cast_fp16")]; + tensor var_4927_cast_fp16 = mul(x = xi_77_cast_fp16, y = rotr_77_cast_fp16)[name = tensor("op_4927_cast_fp16")]; + tensor xoi_77_cast_fp16 = add(x = var_4926_cast_fp16, y = var_4927_cast_fp16)[name = tensor("xoi_77_cast_fp16")]; + tensor out_77_axis_0 = const()[name = tensor("out_77_axis_0"), val = tensor(-1)]; + tensor out_77_cast_fp16 = stack(axis = out_77_axis_0, values = (xor__77_cast_fp16, xoi_77_cast_fp16))[name = tensor("out_77_cast_fp16")]; + tensor var_4950 = const()[name = tensor("op_4950"), val = tensor([1, 16, 32, 2])]; + tensor pairs_79_cast_fp16 = reshape(shape = var_4950, x = x_79_cast_fp16)[name = tensor("pairs_79_cast_fp16")]; + tensor xr_79_begin_0 = const()[name = tensor("xr_79_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_79_end_0 = const()[name = tensor("xr_79_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_79_end_mask_0 = const()[name = tensor("xr_79_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_79_squeeze_mask_0 = const()[name = tensor("xr_79_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_79_cast_fp16 = slice_by_index(begin = xr_79_begin_0, end = xr_79_end_0, end_mask = xr_79_end_mask_0, squeeze_mask = xr_79_squeeze_mask_0, x = pairs_79_cast_fp16)[name = tensor("xr_79_cast_fp16")]; + tensor xi_79_begin_0 = const()[name = tensor("xi_79_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_79_end_0 = const()[name = tensor("xi_79_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_79_end_mask_0 = const()[name = tensor("xi_79_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_79_squeeze_mask_0 = const()[name = tensor("xi_79_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_79_cast_fp16 = slice_by_index(begin = xi_79_begin_0, end = xi_79_end_0, end_mask = xi_79_end_mask_0, squeeze_mask = xi_79_squeeze_mask_0, x = pairs_79_cast_fp16)[name = tensor("xi_79_cast_fp16")]; + tensor var_4958_cast_fp16 = mul(x = xr_79_cast_fp16, y = rotr_77_cast_fp16)[name = tensor("op_4958_cast_fp16")]; + tensor var_4959_cast_fp16 = mul(x = xi_79_cast_fp16, y = roti_77_cast_fp16)[name = tensor("op_4959_cast_fp16")]; + tensor xor__79_cast_fp16 = sub(x = var_4958_cast_fp16, y = var_4959_cast_fp16)[name = tensor("xor__79_cast_fp16")]; + tensor var_4962_cast_fp16 = mul(x = xr_79_cast_fp16, y = roti_77_cast_fp16)[name = tensor("op_4962_cast_fp16")]; + tensor var_4963_cast_fp16 = mul(x = xi_79_cast_fp16, y = rotr_77_cast_fp16)[name = tensor("op_4963_cast_fp16")]; + tensor xoi_79_cast_fp16 = add(x = var_4962_cast_fp16, y = var_4963_cast_fp16)[name = tensor("xoi_79_cast_fp16")]; + tensor out_79_axis_0 = const()[name = tensor("out_79_axis_0"), val = tensor(-1)]; + tensor out_79_cast_fp16 = stack(axis = out_79_axis_0, values = (xor__79_cast_fp16, xoi_79_cast_fp16))[name = tensor("out_79_cast_fp16")]; + tensor _inversed_4975_y_0_to_fp16 = const()[name = tensor("_inversed_4975_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_4975_cast_fp16 = mul(x = position19_to_fp16, y = _inversed_4975_y_0_to_fp16)[name = tensor("_inversed_4975_cast_fp16")]; + tensor var_4976_cast_fp16 = floor(x = _inversed_4975_cast_fp16)[name = tensor("op_4976_cast_fp16")]; + tensor var_4977_to_fp16 = const()[name = tensor("op_4977_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4978_cast_fp16 = mul(x = var_4976_cast_fp16, y = var_4977_to_fp16)[name = tensor("op_4978_cast_fp16")]; + tensor write_idx_39_cast_fp16 = sub(x = position19_to_fp16, y = var_4978_cast_fp16)[name = tensor("write_idx_39_cast_fp16")]; + tensor var_4981_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_39_cast_fp16)[name = tensor("op_4981_cast_fp16")]; + tensor var_4991 = const()[name = tensor("op_4991"), val = tensor([1, 512, 1, 1])]; + tensor onehot_77_to_fp16_dtype_0 = const()[name = tensor("onehot_77_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_4981_cast_fp16_to_fp16 = cast(dtype = onehot_77_to_fp16_dtype_0, x = var_4981_cast_fp16)[name = tensor("cast_424")]; + tensor onehot_79_cast_fp16 = reshape(shape = var_4991, x = var_4981_cast_fp16_to_fp16)[name = tensor("onehot_79_cast_fp16")]; + tensor var_4993_to_fp16 = const()[name = tensor("op_4993_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4995_cast_fp16 = sub(x = var_4993_to_fp16, y = onehot_79_cast_fp16)[name = tensor("op_4995_cast_fp16")]; + tensor k_cache19_to_fp16_dtype_0 = const()[name = tensor("k_cache19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache19_to_fp16 = cast(dtype = k_cache19_to_fp16_dtype_0, x = k_cache19)[name = tensor("cast_423")]; + tensor var_4996_cast_fp16 = mul(x = k_cache19_to_fp16, y = var_4995_cast_fp16)[name = tensor("op_4996_cast_fp16")]; + tensor var_5001 = const()[name = tensor("op_5001"), val = tensor([1, 1, 16, 64])]; + tensor var_5002_cast_fp16 = reshape(shape = var_5001, x = out_79_cast_fp16)[name = tensor("op_5002_cast_fp16")]; + tensor var_5003_cast_fp16 = mul(x = var_5002_cast_fp16, y = onehot_79_cast_fp16)[name = tensor("op_5003_cast_fp16")]; + tensor new_k_cache_39_cast_fp16 = add(x = var_4996_cast_fp16, y = var_5003_cast_fp16)[name = tensor("new_k_cache_39_cast_fp16")]; + tensor new_k_cache_39_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_39_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache19_to_fp16_dtype_0 = const()[name = tensor("v_cache19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache19_to_fp16 = cast(dtype = v_cache19_to_fp16_dtype_0, x = v_cache19)[name = tensor("cast_421")]; + tensor var_5009_cast_fp16 = mul(x = v_cache19_to_fp16, y = var_4995_cast_fp16)[name = tensor("op_5009_cast_fp16")]; + tensor var_5014 = const()[name = tensor("op_5014"), val = tensor([1, 1, 16, 64])]; + tensor var_5015_cast_fp16 = reshape(shape = var_5014, x = v_39_cast_fp16)[name = tensor("op_5015_cast_fp16")]; + tensor var_5016_cast_fp16 = mul(x = var_5015_cast_fp16, y = onehot_79_cast_fp16)[name = tensor("op_5016_cast_fp16")]; + tensor new_v_cache_39_cast_fp16 = add(x = var_5009_cast_fp16, y = var_5016_cast_fp16)[name = tensor("new_v_cache_39_cast_fp16")]; + tensor new_v_cache_39_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_39_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_19_perm_0 = const()[name = tensor("transpose_19_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_5029 = const()[name = tensor("op_5029"), val = tensor([0, 2, 1, 3])]; + tensor var_5035 = const()[name = tensor("op_5035"), val = tensor([1, 16, 1, 64])]; + tensor q4_39_cast_fp16 = reshape(shape = var_5035, x = out_77_cast_fp16)[name = tensor("q4_39_cast_fp16")]; + tensor var_5040_transpose_x_0 = const()[name = tensor("op_5040_transpose_x_0"), val = tensor(false)]; + tensor var_5040_transpose_y_0 = const()[name = tensor("op_5040_transpose_y_0"), val = tensor(false)]; + tensor transpose_19_cast_fp16 = transpose(perm = transpose_19_perm_0, x = new_k_cache_39_cast_fp16)[name = tensor("transpose_38")]; + tensor var_5040_cast_fp16 = matmul(transpose_x = var_5040_transpose_x_0, transpose_y = var_5040_transpose_y_0, x = q4_39_cast_fp16, y = transpose_19_cast_fp16)[name = tensor("op_5040_cast_fp16")]; + tensor var_5041_to_fp16 = const()[name = tensor("op_5041_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_77_cast_fp16 = mul(x = var_5040_cast_fp16, y = var_5041_to_fp16)[name = tensor("scores_77_cast_fp16")]; + tensor var_5045 = const()[name = tensor("op_5045"), val = tensor([1, 1])]; + tensor var_5046_cast_fp16 = reshape(shape = var_5045, x = position19_to_fp16)[name = tensor("op_5046_cast_fp16")]; + tensor var_5047_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_5046_cast_fp16)[name = tensor("op_5047_cast_fp16")]; + tensor mask_39_to_fp16_dtype_0 = const()[name = tensor("mask_39_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_5054_to_fp16 = const()[name = tensor("op_5054_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5047_cast_fp16_to_fp16 = cast(dtype = mask_39_to_fp16_dtype_0, x = var_5047_cast_fp16)[name = tensor("cast_419")]; + tensor var_5055_cast_fp16 = sub(x = var_5047_cast_fp16_to_fp16, y = var_5054_to_fp16)[name = tensor("op_5055_cast_fp16")]; + tensor var_5060 = const()[name = tensor("op_5060"), val = tensor([1, 1, 1, 512])]; + tensor var_5061_cast_fp16 = reshape(shape = var_5060, x = var_5055_cast_fp16)[name = tensor("op_5061_cast_fp16")]; + tensor var_5062_to_fp16 = const()[name = tensor("op_5062_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_5063_cast_fp16 = mul(x = var_5061_cast_fp16, y = var_5062_to_fp16)[name = tensor("op_5063_cast_fp16")]; + tensor scores_79_cast_fp16 = add(x = scores_77_cast_fp16, y = var_5063_cast_fp16)[name = tensor("scores_79_cast_fp16")]; + tensor var_5066 = const()[name = tensor("op_5066"), val = tensor(-1)]; + tensor weights_39_cast_fp16 = softmax(axis = var_5066, x = scores_79_cast_fp16)[name = tensor("weights_39_cast_fp16")]; + tensor attn_39_transpose_x_0 = const()[name = tensor("attn_39_transpose_x_0"), val = tensor(false)]; + tensor attn_39_transpose_y_0 = const()[name = tensor("attn_39_transpose_y_0"), val = tensor(false)]; + tensor values_39_cast_fp16 = transpose(perm = var_5029, x = new_v_cache_39_cast_fp16)[name = tensor("transpose_37")]; + tensor attn_39_cast_fp16 = matmul(transpose_x = attn_39_transpose_x_0, transpose_y = attn_39_transpose_y_0, x = weights_39_cast_fp16, y = values_39_cast_fp16)[name = tensor("attn_39_cast_fp16")]; + tensor var_5074 = const()[name = tensor("op_5074"), val = tensor([0, 2, 1, 3])]; + tensor var_5079 = const()[name = tensor("op_5079"), val = tensor([1, 1, 1024])]; + tensor var_5075_cast_fp16 = transpose(perm = var_5074, x = attn_39_cast_fp16)[name = tensor("transpose_36")]; + tensor input_235_cast_fp16 = reshape(shape = var_5079, x = var_5075_cast_fp16)[name = tensor("input_235_cast_fp16")]; + tensor attn19_out_proj_weight_to_fp16 = const()[name = tensor("attn19_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(484695232)))]; + tensor linear_78_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn19_out_proj_weight_to_fp16, x = input_235_cast_fp16)[name = tensor("linear_78_cast_fp16")]; + tensor var_5085_to_fp16 = const()[name = tensor("op_5085_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5086_cast_fp16 = add(x = position19_to_fp16, y = var_5085_to_fp16)[name = tensor("op_5086_cast_fp16")]; + tensor var_5086_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5086_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_237_cast_fp16 = add(x = input_231_cast_fp16, y = linear_78_cast_fp16)[name = tensor("input_237_cast_fp16")]; + tensor input_239_axes_0 = const()[name = tensor("input_239_axes_0"), val = tensor([-1])]; + tensor norm19_2_weight_to_fp16 = const()[name = tensor("norm19_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486792448)))]; + tensor norm19_2_bias_to_fp16 = const()[name = tensor("norm19_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486794560)))]; + tensor var_5090_to_fp16 = const()[name = tensor("op_5090_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_239_cast_fp16 = layer_norm(axes = input_239_axes_0, beta = norm19_2_bias_to_fp16, epsilon = var_5090_to_fp16, gamma = norm19_2_weight_to_fp16, x = input_237_cast_fp16)[name = tensor("input_239_cast_fp16")]; + tensor linear19_1_weight_to_fp16 = const()[name = tensor("linear19_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486796672)))]; + tensor linear_79_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear19_1_weight_to_fp16, x = input_239_cast_fp16)[name = tensor("linear_79_cast_fp16")]; + tensor input_241_mode_0 = const()[name = tensor("input_241_mode_0"), val = tensor("EXACT")]; + tensor input_241_cast_fp16 = gelu(mode = input_241_mode_0, x = linear_79_cast_fp16)[name = tensor("input_241_cast_fp16")]; + tensor linear19_2_weight_to_fp16 = const()[name = tensor("linear19_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(495185344)))]; + tensor linear_80_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear19_2_weight_to_fp16, x = input_241_cast_fp16)[name = tensor("linear_80_cast_fp16")]; + tensor input_243_cast_fp16 = add(x = input_237_cast_fp16, y = linear_80_cast_fp16)[name = tensor("input_243_cast_fp16")]; + tensor input_245_axes_0 = const()[name = tensor("input_245_axes_0"), val = tensor([-1])]; + tensor norm20_1_weight_to_fp16 = const()[name = tensor("norm20_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(503574016)))]; + tensor norm20_1_bias_to_fp16 = const()[name = tensor("norm20_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(503576128)))]; + tensor var_5107_to_fp16 = const()[name = tensor("op_5107_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_245_cast_fp16 = layer_norm(axes = input_245_axes_0, beta = norm20_1_bias_to_fp16, epsilon = var_5107_to_fp16, gamma = norm20_1_weight_to_fp16, x = input_243_cast_fp16)[name = tensor("input_245_cast_fp16")]; + tensor attn20_in_proj_weight_to_fp16 = const()[name = tensor("attn20_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(503578240)))]; + tensor linear_81_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn20_in_proj_weight_to_fp16, x = input_245_cast_fp16)[name = tensor("linear_81_cast_fp16")]; + tensor var_5120 = const()[name = tensor("op_5120"), val = tensor([1, 3, 16, 64])]; + tensor qkv_41_cast_fp16 = reshape(shape = var_5120, x = linear_81_cast_fp16)[name = tensor("qkv_41_cast_fp16")]; + tensor x_81_begin_0 = const()[name = tensor("x_81_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_81_end_0 = const()[name = tensor("x_81_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_81_end_mask_0 = const()[name = tensor("x_81_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_81_squeeze_mask_0 = const()[name = tensor("x_81_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_81_cast_fp16 = slice_by_index(begin = x_81_begin_0, end = x_81_end_0, end_mask = x_81_end_mask_0, squeeze_mask = x_81_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("x_81_cast_fp16")]; + tensor x_83_begin_0 = const()[name = tensor("x_83_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_83_end_0 = const()[name = tensor("x_83_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_83_end_mask_0 = const()[name = tensor("x_83_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_83_squeeze_mask_0 = const()[name = tensor("x_83_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_83_cast_fp16 = slice_by_index(begin = x_83_begin_0, end = x_83_end_0, end_mask = x_83_end_mask_0, squeeze_mask = x_83_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("x_83_cast_fp16")]; + tensor v_41_begin_0 = const()[name = tensor("v_41_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_41_end_0 = const()[name = tensor("v_41_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_41_end_mask_0 = const()[name = tensor("v_41_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_41_squeeze_mask_0 = const()[name = tensor("v_41_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_41_cast_fp16 = slice_by_index(begin = v_41_begin_0, end = v_41_end_0, end_mask = v_41_end_mask_0, squeeze_mask = v_41_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("v_41_cast_fp16")]; + tensor var_5149 = const()[name = tensor("op_5149"), val = tensor([1, 1, 1])]; + tensor position20_to_fp16_dtype_0 = const()[name = tensor("position20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position20_to_fp16 = cast(dtype = position20_to_fp16_dtype_0, x = position20)[name = tensor("cast_417")]; + tensor var_5150_cast_fp16 = reshape(shape = var_5149, x = position20_to_fp16)[name = tensor("op_5150_cast_fp16")]; + tensor angles_81_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_5150_cast_fp16)[name = tensor("angles_81_cast_fp16")]; + tensor rotr_81_cast_fp16 = cos(x = angles_81_cast_fp16)[name = tensor("rotr_81_cast_fp16")]; + tensor roti_81_cast_fp16 = sin(x = angles_81_cast_fp16)[name = tensor("roti_81_cast_fp16")]; + tensor var_5158 = const()[name = tensor("op_5158"), val = tensor([1, 16, 32, 2])]; + tensor pairs_81_cast_fp16 = reshape(shape = var_5158, x = x_81_cast_fp16)[name = tensor("pairs_81_cast_fp16")]; + tensor xr_81_begin_0 = const()[name = tensor("xr_81_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_81_end_0 = const()[name = tensor("xr_81_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_81_end_mask_0 = const()[name = tensor("xr_81_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_81_squeeze_mask_0 = const()[name = tensor("xr_81_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_81_cast_fp16 = slice_by_index(begin = xr_81_begin_0, end = xr_81_end_0, end_mask = xr_81_end_mask_0, squeeze_mask = xr_81_squeeze_mask_0, x = pairs_81_cast_fp16)[name = tensor("xr_81_cast_fp16")]; + tensor xi_81_begin_0 = const()[name = tensor("xi_81_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_81_end_0 = const()[name = tensor("xi_81_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_81_end_mask_0 = const()[name = tensor("xi_81_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_81_squeeze_mask_0 = const()[name = tensor("xi_81_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_81_cast_fp16 = slice_by_index(begin = xi_81_begin_0, end = xi_81_end_0, end_mask = xi_81_end_mask_0, squeeze_mask = xi_81_squeeze_mask_0, x = pairs_81_cast_fp16)[name = tensor("xi_81_cast_fp16")]; + tensor var_5166_cast_fp16 = mul(x = xr_81_cast_fp16, y = rotr_81_cast_fp16)[name = tensor("op_5166_cast_fp16")]; + tensor var_5167_cast_fp16 = mul(x = xi_81_cast_fp16, y = roti_81_cast_fp16)[name = tensor("op_5167_cast_fp16")]; + tensor xor__81_cast_fp16 = sub(x = var_5166_cast_fp16, y = var_5167_cast_fp16)[name = tensor("xor__81_cast_fp16")]; + tensor var_5170_cast_fp16 = mul(x = xr_81_cast_fp16, y = roti_81_cast_fp16)[name = tensor("op_5170_cast_fp16")]; + tensor var_5171_cast_fp16 = mul(x = xi_81_cast_fp16, y = rotr_81_cast_fp16)[name = tensor("op_5171_cast_fp16")]; + tensor xoi_81_cast_fp16 = add(x = var_5170_cast_fp16, y = var_5171_cast_fp16)[name = tensor("xoi_81_cast_fp16")]; + tensor out_81_axis_0 = const()[name = tensor("out_81_axis_0"), val = tensor(-1)]; + tensor out_81_cast_fp16 = stack(axis = out_81_axis_0, values = (xor__81_cast_fp16, xoi_81_cast_fp16))[name = tensor("out_81_cast_fp16")]; + tensor var_5194 = const()[name = tensor("op_5194"), val = tensor([1, 16, 32, 2])]; + tensor pairs_83_cast_fp16 = reshape(shape = var_5194, x = x_83_cast_fp16)[name = tensor("pairs_83_cast_fp16")]; + tensor xr_83_begin_0 = const()[name = tensor("xr_83_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_83_end_0 = const()[name = tensor("xr_83_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_83_end_mask_0 = const()[name = tensor("xr_83_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_83_squeeze_mask_0 = const()[name = tensor("xr_83_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_83_cast_fp16 = slice_by_index(begin = xr_83_begin_0, end = xr_83_end_0, end_mask = xr_83_end_mask_0, squeeze_mask = xr_83_squeeze_mask_0, x = pairs_83_cast_fp16)[name = tensor("xr_83_cast_fp16")]; + tensor xi_83_begin_0 = const()[name = tensor("xi_83_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_83_end_0 = const()[name = tensor("xi_83_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_83_end_mask_0 = const()[name = tensor("xi_83_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_83_squeeze_mask_0 = const()[name = tensor("xi_83_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_83_cast_fp16 = slice_by_index(begin = xi_83_begin_0, end = xi_83_end_0, end_mask = xi_83_end_mask_0, squeeze_mask = xi_83_squeeze_mask_0, x = pairs_83_cast_fp16)[name = tensor("xi_83_cast_fp16")]; + tensor var_5202_cast_fp16 = mul(x = xr_83_cast_fp16, y = rotr_81_cast_fp16)[name = tensor("op_5202_cast_fp16")]; + tensor var_5203_cast_fp16 = mul(x = xi_83_cast_fp16, y = roti_81_cast_fp16)[name = tensor("op_5203_cast_fp16")]; + tensor xor__83_cast_fp16 = sub(x = var_5202_cast_fp16, y = var_5203_cast_fp16)[name = tensor("xor__83_cast_fp16")]; + tensor var_5206_cast_fp16 = mul(x = xr_83_cast_fp16, y = roti_81_cast_fp16)[name = tensor("op_5206_cast_fp16")]; + tensor var_5207_cast_fp16 = mul(x = xi_83_cast_fp16, y = rotr_81_cast_fp16)[name = tensor("op_5207_cast_fp16")]; + tensor xoi_83_cast_fp16 = add(x = var_5206_cast_fp16, y = var_5207_cast_fp16)[name = tensor("xoi_83_cast_fp16")]; + tensor out_83_axis_0 = const()[name = tensor("out_83_axis_0"), val = tensor(-1)]; + tensor out_83_cast_fp16 = stack(axis = out_83_axis_0, values = (xor__83_cast_fp16, xoi_83_cast_fp16))[name = tensor("out_83_cast_fp16")]; + tensor _inversed_5219_y_0_to_fp16 = const()[name = tensor("_inversed_5219_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5219_cast_fp16 = mul(x = position20_to_fp16, y = _inversed_5219_y_0_to_fp16)[name = tensor("_inversed_5219_cast_fp16")]; + tensor var_5220_cast_fp16 = floor(x = _inversed_5219_cast_fp16)[name = tensor("op_5220_cast_fp16")]; + tensor var_5221_to_fp16 = const()[name = tensor("op_5221_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5222_cast_fp16 = mul(x = var_5220_cast_fp16, y = var_5221_to_fp16)[name = tensor("op_5222_cast_fp16")]; + tensor write_idx_41_cast_fp16 = sub(x = position20_to_fp16, y = var_5222_cast_fp16)[name = tensor("write_idx_41_cast_fp16")]; + tensor var_5225_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_41_cast_fp16)[name = tensor("op_5225_cast_fp16")]; + tensor var_5235 = const()[name = tensor("op_5235"), val = tensor([1, 512, 1, 1])]; + tensor onehot_81_to_fp16_dtype_0 = const()[name = tensor("onehot_81_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_5225_cast_fp16_to_fp16 = cast(dtype = onehot_81_to_fp16_dtype_0, x = var_5225_cast_fp16)[name = tensor("cast_416")]; + tensor onehot_83_cast_fp16 = reshape(shape = var_5235, x = var_5225_cast_fp16_to_fp16)[name = tensor("onehot_83_cast_fp16")]; + tensor var_5237_to_fp16 = const()[name = tensor("op_5237_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5239_cast_fp16 = sub(x = var_5237_to_fp16, y = onehot_83_cast_fp16)[name = tensor("op_5239_cast_fp16")]; + tensor k_cache20_to_fp16_dtype_0 = const()[name = tensor("k_cache20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache20_to_fp16 = cast(dtype = k_cache20_to_fp16_dtype_0, x = k_cache20)[name = tensor("cast_415")]; + tensor var_5240_cast_fp16 = mul(x = k_cache20_to_fp16, y = var_5239_cast_fp16)[name = tensor("op_5240_cast_fp16")]; + tensor var_5245 = const()[name = tensor("op_5245"), val = tensor([1, 1, 16, 64])]; + tensor var_5246_cast_fp16 = reshape(shape = var_5245, x = out_83_cast_fp16)[name = tensor("op_5246_cast_fp16")]; + tensor var_5247_cast_fp16 = mul(x = var_5246_cast_fp16, y = onehot_83_cast_fp16)[name = tensor("op_5247_cast_fp16")]; + tensor new_k_cache_41_cast_fp16 = add(x = var_5240_cast_fp16, y = var_5247_cast_fp16)[name = tensor("new_k_cache_41_cast_fp16")]; + tensor new_k_cache_41_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_41_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache20_to_fp16_dtype_0 = const()[name = tensor("v_cache20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache20_to_fp16 = cast(dtype = v_cache20_to_fp16_dtype_0, x = v_cache20)[name = tensor("cast_413")]; + tensor var_5253_cast_fp16 = mul(x = v_cache20_to_fp16, y = var_5239_cast_fp16)[name = tensor("op_5253_cast_fp16")]; + tensor var_5258 = const()[name = tensor("op_5258"), val = tensor([1, 1, 16, 64])]; + tensor var_5259_cast_fp16 = reshape(shape = var_5258, x = v_41_cast_fp16)[name = tensor("op_5259_cast_fp16")]; + tensor var_5260_cast_fp16 = mul(x = var_5259_cast_fp16, y = onehot_83_cast_fp16)[name = tensor("op_5260_cast_fp16")]; + tensor new_v_cache_41_cast_fp16 = add(x = var_5253_cast_fp16, y = var_5260_cast_fp16)[name = tensor("new_v_cache_41_cast_fp16")]; + tensor new_v_cache_41_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_41_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_20_perm_0 = const()[name = tensor("transpose_20_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_5273 = const()[name = tensor("op_5273"), val = tensor([0, 2, 1, 3])]; + tensor var_5279 = const()[name = tensor("op_5279"), val = tensor([1, 16, 1, 64])]; + tensor q4_41_cast_fp16 = reshape(shape = var_5279, x = out_81_cast_fp16)[name = tensor("q4_41_cast_fp16")]; + tensor var_5284_transpose_x_0 = const()[name = tensor("op_5284_transpose_x_0"), val = tensor(false)]; + tensor var_5284_transpose_y_0 = const()[name = tensor("op_5284_transpose_y_0"), val = tensor(false)]; + tensor transpose_20_cast_fp16 = transpose(perm = transpose_20_perm_0, x = new_k_cache_41_cast_fp16)[name = tensor("transpose_35")]; + tensor var_5284_cast_fp16 = matmul(transpose_x = var_5284_transpose_x_0, transpose_y = var_5284_transpose_y_0, x = q4_41_cast_fp16, y = transpose_20_cast_fp16)[name = tensor("op_5284_cast_fp16")]; + tensor var_5285_to_fp16 = const()[name = tensor("op_5285_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_81_cast_fp16 = mul(x = var_5284_cast_fp16, y = var_5285_to_fp16)[name = tensor("scores_81_cast_fp16")]; + tensor var_5289 = const()[name = tensor("op_5289"), val = tensor([1, 1])]; + tensor var_5290_cast_fp16 = reshape(shape = var_5289, x = position20_to_fp16)[name = tensor("op_5290_cast_fp16")]; + tensor var_5291_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_5290_cast_fp16)[name = tensor("op_5291_cast_fp16")]; + tensor mask_41_to_fp16_dtype_0 = const()[name = tensor("mask_41_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_5298_to_fp16 = const()[name = tensor("op_5298_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5291_cast_fp16_to_fp16 = cast(dtype = mask_41_to_fp16_dtype_0, x = var_5291_cast_fp16)[name = tensor("cast_411")]; + tensor var_5299_cast_fp16 = sub(x = var_5291_cast_fp16_to_fp16, y = var_5298_to_fp16)[name = tensor("op_5299_cast_fp16")]; + tensor var_5304 = const()[name = tensor("op_5304"), val = tensor([1, 1, 1, 512])]; + tensor var_5305_cast_fp16 = reshape(shape = var_5304, x = var_5299_cast_fp16)[name = tensor("op_5305_cast_fp16")]; + tensor var_5306_to_fp16 = const()[name = tensor("op_5306_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_5307_cast_fp16 = mul(x = var_5305_cast_fp16, y = var_5306_to_fp16)[name = tensor("op_5307_cast_fp16")]; + tensor scores_83_cast_fp16 = add(x = scores_81_cast_fp16, y = var_5307_cast_fp16)[name = tensor("scores_83_cast_fp16")]; + tensor var_5310 = const()[name = tensor("op_5310"), val = tensor(-1)]; + tensor weights_41_cast_fp16 = softmax(axis = var_5310, x = scores_83_cast_fp16)[name = tensor("weights_41_cast_fp16")]; + tensor attn_41_transpose_x_0 = const()[name = tensor("attn_41_transpose_x_0"), val = tensor(false)]; + tensor attn_41_transpose_y_0 = const()[name = tensor("attn_41_transpose_y_0"), val = tensor(false)]; + tensor values_41_cast_fp16 = transpose(perm = var_5273, x = new_v_cache_41_cast_fp16)[name = tensor("transpose_34")]; + tensor attn_41_cast_fp16 = matmul(transpose_x = attn_41_transpose_x_0, transpose_y = attn_41_transpose_y_0, x = weights_41_cast_fp16, y = values_41_cast_fp16)[name = tensor("attn_41_cast_fp16")]; + tensor var_5318 = const()[name = tensor("op_5318"), val = tensor([0, 2, 1, 3])]; + tensor var_5323 = const()[name = tensor("op_5323"), val = tensor([1, 1, 1024])]; + tensor var_5319_cast_fp16 = transpose(perm = var_5318, x = attn_41_cast_fp16)[name = tensor("transpose_33")]; + tensor input_247_cast_fp16 = reshape(shape = var_5323, x = var_5319_cast_fp16)[name = tensor("input_247_cast_fp16")]; + tensor attn20_out_proj_weight_to_fp16 = const()[name = tensor("attn20_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(509869760)))]; + tensor linear_82_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn20_out_proj_weight_to_fp16, x = input_247_cast_fp16)[name = tensor("linear_82_cast_fp16")]; + tensor var_5329_to_fp16 = const()[name = tensor("op_5329_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5330_cast_fp16 = add(x = position20_to_fp16, y = var_5329_to_fp16)[name = tensor("op_5330_cast_fp16")]; + tensor var_5330_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5330_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_249_cast_fp16 = add(x = input_243_cast_fp16, y = linear_82_cast_fp16)[name = tensor("input_249_cast_fp16")]; + tensor input_251_axes_0 = const()[name = tensor("input_251_axes_0"), val = tensor([-1])]; + tensor norm20_2_weight_to_fp16 = const()[name = tensor("norm20_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(511966976)))]; + tensor norm20_2_bias_to_fp16 = const()[name = tensor("norm20_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(511969088)))]; + tensor var_5334_to_fp16 = const()[name = tensor("op_5334_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_251_cast_fp16 = layer_norm(axes = input_251_axes_0, beta = norm20_2_bias_to_fp16, epsilon = var_5334_to_fp16, gamma = norm20_2_weight_to_fp16, x = input_249_cast_fp16)[name = tensor("input_251_cast_fp16")]; + tensor linear20_1_weight_to_fp16 = const()[name = tensor("linear20_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(511971200)))]; + tensor linear_83_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear20_1_weight_to_fp16, x = input_251_cast_fp16)[name = tensor("linear_83_cast_fp16")]; + tensor input_253_mode_0 = const()[name = tensor("input_253_mode_0"), val = tensor("EXACT")]; + tensor input_253_cast_fp16 = gelu(mode = input_253_mode_0, x = linear_83_cast_fp16)[name = tensor("input_253_cast_fp16")]; + tensor linear20_2_weight_to_fp16 = const()[name = tensor("linear20_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(520359872)))]; + tensor linear_84_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear20_2_weight_to_fp16, x = input_253_cast_fp16)[name = tensor("linear_84_cast_fp16")]; + tensor input_255_cast_fp16 = add(x = input_249_cast_fp16, y = linear_84_cast_fp16)[name = tensor("input_255_cast_fp16")]; + tensor input_257_axes_0 = const()[name = tensor("input_257_axes_0"), val = tensor([-1])]; + tensor norm21_1_weight_to_fp16 = const()[name = tensor("norm21_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(528748544)))]; + tensor norm21_1_bias_to_fp16 = const()[name = tensor("norm21_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(528750656)))]; + tensor var_5351_to_fp16 = const()[name = tensor("op_5351_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_257_cast_fp16 = layer_norm(axes = input_257_axes_0, beta = norm21_1_bias_to_fp16, epsilon = var_5351_to_fp16, gamma = norm21_1_weight_to_fp16, x = input_255_cast_fp16)[name = tensor("input_257_cast_fp16")]; + tensor attn21_in_proj_weight_to_fp16 = const()[name = tensor("attn21_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(528752768)))]; + tensor linear_85_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn21_in_proj_weight_to_fp16, x = input_257_cast_fp16)[name = tensor("linear_85_cast_fp16")]; + tensor var_5364 = const()[name = tensor("op_5364"), val = tensor([1, 3, 16, 64])]; + tensor qkv_43_cast_fp16 = reshape(shape = var_5364, x = linear_85_cast_fp16)[name = tensor("qkv_43_cast_fp16")]; + tensor x_85_begin_0 = const()[name = tensor("x_85_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_85_end_0 = const()[name = tensor("x_85_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_85_end_mask_0 = const()[name = tensor("x_85_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_85_squeeze_mask_0 = const()[name = tensor("x_85_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_85_cast_fp16 = slice_by_index(begin = x_85_begin_0, end = x_85_end_0, end_mask = x_85_end_mask_0, squeeze_mask = x_85_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("x_85_cast_fp16")]; + tensor x_87_begin_0 = const()[name = tensor("x_87_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_87_end_0 = const()[name = tensor("x_87_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_87_end_mask_0 = const()[name = tensor("x_87_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_87_squeeze_mask_0 = const()[name = tensor("x_87_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_87_cast_fp16 = slice_by_index(begin = x_87_begin_0, end = x_87_end_0, end_mask = x_87_end_mask_0, squeeze_mask = x_87_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("x_87_cast_fp16")]; + tensor v_43_begin_0 = const()[name = tensor("v_43_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_43_end_0 = const()[name = tensor("v_43_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_43_end_mask_0 = const()[name = tensor("v_43_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_43_squeeze_mask_0 = const()[name = tensor("v_43_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_43_cast_fp16 = slice_by_index(begin = v_43_begin_0, end = v_43_end_0, end_mask = v_43_end_mask_0, squeeze_mask = v_43_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("v_43_cast_fp16")]; + tensor var_5393 = const()[name = tensor("op_5393"), val = tensor([1, 1, 1])]; + tensor position21_to_fp16_dtype_0 = const()[name = tensor("position21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position21_to_fp16 = cast(dtype = position21_to_fp16_dtype_0, x = position21)[name = tensor("cast_409")]; + tensor var_5394_cast_fp16 = reshape(shape = var_5393, x = position21_to_fp16)[name = tensor("op_5394_cast_fp16")]; + tensor angles_85_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_5394_cast_fp16)[name = tensor("angles_85_cast_fp16")]; + tensor rotr_85_cast_fp16 = cos(x = angles_85_cast_fp16)[name = tensor("rotr_85_cast_fp16")]; + tensor roti_85_cast_fp16 = sin(x = angles_85_cast_fp16)[name = tensor("roti_85_cast_fp16")]; + tensor var_5402 = const()[name = tensor("op_5402"), val = tensor([1, 16, 32, 2])]; + tensor pairs_85_cast_fp16 = reshape(shape = var_5402, x = x_85_cast_fp16)[name = tensor("pairs_85_cast_fp16")]; + tensor xr_85_begin_0 = const()[name = tensor("xr_85_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_85_end_0 = const()[name = tensor("xr_85_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_85_end_mask_0 = const()[name = tensor("xr_85_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_85_squeeze_mask_0 = const()[name = tensor("xr_85_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_85_cast_fp16 = slice_by_index(begin = xr_85_begin_0, end = xr_85_end_0, end_mask = xr_85_end_mask_0, squeeze_mask = xr_85_squeeze_mask_0, x = pairs_85_cast_fp16)[name = tensor("xr_85_cast_fp16")]; + tensor xi_85_begin_0 = const()[name = tensor("xi_85_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_85_end_0 = const()[name = tensor("xi_85_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_85_end_mask_0 = const()[name = tensor("xi_85_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_85_squeeze_mask_0 = const()[name = tensor("xi_85_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_85_cast_fp16 = slice_by_index(begin = xi_85_begin_0, end = xi_85_end_0, end_mask = xi_85_end_mask_0, squeeze_mask = xi_85_squeeze_mask_0, x = pairs_85_cast_fp16)[name = tensor("xi_85_cast_fp16")]; + tensor var_5410_cast_fp16 = mul(x = xr_85_cast_fp16, y = rotr_85_cast_fp16)[name = tensor("op_5410_cast_fp16")]; + tensor var_5411_cast_fp16 = mul(x = xi_85_cast_fp16, y = roti_85_cast_fp16)[name = tensor("op_5411_cast_fp16")]; + tensor xor__85_cast_fp16 = sub(x = var_5410_cast_fp16, y = var_5411_cast_fp16)[name = tensor("xor__85_cast_fp16")]; + tensor var_5414_cast_fp16 = mul(x = xr_85_cast_fp16, y = roti_85_cast_fp16)[name = tensor("op_5414_cast_fp16")]; + tensor var_5415_cast_fp16 = mul(x = xi_85_cast_fp16, y = rotr_85_cast_fp16)[name = tensor("op_5415_cast_fp16")]; + tensor xoi_85_cast_fp16 = add(x = var_5414_cast_fp16, y = var_5415_cast_fp16)[name = tensor("xoi_85_cast_fp16")]; + tensor out_85_axis_0 = const()[name = tensor("out_85_axis_0"), val = tensor(-1)]; + tensor out_85_cast_fp16 = stack(axis = out_85_axis_0, values = (xor__85_cast_fp16, xoi_85_cast_fp16))[name = tensor("out_85_cast_fp16")]; + tensor var_5438 = const()[name = tensor("op_5438"), val = tensor([1, 16, 32, 2])]; + tensor pairs_87_cast_fp16 = reshape(shape = var_5438, x = x_87_cast_fp16)[name = tensor("pairs_87_cast_fp16")]; + tensor xr_87_begin_0 = const()[name = tensor("xr_87_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_87_end_0 = const()[name = tensor("xr_87_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_87_end_mask_0 = const()[name = tensor("xr_87_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_87_squeeze_mask_0 = const()[name = tensor("xr_87_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_87_cast_fp16 = slice_by_index(begin = xr_87_begin_0, end = xr_87_end_0, end_mask = xr_87_end_mask_0, squeeze_mask = xr_87_squeeze_mask_0, x = pairs_87_cast_fp16)[name = tensor("xr_87_cast_fp16")]; + tensor xi_87_begin_0 = const()[name = tensor("xi_87_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_87_end_0 = const()[name = tensor("xi_87_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_87_end_mask_0 = const()[name = tensor("xi_87_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_87_squeeze_mask_0 = const()[name = tensor("xi_87_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_87_cast_fp16 = slice_by_index(begin = xi_87_begin_0, end = xi_87_end_0, end_mask = xi_87_end_mask_0, squeeze_mask = xi_87_squeeze_mask_0, x = pairs_87_cast_fp16)[name = tensor("xi_87_cast_fp16")]; + tensor var_5446_cast_fp16 = mul(x = xr_87_cast_fp16, y = rotr_85_cast_fp16)[name = tensor("op_5446_cast_fp16")]; + tensor var_5447_cast_fp16 = mul(x = xi_87_cast_fp16, y = roti_85_cast_fp16)[name = tensor("op_5447_cast_fp16")]; + tensor xor__87_cast_fp16 = sub(x = var_5446_cast_fp16, y = var_5447_cast_fp16)[name = tensor("xor__87_cast_fp16")]; + tensor var_5450_cast_fp16 = mul(x = xr_87_cast_fp16, y = roti_85_cast_fp16)[name = tensor("op_5450_cast_fp16")]; + tensor var_5451_cast_fp16 = mul(x = xi_87_cast_fp16, y = rotr_85_cast_fp16)[name = tensor("op_5451_cast_fp16")]; + tensor xoi_87_cast_fp16 = add(x = var_5450_cast_fp16, y = var_5451_cast_fp16)[name = tensor("xoi_87_cast_fp16")]; + tensor out_87_axis_0 = const()[name = tensor("out_87_axis_0"), val = tensor(-1)]; + tensor out_87_cast_fp16 = stack(axis = out_87_axis_0, values = (xor__87_cast_fp16, xoi_87_cast_fp16))[name = tensor("out_87_cast_fp16")]; + tensor _inversed_5463_y_0_to_fp16 = const()[name = tensor("_inversed_5463_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5463_cast_fp16 = mul(x = position21_to_fp16, y = _inversed_5463_y_0_to_fp16)[name = tensor("_inversed_5463_cast_fp16")]; + tensor var_5464_cast_fp16 = floor(x = _inversed_5463_cast_fp16)[name = tensor("op_5464_cast_fp16")]; + tensor var_5465_to_fp16 = const()[name = tensor("op_5465_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5466_cast_fp16 = mul(x = var_5464_cast_fp16, y = var_5465_to_fp16)[name = tensor("op_5466_cast_fp16")]; + tensor write_idx_43_cast_fp16 = sub(x = position21_to_fp16, y = var_5466_cast_fp16)[name = tensor("write_idx_43_cast_fp16")]; + tensor var_5469_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_43_cast_fp16)[name = tensor("op_5469_cast_fp16")]; + tensor var_5479 = const()[name = tensor("op_5479"), val = tensor([1, 512, 1, 1])]; + tensor onehot_85_to_fp16_dtype_0 = const()[name = tensor("onehot_85_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_5469_cast_fp16_to_fp16 = cast(dtype = onehot_85_to_fp16_dtype_0, x = var_5469_cast_fp16)[name = tensor("cast_408")]; + tensor onehot_87_cast_fp16 = reshape(shape = var_5479, x = var_5469_cast_fp16_to_fp16)[name = tensor("onehot_87_cast_fp16")]; + tensor var_5481_to_fp16 = const()[name = tensor("op_5481_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5483_cast_fp16 = sub(x = var_5481_to_fp16, y = onehot_87_cast_fp16)[name = tensor("op_5483_cast_fp16")]; + tensor k_cache21_to_fp16_dtype_0 = const()[name = tensor("k_cache21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache21_to_fp16 = cast(dtype = k_cache21_to_fp16_dtype_0, x = k_cache21)[name = tensor("cast_407")]; + tensor var_5484_cast_fp16 = mul(x = k_cache21_to_fp16, y = var_5483_cast_fp16)[name = tensor("op_5484_cast_fp16")]; + tensor var_5489 = const()[name = tensor("op_5489"), val = tensor([1, 1, 16, 64])]; + tensor var_5490_cast_fp16 = reshape(shape = var_5489, x = out_87_cast_fp16)[name = tensor("op_5490_cast_fp16")]; + tensor var_5491_cast_fp16 = mul(x = var_5490_cast_fp16, y = onehot_87_cast_fp16)[name = tensor("op_5491_cast_fp16")]; + tensor new_k_cache_43_cast_fp16 = add(x = var_5484_cast_fp16, y = var_5491_cast_fp16)[name = tensor("new_k_cache_43_cast_fp16")]; + tensor new_k_cache_43_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_43_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache21_to_fp16_dtype_0 = const()[name = tensor("v_cache21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache21_to_fp16 = cast(dtype = v_cache21_to_fp16_dtype_0, x = v_cache21)[name = tensor("cast_405")]; + tensor var_5497_cast_fp16 = mul(x = v_cache21_to_fp16, y = var_5483_cast_fp16)[name = tensor("op_5497_cast_fp16")]; + tensor var_5502 = const()[name = tensor("op_5502"), val = tensor([1, 1, 16, 64])]; + tensor var_5503_cast_fp16 = reshape(shape = var_5502, x = v_43_cast_fp16)[name = tensor("op_5503_cast_fp16")]; + tensor var_5504_cast_fp16 = mul(x = var_5503_cast_fp16, y = onehot_87_cast_fp16)[name = tensor("op_5504_cast_fp16")]; + tensor new_v_cache_43_cast_fp16 = add(x = var_5497_cast_fp16, y = var_5504_cast_fp16)[name = tensor("new_v_cache_43_cast_fp16")]; + tensor new_v_cache_43_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_43_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_21_perm_0 = const()[name = tensor("transpose_21_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_5517 = const()[name = tensor("op_5517"), val = tensor([0, 2, 1, 3])]; + tensor var_5523 = const()[name = tensor("op_5523"), val = tensor([1, 16, 1, 64])]; + tensor q4_43_cast_fp16 = reshape(shape = var_5523, x = out_85_cast_fp16)[name = tensor("q4_43_cast_fp16")]; + tensor var_5528_transpose_x_0 = const()[name = tensor("op_5528_transpose_x_0"), val = tensor(false)]; + tensor var_5528_transpose_y_0 = const()[name = tensor("op_5528_transpose_y_0"), val = tensor(false)]; + tensor transpose_21_cast_fp16 = transpose(perm = transpose_21_perm_0, x = new_k_cache_43_cast_fp16)[name = tensor("transpose_32")]; + tensor var_5528_cast_fp16 = matmul(transpose_x = var_5528_transpose_x_0, transpose_y = var_5528_transpose_y_0, x = q4_43_cast_fp16, y = transpose_21_cast_fp16)[name = tensor("op_5528_cast_fp16")]; + tensor var_5529_to_fp16 = const()[name = tensor("op_5529_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_85_cast_fp16 = mul(x = var_5528_cast_fp16, y = var_5529_to_fp16)[name = tensor("scores_85_cast_fp16")]; + tensor var_5533 = const()[name = tensor("op_5533"), val = tensor([1, 1])]; + tensor var_5534_cast_fp16 = reshape(shape = var_5533, x = position21_to_fp16)[name = tensor("op_5534_cast_fp16")]; + tensor var_5535_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_5534_cast_fp16)[name = tensor("op_5535_cast_fp16")]; + tensor mask_43_to_fp16_dtype_0 = const()[name = tensor("mask_43_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_5542_to_fp16 = const()[name = tensor("op_5542_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5535_cast_fp16_to_fp16 = cast(dtype = mask_43_to_fp16_dtype_0, x = var_5535_cast_fp16)[name = tensor("cast_403")]; + tensor var_5543_cast_fp16 = sub(x = var_5535_cast_fp16_to_fp16, y = var_5542_to_fp16)[name = tensor("op_5543_cast_fp16")]; + tensor var_5548 = const()[name = tensor("op_5548"), val = tensor([1, 1, 1, 512])]; + tensor var_5549_cast_fp16 = reshape(shape = var_5548, x = var_5543_cast_fp16)[name = tensor("op_5549_cast_fp16")]; + tensor var_5550_to_fp16 = const()[name = tensor("op_5550_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_5551_cast_fp16 = mul(x = var_5549_cast_fp16, y = var_5550_to_fp16)[name = tensor("op_5551_cast_fp16")]; + tensor scores_87_cast_fp16 = add(x = scores_85_cast_fp16, y = var_5551_cast_fp16)[name = tensor("scores_87_cast_fp16")]; + tensor var_5554 = const()[name = tensor("op_5554"), val = tensor(-1)]; + tensor weights_43_cast_fp16 = softmax(axis = var_5554, x = scores_87_cast_fp16)[name = tensor("weights_43_cast_fp16")]; + tensor attn_43_transpose_x_0 = const()[name = tensor("attn_43_transpose_x_0"), val = tensor(false)]; + tensor attn_43_transpose_y_0 = const()[name = tensor("attn_43_transpose_y_0"), val = tensor(false)]; + tensor values_43_cast_fp16 = transpose(perm = var_5517, x = new_v_cache_43_cast_fp16)[name = tensor("transpose_31")]; + tensor attn_43_cast_fp16 = matmul(transpose_x = attn_43_transpose_x_0, transpose_y = attn_43_transpose_y_0, x = weights_43_cast_fp16, y = values_43_cast_fp16)[name = tensor("attn_43_cast_fp16")]; + tensor var_5562 = const()[name = tensor("op_5562"), val = tensor([0, 2, 1, 3])]; + tensor var_5567 = const()[name = tensor("op_5567"), val = tensor([1, 1, 1024])]; + tensor var_5563_cast_fp16 = transpose(perm = var_5562, x = attn_43_cast_fp16)[name = tensor("transpose_30")]; + tensor input_259_cast_fp16 = reshape(shape = var_5567, x = var_5563_cast_fp16)[name = tensor("input_259_cast_fp16")]; + tensor attn21_out_proj_weight_to_fp16 = const()[name = tensor("attn21_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(535044288)))]; + tensor linear_86_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn21_out_proj_weight_to_fp16, x = input_259_cast_fp16)[name = tensor("linear_86_cast_fp16")]; + tensor var_5573_to_fp16 = const()[name = tensor("op_5573_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5574_cast_fp16 = add(x = position21_to_fp16, y = var_5573_to_fp16)[name = tensor("op_5574_cast_fp16")]; + tensor var_5574_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5574_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_261_cast_fp16 = add(x = input_255_cast_fp16, y = linear_86_cast_fp16)[name = tensor("input_261_cast_fp16")]; + tensor input_263_axes_0 = const()[name = tensor("input_263_axes_0"), val = tensor([-1])]; + tensor norm21_2_weight_to_fp16 = const()[name = tensor("norm21_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(537141504)))]; + tensor norm21_2_bias_to_fp16 = const()[name = tensor("norm21_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(537143616)))]; + tensor var_5578_to_fp16 = const()[name = tensor("op_5578_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_263_cast_fp16 = layer_norm(axes = input_263_axes_0, beta = norm21_2_bias_to_fp16, epsilon = var_5578_to_fp16, gamma = norm21_2_weight_to_fp16, x = input_261_cast_fp16)[name = tensor("input_263_cast_fp16")]; + tensor linear21_1_weight_to_fp16 = const()[name = tensor("linear21_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(537145728)))]; + tensor linear_87_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear21_1_weight_to_fp16, x = input_263_cast_fp16)[name = tensor("linear_87_cast_fp16")]; + tensor input_265_mode_0 = const()[name = tensor("input_265_mode_0"), val = tensor("EXACT")]; + tensor input_265_cast_fp16 = gelu(mode = input_265_mode_0, x = linear_87_cast_fp16)[name = tensor("input_265_cast_fp16")]; + tensor linear21_2_weight_to_fp16 = const()[name = tensor("linear21_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(545534400)))]; + tensor linear_88_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear21_2_weight_to_fp16, x = input_265_cast_fp16)[name = tensor("linear_88_cast_fp16")]; + tensor input_267_cast_fp16 = add(x = input_261_cast_fp16, y = linear_88_cast_fp16)[name = tensor("input_267_cast_fp16")]; + tensor input_269_axes_0 = const()[name = tensor("input_269_axes_0"), val = tensor([-1])]; + tensor norm22_1_weight_to_fp16 = const()[name = tensor("norm22_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(553923072)))]; + tensor norm22_1_bias_to_fp16 = const()[name = tensor("norm22_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(553925184)))]; + tensor var_5595_to_fp16 = const()[name = tensor("op_5595_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_269_cast_fp16 = layer_norm(axes = input_269_axes_0, beta = norm22_1_bias_to_fp16, epsilon = var_5595_to_fp16, gamma = norm22_1_weight_to_fp16, x = input_267_cast_fp16)[name = tensor("input_269_cast_fp16")]; + tensor attn22_in_proj_weight_to_fp16 = const()[name = tensor("attn22_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(553927296)))]; + tensor linear_89_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn22_in_proj_weight_to_fp16, x = input_269_cast_fp16)[name = tensor("linear_89_cast_fp16")]; + tensor var_5608 = const()[name = tensor("op_5608"), val = tensor([1, 3, 16, 64])]; + tensor qkv_45_cast_fp16 = reshape(shape = var_5608, x = linear_89_cast_fp16)[name = tensor("qkv_45_cast_fp16")]; + tensor x_89_begin_0 = const()[name = tensor("x_89_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_89_end_0 = const()[name = tensor("x_89_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_89_end_mask_0 = const()[name = tensor("x_89_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_89_squeeze_mask_0 = const()[name = tensor("x_89_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_89_cast_fp16 = slice_by_index(begin = x_89_begin_0, end = x_89_end_0, end_mask = x_89_end_mask_0, squeeze_mask = x_89_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("x_89_cast_fp16")]; + tensor x_91_begin_0 = const()[name = tensor("x_91_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_91_end_0 = const()[name = tensor("x_91_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_91_end_mask_0 = const()[name = tensor("x_91_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_91_squeeze_mask_0 = const()[name = tensor("x_91_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_91_cast_fp16 = slice_by_index(begin = x_91_begin_0, end = x_91_end_0, end_mask = x_91_end_mask_0, squeeze_mask = x_91_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("x_91_cast_fp16")]; + tensor v_45_begin_0 = const()[name = tensor("v_45_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_45_end_0 = const()[name = tensor("v_45_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_45_end_mask_0 = const()[name = tensor("v_45_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_45_squeeze_mask_0 = const()[name = tensor("v_45_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_45_cast_fp16 = slice_by_index(begin = v_45_begin_0, end = v_45_end_0, end_mask = v_45_end_mask_0, squeeze_mask = v_45_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("v_45_cast_fp16")]; + tensor var_5637 = const()[name = tensor("op_5637"), val = tensor([1, 1, 1])]; + tensor position22_to_fp16_dtype_0 = const()[name = tensor("position22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position22_to_fp16 = cast(dtype = position22_to_fp16_dtype_0, x = position22)[name = tensor("cast_401")]; + tensor var_5638_cast_fp16 = reshape(shape = var_5637, x = position22_to_fp16)[name = tensor("op_5638_cast_fp16")]; + tensor angles_89_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_5638_cast_fp16)[name = tensor("angles_89_cast_fp16")]; + tensor rotr_89_cast_fp16 = cos(x = angles_89_cast_fp16)[name = tensor("rotr_89_cast_fp16")]; + tensor roti_89_cast_fp16 = sin(x = angles_89_cast_fp16)[name = tensor("roti_89_cast_fp16")]; + tensor var_5646 = const()[name = tensor("op_5646"), val = tensor([1, 16, 32, 2])]; + tensor pairs_89_cast_fp16 = reshape(shape = var_5646, x = x_89_cast_fp16)[name = tensor("pairs_89_cast_fp16")]; + tensor xr_89_begin_0 = const()[name = tensor("xr_89_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_89_end_0 = const()[name = tensor("xr_89_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_89_end_mask_0 = const()[name = tensor("xr_89_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_89_squeeze_mask_0 = const()[name = tensor("xr_89_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_89_cast_fp16 = slice_by_index(begin = xr_89_begin_0, end = xr_89_end_0, end_mask = xr_89_end_mask_0, squeeze_mask = xr_89_squeeze_mask_0, x = pairs_89_cast_fp16)[name = tensor("xr_89_cast_fp16")]; + tensor xi_89_begin_0 = const()[name = tensor("xi_89_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_89_end_0 = const()[name = tensor("xi_89_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_89_end_mask_0 = const()[name = tensor("xi_89_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_89_squeeze_mask_0 = const()[name = tensor("xi_89_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_89_cast_fp16 = slice_by_index(begin = xi_89_begin_0, end = xi_89_end_0, end_mask = xi_89_end_mask_0, squeeze_mask = xi_89_squeeze_mask_0, x = pairs_89_cast_fp16)[name = tensor("xi_89_cast_fp16")]; + tensor var_5654_cast_fp16 = mul(x = xr_89_cast_fp16, y = rotr_89_cast_fp16)[name = tensor("op_5654_cast_fp16")]; + tensor var_5655_cast_fp16 = mul(x = xi_89_cast_fp16, y = roti_89_cast_fp16)[name = tensor("op_5655_cast_fp16")]; + tensor xor__89_cast_fp16 = sub(x = var_5654_cast_fp16, y = var_5655_cast_fp16)[name = tensor("xor__89_cast_fp16")]; + tensor var_5658_cast_fp16 = mul(x = xr_89_cast_fp16, y = roti_89_cast_fp16)[name = tensor("op_5658_cast_fp16")]; + tensor var_5659_cast_fp16 = mul(x = xi_89_cast_fp16, y = rotr_89_cast_fp16)[name = tensor("op_5659_cast_fp16")]; + tensor xoi_89_cast_fp16 = add(x = var_5658_cast_fp16, y = var_5659_cast_fp16)[name = tensor("xoi_89_cast_fp16")]; + tensor out_89_axis_0 = const()[name = tensor("out_89_axis_0"), val = tensor(-1)]; + tensor out_89_cast_fp16 = stack(axis = out_89_axis_0, values = (xor__89_cast_fp16, xoi_89_cast_fp16))[name = tensor("out_89_cast_fp16")]; + tensor var_5682 = const()[name = tensor("op_5682"), val = tensor([1, 16, 32, 2])]; + tensor pairs_91_cast_fp16 = reshape(shape = var_5682, x = x_91_cast_fp16)[name = tensor("pairs_91_cast_fp16")]; + tensor xr_91_begin_0 = const()[name = tensor("xr_91_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_91_end_0 = const()[name = tensor("xr_91_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_91_end_mask_0 = const()[name = tensor("xr_91_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_91_squeeze_mask_0 = const()[name = tensor("xr_91_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_91_cast_fp16 = slice_by_index(begin = xr_91_begin_0, end = xr_91_end_0, end_mask = xr_91_end_mask_0, squeeze_mask = xr_91_squeeze_mask_0, x = pairs_91_cast_fp16)[name = tensor("xr_91_cast_fp16")]; + tensor xi_91_begin_0 = const()[name = tensor("xi_91_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_91_end_0 = const()[name = tensor("xi_91_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_91_end_mask_0 = const()[name = tensor("xi_91_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_91_squeeze_mask_0 = const()[name = tensor("xi_91_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_91_cast_fp16 = slice_by_index(begin = xi_91_begin_0, end = xi_91_end_0, end_mask = xi_91_end_mask_0, squeeze_mask = xi_91_squeeze_mask_0, x = pairs_91_cast_fp16)[name = tensor("xi_91_cast_fp16")]; + tensor var_5690_cast_fp16 = mul(x = xr_91_cast_fp16, y = rotr_89_cast_fp16)[name = tensor("op_5690_cast_fp16")]; + tensor var_5691_cast_fp16 = mul(x = xi_91_cast_fp16, y = roti_89_cast_fp16)[name = tensor("op_5691_cast_fp16")]; + tensor xor__91_cast_fp16 = sub(x = var_5690_cast_fp16, y = var_5691_cast_fp16)[name = tensor("xor__91_cast_fp16")]; + tensor var_5694_cast_fp16 = mul(x = xr_91_cast_fp16, y = roti_89_cast_fp16)[name = tensor("op_5694_cast_fp16")]; + tensor var_5695_cast_fp16 = mul(x = xi_91_cast_fp16, y = rotr_89_cast_fp16)[name = tensor("op_5695_cast_fp16")]; + tensor xoi_91_cast_fp16 = add(x = var_5694_cast_fp16, y = var_5695_cast_fp16)[name = tensor("xoi_91_cast_fp16")]; + tensor out_91_axis_0 = const()[name = tensor("out_91_axis_0"), val = tensor(-1)]; + tensor out_91_cast_fp16 = stack(axis = out_91_axis_0, values = (xor__91_cast_fp16, xoi_91_cast_fp16))[name = tensor("out_91_cast_fp16")]; + tensor _inversed_5707_y_0_to_fp16 = const()[name = tensor("_inversed_5707_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5707_cast_fp16 = mul(x = position22_to_fp16, y = _inversed_5707_y_0_to_fp16)[name = tensor("_inversed_5707_cast_fp16")]; + tensor var_5708_cast_fp16 = floor(x = _inversed_5707_cast_fp16)[name = tensor("op_5708_cast_fp16")]; + tensor var_5709_to_fp16 = const()[name = tensor("op_5709_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5710_cast_fp16 = mul(x = var_5708_cast_fp16, y = var_5709_to_fp16)[name = tensor("op_5710_cast_fp16")]; + tensor write_idx_45_cast_fp16 = sub(x = position22_to_fp16, y = var_5710_cast_fp16)[name = tensor("write_idx_45_cast_fp16")]; + tensor var_5713_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_45_cast_fp16)[name = tensor("op_5713_cast_fp16")]; + tensor var_5723 = const()[name = tensor("op_5723"), val = tensor([1, 512, 1, 1])]; + tensor onehot_89_to_fp16_dtype_0 = const()[name = tensor("onehot_89_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_5713_cast_fp16_to_fp16 = cast(dtype = onehot_89_to_fp16_dtype_0, x = var_5713_cast_fp16)[name = tensor("cast_400")]; + tensor onehot_91_cast_fp16 = reshape(shape = var_5723, x = var_5713_cast_fp16_to_fp16)[name = tensor("onehot_91_cast_fp16")]; + tensor var_5725_to_fp16 = const()[name = tensor("op_5725_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5727_cast_fp16 = sub(x = var_5725_to_fp16, y = onehot_91_cast_fp16)[name = tensor("op_5727_cast_fp16")]; + tensor k_cache22_to_fp16_dtype_0 = const()[name = tensor("k_cache22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache22_to_fp16 = cast(dtype = k_cache22_to_fp16_dtype_0, x = k_cache22)[name = tensor("cast_399")]; + tensor var_5728_cast_fp16 = mul(x = k_cache22_to_fp16, y = var_5727_cast_fp16)[name = tensor("op_5728_cast_fp16")]; + tensor var_5733 = const()[name = tensor("op_5733"), val = tensor([1, 1, 16, 64])]; + tensor var_5734_cast_fp16 = reshape(shape = var_5733, x = out_91_cast_fp16)[name = tensor("op_5734_cast_fp16")]; + tensor var_5735_cast_fp16 = mul(x = var_5734_cast_fp16, y = onehot_91_cast_fp16)[name = tensor("op_5735_cast_fp16")]; + tensor new_k_cache_45_cast_fp16 = add(x = var_5728_cast_fp16, y = var_5735_cast_fp16)[name = tensor("new_k_cache_45_cast_fp16")]; + tensor new_k_cache_45_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_45_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache22_to_fp16_dtype_0 = const()[name = tensor("v_cache22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache22_to_fp16 = cast(dtype = v_cache22_to_fp16_dtype_0, x = v_cache22)[name = tensor("cast_397")]; + tensor var_5741_cast_fp16 = mul(x = v_cache22_to_fp16, y = var_5727_cast_fp16)[name = tensor("op_5741_cast_fp16")]; + tensor var_5746 = const()[name = tensor("op_5746"), val = tensor([1, 1, 16, 64])]; + tensor var_5747_cast_fp16 = reshape(shape = var_5746, x = v_45_cast_fp16)[name = tensor("op_5747_cast_fp16")]; + tensor var_5748_cast_fp16 = mul(x = var_5747_cast_fp16, y = onehot_91_cast_fp16)[name = tensor("op_5748_cast_fp16")]; + tensor new_v_cache_45_cast_fp16 = add(x = var_5741_cast_fp16, y = var_5748_cast_fp16)[name = tensor("new_v_cache_45_cast_fp16")]; + tensor new_v_cache_45_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_45_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_22_perm_0 = const()[name = tensor("transpose_22_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_5761 = const()[name = tensor("op_5761"), val = tensor([0, 2, 1, 3])]; + tensor var_5767 = const()[name = tensor("op_5767"), val = tensor([1, 16, 1, 64])]; + tensor q4_45_cast_fp16 = reshape(shape = var_5767, x = out_89_cast_fp16)[name = tensor("q4_45_cast_fp16")]; + tensor var_5772_transpose_x_0 = const()[name = tensor("op_5772_transpose_x_0"), val = tensor(false)]; + tensor var_5772_transpose_y_0 = const()[name = tensor("op_5772_transpose_y_0"), val = tensor(false)]; + tensor transpose_22_cast_fp16 = transpose(perm = transpose_22_perm_0, x = new_k_cache_45_cast_fp16)[name = tensor("transpose_29")]; + tensor var_5772_cast_fp16 = matmul(transpose_x = var_5772_transpose_x_0, transpose_y = var_5772_transpose_y_0, x = q4_45_cast_fp16, y = transpose_22_cast_fp16)[name = tensor("op_5772_cast_fp16")]; + tensor var_5773_to_fp16 = const()[name = tensor("op_5773_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_89_cast_fp16 = mul(x = var_5772_cast_fp16, y = var_5773_to_fp16)[name = tensor("scores_89_cast_fp16")]; + tensor var_5777 = const()[name = tensor("op_5777"), val = tensor([1, 1])]; + tensor var_5778_cast_fp16 = reshape(shape = var_5777, x = position22_to_fp16)[name = tensor("op_5778_cast_fp16")]; + tensor var_5779_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_5778_cast_fp16)[name = tensor("op_5779_cast_fp16")]; + tensor mask_45_to_fp16_dtype_0 = const()[name = tensor("mask_45_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_5786_to_fp16 = const()[name = tensor("op_5786_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5779_cast_fp16_to_fp16 = cast(dtype = mask_45_to_fp16_dtype_0, x = var_5779_cast_fp16)[name = tensor("cast_395")]; + tensor var_5787_cast_fp16 = sub(x = var_5779_cast_fp16_to_fp16, y = var_5786_to_fp16)[name = tensor("op_5787_cast_fp16")]; + tensor var_5792 = const()[name = tensor("op_5792"), val = tensor([1, 1, 1, 512])]; + tensor var_5793_cast_fp16 = reshape(shape = var_5792, x = var_5787_cast_fp16)[name = tensor("op_5793_cast_fp16")]; + tensor var_5794_to_fp16 = const()[name = tensor("op_5794_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_5795_cast_fp16 = mul(x = var_5793_cast_fp16, y = var_5794_to_fp16)[name = tensor("op_5795_cast_fp16")]; + tensor scores_91_cast_fp16 = add(x = scores_89_cast_fp16, y = var_5795_cast_fp16)[name = tensor("scores_91_cast_fp16")]; + tensor var_5798 = const()[name = tensor("op_5798"), val = tensor(-1)]; + tensor weights_45_cast_fp16 = softmax(axis = var_5798, x = scores_91_cast_fp16)[name = tensor("weights_45_cast_fp16")]; + tensor attn_45_transpose_x_0 = const()[name = tensor("attn_45_transpose_x_0"), val = tensor(false)]; + tensor attn_45_transpose_y_0 = const()[name = tensor("attn_45_transpose_y_0"), val = tensor(false)]; + tensor values_45_cast_fp16 = transpose(perm = var_5761, x = new_v_cache_45_cast_fp16)[name = tensor("transpose_28")]; + tensor attn_45_cast_fp16 = matmul(transpose_x = attn_45_transpose_x_0, transpose_y = attn_45_transpose_y_0, x = weights_45_cast_fp16, y = values_45_cast_fp16)[name = tensor("attn_45_cast_fp16")]; + tensor var_5806 = const()[name = tensor("op_5806"), val = tensor([0, 2, 1, 3])]; + tensor var_5811 = const()[name = tensor("op_5811"), val = tensor([1, 1, 1024])]; + tensor var_5807_cast_fp16 = transpose(perm = var_5806, x = attn_45_cast_fp16)[name = tensor("transpose_27")]; + tensor input_271_cast_fp16 = reshape(shape = var_5811, x = var_5807_cast_fp16)[name = tensor("input_271_cast_fp16")]; + tensor attn22_out_proj_weight_to_fp16 = const()[name = tensor("attn22_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(560218816)))]; + tensor linear_90_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn22_out_proj_weight_to_fp16, x = input_271_cast_fp16)[name = tensor("linear_90_cast_fp16")]; + tensor var_5817_to_fp16 = const()[name = tensor("op_5817_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5818_cast_fp16 = add(x = position22_to_fp16, y = var_5817_to_fp16)[name = tensor("op_5818_cast_fp16")]; + tensor var_5818_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5818_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_273_cast_fp16 = add(x = input_267_cast_fp16, y = linear_90_cast_fp16)[name = tensor("input_273_cast_fp16")]; + tensor input_275_axes_0 = const()[name = tensor("input_275_axes_0"), val = tensor([-1])]; + tensor norm22_2_weight_to_fp16 = const()[name = tensor("norm22_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(562316032)))]; + tensor norm22_2_bias_to_fp16 = const()[name = tensor("norm22_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(562318144)))]; + tensor var_5822_to_fp16 = const()[name = tensor("op_5822_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_275_cast_fp16 = layer_norm(axes = input_275_axes_0, beta = norm22_2_bias_to_fp16, epsilon = var_5822_to_fp16, gamma = norm22_2_weight_to_fp16, x = input_273_cast_fp16)[name = tensor("input_275_cast_fp16")]; + tensor linear22_1_weight_to_fp16 = const()[name = tensor("linear22_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(562320256)))]; + tensor linear_91_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear22_1_weight_to_fp16, x = input_275_cast_fp16)[name = tensor("linear_91_cast_fp16")]; + tensor input_277_mode_0 = const()[name = tensor("input_277_mode_0"), val = tensor("EXACT")]; + tensor input_277_cast_fp16 = gelu(mode = input_277_mode_0, x = linear_91_cast_fp16)[name = tensor("input_277_cast_fp16")]; + tensor linear22_2_weight_to_fp16 = const()[name = tensor("linear22_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(570708928)))]; + tensor linear_92_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear22_2_weight_to_fp16, x = input_277_cast_fp16)[name = tensor("linear_92_cast_fp16")]; + tensor input_279_cast_fp16 = add(x = input_273_cast_fp16, y = linear_92_cast_fp16)[name = tensor("input_279_cast_fp16")]; + tensor input_281_axes_0 = const()[name = tensor("input_281_axes_0"), val = tensor([-1])]; + tensor norm23_1_weight_to_fp16 = const()[name = tensor("norm23_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(579097600)))]; + tensor norm23_1_bias_to_fp16 = const()[name = tensor("norm23_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(579099712)))]; + tensor var_5839_to_fp16 = const()[name = tensor("op_5839_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_281_cast_fp16 = layer_norm(axes = input_281_axes_0, beta = norm23_1_bias_to_fp16, epsilon = var_5839_to_fp16, gamma = norm23_1_weight_to_fp16, x = input_279_cast_fp16)[name = tensor("input_281_cast_fp16")]; + tensor attn23_in_proj_weight_to_fp16 = const()[name = tensor("attn23_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(579101824)))]; + tensor linear_93_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn23_in_proj_weight_to_fp16, x = input_281_cast_fp16)[name = tensor("linear_93_cast_fp16")]; + tensor var_5852 = const()[name = tensor("op_5852"), val = tensor([1, 3, 16, 64])]; + tensor qkv_cast_fp16 = reshape(shape = var_5852, x = linear_93_cast_fp16)[name = tensor("qkv_cast_fp16")]; + tensor x_93_begin_0 = const()[name = tensor("x_93_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_93_end_0 = const()[name = tensor("x_93_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_93_end_mask_0 = const()[name = tensor("x_93_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_93_squeeze_mask_0 = const()[name = tensor("x_93_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_93_cast_fp16 = slice_by_index(begin = x_93_begin_0, end = x_93_end_0, end_mask = x_93_end_mask_0, squeeze_mask = x_93_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("x_93_cast_fp16")]; + tensor x_begin_0 = const()[name = tensor("x_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_end_0 = const()[name = tensor("x_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_end_mask_0 = const()[name = tensor("x_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_squeeze_mask_0 = const()[name = tensor("x_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_cast_fp16 = slice_by_index(begin = x_begin_0, end = x_end_0, end_mask = x_end_mask_0, squeeze_mask = x_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("x_cast_fp16")]; + tensor v_begin_0 = const()[name = tensor("v_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_end_0 = const()[name = tensor("v_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_end_mask_0 = const()[name = tensor("v_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_squeeze_mask_0 = const()[name = tensor("v_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_cast_fp16 = slice_by_index(begin = v_begin_0, end = v_end_0, end_mask = v_end_mask_0, squeeze_mask = v_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("v_cast_fp16")]; + tensor var_5881 = const()[name = tensor("op_5881"), val = tensor([1, 1, 1])]; + tensor position23_to_fp16_dtype_0 = const()[name = tensor("position23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position23_to_fp16 = cast(dtype = position23_to_fp16_dtype_0, x = position23)[name = tensor("cast_393")]; + tensor var_5882_cast_fp16 = reshape(shape = var_5881, x = position23_to_fp16)[name = tensor("op_5882_cast_fp16")]; + tensor angles_93_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_5882_cast_fp16)[name = tensor("angles_93_cast_fp16")]; + tensor rotr_93_cast_fp16 = cos(x = angles_93_cast_fp16)[name = tensor("rotr_93_cast_fp16")]; + tensor roti_93_cast_fp16 = sin(x = angles_93_cast_fp16)[name = tensor("roti_93_cast_fp16")]; + tensor var_5890 = const()[name = tensor("op_5890"), val = tensor([1, 16, 32, 2])]; + tensor pairs_93_cast_fp16 = reshape(shape = var_5890, x = x_93_cast_fp16)[name = tensor("pairs_93_cast_fp16")]; + tensor xr_93_begin_0 = const()[name = tensor("xr_93_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_93_end_0 = const()[name = tensor("xr_93_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_93_end_mask_0 = const()[name = tensor("xr_93_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_93_squeeze_mask_0 = const()[name = tensor("xr_93_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_93_cast_fp16 = slice_by_index(begin = xr_93_begin_0, end = xr_93_end_0, end_mask = xr_93_end_mask_0, squeeze_mask = xr_93_squeeze_mask_0, x = pairs_93_cast_fp16)[name = tensor("xr_93_cast_fp16")]; + tensor xi_93_begin_0 = const()[name = tensor("xi_93_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_93_end_0 = const()[name = tensor("xi_93_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_93_end_mask_0 = const()[name = tensor("xi_93_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_93_squeeze_mask_0 = const()[name = tensor("xi_93_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_93_cast_fp16 = slice_by_index(begin = xi_93_begin_0, end = xi_93_end_0, end_mask = xi_93_end_mask_0, squeeze_mask = xi_93_squeeze_mask_0, x = pairs_93_cast_fp16)[name = tensor("xi_93_cast_fp16")]; + tensor var_5898_cast_fp16 = mul(x = xr_93_cast_fp16, y = rotr_93_cast_fp16)[name = tensor("op_5898_cast_fp16")]; + tensor var_5899_cast_fp16 = mul(x = xi_93_cast_fp16, y = roti_93_cast_fp16)[name = tensor("op_5899_cast_fp16")]; + tensor xor__93_cast_fp16 = sub(x = var_5898_cast_fp16, y = var_5899_cast_fp16)[name = tensor("xor__93_cast_fp16")]; + tensor var_5902_cast_fp16 = mul(x = xr_93_cast_fp16, y = roti_93_cast_fp16)[name = tensor("op_5902_cast_fp16")]; + tensor var_5903_cast_fp16 = mul(x = xi_93_cast_fp16, y = rotr_93_cast_fp16)[name = tensor("op_5903_cast_fp16")]; + tensor xoi_93_cast_fp16 = add(x = var_5902_cast_fp16, y = var_5903_cast_fp16)[name = tensor("xoi_93_cast_fp16")]; + tensor out_93_axis_0 = const()[name = tensor("out_93_axis_0"), val = tensor(-1)]; + tensor out_93_cast_fp16 = stack(axis = out_93_axis_0, values = (xor__93_cast_fp16, xoi_93_cast_fp16))[name = tensor("out_93_cast_fp16")]; + tensor var_5926 = const()[name = tensor("op_5926"), val = tensor([1, 16, 32, 2])]; + tensor pairs_cast_fp16 = reshape(shape = var_5926, x = x_cast_fp16)[name = tensor("pairs_cast_fp16")]; + tensor xr_begin_0 = const()[name = tensor("xr_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_end_0 = const()[name = tensor("xr_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_end_mask_0 = const()[name = tensor("xr_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_squeeze_mask_0 = const()[name = tensor("xr_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_cast_fp16 = slice_by_index(begin = xr_begin_0, end = xr_end_0, end_mask = xr_end_mask_0, squeeze_mask = xr_squeeze_mask_0, x = pairs_cast_fp16)[name = tensor("xr_cast_fp16")]; + tensor xi_begin_0 = const()[name = tensor("xi_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_end_0 = const()[name = tensor("xi_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_end_mask_0 = const()[name = tensor("xi_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_squeeze_mask_0 = const()[name = tensor("xi_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_cast_fp16 = slice_by_index(begin = xi_begin_0, end = xi_end_0, end_mask = xi_end_mask_0, squeeze_mask = xi_squeeze_mask_0, x = pairs_cast_fp16)[name = tensor("xi_cast_fp16")]; + tensor var_5934_cast_fp16 = mul(x = xr_cast_fp16, y = rotr_93_cast_fp16)[name = tensor("op_5934_cast_fp16")]; + tensor var_5935_cast_fp16 = mul(x = xi_cast_fp16, y = roti_93_cast_fp16)[name = tensor("op_5935_cast_fp16")]; + tensor xor__cast_fp16 = sub(x = var_5934_cast_fp16, y = var_5935_cast_fp16)[name = tensor("xor__cast_fp16")]; + tensor var_5938_cast_fp16 = mul(x = xr_cast_fp16, y = roti_93_cast_fp16)[name = tensor("op_5938_cast_fp16")]; + tensor var_5939_cast_fp16 = mul(x = xi_cast_fp16, y = rotr_93_cast_fp16)[name = tensor("op_5939_cast_fp16")]; + tensor xoi_cast_fp16 = add(x = var_5938_cast_fp16, y = var_5939_cast_fp16)[name = tensor("xoi_cast_fp16")]; + tensor out_axis_0 = const()[name = tensor("out_axis_0"), val = tensor(-1)]; + tensor out_cast_fp16 = stack(axis = out_axis_0, values = (xor__cast_fp16, xoi_cast_fp16))[name = tensor("out_cast_fp16")]; + tensor _inversed_5951_y_0_to_fp16 = const()[name = tensor("_inversed_5951_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5951_cast_fp16 = mul(x = position23_to_fp16, y = _inversed_5951_y_0_to_fp16)[name = tensor("_inversed_5951_cast_fp16")]; + tensor var_5952_cast_fp16 = floor(x = _inversed_5951_cast_fp16)[name = tensor("op_5952_cast_fp16")]; + tensor var_5953_to_fp16 = const()[name = tensor("op_5953_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5954_cast_fp16 = mul(x = var_5952_cast_fp16, y = var_5953_to_fp16)[name = tensor("op_5954_cast_fp16")]; + tensor write_idx_cast_fp16 = sub(x = position23_to_fp16, y = var_5954_cast_fp16)[name = tensor("write_idx_cast_fp16")]; + tensor var_5957_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_cast_fp16)[name = tensor("op_5957_cast_fp16")]; + tensor var_5967 = const()[name = tensor("op_5967"), val = tensor([1, 512, 1, 1])]; + tensor onehot_93_to_fp16_dtype_0 = const()[name = tensor("onehot_93_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_5957_cast_fp16_to_fp16 = cast(dtype = onehot_93_to_fp16_dtype_0, x = var_5957_cast_fp16)[name = tensor("cast_392")]; + tensor onehot_cast_fp16 = reshape(shape = var_5967, x = var_5957_cast_fp16_to_fp16)[name = tensor("onehot_cast_fp16")]; + tensor var_5969_to_fp16 = const()[name = tensor("op_5969_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5971_cast_fp16 = sub(x = var_5969_to_fp16, y = onehot_cast_fp16)[name = tensor("op_5971_cast_fp16")]; + tensor k_cache23_to_fp16_dtype_0 = const()[name = tensor("k_cache23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache23_to_fp16 = cast(dtype = k_cache23_to_fp16_dtype_0, x = k_cache23)[name = tensor("cast_391")]; + tensor var_5972_cast_fp16 = mul(x = k_cache23_to_fp16, y = var_5971_cast_fp16)[name = tensor("op_5972_cast_fp16")]; + tensor var_5977 = const()[name = tensor("op_5977"), val = tensor([1, 1, 16, 64])]; + tensor var_5978_cast_fp16 = reshape(shape = var_5977, x = out_cast_fp16)[name = tensor("op_5978_cast_fp16")]; + tensor var_5979_cast_fp16 = mul(x = var_5978_cast_fp16, y = onehot_cast_fp16)[name = tensor("op_5979_cast_fp16")]; + tensor new_k_cache_cast_fp16 = add(x = var_5972_cast_fp16, y = var_5979_cast_fp16)[name = tensor("new_k_cache_cast_fp16")]; + tensor new_k_cache_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache23_to_fp16_dtype_0 = const()[name = tensor("v_cache23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache23_to_fp16 = cast(dtype = v_cache23_to_fp16_dtype_0, x = v_cache23)[name = tensor("cast_389")]; + tensor var_5985_cast_fp16 = mul(x = v_cache23_to_fp16, y = var_5971_cast_fp16)[name = tensor("op_5985_cast_fp16")]; + tensor var_5990 = const()[name = tensor("op_5990"), val = tensor([1, 1, 16, 64])]; + tensor var_5991_cast_fp16 = reshape(shape = var_5990, x = v_cast_fp16)[name = tensor("op_5991_cast_fp16")]; + tensor var_5992_cast_fp16 = mul(x = var_5991_cast_fp16, y = onehot_cast_fp16)[name = tensor("op_5992_cast_fp16")]; + tensor new_v_cache_cast_fp16 = add(x = var_5985_cast_fp16, y = var_5992_cast_fp16)[name = tensor("new_v_cache_cast_fp16")]; + tensor new_v_cache_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_23_perm_0 = const()[name = tensor("transpose_23_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_6005 = const()[name = tensor("op_6005"), val = tensor([0, 2, 1, 3])]; + tensor var_6011 = const()[name = tensor("op_6011"), val = tensor([1, 16, 1, 64])]; + tensor q4_cast_fp16 = reshape(shape = var_6011, x = out_93_cast_fp16)[name = tensor("q4_cast_fp16")]; + tensor var_6016_transpose_x_0 = const()[name = tensor("op_6016_transpose_x_0"), val = tensor(false)]; + tensor var_6016_transpose_y_0 = const()[name = tensor("op_6016_transpose_y_0"), val = tensor(false)]; + tensor transpose_23_cast_fp16 = transpose(perm = transpose_23_perm_0, x = new_k_cache_cast_fp16)[name = tensor("transpose_26")]; + tensor var_6016_cast_fp16 = matmul(transpose_x = var_6016_transpose_x_0, transpose_y = var_6016_transpose_y_0, x = q4_cast_fp16, y = transpose_23_cast_fp16)[name = tensor("op_6016_cast_fp16")]; + tensor var_6017_to_fp16 = const()[name = tensor("op_6017_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_93_cast_fp16 = mul(x = var_6016_cast_fp16, y = var_6017_to_fp16)[name = tensor("scores_93_cast_fp16")]; + tensor var_6021 = const()[name = tensor("op_6021"), val = tensor([1, 1])]; + tensor var_6022_cast_fp16 = reshape(shape = var_6021, x = position23_to_fp16)[name = tensor("op_6022_cast_fp16")]; + tensor var_6023_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_6022_cast_fp16)[name = tensor("op_6023_cast_fp16")]; + tensor mask_to_fp16_dtype_0 = const()[name = tensor("mask_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_6030_to_fp16 = const()[name = tensor("op_6030_to_fp16"), val = tensor(0x1p+0)]; + tensor var_6023_cast_fp16_to_fp16 = cast(dtype = mask_to_fp16_dtype_0, x = var_6023_cast_fp16)[name = tensor("cast_387")]; + tensor var_6031_cast_fp16 = sub(x = var_6023_cast_fp16_to_fp16, y = var_6030_to_fp16)[name = tensor("op_6031_cast_fp16")]; + tensor var_6036 = const()[name = tensor("op_6036"), val = tensor([1, 1, 1, 512])]; + tensor var_6037_cast_fp16 = reshape(shape = var_6036, x = var_6031_cast_fp16)[name = tensor("op_6037_cast_fp16")]; + tensor var_6038_to_fp16 = const()[name = tensor("op_6038_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_6039_cast_fp16 = mul(x = var_6037_cast_fp16, y = var_6038_to_fp16)[name = tensor("op_6039_cast_fp16")]; + tensor scores_cast_fp16 = add(x = scores_93_cast_fp16, y = var_6039_cast_fp16)[name = tensor("scores_cast_fp16")]; + tensor var_6042 = const()[name = tensor("op_6042"), val = tensor(-1)]; + tensor weights_cast_fp16 = softmax(axis = var_6042, x = scores_cast_fp16)[name = tensor("weights_cast_fp16")]; + tensor attn_transpose_x_0 = const()[name = tensor("attn_transpose_x_0"), val = tensor(false)]; + tensor attn_transpose_y_0 = const()[name = tensor("attn_transpose_y_0"), val = tensor(false)]; + tensor values_cast_fp16 = transpose(perm = var_6005, x = new_v_cache_cast_fp16)[name = tensor("transpose_25")]; + tensor attn_cast_fp16 = matmul(transpose_x = attn_transpose_x_0, transpose_y = attn_transpose_y_0, x = weights_cast_fp16, y = values_cast_fp16)[name = tensor("attn_cast_fp16")]; + tensor var_6050 = const()[name = tensor("op_6050"), val = tensor([0, 2, 1, 3])]; + tensor var_6055 = const()[name = tensor("op_6055"), val = tensor([1, 1, 1024])]; + tensor var_6051_cast_fp16 = transpose(perm = var_6050, x = attn_cast_fp16)[name = tensor("transpose_24")]; + tensor input_283_cast_fp16 = reshape(shape = var_6055, x = var_6051_cast_fp16)[name = tensor("input_283_cast_fp16")]; + tensor attn23_out_proj_weight_to_fp16 = const()[name = tensor("attn23_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(585393344)))]; + tensor linear_94_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn23_out_proj_weight_to_fp16, x = input_283_cast_fp16)[name = tensor("linear_94_cast_fp16")]; + tensor var_6061_to_fp16 = const()[name = tensor("op_6061_to_fp16"), val = tensor(0x1p+0)]; + tensor var_6062_cast_fp16 = add(x = position23_to_fp16, y = var_6061_to_fp16)[name = tensor("op_6062_cast_fp16")]; + tensor var_6062_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_6062_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_285_cast_fp16 = add(x = input_279_cast_fp16, y = linear_94_cast_fp16)[name = tensor("input_285_cast_fp16")]; + tensor input_287_axes_0 = const()[name = tensor("input_287_axes_0"), val = tensor([-1])]; + tensor norm23_2_weight_to_fp16 = const()[name = tensor("norm23_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(587490560)))]; + tensor norm23_2_bias_to_fp16 = const()[name = tensor("norm23_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(587492672)))]; + tensor var_6066_to_fp16 = const()[name = tensor("op_6066_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_287_cast_fp16 = layer_norm(axes = input_287_axes_0, beta = norm23_2_bias_to_fp16, epsilon = var_6066_to_fp16, gamma = norm23_2_weight_to_fp16, x = input_285_cast_fp16)[name = tensor("input_287_cast_fp16")]; + tensor linear23_1_weight_to_fp16 = const()[name = tensor("linear23_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(587494784)))]; + tensor linear_95_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear23_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_95_cast_fp16")]; + tensor input_289_mode_0 = const()[name = tensor("input_289_mode_0"), val = tensor("EXACT")]; + tensor input_289_cast_fp16 = gelu(mode = input_289_mode_0, x = linear_95_cast_fp16)[name = tensor("input_289_cast_fp16")]; + tensor linear23_2_weight_to_fp16 = const()[name = tensor("linear23_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(595883456)))]; + tensor linear_96_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear23_2_weight_to_fp16, x = input_289_cast_fp16)[name = tensor("linear_96_cast_fp16")]; + tensor input_291_cast_fp16 = add(x = input_285_cast_fp16, y = linear_96_cast_fp16)[name = tensor("input_291_cast_fp16")]; + tensor input_axes_0 = const()[name = tensor("input_axes_0"), val = tensor([-1])]; + tensor out_norm_weight_to_fp16 = const()[name = tensor("out_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(604272128)))]; + tensor out_norm_bias_to_fp16 = const()[name = tensor("out_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(604274240)))]; + tensor var_6083_to_fp16 = const()[name = tensor("op_6083_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_cast_fp16 = layer_norm(axes = input_axes_0, beta = out_norm_bias_to_fp16, epsilon = var_6083_to_fp16, gamma = out_norm_weight_to_fp16, x = input_291_cast_fp16)[name = tensor("input_cast_fp16")]; + tensor input_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("input_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor out_eos_weight_to_fp16 = const()[name = tensor("out_eos_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(604276352)))]; + tensor out_eos_bias_to_fp16 = const()[name = tensor("out_eos_bias_to_fp16"), val = tensor([-0x1.8ap-3])]; + tensor linear_97_cast_fp16 = linear(bias = out_eos_bias_to_fp16, weight = out_eos_weight_to_fp16, x = input_cast_fp16)[name = tensor("linear_97_cast_fp16")]; + tensor linear_97_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("linear_97_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor is_eos = cast(dtype = linear_97_cast_fp16_to_fp32_dtype_0, x = linear_97_cast_fp16)[name = tensor("cast_384")]; + tensor transformer_out = cast(dtype = input_cast_fp16_to_fp32_dtype_0, x = input_cast_fp16)[name = tensor("cast_385")]; + tensor new_position23 = cast(dtype = var_6062_cast_fp16_to_fp32_dtype_0, x = var_6062_cast_fp16)[name = tensor("cast_386")]; + tensor new_v_cache23 = cast(dtype = new_v_cache_cast_fp16_to_fp32_dtype_0, x = new_v_cache_cast_fp16)[name = tensor("cast_388")]; + tensor new_k_cache23 = cast(dtype = new_k_cache_cast_fp16_to_fp32_dtype_0, x = new_k_cache_cast_fp16)[name = tensor("cast_390")]; + tensor new_position22 = cast(dtype = var_5818_cast_fp16_to_fp32_dtype_0, x = var_5818_cast_fp16)[name = tensor("cast_394")]; + tensor new_v_cache22 = cast(dtype = new_v_cache_45_cast_fp16_to_fp32_dtype_0, x = new_v_cache_45_cast_fp16)[name = tensor("cast_396")]; + tensor new_k_cache22 = cast(dtype = new_k_cache_45_cast_fp16_to_fp32_dtype_0, x = new_k_cache_45_cast_fp16)[name = tensor("cast_398")]; + tensor new_position21 = cast(dtype = var_5574_cast_fp16_to_fp32_dtype_0, x = var_5574_cast_fp16)[name = tensor("cast_402")]; + tensor new_v_cache21 = cast(dtype = new_v_cache_43_cast_fp16_to_fp32_dtype_0, x = new_v_cache_43_cast_fp16)[name = tensor("cast_404")]; + tensor new_k_cache21 = cast(dtype = new_k_cache_43_cast_fp16_to_fp32_dtype_0, x = new_k_cache_43_cast_fp16)[name = tensor("cast_406")]; + tensor new_position20 = cast(dtype = var_5330_cast_fp16_to_fp32_dtype_0, x = var_5330_cast_fp16)[name = tensor("cast_410")]; + tensor new_v_cache20 = cast(dtype = new_v_cache_41_cast_fp16_to_fp32_dtype_0, x = new_v_cache_41_cast_fp16)[name = tensor("cast_412")]; + tensor new_k_cache20 = cast(dtype = new_k_cache_41_cast_fp16_to_fp32_dtype_0, x = new_k_cache_41_cast_fp16)[name = tensor("cast_414")]; + tensor new_position19 = cast(dtype = var_5086_cast_fp16_to_fp32_dtype_0, x = var_5086_cast_fp16)[name = tensor("cast_418")]; + tensor new_v_cache19 = cast(dtype = new_v_cache_39_cast_fp16_to_fp32_dtype_0, x = new_v_cache_39_cast_fp16)[name = tensor("cast_420")]; + tensor new_k_cache19 = cast(dtype = new_k_cache_39_cast_fp16_to_fp32_dtype_0, x = new_k_cache_39_cast_fp16)[name = tensor("cast_422")]; + tensor new_position18 = cast(dtype = var_4842_cast_fp16_to_fp32_dtype_0, x = var_4842_cast_fp16)[name = tensor("cast_426")]; + tensor new_v_cache18 = cast(dtype = new_v_cache_37_cast_fp16_to_fp32_dtype_0, x = new_v_cache_37_cast_fp16)[name = tensor("cast_428")]; + tensor new_k_cache18 = cast(dtype = new_k_cache_37_cast_fp16_to_fp32_dtype_0, x = new_k_cache_37_cast_fp16)[name = tensor("cast_430")]; + tensor new_position17 = cast(dtype = var_4598_cast_fp16_to_fp32_dtype_0, x = var_4598_cast_fp16)[name = tensor("cast_434")]; + tensor new_v_cache17 = cast(dtype = new_v_cache_35_cast_fp16_to_fp32_dtype_0, x = new_v_cache_35_cast_fp16)[name = tensor("cast_436")]; + tensor new_k_cache17 = cast(dtype = new_k_cache_35_cast_fp16_to_fp32_dtype_0, x = new_k_cache_35_cast_fp16)[name = tensor("cast_438")]; + tensor new_position16 = cast(dtype = var_4354_cast_fp16_to_fp32_dtype_0, x = var_4354_cast_fp16)[name = tensor("cast_442")]; + tensor new_v_cache16 = cast(dtype = new_v_cache_33_cast_fp16_to_fp32_dtype_0, x = new_v_cache_33_cast_fp16)[name = tensor("cast_444")]; + tensor new_k_cache16 = cast(dtype = new_k_cache_33_cast_fp16_to_fp32_dtype_0, x = new_k_cache_33_cast_fp16)[name = tensor("cast_446")]; + tensor new_position15 = cast(dtype = var_4110_cast_fp16_to_fp32_dtype_0, x = var_4110_cast_fp16)[name = tensor("cast_450")]; + tensor new_v_cache15 = cast(dtype = new_v_cache_31_cast_fp16_to_fp32_dtype_0, x = new_v_cache_31_cast_fp16)[name = tensor("cast_452")]; + tensor new_k_cache15 = cast(dtype = new_k_cache_31_cast_fp16_to_fp32_dtype_0, x = new_k_cache_31_cast_fp16)[name = tensor("cast_454")]; + tensor new_position14 = cast(dtype = var_3866_cast_fp16_to_fp32_dtype_0, x = var_3866_cast_fp16)[name = tensor("cast_458")]; + tensor new_v_cache14 = cast(dtype = new_v_cache_29_cast_fp16_to_fp32_dtype_0, x = new_v_cache_29_cast_fp16)[name = tensor("cast_460")]; + tensor new_k_cache14 = cast(dtype = new_k_cache_29_cast_fp16_to_fp32_dtype_0, x = new_k_cache_29_cast_fp16)[name = tensor("cast_462")]; + tensor new_position13 = cast(dtype = var_3622_cast_fp16_to_fp32_dtype_0, x = var_3622_cast_fp16)[name = tensor("cast_466")]; + tensor new_v_cache13 = cast(dtype = new_v_cache_27_cast_fp16_to_fp32_dtype_0, x = new_v_cache_27_cast_fp16)[name = tensor("cast_468")]; + tensor new_k_cache13 = cast(dtype = new_k_cache_27_cast_fp16_to_fp32_dtype_0, x = new_k_cache_27_cast_fp16)[name = tensor("cast_470")]; + tensor new_position12 = cast(dtype = var_3378_cast_fp16_to_fp32_dtype_0, x = var_3378_cast_fp16)[name = tensor("cast_474")]; + tensor new_v_cache12 = cast(dtype = new_v_cache_25_cast_fp16_to_fp32_dtype_0, x = new_v_cache_25_cast_fp16)[name = tensor("cast_476")]; + tensor new_k_cache12 = cast(dtype = new_k_cache_25_cast_fp16_to_fp32_dtype_0, x = new_k_cache_25_cast_fp16)[name = tensor("cast_478")]; + tensor new_position11 = cast(dtype = var_3134_cast_fp16_to_fp32_dtype_0, x = var_3134_cast_fp16)[name = tensor("cast_482")]; + tensor new_v_cache11 = cast(dtype = new_v_cache_23_cast_fp16_to_fp32_dtype_0, x = new_v_cache_23_cast_fp16)[name = tensor("cast_484")]; + tensor new_k_cache11 = cast(dtype = new_k_cache_23_cast_fp16_to_fp32_dtype_0, x = new_k_cache_23_cast_fp16)[name = tensor("cast_486")]; + tensor new_position10 = cast(dtype = var_2890_cast_fp16_to_fp32_dtype_0, x = var_2890_cast_fp16)[name = tensor("cast_490")]; + tensor new_v_cache10 = cast(dtype = new_v_cache_21_cast_fp16_to_fp32_dtype_0, x = new_v_cache_21_cast_fp16)[name = tensor("cast_492")]; + tensor new_k_cache10 = cast(dtype = new_k_cache_21_cast_fp16_to_fp32_dtype_0, x = new_k_cache_21_cast_fp16)[name = tensor("cast_494")]; + tensor new_position9 = cast(dtype = var_2646_cast_fp16_to_fp32_dtype_0, x = var_2646_cast_fp16)[name = tensor("cast_498")]; + tensor new_v_cache9 = cast(dtype = new_v_cache_19_cast_fp16_to_fp32_dtype_0, x = new_v_cache_19_cast_fp16)[name = tensor("cast_500")]; + tensor new_k_cache9 = cast(dtype = new_k_cache_19_cast_fp16_to_fp32_dtype_0, x = new_k_cache_19_cast_fp16)[name = tensor("cast_502")]; + tensor new_position8 = cast(dtype = var_2402_cast_fp16_to_fp32_dtype_0, x = var_2402_cast_fp16)[name = tensor("cast_506")]; + tensor new_v_cache8 = cast(dtype = new_v_cache_17_cast_fp16_to_fp32_dtype_0, x = new_v_cache_17_cast_fp16)[name = tensor("cast_508")]; + tensor new_k_cache8 = cast(dtype = new_k_cache_17_cast_fp16_to_fp32_dtype_0, x = new_k_cache_17_cast_fp16)[name = tensor("cast_510")]; + tensor new_position7 = cast(dtype = var_2158_cast_fp16_to_fp32_dtype_0, x = var_2158_cast_fp16)[name = tensor("cast_514")]; + tensor new_v_cache7 = cast(dtype = new_v_cache_15_cast_fp16_to_fp32_dtype_0, x = new_v_cache_15_cast_fp16)[name = tensor("cast_516")]; + tensor new_k_cache7 = cast(dtype = new_k_cache_15_cast_fp16_to_fp32_dtype_0, x = new_k_cache_15_cast_fp16)[name = tensor("cast_518")]; + tensor new_position6 = cast(dtype = var_1914_cast_fp16_to_fp32_dtype_0, x = var_1914_cast_fp16)[name = tensor("cast_522")]; + tensor new_v_cache6 = cast(dtype = new_v_cache_13_cast_fp16_to_fp32_dtype_0, x = new_v_cache_13_cast_fp16)[name = tensor("cast_524")]; + tensor new_k_cache6 = cast(dtype = new_k_cache_13_cast_fp16_to_fp32_dtype_0, x = new_k_cache_13_cast_fp16)[name = tensor("cast_526")]; + tensor new_position5 = cast(dtype = var_1670_cast_fp16_to_fp32_dtype_0, x = var_1670_cast_fp16)[name = tensor("cast_530")]; + tensor new_v_cache5 = cast(dtype = new_v_cache_11_cast_fp16_to_fp32_dtype_0, x = new_v_cache_11_cast_fp16)[name = tensor("cast_532")]; + tensor new_k_cache5 = cast(dtype = new_k_cache_11_cast_fp16_to_fp32_dtype_0, x = new_k_cache_11_cast_fp16)[name = tensor("cast_534")]; + tensor new_position4 = cast(dtype = var_1426_cast_fp16_to_fp32_dtype_0, x = var_1426_cast_fp16)[name = tensor("cast_538")]; + tensor new_v_cache4 = cast(dtype = new_v_cache_9_cast_fp16_to_fp32_dtype_0, x = new_v_cache_9_cast_fp16)[name = tensor("cast_540")]; + tensor new_k_cache4 = cast(dtype = new_k_cache_9_cast_fp16_to_fp32_dtype_0, x = new_k_cache_9_cast_fp16)[name = tensor("cast_542")]; + tensor new_position3 = cast(dtype = var_1182_cast_fp16_to_fp32_dtype_0, x = var_1182_cast_fp16)[name = tensor("cast_546")]; + tensor new_v_cache3 = cast(dtype = new_v_cache_7_cast_fp16_to_fp32_dtype_0, x = new_v_cache_7_cast_fp16)[name = tensor("cast_548")]; + tensor new_k_cache3 = cast(dtype = new_k_cache_7_cast_fp16_to_fp32_dtype_0, x = new_k_cache_7_cast_fp16)[name = tensor("cast_550")]; + tensor new_position2 = cast(dtype = var_938_cast_fp16_to_fp32_dtype_0, x = var_938_cast_fp16)[name = tensor("cast_554")]; + tensor new_v_cache2 = cast(dtype = new_v_cache_5_cast_fp16_to_fp32_dtype_0, x = new_v_cache_5_cast_fp16)[name = tensor("cast_556")]; + tensor new_k_cache2 = cast(dtype = new_k_cache_5_cast_fp16_to_fp32_dtype_0, x = new_k_cache_5_cast_fp16)[name = tensor("cast_558")]; + tensor new_position1 = cast(dtype = var_694_cast_fp16_to_fp32_dtype_0, x = var_694_cast_fp16)[name = tensor("cast_562")]; + tensor new_v_cache1 = cast(dtype = new_v_cache_3_cast_fp16_to_fp32_dtype_0, x = new_v_cache_3_cast_fp16)[name = tensor("cast_564")]; + tensor new_k_cache1 = cast(dtype = new_k_cache_3_cast_fp16_to_fp32_dtype_0, x = new_k_cache_3_cast_fp16)[name = tensor("cast_566")]; + tensor new_position0 = cast(dtype = var_450_cast_fp16_to_fp32_dtype_0, x = var_450_cast_fp16)[name = tensor("cast_570")]; + tensor new_v_cache0 = cast(dtype = new_v_cache_1_cast_fp16_to_fp32_dtype_0, x = new_v_cache_1_cast_fp16)[name = tensor("cast_572")]; + tensor new_k_cache0 = cast(dtype = new_k_cache_1_cast_fp16_to_fp32_dtype_0, x = new_k_cache_1_cast_fp16)[name = tensor("cast_574")]; + } -> (transformer_out, is_eos, new_k_cache0, new_v_cache0, new_position0, new_k_cache1, new_v_cache1, new_position1, new_k_cache2, new_v_cache2, new_position2, new_k_cache3, new_v_cache3, new_position3, new_k_cache4, new_v_cache4, new_position4, new_k_cache5, new_v_cache5, new_position5, new_k_cache6, new_v_cache6, new_position6, new_k_cache7, new_v_cache7, new_position7, new_k_cache8, new_v_cache8, new_position8, new_k_cache9, new_v_cache9, new_position9, new_k_cache10, new_v_cache10, new_position10, new_k_cache11, new_v_cache11, new_position11, new_k_cache12, new_v_cache12, new_position12, new_k_cache13, new_v_cache13, new_position13, new_k_cache14, new_v_cache14, new_position14, new_k_cache15, new_v_cache15, new_position15, new_k_cache16, new_v_cache16, new_position16, new_k_cache17, new_v_cache17, new_position17, new_k_cache18, new_v_cache18, new_position18, new_k_cache19, new_v_cache19, new_position19, new_k_cache20, new_v_cache20, new_position20, new_k_cache21, new_v_cache21, new_position21, new_k_cache22, new_v_cache22, new_position22, new_k_cache23, new_v_cache23, new_position23); +} \ No newline at end of file diff --git a/v2.1/french_24l/flowlm_step_ane.mlmodelc/weights/weight.bin b/v2.1/french_24l/flowlm_step_ane.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..44b65b80a850121a0963a2bcece5d40ade315ef0 --- /dev/null +++ b/v2.1/french_24l/flowlm_step_ane.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:77dd60f8624625cfe07d36c7b12c679ed58f40e49be660db253a5254bb77c4e3 +size 604278464 diff --git a/v2.1/french_24l/flowlm_step_ane.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/french_24l/flowlm_step_ane.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..15e5a45cd11ae52c808e71392ebae9690284dad1 --- /dev/null +++ b/v2.1/french_24l/flowlm_step_ane.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:26622dbef2bd92f672b9afcf451f7b87d3e5a0523333d89bbeffab4339de186e +size 545925 diff --git a/v2.1/french_24l/flowlm_step_ane.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/french_24l/flowlm_step_ane.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..44b65b80a850121a0963a2bcece5d40ade315ef0 --- /dev/null +++ b/v2.1/french_24l/flowlm_step_ane.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:77dd60f8624625cfe07d36c7b12c679ed58f40e49be660db253a5254bb77c4e3 +size 604278464 diff --git a/v2.1/french_24l/flowlm_step_ane.mlpackage/Manifest.json b/v2.1/french_24l/flowlm_step_ane.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..2a9609b68838ac8c3f9e8d57c0b0861fc0043a5a --- /dev/null +++ b/v2.1/french_24l/flowlm_step_ane.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "35F5846E-F16E-4B6A-A1A4-435EAB97231F": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "51DAF9A2-F777-49C1-8A3A-833E84ECFBC4": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "35F5846E-F16E-4B6A-A1A4-435EAB97231F" +} diff --git a/v2.1/german/cond_prefill_ane.mlmodelc/analytics/coremldata.bin b/v2.1/german/cond_prefill_ane.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..f8347ed2f26a7446887f1e3a57e2477166e57056 --- /dev/null +++ b/v2.1/german/cond_prefill_ane.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ae7483ff54c8e37ba86a3548fd4aa89e10a99ec314fce1849e5c9f35030b0258 +size 243 diff --git a/v2.1/german/cond_prefill_ane.mlmodelc/coremldata.bin b/v2.1/german/cond_prefill_ane.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..64eeaedd93af76516cc0c238060aa0461d5bc909 --- /dev/null +++ b/v2.1/german/cond_prefill_ane.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7632fcfcdd94ec27c8067bf44ba3ee3436f258e8e84326513e1fa82e6bbb9827 +size 1388 diff --git a/v2.1/german/cond_prefill_ane.mlmodelc/model.mil b/v2.1/german/cond_prefill_ane.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..2aef44ccd7c895876ef314e46a742909df52fe28 --- /dev/null +++ b/v2.1/german/cond_prefill_ane.mlmodelc/model.mil @@ -0,0 +1,804 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.9.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor conditioning, tensor k_cache0, tensor k_cache1, tensor k_cache2, tensor k_cache3, tensor k_cache4, tensor k_cache5, tensor position0, tensor position1, tensor position2, tensor position3, tensor position4, tensor position5, tensor v_cache0, tensor v_cache1, tensor v_cache2, tensor v_cache3, tensor v_cache4, tensor v_cache5, tensor valid_len) { + tensor token_indices_to_fp16 = const()[name = tensor("token_indices_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor position0_to_fp16_dtype_0 = const()[name = tensor("position0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position0_to_fp16 = cast(dtype = position0_to_fp16_dtype_0, x = position0)[name = tensor("cast_128")]; + tensor ts_cast_fp16 = add(x = token_indices_to_fp16, y = position0_to_fp16)[name = tensor("ts_cast_fp16")]; + tensor var_69 = const()[name = tensor("op_69"), val = tensor([256, 1, 1])]; + tensor var_70_cast_fp16 = reshape(shape = var_69, x = ts_cast_fp16)[name = tensor("op_70_cast_fp16")]; + tensor var_65_to_fp16 = const()[name = tensor("op_65_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(640)))]; + tensor angles_cast_fp16 = mul(x = var_65_to_fp16, y = var_70_cast_fp16)[name = tensor("angles_cast_fp16")]; + tensor rotr_cast_fp16 = cos(x = angles_cast_fp16)[name = tensor("rotr_cast_fp16")]; + tensor roti_cast_fp16 = sin(x = angles_cast_fp16)[name = tensor("roti_cast_fp16")]; + tensor _inversed_75_y_0_to_fp16 = const()[name = tensor("_inversed_75_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_75_cast_fp16 = mul(x = ts_cast_fp16, y = _inversed_75_y_0_to_fp16)[name = tensor("_inversed_75_cast_fp16")]; + tensor var_76_cast_fp16 = floor(x = _inversed_75_cast_fp16)[name = tensor("op_76_cast_fp16")]; + tensor var_77_to_fp16 = const()[name = tensor("op_77_to_fp16"), val = tensor(0x1p+9)]; + tensor var_78_cast_fp16 = mul(x = var_76_cast_fp16, y = var_77_to_fp16)[name = tensor("op_78_cast_fp16")]; + tensor write_idx_1_cast_fp16 = sub(x = ts_cast_fp16, y = var_78_cast_fp16)[name = tensor("write_idx_1_cast_fp16")]; + tensor valid_len_to_fp16_dtype_0 = const()[name = tensor("valid_len_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor valid_len_to_fp16 = cast(dtype = valid_len_to_fp16_dtype_0, x = valid_len)[name = tensor("cast_127")]; + tensor var_84_cast_fp16 = less(x = token_indices_to_fp16, y = valid_len_to_fp16)[name = tensor("op_84_cast_fp16")]; + tensor token_valid_to_fp16_dtype_0 = const()[name = tensor("token_valid_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_84_cast_fp16_to_fp16 = cast(dtype = token_valid_to_fp16_dtype_0, x = var_84_cast_fp16)[name = tensor("cast_126")]; + tensor var_90_cast_fp16 = mul(x = write_idx_1_cast_fp16, y = var_84_cast_fp16_to_fp16)[name = tensor("op_90_cast_fp16")]; + tensor var_91_to_fp16 = const()[name = tensor("op_91_to_fp16"), val = tensor(0x1p+0)]; + tensor var_93_cast_fp16 = sub(x = var_91_to_fp16, y = var_84_cast_fp16_to_fp16)[name = tensor("op_93_cast_fp16")]; + tensor var_94_to_fp16 = const()[name = tensor("op_94_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_95_cast_fp16 = mul(x = var_93_cast_fp16, y = var_94_to_fp16)[name = tensor("op_95_cast_fp16")]; + tensor write_idx_cast_fp16 = add(x = var_90_cast_fp16, y = var_95_cast_fp16)[name = tensor("write_idx_cast_fp16")]; + tensor var_100 = const()[name = tensor("op_100"), val = tensor([256, 1])]; + tensor var_101_cast_fp16 = reshape(shape = var_100, x = write_idx_cast_fp16)[name = tensor("op_101_cast_fp16")]; + tensor var_105_to_fp16 = const()[name = tensor("op_105_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(768)))]; + tensor var_106_cast_fp16 = equal(x = var_101_cast_fp16, y = var_105_to_fp16)[name = tensor("op_106_cast_fp16")]; + tensor var_116_axes_0 = const()[name = tensor("op_116_axes_0"), val = tensor([0])]; + tensor var_116_keep_dims_0 = const()[name = tensor("op_116_keep_dims_0"), val = tensor(false)]; + tensor assign_to_fp16_dtype_0 = const()[name = tensor("assign_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_106_cast_fp16_to_fp16 = cast(dtype = assign_to_fp16_dtype_0, x = var_106_cast_fp16)[name = tensor("cast_125")]; + tensor var_116_cast_fp16 = reduce_sum(axes = var_116_axes_0, keep_dims = var_116_keep_dims_0, x = var_106_cast_fp16_to_fp16)[name = tensor("op_116_cast_fp16")]; + tensor const_0_to_fp16 = const()[name = tensor("const_0_to_fp16"), val = tensor(-inf)]; + tensor var_118_to_fp16 = const()[name = tensor("op_118_to_fp16"), val = tensor(0x1p+0)]; + tensor clip_0_cast_fp16 = clip(alpha = const_0_to_fp16, beta = var_118_to_fp16, x = var_116_cast_fp16)[name = tensor("clip_0_cast_fp16")]; + tensor var_122 = const()[name = tensor("op_122"), val = tensor([1, 1])]; + tensor var_123_cast_fp16 = reshape(shape = var_122, x = position0_to_fp16)[name = tensor("op_123_cast_fp16")]; + tensor var_127_to_fp16 = const()[name = tensor("op_127_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1856)))]; + tensor q_positions_cast_fp16 = add(x = var_123_cast_fp16, y = var_127_to_fp16)[name = tensor("q_positions_cast_fp16")]; + tensor var_135_cast_fp16 = add(x = position0_to_fp16, y = valid_len_to_fp16)[name = tensor("op_135_cast_fp16")]; + tensor var_138 = const()[name = tensor("op_138"), val = tensor([1, 1])]; + tensor valid_end_cast_fp16 = reshape(shape = var_138, x = var_135_cast_fp16)[name = tensor("valid_end_cast_fp16")]; + tensor var_140_cast_fp16 = less(x = var_105_to_fp16, y = valid_end_cast_fp16)[name = tensor("op_140_cast_fp16")]; + tensor var_141_cast_fp16 = less_equal(x = var_105_to_fp16, y = q_positions_cast_fp16)[name = tensor("op_141_cast_fp16")]; + tensor var_142 = logical_and(x = var_140_cast_fp16, y = var_141_cast_fp16)[name = tensor("op_142")]; + tensor mask_to_fp16_dtype_0 = const()[name = tensor("mask_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_149_to_fp16 = const()[name = tensor("op_149_to_fp16"), val = tensor(0x1p+0)]; + tensor var_142_to_fp16 = cast(dtype = mask_to_fp16_dtype_0, x = var_142)[name = tensor("cast_124")]; + tensor var_150_cast_fp16 = sub(x = var_142_to_fp16, y = var_149_to_fp16)[name = tensor("op_150_cast_fp16")]; + tensor var_151_to_fp16 = const()[name = tensor("op_151_to_fp16"), val = tensor(0x1.388p+13)]; + tensor mask_bias_cast_fp16 = mul(x = var_150_cast_fp16, y = var_151_to_fp16)[name = tensor("mask_bias_cast_fp16")]; + tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; + tensor conditioning_to_fp16_dtype_0 = const()[name = tensor("conditioning_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor norm0_1_weight_to_fp16 = const()[name = tensor("norm0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2432)))]; + tensor norm0_1_bias_to_fp16 = const()[name = tensor("norm0_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4544)))]; + tensor var_154_to_fp16 = const()[name = tensor("op_154_to_fp16"), val = tensor(0x1.5p-17)]; + tensor conditioning_to_fp16 = cast(dtype = conditioning_to_fp16_dtype_0, x = conditioning)[name = tensor("cast_123")]; + tensor input_3_cast_fp16 = layer_norm(axes = input_3_axes_0, beta = norm0_1_bias_to_fp16, epsilon = var_154_to_fp16, gamma = norm0_1_weight_to_fp16, x = conditioning_to_fp16)[name = tensor("input_3_cast_fp16")]; + tensor attn0_in_proj_weight_to_fp16 = const()[name = tensor("attn0_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6656)))]; + tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6298176)))]; + tensor linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn0_in_proj_weight_to_fp16, x = input_3_cast_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor var_168 = const()[name = tensor("op_168"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_1_cast_fp16 = reshape(shape = var_168, x = linear_0_cast_fp16)[name = tensor("qkv_1_cast_fp16")]; + tensor x_1_begin_0 = const()[name = tensor("x_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_1_end_0 = const()[name = tensor("x_1_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_1_end_mask_0 = const()[name = tensor("x_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_1_squeeze_mask_0 = const()[name = tensor("x_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_1_cast_fp16 = slice_by_index(begin = x_1_begin_0, end = x_1_end_0, end_mask = x_1_end_mask_0, squeeze_mask = x_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("x_1_cast_fp16")]; + tensor x_3_begin_0 = const()[name = tensor("x_3_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_3_end_0 = const()[name = tensor("x_3_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_3_end_mask_0 = const()[name = tensor("x_3_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_3_squeeze_mask_0 = const()[name = tensor("x_3_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_3_cast_fp16 = slice_by_index(begin = x_3_begin_0, end = x_3_end_0, end_mask = x_3_end_mask_0, squeeze_mask = x_3_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("x_3_cast_fp16")]; + tensor v_1_begin_0 = const()[name = tensor("v_1_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_1_end_0 = const()[name = tensor("v_1_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_1_end_mask_0 = const()[name = tensor("v_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_1_squeeze_mask_0 = const()[name = tensor("v_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_1_cast_fp16 = slice_by_index(begin = v_1_begin_0, end = v_1_end_0, end_mask = v_1_end_mask_0, squeeze_mask = v_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("v_1_cast_fp16")]; + tensor var_213 = const()[name = tensor("op_213"), val = tensor([256, 16, 32, 2])]; + tensor pairs_1_cast_fp16 = reshape(shape = var_213, x = x_1_cast_fp16)[name = tensor("pairs_1_cast_fp16")]; + tensor xr_1_begin_0 = const()[name = tensor("xr_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_1_end_0 = const()[name = tensor("xr_1_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_1_end_mask_0 = const()[name = tensor("xr_1_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_1_squeeze_mask_0 = const()[name = tensor("xr_1_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_1_cast_fp16 = slice_by_index(begin = xr_1_begin_0, end = xr_1_end_0, end_mask = xr_1_end_mask_0, squeeze_mask = xr_1_squeeze_mask_0, x = pairs_1_cast_fp16)[name = tensor("xr_1_cast_fp16")]; + tensor xi_1_begin_0 = const()[name = tensor("xi_1_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_1_end_0 = const()[name = tensor("xi_1_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_1_end_mask_0 = const()[name = tensor("xi_1_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_1_squeeze_mask_0 = const()[name = tensor("xi_1_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_1_cast_fp16 = slice_by_index(begin = xi_1_begin_0, end = xi_1_end_0, end_mask = xi_1_end_mask_0, squeeze_mask = xi_1_squeeze_mask_0, x = pairs_1_cast_fp16)[name = tensor("xi_1_cast_fp16")]; + tensor var_221_cast_fp16 = mul(x = xr_1_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_221_cast_fp16")]; + tensor var_222_cast_fp16 = mul(x = xi_1_cast_fp16, y = roti_cast_fp16)[name = tensor("op_222_cast_fp16")]; + tensor xor__1_cast_fp16 = sub(x = var_221_cast_fp16, y = var_222_cast_fp16)[name = tensor("xor__1_cast_fp16")]; + tensor var_225_cast_fp16 = mul(x = xr_1_cast_fp16, y = roti_cast_fp16)[name = tensor("op_225_cast_fp16")]; + tensor var_226_cast_fp16 = mul(x = xi_1_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_226_cast_fp16")]; + tensor xoi_1_cast_fp16 = add(x = var_225_cast_fp16, y = var_226_cast_fp16)[name = tensor("xoi_1_cast_fp16")]; + tensor out_1_axis_0 = const()[name = tensor("out_1_axis_0"), val = tensor(-1)]; + tensor out_1_cast_fp16 = stack(axis = out_1_axis_0, values = (xor__1_cast_fp16, xoi_1_cast_fp16))[name = tensor("out_1_cast_fp16")]; + tensor var_236 = const()[name = tensor("op_236"), val = tensor([1, 256, 16, 64])]; + tensor q_1_cast_fp16 = reshape(shape = var_236, x = out_1_cast_fp16)[name = tensor("q_1_cast_fp16")]; + tensor var_242 = const()[name = tensor("op_242"), val = tensor([256, 16, 32, 2])]; + tensor pairs_3_cast_fp16 = reshape(shape = var_242, x = x_3_cast_fp16)[name = tensor("pairs_3_cast_fp16")]; + tensor xr_3_begin_0 = const()[name = tensor("xr_3_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_3_end_0 = const()[name = tensor("xr_3_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_3_end_mask_0 = const()[name = tensor("xr_3_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_3_squeeze_mask_0 = const()[name = tensor("xr_3_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_3_cast_fp16 = slice_by_index(begin = xr_3_begin_0, end = xr_3_end_0, end_mask = xr_3_end_mask_0, squeeze_mask = xr_3_squeeze_mask_0, x = pairs_3_cast_fp16)[name = tensor("xr_3_cast_fp16")]; + tensor xi_3_begin_0 = const()[name = tensor("xi_3_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_3_end_0 = const()[name = tensor("xi_3_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_3_end_mask_0 = const()[name = tensor("xi_3_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_3_squeeze_mask_0 = const()[name = tensor("xi_3_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_3_cast_fp16 = slice_by_index(begin = xi_3_begin_0, end = xi_3_end_0, end_mask = xi_3_end_mask_0, squeeze_mask = xi_3_squeeze_mask_0, x = pairs_3_cast_fp16)[name = tensor("xi_3_cast_fp16")]; + tensor var_250_cast_fp16 = mul(x = xr_3_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_250_cast_fp16")]; + tensor var_251_cast_fp16 = mul(x = xi_3_cast_fp16, y = roti_cast_fp16)[name = tensor("op_251_cast_fp16")]; + tensor xor__3_cast_fp16 = sub(x = var_250_cast_fp16, y = var_251_cast_fp16)[name = tensor("xor__3_cast_fp16")]; + tensor var_254_cast_fp16 = mul(x = xr_3_cast_fp16, y = roti_cast_fp16)[name = tensor("op_254_cast_fp16")]; + tensor var_255_cast_fp16 = mul(x = xi_3_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_255_cast_fp16")]; + tensor xoi_3_cast_fp16 = add(x = var_254_cast_fp16, y = var_255_cast_fp16)[name = tensor("xoi_3_cast_fp16")]; + tensor out_3_axis_0 = const()[name = tensor("out_3_axis_0"), val = tensor(-1)]; + tensor out_3_cast_fp16 = stack(axis = out_3_axis_0, values = (xor__3_cast_fp16, xoi_3_cast_fp16))[name = tensor("out_3_cast_fp16")]; + tensor var_272 = const()[name = tensor("op_272"), val = tensor([256, 1024])]; + tensor var_273_cast_fp16 = reshape(shape = var_272, x = out_3_cast_fp16)[name = tensor("op_273_cast_fp16")]; + tensor written_k_1_transpose_x_1 = const()[name = tensor("written_k_1_transpose_x_1"), val = tensor(true)]; + tensor written_k_1_transpose_y_1 = const()[name = tensor("written_k_1_transpose_y_1"), val = tensor(false)]; + tensor written_k_1_cast_fp16 = matmul(transpose_x = written_k_1_transpose_x_1, transpose_y = written_k_1_transpose_y_1, x = var_106_cast_fp16_to_fp16, y = var_273_cast_fp16)[name = tensor("written_k_1_cast_fp16")]; + tensor var_280 = const()[name = tensor("op_280"), val = tensor([256, 1024])]; + tensor var_281_cast_fp16 = reshape(shape = var_280, x = v_1_cast_fp16)[name = tensor("op_281_cast_fp16")]; + tensor written_v_1_transpose_x_1 = const()[name = tensor("written_v_1_transpose_x_1"), val = tensor(true)]; + tensor written_v_1_transpose_y_1 = const()[name = tensor("written_v_1_transpose_y_1"), val = tensor(false)]; + tensor written_v_1_cast_fp16 = matmul(transpose_x = written_v_1_transpose_x_1, transpose_y = written_v_1_transpose_y_1, x = var_106_cast_fp16_to_fp16, y = var_281_cast_fp16)[name = tensor("written_v_1_cast_fp16")]; + tensor var_283_to_fp16 = const()[name = tensor("op_283_to_fp16"), val = tensor(0x1p+0)]; + tensor var_285_cast_fp16 = sub(x = var_283_to_fp16, y = clip_0_cast_fp16)[name = tensor("op_285_cast_fp16")]; + tensor var_290 = const()[name = tensor("op_290"), val = tensor([1, 512, 1, 1])]; + tensor keep_1_cast_fp16 = reshape(shape = var_290, x = var_285_cast_fp16)[name = tensor("keep_1_cast_fp16")]; + tensor k_cache0_to_fp16_dtype_0 = const()[name = tensor("k_cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache0_to_fp16 = cast(dtype = k_cache0_to_fp16_dtype_0, x = k_cache0)[name = tensor("cast_122")]; + tensor var_292_cast_fp16 = mul(x = k_cache0_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_292_cast_fp16")]; + tensor var_297 = const()[name = tensor("op_297"), val = tensor([1, 512, 16, 64])]; + tensor var_298_cast_fp16 = reshape(shape = var_297, x = written_k_1_cast_fp16)[name = tensor("op_298_cast_fp16")]; + tensor new_k_cache_1_cast_fp16 = add(x = var_292_cast_fp16, y = var_298_cast_fp16)[name = tensor("new_k_cache_1_cast_fp16")]; + tensor new_k_cache_1_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_1_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache0_to_fp16_dtype_0 = const()[name = tensor("v_cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache0_to_fp16 = cast(dtype = v_cache0_to_fp16_dtype_0, x = v_cache0)[name = tensor("cast_120")]; + tensor var_301_cast_fp16 = mul(x = v_cache0_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_301_cast_fp16")]; + tensor var_306 = const()[name = tensor("op_306"), val = tensor([1, 512, 16, 64])]; + tensor var_307_cast_fp16 = reshape(shape = var_306, x = written_v_1_cast_fp16)[name = tensor("op_307_cast_fp16")]; + tensor new_v_cache_1_cast_fp16 = add(x = var_301_cast_fp16, y = var_307_cast_fp16)[name = tensor("new_v_cache_1_cast_fp16")]; + tensor new_v_cache_1_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_1_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_320 = const()[name = tensor("op_320"), val = tensor([0, 2, 1, 3])]; + tensor var_331_transpose_x_0 = const()[name = tensor("op_331_transpose_x_0"), val = tensor(false)]; + tensor var_331_transpose_y_0 = const()[name = tensor("op_331_transpose_y_0"), val = tensor(false)]; + tensor transpose_15_perm_0 = const()[name = tensor("transpose_15_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_16_perm_0 = const()[name = tensor("transpose_16_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_16 = transpose(perm = transpose_16_perm_0, x = new_k_cache_1_cast_fp16)[name = tensor("transpose_42")]; + tensor transpose_15 = transpose(perm = transpose_15_perm_0, x = q_1_cast_fp16)[name = tensor("transpose_43")]; + tensor var_331_cast_fp16 = matmul(transpose_x = var_331_transpose_x_0, transpose_y = var_331_transpose_y_0, x = transpose_15, y = transpose_16)[name = tensor("op_331_cast_fp16")]; + tensor var_332_to_fp16 = const()[name = tensor("op_332_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_1_cast_fp16 = mul(x = var_331_cast_fp16, y = var_332_to_fp16)[name = tensor("scores_1_cast_fp16")]; + tensor var_338 = const()[name = tensor("op_338"), val = tensor([1, 1, 256, 512])]; + tensor var_339_cast_fp16 = reshape(shape = var_338, x = mask_bias_cast_fp16)[name = tensor("op_339_cast_fp16")]; + tensor scores_3_cast_fp16 = add(x = scores_1_cast_fp16, y = var_339_cast_fp16)[name = tensor("scores_3_cast_fp16")]; + tensor var_342 = const()[name = tensor("op_342"), val = tensor(-1)]; + tensor weights_1_cast_fp16 = softmax(axis = var_342, x = scores_3_cast_fp16)[name = tensor("weights_1_cast_fp16")]; + tensor attn_1_transpose_x_0 = const()[name = tensor("attn_1_transpose_x_0"), val = tensor(false)]; + tensor attn_1_transpose_y_0 = const()[name = tensor("attn_1_transpose_y_0"), val = tensor(false)]; + tensor values_1_cast_fp16 = transpose(perm = var_320, x = new_v_cache_1_cast_fp16)[name = tensor("transpose_44")]; + tensor attn_1_cast_fp16 = matmul(transpose_x = attn_1_transpose_x_0, transpose_y = attn_1_transpose_y_0, x = weights_1_cast_fp16, y = values_1_cast_fp16)[name = tensor("attn_1_cast_fp16")]; + tensor var_350 = const()[name = tensor("op_350"), val = tensor([0, 2, 1, 3])]; + tensor var_355 = const()[name = tensor("op_355"), val = tensor([1, 256, 1024])]; + tensor var_351_cast_fp16 = transpose(perm = var_350, x = attn_1_cast_fp16)[name = tensor("transpose_41")]; + tensor input_5_cast_fp16 = reshape(shape = var_355, x = var_351_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor attn0_out_proj_weight_to_fp16 = const()[name = tensor("attn0_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6304384)))]; + tensor linear_1_bias_0_to_fp16 = const()[name = tensor("linear_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401600)))]; + tensor linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn0_out_proj_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("linear_1_cast_fp16")]; + tensor var_364_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_364_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_7_cast_fp16 = add(x = conditioning_to_fp16, y = linear_1_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor input_9_axes_0 = const()[name = tensor("input_9_axes_0"), val = tensor([-1])]; + tensor norm0_2_weight_to_fp16 = const()[name = tensor("norm0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8403712)))]; + tensor norm0_2_bias_to_fp16 = const()[name = tensor("norm0_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8405824)))]; + tensor var_368_to_fp16 = const()[name = tensor("op_368_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_9_cast_fp16 = layer_norm(axes = input_9_axes_0, beta = norm0_2_bias_to_fp16, epsilon = var_368_to_fp16, gamma = norm0_2_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor linear0_1_weight_to_fp16 = const()[name = tensor("linear0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8407936)))]; + tensor linear_2_bias_0_to_fp16 = const()[name = tensor("linear_2_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16796608)))]; + tensor linear_2_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear0_1_weight_to_fp16, x = input_9_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor input_11_mode_0 = const()[name = tensor("input_11_mode_0"), val = tensor("EXACT")]; + tensor input_11_cast_fp16 = gelu(mode = input_11_mode_0, x = linear_2_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor linear0_2_weight_to_fp16 = const()[name = tensor("linear0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16804864)))]; + tensor linear_3_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear0_2_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("linear_3_cast_fp16")]; + tensor input_13_cast_fp16 = add(x = input_7_cast_fp16, y = linear_3_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor input_15_axes_0 = const()[name = tensor("input_15_axes_0"), val = tensor([-1])]; + tensor norm1_1_weight_to_fp16 = const()[name = tensor("norm1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25193536)))]; + tensor norm1_1_bias_to_fp16 = const()[name = tensor("norm1_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25195648)))]; + tensor var_385_to_fp16 = const()[name = tensor("op_385_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_15_cast_fp16 = layer_norm(axes = input_15_axes_0, beta = norm1_1_bias_to_fp16, epsilon = var_385_to_fp16, gamma = norm1_1_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor attn1_in_proj_weight_to_fp16 = const()[name = tensor("attn1_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25197760)))]; + tensor linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn1_in_proj_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor var_399 = const()[name = tensor("op_399"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_3_cast_fp16 = reshape(shape = var_399, x = linear_4_cast_fp16)[name = tensor("qkv_3_cast_fp16")]; + tensor x_5_begin_0 = const()[name = tensor("x_5_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_5_end_0 = const()[name = tensor("x_5_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_5_end_mask_0 = const()[name = tensor("x_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_5_squeeze_mask_0 = const()[name = tensor("x_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_5_cast_fp16 = slice_by_index(begin = x_5_begin_0, end = x_5_end_0, end_mask = x_5_end_mask_0, squeeze_mask = x_5_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("x_5_cast_fp16")]; + tensor x_7_begin_0 = const()[name = tensor("x_7_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_7_end_0 = const()[name = tensor("x_7_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_7_end_mask_0 = const()[name = tensor("x_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_7_squeeze_mask_0 = const()[name = tensor("x_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_7_cast_fp16 = slice_by_index(begin = x_7_begin_0, end = x_7_end_0, end_mask = x_7_end_mask_0, squeeze_mask = x_7_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("x_7_cast_fp16")]; + tensor v_3_begin_0 = const()[name = tensor("v_3_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_3_end_0 = const()[name = tensor("v_3_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_3_end_mask_0 = const()[name = tensor("v_3_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_3_squeeze_mask_0 = const()[name = tensor("v_3_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_3_cast_fp16 = slice_by_index(begin = v_3_begin_0, end = v_3_end_0, end_mask = v_3_end_mask_0, squeeze_mask = v_3_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("v_3_cast_fp16")]; + tensor var_444 = const()[name = tensor("op_444"), val = tensor([256, 16, 32, 2])]; + tensor pairs_5_cast_fp16 = reshape(shape = var_444, x = x_5_cast_fp16)[name = tensor("pairs_5_cast_fp16")]; + tensor xr_5_begin_0 = const()[name = tensor("xr_5_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_5_end_0 = const()[name = tensor("xr_5_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_5_end_mask_0 = const()[name = tensor("xr_5_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_5_squeeze_mask_0 = const()[name = tensor("xr_5_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_5_cast_fp16 = slice_by_index(begin = xr_5_begin_0, end = xr_5_end_0, end_mask = xr_5_end_mask_0, squeeze_mask = xr_5_squeeze_mask_0, x = pairs_5_cast_fp16)[name = tensor("xr_5_cast_fp16")]; + tensor xi_5_begin_0 = const()[name = tensor("xi_5_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_5_end_0 = const()[name = tensor("xi_5_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_5_end_mask_0 = const()[name = tensor("xi_5_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_5_squeeze_mask_0 = const()[name = tensor("xi_5_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_5_cast_fp16 = slice_by_index(begin = xi_5_begin_0, end = xi_5_end_0, end_mask = xi_5_end_mask_0, squeeze_mask = xi_5_squeeze_mask_0, x = pairs_5_cast_fp16)[name = tensor("xi_5_cast_fp16")]; + tensor var_452_cast_fp16 = mul(x = xr_5_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_452_cast_fp16")]; + tensor var_453_cast_fp16 = mul(x = xi_5_cast_fp16, y = roti_cast_fp16)[name = tensor("op_453_cast_fp16")]; + tensor xor__5_cast_fp16 = sub(x = var_452_cast_fp16, y = var_453_cast_fp16)[name = tensor("xor__5_cast_fp16")]; + tensor var_456_cast_fp16 = mul(x = xr_5_cast_fp16, y = roti_cast_fp16)[name = tensor("op_456_cast_fp16")]; + tensor var_457_cast_fp16 = mul(x = xi_5_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_457_cast_fp16")]; + tensor xoi_5_cast_fp16 = add(x = var_456_cast_fp16, y = var_457_cast_fp16)[name = tensor("xoi_5_cast_fp16")]; + tensor out_5_axis_0 = const()[name = tensor("out_5_axis_0"), val = tensor(-1)]; + tensor out_5_cast_fp16 = stack(axis = out_5_axis_0, values = (xor__5_cast_fp16, xoi_5_cast_fp16))[name = tensor("out_5_cast_fp16")]; + tensor var_467 = const()[name = tensor("op_467"), val = tensor([1, 256, 16, 64])]; + tensor q_3_cast_fp16 = reshape(shape = var_467, x = out_5_cast_fp16)[name = tensor("q_3_cast_fp16")]; + tensor var_473 = const()[name = tensor("op_473"), val = tensor([256, 16, 32, 2])]; + tensor pairs_7_cast_fp16 = reshape(shape = var_473, x = x_7_cast_fp16)[name = tensor("pairs_7_cast_fp16")]; + tensor xr_7_begin_0 = const()[name = tensor("xr_7_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_7_end_0 = const()[name = tensor("xr_7_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_7_end_mask_0 = const()[name = tensor("xr_7_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_7_squeeze_mask_0 = const()[name = tensor("xr_7_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_7_cast_fp16 = slice_by_index(begin = xr_7_begin_0, end = xr_7_end_0, end_mask = xr_7_end_mask_0, squeeze_mask = xr_7_squeeze_mask_0, x = pairs_7_cast_fp16)[name = tensor("xr_7_cast_fp16")]; + tensor xi_7_begin_0 = const()[name = tensor("xi_7_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_7_end_0 = const()[name = tensor("xi_7_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_7_end_mask_0 = const()[name = tensor("xi_7_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_7_squeeze_mask_0 = const()[name = tensor("xi_7_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_7_cast_fp16 = slice_by_index(begin = xi_7_begin_0, end = xi_7_end_0, end_mask = xi_7_end_mask_0, squeeze_mask = xi_7_squeeze_mask_0, x = pairs_7_cast_fp16)[name = tensor("xi_7_cast_fp16")]; + tensor var_481_cast_fp16 = mul(x = xr_7_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_481_cast_fp16")]; + tensor var_482_cast_fp16 = mul(x = xi_7_cast_fp16, y = roti_cast_fp16)[name = tensor("op_482_cast_fp16")]; + tensor xor__7_cast_fp16 = sub(x = var_481_cast_fp16, y = var_482_cast_fp16)[name = tensor("xor__7_cast_fp16")]; + tensor var_485_cast_fp16 = mul(x = xr_7_cast_fp16, y = roti_cast_fp16)[name = tensor("op_485_cast_fp16")]; + tensor var_486_cast_fp16 = mul(x = xi_7_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_486_cast_fp16")]; + tensor xoi_7_cast_fp16 = add(x = var_485_cast_fp16, y = var_486_cast_fp16)[name = tensor("xoi_7_cast_fp16")]; + tensor out_7_axis_0 = const()[name = tensor("out_7_axis_0"), val = tensor(-1)]; + tensor out_7_cast_fp16 = stack(axis = out_7_axis_0, values = (xor__7_cast_fp16, xoi_7_cast_fp16))[name = tensor("out_7_cast_fp16")]; + tensor var_503 = const()[name = tensor("op_503"), val = tensor([256, 1024])]; + tensor var_504_cast_fp16 = reshape(shape = var_503, x = out_7_cast_fp16)[name = tensor("op_504_cast_fp16")]; + tensor written_k_3_transpose_x_1 = const()[name = tensor("written_k_3_transpose_x_1"), val = tensor(true)]; + tensor written_k_3_transpose_y_1 = const()[name = tensor("written_k_3_transpose_y_1"), val = tensor(false)]; + tensor written_k_3_cast_fp16 = matmul(transpose_x = written_k_3_transpose_x_1, transpose_y = written_k_3_transpose_y_1, x = var_106_cast_fp16_to_fp16, y = var_504_cast_fp16)[name = tensor("written_k_3_cast_fp16")]; + tensor var_511 = const()[name = tensor("op_511"), val = tensor([256, 1024])]; + tensor var_512_cast_fp16 = reshape(shape = var_511, x = v_3_cast_fp16)[name = tensor("op_512_cast_fp16")]; + tensor written_v_3_transpose_x_1 = const()[name = tensor("written_v_3_transpose_x_1"), val = tensor(true)]; + tensor written_v_3_transpose_y_1 = const()[name = tensor("written_v_3_transpose_y_1"), val = tensor(false)]; + tensor written_v_3_cast_fp16 = matmul(transpose_x = written_v_3_transpose_x_1, transpose_y = written_v_3_transpose_y_1, x = var_106_cast_fp16_to_fp16, y = var_512_cast_fp16)[name = tensor("written_v_3_cast_fp16")]; + tensor k_cache1_to_fp16_dtype_0 = const()[name = tensor("k_cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache1_to_fp16 = cast(dtype = k_cache1_to_fp16_dtype_0, x = k_cache1)[name = tensor("cast_117")]; + tensor var_523_cast_fp16 = mul(x = k_cache1_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_523_cast_fp16")]; + tensor var_528 = const()[name = tensor("op_528"), val = tensor([1, 512, 16, 64])]; + tensor var_529_cast_fp16 = reshape(shape = var_528, x = written_k_3_cast_fp16)[name = tensor("op_529_cast_fp16")]; + tensor new_k_cache_3_cast_fp16 = add(x = var_523_cast_fp16, y = var_529_cast_fp16)[name = tensor("new_k_cache_3_cast_fp16")]; + tensor new_k_cache_3_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_3_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache1_to_fp16_dtype_0 = const()[name = tensor("v_cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache1_to_fp16 = cast(dtype = v_cache1_to_fp16_dtype_0, x = v_cache1)[name = tensor("cast_115")]; + tensor var_532_cast_fp16 = mul(x = v_cache1_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_532_cast_fp16")]; + tensor var_537 = const()[name = tensor("op_537"), val = tensor([1, 512, 16, 64])]; + tensor var_538_cast_fp16 = reshape(shape = var_537, x = written_v_3_cast_fp16)[name = tensor("op_538_cast_fp16")]; + tensor new_v_cache_3_cast_fp16 = add(x = var_532_cast_fp16, y = var_538_cast_fp16)[name = tensor("new_v_cache_3_cast_fp16")]; + tensor new_v_cache_3_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_3_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_551 = const()[name = tensor("op_551"), val = tensor([0, 2, 1, 3])]; + tensor var_562_transpose_x_0 = const()[name = tensor("op_562_transpose_x_0"), val = tensor(false)]; + tensor var_562_transpose_y_0 = const()[name = tensor("op_562_transpose_y_0"), val = tensor(false)]; + tensor transpose_17_perm_0 = const()[name = tensor("transpose_17_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_18_perm_0 = const()[name = tensor("transpose_18_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_18 = transpose(perm = transpose_18_perm_0, x = new_k_cache_3_cast_fp16)[name = tensor("transpose_38")]; + tensor transpose_17 = transpose(perm = transpose_17_perm_0, x = q_3_cast_fp16)[name = tensor("transpose_39")]; + tensor var_562_cast_fp16 = matmul(transpose_x = var_562_transpose_x_0, transpose_y = var_562_transpose_y_0, x = transpose_17, y = transpose_18)[name = tensor("op_562_cast_fp16")]; + tensor var_563_to_fp16 = const()[name = tensor("op_563_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_5_cast_fp16 = mul(x = var_562_cast_fp16, y = var_563_to_fp16)[name = tensor("scores_5_cast_fp16")]; + tensor scores_7_cast_fp16 = add(x = scores_5_cast_fp16, y = var_339_cast_fp16)[name = tensor("scores_7_cast_fp16")]; + tensor var_573 = const()[name = tensor("op_573"), val = tensor(-1)]; + tensor weights_3_cast_fp16 = softmax(axis = var_573, x = scores_7_cast_fp16)[name = tensor("weights_3_cast_fp16")]; + tensor attn_3_transpose_x_0 = const()[name = tensor("attn_3_transpose_x_0"), val = tensor(false)]; + tensor attn_3_transpose_y_0 = const()[name = tensor("attn_3_transpose_y_0"), val = tensor(false)]; + tensor values_3_cast_fp16 = transpose(perm = var_551, x = new_v_cache_3_cast_fp16)[name = tensor("transpose_40")]; + tensor attn_3_cast_fp16 = matmul(transpose_x = attn_3_transpose_x_0, transpose_y = attn_3_transpose_y_0, x = weights_3_cast_fp16, y = values_3_cast_fp16)[name = tensor("attn_3_cast_fp16")]; + tensor var_581 = const()[name = tensor("op_581"), val = tensor([0, 2, 1, 3])]; + tensor var_586 = const()[name = tensor("op_586"), val = tensor([1, 256, 1024])]; + tensor var_582_cast_fp16 = transpose(perm = var_581, x = attn_3_cast_fp16)[name = tensor("transpose_37")]; + tensor input_17_cast_fp16 = reshape(shape = var_586, x = var_582_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor attn1_out_proj_weight_to_fp16 = const()[name = tensor("attn1_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31489280)))]; + tensor linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn1_out_proj_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor position1_to_fp16_dtype_0 = const()[name = tensor("position1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position1_to_fp16 = cast(dtype = position1_to_fp16_dtype_0, x = position1)[name = tensor("cast_113")]; + tensor var_595_cast_fp16 = add(x = position1_to_fp16, y = valid_len_to_fp16)[name = tensor("op_595_cast_fp16")]; + tensor var_595_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_595_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_19_cast_fp16 = add(x = input_13_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor input_21_axes_0 = const()[name = tensor("input_21_axes_0"), val = tensor([-1])]; + tensor norm1_2_weight_to_fp16 = const()[name = tensor("norm1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33586496)))]; + tensor norm1_2_bias_to_fp16 = const()[name = tensor("norm1_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33588608)))]; + tensor var_599_to_fp16 = const()[name = tensor("op_599_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_21_cast_fp16 = layer_norm(axes = input_21_axes_0, beta = norm1_2_bias_to_fp16, epsilon = var_599_to_fp16, gamma = norm1_2_weight_to_fp16, x = input_19_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor linear1_1_weight_to_fp16 = const()[name = tensor("linear1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33590720)))]; + tensor linear_6_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear1_1_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor input_23_mode_0 = const()[name = tensor("input_23_mode_0"), val = tensor("EXACT")]; + tensor input_23_cast_fp16 = gelu(mode = input_23_mode_0, x = linear_6_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor linear1_2_weight_to_fp16 = const()[name = tensor("linear1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41979392)))]; + tensor linear_7_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear1_2_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_25_cast_fp16 = add(x = input_19_cast_fp16, y = linear_7_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor input_27_axes_0 = const()[name = tensor("input_27_axes_0"), val = tensor([-1])]; + tensor norm2_1_weight_to_fp16 = const()[name = tensor("norm2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50368064)))]; + tensor norm2_1_bias_to_fp16 = const()[name = tensor("norm2_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50370176)))]; + tensor var_616_to_fp16 = const()[name = tensor("op_616_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = norm2_1_bias_to_fp16, epsilon = var_616_to_fp16, gamma = norm2_1_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor attn2_in_proj_weight_to_fp16 = const()[name = tensor("attn2_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50372288)))]; + tensor linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn2_in_proj_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor var_630 = const()[name = tensor("op_630"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_5_cast_fp16 = reshape(shape = var_630, x = linear_8_cast_fp16)[name = tensor("qkv_5_cast_fp16")]; + tensor x_9_begin_0 = const()[name = tensor("x_9_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_9_end_0 = const()[name = tensor("x_9_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_9_end_mask_0 = const()[name = tensor("x_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_9_squeeze_mask_0 = const()[name = tensor("x_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_9_cast_fp16 = slice_by_index(begin = x_9_begin_0, end = x_9_end_0, end_mask = x_9_end_mask_0, squeeze_mask = x_9_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("x_9_cast_fp16")]; + tensor x_11_begin_0 = const()[name = tensor("x_11_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_11_end_0 = const()[name = tensor("x_11_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_11_end_mask_0 = const()[name = tensor("x_11_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_11_squeeze_mask_0 = const()[name = tensor("x_11_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_11_cast_fp16 = slice_by_index(begin = x_11_begin_0, end = x_11_end_0, end_mask = x_11_end_mask_0, squeeze_mask = x_11_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("x_11_cast_fp16")]; + tensor v_5_begin_0 = const()[name = tensor("v_5_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_5_end_0 = const()[name = tensor("v_5_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_5_end_mask_0 = const()[name = tensor("v_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_5_squeeze_mask_0 = const()[name = tensor("v_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_5_cast_fp16 = slice_by_index(begin = v_5_begin_0, end = v_5_end_0, end_mask = v_5_end_mask_0, squeeze_mask = v_5_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("v_5_cast_fp16")]; + tensor var_675 = const()[name = tensor("op_675"), val = tensor([256, 16, 32, 2])]; + tensor pairs_9_cast_fp16 = reshape(shape = var_675, x = x_9_cast_fp16)[name = tensor("pairs_9_cast_fp16")]; + tensor xr_9_begin_0 = const()[name = tensor("xr_9_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_9_end_0 = const()[name = tensor("xr_9_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_9_end_mask_0 = const()[name = tensor("xr_9_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_9_squeeze_mask_0 = const()[name = tensor("xr_9_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_9_cast_fp16 = slice_by_index(begin = xr_9_begin_0, end = xr_9_end_0, end_mask = xr_9_end_mask_0, squeeze_mask = xr_9_squeeze_mask_0, x = pairs_9_cast_fp16)[name = tensor("xr_9_cast_fp16")]; + tensor xi_9_begin_0 = const()[name = tensor("xi_9_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_9_end_0 = const()[name = tensor("xi_9_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_9_end_mask_0 = const()[name = tensor("xi_9_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_9_squeeze_mask_0 = const()[name = tensor("xi_9_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_9_cast_fp16 = slice_by_index(begin = xi_9_begin_0, end = xi_9_end_0, end_mask = xi_9_end_mask_0, squeeze_mask = xi_9_squeeze_mask_0, x = pairs_9_cast_fp16)[name = tensor("xi_9_cast_fp16")]; + tensor var_683_cast_fp16 = mul(x = xr_9_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_683_cast_fp16")]; + tensor var_684_cast_fp16 = mul(x = xi_9_cast_fp16, y = roti_cast_fp16)[name = tensor("op_684_cast_fp16")]; + tensor xor__9_cast_fp16 = sub(x = var_683_cast_fp16, y = var_684_cast_fp16)[name = tensor("xor__9_cast_fp16")]; + tensor var_687_cast_fp16 = mul(x = xr_9_cast_fp16, y = roti_cast_fp16)[name = tensor("op_687_cast_fp16")]; + tensor var_688_cast_fp16 = mul(x = xi_9_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_688_cast_fp16")]; + tensor xoi_9_cast_fp16 = add(x = var_687_cast_fp16, y = var_688_cast_fp16)[name = tensor("xoi_9_cast_fp16")]; + tensor out_9_axis_0 = const()[name = tensor("out_9_axis_0"), val = tensor(-1)]; + tensor out_9_cast_fp16 = stack(axis = out_9_axis_0, values = (xor__9_cast_fp16, xoi_9_cast_fp16))[name = tensor("out_9_cast_fp16")]; + tensor var_698 = const()[name = tensor("op_698"), val = tensor([1, 256, 16, 64])]; + tensor q_5_cast_fp16 = reshape(shape = var_698, x = out_9_cast_fp16)[name = tensor("q_5_cast_fp16")]; + tensor var_704 = const()[name = tensor("op_704"), val = tensor([256, 16, 32, 2])]; + tensor pairs_11_cast_fp16 = reshape(shape = var_704, x = x_11_cast_fp16)[name = tensor("pairs_11_cast_fp16")]; + tensor xr_11_begin_0 = const()[name = tensor("xr_11_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_11_end_0 = const()[name = tensor("xr_11_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_11_end_mask_0 = const()[name = tensor("xr_11_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_11_squeeze_mask_0 = const()[name = tensor("xr_11_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_11_cast_fp16 = slice_by_index(begin = xr_11_begin_0, end = xr_11_end_0, end_mask = xr_11_end_mask_0, squeeze_mask = xr_11_squeeze_mask_0, x = pairs_11_cast_fp16)[name = tensor("xr_11_cast_fp16")]; + tensor xi_11_begin_0 = const()[name = tensor("xi_11_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_11_end_0 = const()[name = tensor("xi_11_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_11_end_mask_0 = const()[name = tensor("xi_11_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_11_squeeze_mask_0 = const()[name = tensor("xi_11_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_11_cast_fp16 = slice_by_index(begin = xi_11_begin_0, end = xi_11_end_0, end_mask = xi_11_end_mask_0, squeeze_mask = xi_11_squeeze_mask_0, x = pairs_11_cast_fp16)[name = tensor("xi_11_cast_fp16")]; + tensor var_712_cast_fp16 = mul(x = xr_11_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_712_cast_fp16")]; + tensor var_713_cast_fp16 = mul(x = xi_11_cast_fp16, y = roti_cast_fp16)[name = tensor("op_713_cast_fp16")]; + tensor xor__11_cast_fp16 = sub(x = var_712_cast_fp16, y = var_713_cast_fp16)[name = tensor("xor__11_cast_fp16")]; + tensor var_716_cast_fp16 = mul(x = xr_11_cast_fp16, y = roti_cast_fp16)[name = tensor("op_716_cast_fp16")]; + tensor var_717_cast_fp16 = mul(x = xi_11_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_717_cast_fp16")]; + tensor xoi_11_cast_fp16 = add(x = var_716_cast_fp16, y = var_717_cast_fp16)[name = tensor("xoi_11_cast_fp16")]; + tensor out_11_axis_0 = const()[name = tensor("out_11_axis_0"), val = tensor(-1)]; + tensor out_11_cast_fp16 = stack(axis = out_11_axis_0, values = (xor__11_cast_fp16, xoi_11_cast_fp16))[name = tensor("out_11_cast_fp16")]; + tensor var_734 = const()[name = tensor("op_734"), val = tensor([256, 1024])]; + tensor var_735_cast_fp16 = reshape(shape = var_734, x = out_11_cast_fp16)[name = tensor("op_735_cast_fp16")]; + tensor written_k_5_transpose_x_1 = const()[name = tensor("written_k_5_transpose_x_1"), val = tensor(true)]; + tensor written_k_5_transpose_y_1 = const()[name = tensor("written_k_5_transpose_y_1"), val = tensor(false)]; + tensor written_k_5_cast_fp16 = matmul(transpose_x = written_k_5_transpose_x_1, transpose_y = written_k_5_transpose_y_1, x = var_106_cast_fp16_to_fp16, y = var_735_cast_fp16)[name = tensor("written_k_5_cast_fp16")]; + tensor var_742 = const()[name = tensor("op_742"), val = tensor([256, 1024])]; + tensor var_743_cast_fp16 = reshape(shape = var_742, x = v_5_cast_fp16)[name = tensor("op_743_cast_fp16")]; + tensor written_v_5_transpose_x_1 = const()[name = tensor("written_v_5_transpose_x_1"), val = tensor(true)]; + tensor written_v_5_transpose_y_1 = const()[name = tensor("written_v_5_transpose_y_1"), val = tensor(false)]; + tensor written_v_5_cast_fp16 = matmul(transpose_x = written_v_5_transpose_x_1, transpose_y = written_v_5_transpose_y_1, x = var_106_cast_fp16_to_fp16, y = var_743_cast_fp16)[name = tensor("written_v_5_cast_fp16")]; + tensor k_cache2_to_fp16_dtype_0 = const()[name = tensor("k_cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache2_to_fp16 = cast(dtype = k_cache2_to_fp16_dtype_0, x = k_cache2)[name = tensor("cast_111")]; + tensor var_754_cast_fp16 = mul(x = k_cache2_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_754_cast_fp16")]; + tensor var_759 = const()[name = tensor("op_759"), val = tensor([1, 512, 16, 64])]; + tensor var_760_cast_fp16 = reshape(shape = var_759, x = written_k_5_cast_fp16)[name = tensor("op_760_cast_fp16")]; + tensor new_k_cache_5_cast_fp16 = add(x = var_754_cast_fp16, y = var_760_cast_fp16)[name = tensor("new_k_cache_5_cast_fp16")]; + tensor new_k_cache_5_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_5_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache2_to_fp16_dtype_0 = const()[name = tensor("v_cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache2_to_fp16 = cast(dtype = v_cache2_to_fp16_dtype_0, x = v_cache2)[name = tensor("cast_109")]; + tensor var_763_cast_fp16 = mul(x = v_cache2_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_763_cast_fp16")]; + tensor var_768 = const()[name = tensor("op_768"), val = tensor([1, 512, 16, 64])]; + tensor var_769_cast_fp16 = reshape(shape = var_768, x = written_v_5_cast_fp16)[name = tensor("op_769_cast_fp16")]; + tensor new_v_cache_5_cast_fp16 = add(x = var_763_cast_fp16, y = var_769_cast_fp16)[name = tensor("new_v_cache_5_cast_fp16")]; + tensor new_v_cache_5_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_5_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_782 = const()[name = tensor("op_782"), val = tensor([0, 2, 1, 3])]; + tensor var_793_transpose_x_0 = const()[name = tensor("op_793_transpose_x_0"), val = tensor(false)]; + tensor var_793_transpose_y_0 = const()[name = tensor("op_793_transpose_y_0"), val = tensor(false)]; + tensor transpose_19_perm_0 = const()[name = tensor("transpose_19_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_20_perm_0 = const()[name = tensor("transpose_20_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_20 = transpose(perm = transpose_20_perm_0, x = new_k_cache_5_cast_fp16)[name = tensor("transpose_34")]; + tensor transpose_19 = transpose(perm = transpose_19_perm_0, x = q_5_cast_fp16)[name = tensor("transpose_35")]; + tensor var_793_cast_fp16 = matmul(transpose_x = var_793_transpose_x_0, transpose_y = var_793_transpose_y_0, x = transpose_19, y = transpose_20)[name = tensor("op_793_cast_fp16")]; + tensor var_794_to_fp16 = const()[name = tensor("op_794_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_9_cast_fp16 = mul(x = var_793_cast_fp16, y = var_794_to_fp16)[name = tensor("scores_9_cast_fp16")]; + tensor scores_11_cast_fp16 = add(x = scores_9_cast_fp16, y = var_339_cast_fp16)[name = tensor("scores_11_cast_fp16")]; + tensor var_804 = const()[name = tensor("op_804"), val = tensor(-1)]; + tensor weights_5_cast_fp16 = softmax(axis = var_804, x = scores_11_cast_fp16)[name = tensor("weights_5_cast_fp16")]; + tensor attn_5_transpose_x_0 = const()[name = tensor("attn_5_transpose_x_0"), val = tensor(false)]; + tensor attn_5_transpose_y_0 = const()[name = tensor("attn_5_transpose_y_0"), val = tensor(false)]; + tensor values_5_cast_fp16 = transpose(perm = var_782, x = new_v_cache_5_cast_fp16)[name = tensor("transpose_36")]; + tensor attn_5_cast_fp16 = matmul(transpose_x = attn_5_transpose_x_0, transpose_y = attn_5_transpose_y_0, x = weights_5_cast_fp16, y = values_5_cast_fp16)[name = tensor("attn_5_cast_fp16")]; + tensor var_812 = const()[name = tensor("op_812"), val = tensor([0, 2, 1, 3])]; + tensor var_817 = const()[name = tensor("op_817"), val = tensor([1, 256, 1024])]; + tensor var_813_cast_fp16 = transpose(perm = var_812, x = attn_5_cast_fp16)[name = tensor("transpose_33")]; + tensor input_29_cast_fp16 = reshape(shape = var_817, x = var_813_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor attn2_out_proj_weight_to_fp16 = const()[name = tensor("attn2_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(56663808)))]; + tensor linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn2_out_proj_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor position2_to_fp16_dtype_0 = const()[name = tensor("position2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position2_to_fp16 = cast(dtype = position2_to_fp16_dtype_0, x = position2)[name = tensor("cast_107")]; + tensor var_826_cast_fp16 = add(x = position2_to_fp16, y = valid_len_to_fp16)[name = tensor("op_826_cast_fp16")]; + tensor var_826_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_826_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_31_cast_fp16 = add(x = input_25_cast_fp16, y = linear_9_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; + tensor norm2_2_weight_to_fp16 = const()[name = tensor("norm2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58761024)))]; + tensor norm2_2_bias_to_fp16 = const()[name = tensor("norm2_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58763136)))]; + tensor var_830_to_fp16 = const()[name = tensor("op_830_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_33_cast_fp16 = layer_norm(axes = input_33_axes_0, beta = norm2_2_bias_to_fp16, epsilon = var_830_to_fp16, gamma = norm2_2_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor linear2_1_weight_to_fp16 = const()[name = tensor("linear2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58765248)))]; + tensor linear_10_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear2_1_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor input_35_mode_0 = const()[name = tensor("input_35_mode_0"), val = tensor("EXACT")]; + tensor input_35_cast_fp16 = gelu(mode = input_35_mode_0, x = linear_10_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor linear2_2_weight_to_fp16 = const()[name = tensor("linear2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67153920)))]; + tensor linear_11_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear2_2_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor input_37_cast_fp16 = add(x = input_31_cast_fp16, y = linear_11_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor input_39_axes_0 = const()[name = tensor("input_39_axes_0"), val = tensor([-1])]; + tensor norm3_1_weight_to_fp16 = const()[name = tensor("norm3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75542592)))]; + tensor norm3_1_bias_to_fp16 = const()[name = tensor("norm3_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75544704)))]; + tensor var_847_to_fp16 = const()[name = tensor("op_847_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_39_cast_fp16 = layer_norm(axes = input_39_axes_0, beta = norm3_1_bias_to_fp16, epsilon = var_847_to_fp16, gamma = norm3_1_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor attn3_in_proj_weight_to_fp16 = const()[name = tensor("attn3_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75546816)))]; + tensor linear_12_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn3_in_proj_weight_to_fp16, x = input_39_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor var_861 = const()[name = tensor("op_861"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_7_cast_fp16 = reshape(shape = var_861, x = linear_12_cast_fp16)[name = tensor("qkv_7_cast_fp16")]; + tensor x_13_begin_0 = const()[name = tensor("x_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_13_end_0 = const()[name = tensor("x_13_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_13_end_mask_0 = const()[name = tensor("x_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_13_squeeze_mask_0 = const()[name = tensor("x_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_13_cast_fp16 = slice_by_index(begin = x_13_begin_0, end = x_13_end_0, end_mask = x_13_end_mask_0, squeeze_mask = x_13_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("x_13_cast_fp16")]; + tensor x_15_begin_0 = const()[name = tensor("x_15_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_15_end_0 = const()[name = tensor("x_15_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_15_end_mask_0 = const()[name = tensor("x_15_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_15_squeeze_mask_0 = const()[name = tensor("x_15_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_15_cast_fp16 = slice_by_index(begin = x_15_begin_0, end = x_15_end_0, end_mask = x_15_end_mask_0, squeeze_mask = x_15_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("x_15_cast_fp16")]; + tensor v_7_begin_0 = const()[name = tensor("v_7_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_7_end_0 = const()[name = tensor("v_7_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_7_end_mask_0 = const()[name = tensor("v_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_7_squeeze_mask_0 = const()[name = tensor("v_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_7_cast_fp16 = slice_by_index(begin = v_7_begin_0, end = v_7_end_0, end_mask = v_7_end_mask_0, squeeze_mask = v_7_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("v_7_cast_fp16")]; + tensor var_906 = const()[name = tensor("op_906"), val = tensor([256, 16, 32, 2])]; + tensor pairs_13_cast_fp16 = reshape(shape = var_906, x = x_13_cast_fp16)[name = tensor("pairs_13_cast_fp16")]; + tensor xr_13_begin_0 = const()[name = tensor("xr_13_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_13_end_0 = const()[name = tensor("xr_13_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_13_end_mask_0 = const()[name = tensor("xr_13_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_13_squeeze_mask_0 = const()[name = tensor("xr_13_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_13_cast_fp16 = slice_by_index(begin = xr_13_begin_0, end = xr_13_end_0, end_mask = xr_13_end_mask_0, squeeze_mask = xr_13_squeeze_mask_0, x = pairs_13_cast_fp16)[name = tensor("xr_13_cast_fp16")]; + tensor xi_13_begin_0 = const()[name = tensor("xi_13_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_13_end_0 = const()[name = tensor("xi_13_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_13_end_mask_0 = const()[name = tensor("xi_13_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_13_squeeze_mask_0 = const()[name = tensor("xi_13_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_13_cast_fp16 = slice_by_index(begin = xi_13_begin_0, end = xi_13_end_0, end_mask = xi_13_end_mask_0, squeeze_mask = xi_13_squeeze_mask_0, x = pairs_13_cast_fp16)[name = tensor("xi_13_cast_fp16")]; + tensor var_914_cast_fp16 = mul(x = xr_13_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_914_cast_fp16")]; + tensor var_915_cast_fp16 = mul(x = xi_13_cast_fp16, y = roti_cast_fp16)[name = tensor("op_915_cast_fp16")]; + tensor xor__13_cast_fp16 = sub(x = var_914_cast_fp16, y = var_915_cast_fp16)[name = tensor("xor__13_cast_fp16")]; + tensor var_918_cast_fp16 = mul(x = xr_13_cast_fp16, y = roti_cast_fp16)[name = tensor("op_918_cast_fp16")]; + tensor var_919_cast_fp16 = mul(x = xi_13_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_919_cast_fp16")]; + tensor xoi_13_cast_fp16 = add(x = var_918_cast_fp16, y = var_919_cast_fp16)[name = tensor("xoi_13_cast_fp16")]; + tensor out_13_axis_0 = const()[name = tensor("out_13_axis_0"), val = tensor(-1)]; + tensor out_13_cast_fp16 = stack(axis = out_13_axis_0, values = (xor__13_cast_fp16, xoi_13_cast_fp16))[name = tensor("out_13_cast_fp16")]; + tensor var_929 = const()[name = tensor("op_929"), val = tensor([1, 256, 16, 64])]; + tensor q_7_cast_fp16 = reshape(shape = var_929, x = out_13_cast_fp16)[name = tensor("q_7_cast_fp16")]; + tensor var_935 = const()[name = tensor("op_935"), val = tensor([256, 16, 32, 2])]; + tensor pairs_15_cast_fp16 = reshape(shape = var_935, x = x_15_cast_fp16)[name = tensor("pairs_15_cast_fp16")]; + tensor xr_15_begin_0 = const()[name = tensor("xr_15_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_15_end_0 = const()[name = tensor("xr_15_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_15_end_mask_0 = const()[name = tensor("xr_15_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_15_squeeze_mask_0 = const()[name = tensor("xr_15_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_15_cast_fp16 = slice_by_index(begin = xr_15_begin_0, end = xr_15_end_0, end_mask = xr_15_end_mask_0, squeeze_mask = xr_15_squeeze_mask_0, x = pairs_15_cast_fp16)[name = tensor("xr_15_cast_fp16")]; + tensor xi_15_begin_0 = const()[name = tensor("xi_15_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_15_end_0 = const()[name = tensor("xi_15_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_15_end_mask_0 = const()[name = tensor("xi_15_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_15_squeeze_mask_0 = const()[name = tensor("xi_15_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_15_cast_fp16 = slice_by_index(begin = xi_15_begin_0, end = xi_15_end_0, end_mask = xi_15_end_mask_0, squeeze_mask = xi_15_squeeze_mask_0, x = pairs_15_cast_fp16)[name = tensor("xi_15_cast_fp16")]; + tensor var_943_cast_fp16 = mul(x = xr_15_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_943_cast_fp16")]; + tensor var_944_cast_fp16 = mul(x = xi_15_cast_fp16, y = roti_cast_fp16)[name = tensor("op_944_cast_fp16")]; + tensor xor__15_cast_fp16 = sub(x = var_943_cast_fp16, y = var_944_cast_fp16)[name = tensor("xor__15_cast_fp16")]; + tensor var_947_cast_fp16 = mul(x = xr_15_cast_fp16, y = roti_cast_fp16)[name = tensor("op_947_cast_fp16")]; + tensor var_948_cast_fp16 = mul(x = xi_15_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_948_cast_fp16")]; + tensor xoi_15_cast_fp16 = add(x = var_947_cast_fp16, y = var_948_cast_fp16)[name = tensor("xoi_15_cast_fp16")]; + tensor out_15_axis_0 = const()[name = tensor("out_15_axis_0"), val = tensor(-1)]; + tensor out_15_cast_fp16 = stack(axis = out_15_axis_0, values = (xor__15_cast_fp16, xoi_15_cast_fp16))[name = tensor("out_15_cast_fp16")]; + tensor var_965 = const()[name = tensor("op_965"), val = tensor([256, 1024])]; + tensor var_966_cast_fp16 = reshape(shape = var_965, x = out_15_cast_fp16)[name = tensor("op_966_cast_fp16")]; + tensor written_k_7_transpose_x_1 = const()[name = tensor("written_k_7_transpose_x_1"), val = tensor(true)]; + tensor written_k_7_transpose_y_1 = const()[name = tensor("written_k_7_transpose_y_1"), val = tensor(false)]; + tensor written_k_7_cast_fp16 = matmul(transpose_x = written_k_7_transpose_x_1, transpose_y = written_k_7_transpose_y_1, x = var_106_cast_fp16_to_fp16, y = var_966_cast_fp16)[name = tensor("written_k_7_cast_fp16")]; + tensor var_973 = const()[name = tensor("op_973"), val = tensor([256, 1024])]; + tensor var_974_cast_fp16 = reshape(shape = var_973, x = v_7_cast_fp16)[name = tensor("op_974_cast_fp16")]; + tensor written_v_7_transpose_x_1 = const()[name = tensor("written_v_7_transpose_x_1"), val = tensor(true)]; + tensor written_v_7_transpose_y_1 = const()[name = tensor("written_v_7_transpose_y_1"), val = tensor(false)]; + tensor written_v_7_cast_fp16 = matmul(transpose_x = written_v_7_transpose_x_1, transpose_y = written_v_7_transpose_y_1, x = var_106_cast_fp16_to_fp16, y = var_974_cast_fp16)[name = tensor("written_v_7_cast_fp16")]; + tensor k_cache3_to_fp16_dtype_0 = const()[name = tensor("k_cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache3_to_fp16 = cast(dtype = k_cache3_to_fp16_dtype_0, x = k_cache3)[name = tensor("cast_105")]; + tensor var_985_cast_fp16 = mul(x = k_cache3_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_985_cast_fp16")]; + tensor var_990 = const()[name = tensor("op_990"), val = tensor([1, 512, 16, 64])]; + tensor var_991_cast_fp16 = reshape(shape = var_990, x = written_k_7_cast_fp16)[name = tensor("op_991_cast_fp16")]; + tensor new_k_cache_7_cast_fp16 = add(x = var_985_cast_fp16, y = var_991_cast_fp16)[name = tensor("new_k_cache_7_cast_fp16")]; + tensor new_k_cache_7_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_7_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache3_to_fp16_dtype_0 = const()[name = tensor("v_cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache3_to_fp16 = cast(dtype = v_cache3_to_fp16_dtype_0, x = v_cache3)[name = tensor("cast_103")]; + tensor var_994_cast_fp16 = mul(x = v_cache3_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_994_cast_fp16")]; + tensor var_999 = const()[name = tensor("op_999"), val = tensor([1, 512, 16, 64])]; + tensor var_1000_cast_fp16 = reshape(shape = var_999, x = written_v_7_cast_fp16)[name = tensor("op_1000_cast_fp16")]; + tensor new_v_cache_7_cast_fp16 = add(x = var_994_cast_fp16, y = var_1000_cast_fp16)[name = tensor("new_v_cache_7_cast_fp16")]; + tensor new_v_cache_7_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_7_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_1013 = const()[name = tensor("op_1013"), val = tensor([0, 2, 1, 3])]; + tensor var_1024_transpose_x_0 = const()[name = tensor("op_1024_transpose_x_0"), val = tensor(false)]; + tensor var_1024_transpose_y_0 = const()[name = tensor("op_1024_transpose_y_0"), val = tensor(false)]; + tensor transpose_21_perm_0 = const()[name = tensor("transpose_21_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_22_perm_0 = const()[name = tensor("transpose_22_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_22 = transpose(perm = transpose_22_perm_0, x = new_k_cache_7_cast_fp16)[name = tensor("transpose_30")]; + tensor transpose_21 = transpose(perm = transpose_21_perm_0, x = q_7_cast_fp16)[name = tensor("transpose_31")]; + tensor var_1024_cast_fp16 = matmul(transpose_x = var_1024_transpose_x_0, transpose_y = var_1024_transpose_y_0, x = transpose_21, y = transpose_22)[name = tensor("op_1024_cast_fp16")]; + tensor var_1025_to_fp16 = const()[name = tensor("op_1025_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_13_cast_fp16 = mul(x = var_1024_cast_fp16, y = var_1025_to_fp16)[name = tensor("scores_13_cast_fp16")]; + tensor scores_15_cast_fp16 = add(x = scores_13_cast_fp16, y = var_339_cast_fp16)[name = tensor("scores_15_cast_fp16")]; + tensor var_1035 = const()[name = tensor("op_1035"), val = tensor(-1)]; + tensor weights_7_cast_fp16 = softmax(axis = var_1035, x = scores_15_cast_fp16)[name = tensor("weights_7_cast_fp16")]; + tensor attn_7_transpose_x_0 = const()[name = tensor("attn_7_transpose_x_0"), val = tensor(false)]; + tensor attn_7_transpose_y_0 = const()[name = tensor("attn_7_transpose_y_0"), val = tensor(false)]; + tensor values_7_cast_fp16 = transpose(perm = var_1013, x = new_v_cache_7_cast_fp16)[name = tensor("transpose_32")]; + tensor attn_7_cast_fp16 = matmul(transpose_x = attn_7_transpose_x_0, transpose_y = attn_7_transpose_y_0, x = weights_7_cast_fp16, y = values_7_cast_fp16)[name = tensor("attn_7_cast_fp16")]; + tensor var_1043 = const()[name = tensor("op_1043"), val = tensor([0, 2, 1, 3])]; + tensor var_1048 = const()[name = tensor("op_1048"), val = tensor([1, 256, 1024])]; + tensor var_1044_cast_fp16 = transpose(perm = var_1043, x = attn_7_cast_fp16)[name = tensor("transpose_29")]; + tensor input_41_cast_fp16 = reshape(shape = var_1048, x = var_1044_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor attn3_out_proj_weight_to_fp16 = const()[name = tensor("attn3_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(81838336)))]; + tensor linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn3_out_proj_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor position3_to_fp16_dtype_0 = const()[name = tensor("position3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position3_to_fp16 = cast(dtype = position3_to_fp16_dtype_0, x = position3)[name = tensor("cast_101")]; + tensor var_1057_cast_fp16 = add(x = position3_to_fp16, y = valid_len_to_fp16)[name = tensor("op_1057_cast_fp16")]; + tensor var_1057_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1057_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_43_cast_fp16 = add(x = input_37_cast_fp16, y = linear_13_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; + tensor norm3_2_weight_to_fp16 = const()[name = tensor("norm3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(83935552)))]; + tensor norm3_2_bias_to_fp16 = const()[name = tensor("norm3_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(83937664)))]; + tensor var_1061_to_fp16 = const()[name = tensor("op_1061_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_45_cast_fp16 = layer_norm(axes = input_45_axes_0, beta = norm3_2_bias_to_fp16, epsilon = var_1061_to_fp16, gamma = norm3_2_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor linear3_1_weight_to_fp16 = const()[name = tensor("linear3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(83939776)))]; + tensor linear_14_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear3_1_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor input_47_mode_0 = const()[name = tensor("input_47_mode_0"), val = tensor("EXACT")]; + tensor input_47_cast_fp16 = gelu(mode = input_47_mode_0, x = linear_14_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor linear3_2_weight_to_fp16 = const()[name = tensor("linear3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(92328448)))]; + tensor linear_15_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear3_2_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor input_49_cast_fp16 = add(x = input_43_cast_fp16, y = linear_15_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor input_51_axes_0 = const()[name = tensor("input_51_axes_0"), val = tensor([-1])]; + tensor norm4_1_weight_to_fp16 = const()[name = tensor("norm4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100717120)))]; + tensor norm4_1_bias_to_fp16 = const()[name = tensor("norm4_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100719232)))]; + tensor var_1078_to_fp16 = const()[name = tensor("op_1078_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_51_cast_fp16 = layer_norm(axes = input_51_axes_0, beta = norm4_1_bias_to_fp16, epsilon = var_1078_to_fp16, gamma = norm4_1_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor attn4_in_proj_weight_to_fp16 = const()[name = tensor("attn4_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100721344)))]; + tensor linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn4_in_proj_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor var_1092 = const()[name = tensor("op_1092"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_9_cast_fp16 = reshape(shape = var_1092, x = linear_16_cast_fp16)[name = tensor("qkv_9_cast_fp16")]; + tensor x_17_begin_0 = const()[name = tensor("x_17_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_17_end_0 = const()[name = tensor("x_17_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_17_end_mask_0 = const()[name = tensor("x_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_17_squeeze_mask_0 = const()[name = tensor("x_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_17_cast_fp16 = slice_by_index(begin = x_17_begin_0, end = x_17_end_0, end_mask = x_17_end_mask_0, squeeze_mask = x_17_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("x_17_cast_fp16")]; + tensor x_19_begin_0 = const()[name = tensor("x_19_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_19_end_0 = const()[name = tensor("x_19_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_19_end_mask_0 = const()[name = tensor("x_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_19_squeeze_mask_0 = const()[name = tensor("x_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_19_cast_fp16 = slice_by_index(begin = x_19_begin_0, end = x_19_end_0, end_mask = x_19_end_mask_0, squeeze_mask = x_19_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("x_19_cast_fp16")]; + tensor v_9_begin_0 = const()[name = tensor("v_9_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_9_end_0 = const()[name = tensor("v_9_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_9_end_mask_0 = const()[name = tensor("v_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_9_squeeze_mask_0 = const()[name = tensor("v_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_9_cast_fp16 = slice_by_index(begin = v_9_begin_0, end = v_9_end_0, end_mask = v_9_end_mask_0, squeeze_mask = v_9_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("v_9_cast_fp16")]; + tensor var_1137 = const()[name = tensor("op_1137"), val = tensor([256, 16, 32, 2])]; + tensor pairs_17_cast_fp16 = reshape(shape = var_1137, x = x_17_cast_fp16)[name = tensor("pairs_17_cast_fp16")]; + tensor xr_17_begin_0 = const()[name = tensor("xr_17_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_17_end_0 = const()[name = tensor("xr_17_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_17_end_mask_0 = const()[name = tensor("xr_17_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_17_squeeze_mask_0 = const()[name = tensor("xr_17_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_17_cast_fp16 = slice_by_index(begin = xr_17_begin_0, end = xr_17_end_0, end_mask = xr_17_end_mask_0, squeeze_mask = xr_17_squeeze_mask_0, x = pairs_17_cast_fp16)[name = tensor("xr_17_cast_fp16")]; + tensor xi_17_begin_0 = const()[name = tensor("xi_17_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_17_end_0 = const()[name = tensor("xi_17_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_17_end_mask_0 = const()[name = tensor("xi_17_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_17_squeeze_mask_0 = const()[name = tensor("xi_17_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_17_cast_fp16 = slice_by_index(begin = xi_17_begin_0, end = xi_17_end_0, end_mask = xi_17_end_mask_0, squeeze_mask = xi_17_squeeze_mask_0, x = pairs_17_cast_fp16)[name = tensor("xi_17_cast_fp16")]; + tensor var_1145_cast_fp16 = mul(x = xr_17_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1145_cast_fp16")]; + tensor var_1146_cast_fp16 = mul(x = xi_17_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1146_cast_fp16")]; + tensor xor__17_cast_fp16 = sub(x = var_1145_cast_fp16, y = var_1146_cast_fp16)[name = tensor("xor__17_cast_fp16")]; + tensor var_1149_cast_fp16 = mul(x = xr_17_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1149_cast_fp16")]; + tensor var_1150_cast_fp16 = mul(x = xi_17_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1150_cast_fp16")]; + tensor xoi_17_cast_fp16 = add(x = var_1149_cast_fp16, y = var_1150_cast_fp16)[name = tensor("xoi_17_cast_fp16")]; + tensor out_17_axis_0 = const()[name = tensor("out_17_axis_0"), val = tensor(-1)]; + tensor out_17_cast_fp16 = stack(axis = out_17_axis_0, values = (xor__17_cast_fp16, xoi_17_cast_fp16))[name = tensor("out_17_cast_fp16")]; + tensor var_1160 = const()[name = tensor("op_1160"), val = tensor([1, 256, 16, 64])]; + tensor q_9_cast_fp16 = reshape(shape = var_1160, x = out_17_cast_fp16)[name = tensor("q_9_cast_fp16")]; + tensor var_1166 = const()[name = tensor("op_1166"), val = tensor([256, 16, 32, 2])]; + tensor pairs_19_cast_fp16 = reshape(shape = var_1166, x = x_19_cast_fp16)[name = tensor("pairs_19_cast_fp16")]; + tensor xr_19_begin_0 = const()[name = tensor("xr_19_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_19_end_0 = const()[name = tensor("xr_19_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_19_end_mask_0 = const()[name = tensor("xr_19_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_19_squeeze_mask_0 = const()[name = tensor("xr_19_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_19_cast_fp16 = slice_by_index(begin = xr_19_begin_0, end = xr_19_end_0, end_mask = xr_19_end_mask_0, squeeze_mask = xr_19_squeeze_mask_0, x = pairs_19_cast_fp16)[name = tensor("xr_19_cast_fp16")]; + tensor xi_19_begin_0 = const()[name = tensor("xi_19_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_19_end_0 = const()[name = tensor("xi_19_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_19_end_mask_0 = const()[name = tensor("xi_19_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_19_squeeze_mask_0 = const()[name = tensor("xi_19_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_19_cast_fp16 = slice_by_index(begin = xi_19_begin_0, end = xi_19_end_0, end_mask = xi_19_end_mask_0, squeeze_mask = xi_19_squeeze_mask_0, x = pairs_19_cast_fp16)[name = tensor("xi_19_cast_fp16")]; + tensor var_1174_cast_fp16 = mul(x = xr_19_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1174_cast_fp16")]; + tensor var_1175_cast_fp16 = mul(x = xi_19_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1175_cast_fp16")]; + tensor xor__19_cast_fp16 = sub(x = var_1174_cast_fp16, y = var_1175_cast_fp16)[name = tensor("xor__19_cast_fp16")]; + tensor var_1178_cast_fp16 = mul(x = xr_19_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1178_cast_fp16")]; + tensor var_1179_cast_fp16 = mul(x = xi_19_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1179_cast_fp16")]; + tensor xoi_19_cast_fp16 = add(x = var_1178_cast_fp16, y = var_1179_cast_fp16)[name = tensor("xoi_19_cast_fp16")]; + tensor out_19_axis_0 = const()[name = tensor("out_19_axis_0"), val = tensor(-1)]; + tensor out_19_cast_fp16 = stack(axis = out_19_axis_0, values = (xor__19_cast_fp16, xoi_19_cast_fp16))[name = tensor("out_19_cast_fp16")]; + tensor var_1196 = const()[name = tensor("op_1196"), val = tensor([256, 1024])]; + tensor var_1197_cast_fp16 = reshape(shape = var_1196, x = out_19_cast_fp16)[name = tensor("op_1197_cast_fp16")]; + tensor written_k_9_transpose_x_1 = const()[name = tensor("written_k_9_transpose_x_1"), val = tensor(true)]; + tensor written_k_9_transpose_y_1 = const()[name = tensor("written_k_9_transpose_y_1"), val = tensor(false)]; + tensor written_k_9_cast_fp16 = matmul(transpose_x = written_k_9_transpose_x_1, transpose_y = written_k_9_transpose_y_1, x = var_106_cast_fp16_to_fp16, y = var_1197_cast_fp16)[name = tensor("written_k_9_cast_fp16")]; + tensor var_1204 = const()[name = tensor("op_1204"), val = tensor([256, 1024])]; + tensor var_1205_cast_fp16 = reshape(shape = var_1204, x = v_9_cast_fp16)[name = tensor("op_1205_cast_fp16")]; + tensor written_v_9_transpose_x_1 = const()[name = tensor("written_v_9_transpose_x_1"), val = tensor(true)]; + tensor written_v_9_transpose_y_1 = const()[name = tensor("written_v_9_transpose_y_1"), val = tensor(false)]; + tensor written_v_9_cast_fp16 = matmul(transpose_x = written_v_9_transpose_x_1, transpose_y = written_v_9_transpose_y_1, x = var_106_cast_fp16_to_fp16, y = var_1205_cast_fp16)[name = tensor("written_v_9_cast_fp16")]; + tensor k_cache4_to_fp16_dtype_0 = const()[name = tensor("k_cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache4_to_fp16 = cast(dtype = k_cache4_to_fp16_dtype_0, x = k_cache4)[name = tensor("cast_99")]; + tensor var_1216_cast_fp16 = mul(x = k_cache4_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_1216_cast_fp16")]; + tensor var_1221 = const()[name = tensor("op_1221"), val = tensor([1, 512, 16, 64])]; + tensor var_1222_cast_fp16 = reshape(shape = var_1221, x = written_k_9_cast_fp16)[name = tensor("op_1222_cast_fp16")]; + tensor new_k_cache_9_cast_fp16 = add(x = var_1216_cast_fp16, y = var_1222_cast_fp16)[name = tensor("new_k_cache_9_cast_fp16")]; + tensor new_k_cache_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache4_to_fp16_dtype_0 = const()[name = tensor("v_cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache4_to_fp16 = cast(dtype = v_cache4_to_fp16_dtype_0, x = v_cache4)[name = tensor("cast_97")]; + tensor var_1225_cast_fp16 = mul(x = v_cache4_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_1225_cast_fp16")]; + tensor var_1230 = const()[name = tensor("op_1230"), val = tensor([1, 512, 16, 64])]; + tensor var_1231_cast_fp16 = reshape(shape = var_1230, x = written_v_9_cast_fp16)[name = tensor("op_1231_cast_fp16")]; + tensor new_v_cache_9_cast_fp16 = add(x = var_1225_cast_fp16, y = var_1231_cast_fp16)[name = tensor("new_v_cache_9_cast_fp16")]; + tensor new_v_cache_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_1244 = const()[name = tensor("op_1244"), val = tensor([0, 2, 1, 3])]; + tensor var_1255_transpose_x_0 = const()[name = tensor("op_1255_transpose_x_0"), val = tensor(false)]; + tensor var_1255_transpose_y_0 = const()[name = tensor("op_1255_transpose_y_0"), val = tensor(false)]; + tensor transpose_23_perm_0 = const()[name = tensor("transpose_23_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_24_perm_0 = const()[name = tensor("transpose_24_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_24 = transpose(perm = transpose_24_perm_0, x = new_k_cache_9_cast_fp16)[name = tensor("transpose_26")]; + tensor transpose_23 = transpose(perm = transpose_23_perm_0, x = q_9_cast_fp16)[name = tensor("transpose_27")]; + tensor var_1255_cast_fp16 = matmul(transpose_x = var_1255_transpose_x_0, transpose_y = var_1255_transpose_y_0, x = transpose_23, y = transpose_24)[name = tensor("op_1255_cast_fp16")]; + tensor var_1256_to_fp16 = const()[name = tensor("op_1256_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_17_cast_fp16 = mul(x = var_1255_cast_fp16, y = var_1256_to_fp16)[name = tensor("scores_17_cast_fp16")]; + tensor scores_19_cast_fp16 = add(x = scores_17_cast_fp16, y = var_339_cast_fp16)[name = tensor("scores_19_cast_fp16")]; + tensor var_1266 = const()[name = tensor("op_1266"), val = tensor(-1)]; + tensor weights_9_cast_fp16 = softmax(axis = var_1266, x = scores_19_cast_fp16)[name = tensor("weights_9_cast_fp16")]; + tensor attn_9_transpose_x_0 = const()[name = tensor("attn_9_transpose_x_0"), val = tensor(false)]; + tensor attn_9_transpose_y_0 = const()[name = tensor("attn_9_transpose_y_0"), val = tensor(false)]; + tensor values_9_cast_fp16 = transpose(perm = var_1244, x = new_v_cache_9_cast_fp16)[name = tensor("transpose_28")]; + tensor attn_9_cast_fp16 = matmul(transpose_x = attn_9_transpose_x_0, transpose_y = attn_9_transpose_y_0, x = weights_9_cast_fp16, y = values_9_cast_fp16)[name = tensor("attn_9_cast_fp16")]; + tensor var_1274 = const()[name = tensor("op_1274"), val = tensor([0, 2, 1, 3])]; + tensor var_1279 = const()[name = tensor("op_1279"), val = tensor([1, 256, 1024])]; + tensor var_1275_cast_fp16 = transpose(perm = var_1274, x = attn_9_cast_fp16)[name = tensor("transpose_25")]; + tensor input_53_cast_fp16 = reshape(shape = var_1279, x = var_1275_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor attn4_out_proj_weight_to_fp16 = const()[name = tensor("attn4_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(107012864)))]; + tensor linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn4_out_proj_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor position4_to_fp16_dtype_0 = const()[name = tensor("position4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position4_to_fp16 = cast(dtype = position4_to_fp16_dtype_0, x = position4)[name = tensor("cast_95")]; + tensor var_1288_cast_fp16 = add(x = position4_to_fp16, y = valid_len_to_fp16)[name = tensor("op_1288_cast_fp16")]; + tensor var_1288_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1288_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_55_cast_fp16 = add(x = input_49_cast_fp16, y = linear_17_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor input_57_axes_0 = const()[name = tensor("input_57_axes_0"), val = tensor([-1])]; + tensor norm4_2_weight_to_fp16 = const()[name = tensor("norm4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109110080)))]; + tensor norm4_2_bias_to_fp16 = const()[name = tensor("norm4_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109112192)))]; + tensor var_1292_to_fp16 = const()[name = tensor("op_1292_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_57_cast_fp16 = layer_norm(axes = input_57_axes_0, beta = norm4_2_bias_to_fp16, epsilon = var_1292_to_fp16, gamma = norm4_2_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor linear4_1_weight_to_fp16 = const()[name = tensor("linear4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109114304)))]; + tensor linear_18_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear4_1_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor input_59_mode_0 = const()[name = tensor("input_59_mode_0"), val = tensor("EXACT")]; + tensor input_59_cast_fp16 = gelu(mode = input_59_mode_0, x = linear_18_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor linear4_2_weight_to_fp16 = const()[name = tensor("linear4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(117502976)))]; + tensor linear_19_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear4_2_weight_to_fp16, x = input_59_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_61_cast_fp16 = add(x = input_55_cast_fp16, y = linear_19_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor input_63_axes_0 = const()[name = tensor("input_63_axes_0"), val = tensor([-1])]; + tensor norm5_1_weight_to_fp16 = const()[name = tensor("norm5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125891648)))]; + tensor norm5_1_bias_to_fp16 = const()[name = tensor("norm5_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125893760)))]; + tensor var_1309_to_fp16 = const()[name = tensor("op_1309_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_63_cast_fp16 = layer_norm(axes = input_63_axes_0, beta = norm5_1_bias_to_fp16, epsilon = var_1309_to_fp16, gamma = norm5_1_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor attn5_in_proj_weight_to_fp16 = const()[name = tensor("attn5_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125895872)))]; + tensor linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn5_in_proj_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor var_1323 = const()[name = tensor("op_1323"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_cast_fp16 = reshape(shape = var_1323, x = linear_20_cast_fp16)[name = tensor("qkv_cast_fp16")]; + tensor x_begin_0 = const()[name = tensor("x_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_end_0 = const()[name = tensor("x_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_end_mask_0 = const()[name = tensor("x_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_squeeze_mask_0 = const()[name = tensor("x_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_cast_fp16 = slice_by_index(begin = x_begin_0, end = x_end_0, end_mask = x_end_mask_0, squeeze_mask = x_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("x_cast_fp16")]; + tensor v_begin_0 = const()[name = tensor("v_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_end_0 = const()[name = tensor("v_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_end_mask_0 = const()[name = tensor("v_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_squeeze_mask_0 = const()[name = tensor("v_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_cast_fp16 = slice_by_index(begin = v_begin_0, end = v_end_0, end_mask = v_end_mask_0, squeeze_mask = v_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("v_cast_fp16")]; + tensor var_1355 = const()[name = tensor("op_1355"), val = tensor([256, 16, 32, 2])]; + tensor pairs_cast_fp16 = reshape(shape = var_1355, x = x_cast_fp16)[name = tensor("pairs_cast_fp16")]; + tensor xr_begin_0 = const()[name = tensor("xr_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_end_0 = const()[name = tensor("xr_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_end_mask_0 = const()[name = tensor("xr_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_squeeze_mask_0 = const()[name = tensor("xr_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_cast_fp16 = slice_by_index(begin = xr_begin_0, end = xr_end_0, end_mask = xr_end_mask_0, squeeze_mask = xr_squeeze_mask_0, x = pairs_cast_fp16)[name = tensor("xr_cast_fp16")]; + tensor xi_begin_0 = const()[name = tensor("xi_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_end_0 = const()[name = tensor("xi_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_end_mask_0 = const()[name = tensor("xi_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_squeeze_mask_0 = const()[name = tensor("xi_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_cast_fp16 = slice_by_index(begin = xi_begin_0, end = xi_end_0, end_mask = xi_end_mask_0, squeeze_mask = xi_squeeze_mask_0, x = pairs_cast_fp16)[name = tensor("xi_cast_fp16")]; + tensor var_1363_cast_fp16 = mul(x = xr_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1363_cast_fp16")]; + tensor var_1364_cast_fp16 = mul(x = xi_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1364_cast_fp16")]; + tensor xor__cast_fp16 = sub(x = var_1363_cast_fp16, y = var_1364_cast_fp16)[name = tensor("xor__cast_fp16")]; + tensor var_1367_cast_fp16 = mul(x = xr_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1367_cast_fp16")]; + tensor var_1368_cast_fp16 = mul(x = xi_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1368_cast_fp16")]; + tensor xoi_cast_fp16 = add(x = var_1367_cast_fp16, y = var_1368_cast_fp16)[name = tensor("xoi_cast_fp16")]; + tensor out_axis_0 = const()[name = tensor("out_axis_0"), val = tensor(-1)]; + tensor out_cast_fp16 = stack(axis = out_axis_0, values = (xor__cast_fp16, xoi_cast_fp16))[name = tensor("out_cast_fp16")]; + tensor var_1385 = const()[name = tensor("op_1385"), val = tensor([256, 1024])]; + tensor var_1386_cast_fp16 = reshape(shape = var_1385, x = out_cast_fp16)[name = tensor("op_1386_cast_fp16")]; + tensor written_k_transpose_x_1 = const()[name = tensor("written_k_transpose_x_1"), val = tensor(true)]; + tensor written_k_transpose_y_1 = const()[name = tensor("written_k_transpose_y_1"), val = tensor(false)]; + tensor written_k_cast_fp16 = matmul(transpose_x = written_k_transpose_x_1, transpose_y = written_k_transpose_y_1, x = var_106_cast_fp16_to_fp16, y = var_1386_cast_fp16)[name = tensor("written_k_cast_fp16")]; + tensor var_1393 = const()[name = tensor("op_1393"), val = tensor([256, 1024])]; + tensor var_1394_cast_fp16 = reshape(shape = var_1393, x = v_cast_fp16)[name = tensor("op_1394_cast_fp16")]; + tensor written_v_transpose_x_1 = const()[name = tensor("written_v_transpose_x_1"), val = tensor(true)]; + tensor written_v_transpose_y_1 = const()[name = tensor("written_v_transpose_y_1"), val = tensor(false)]; + tensor written_v_cast_fp16 = matmul(transpose_x = written_v_transpose_x_1, transpose_y = written_v_transpose_y_1, x = var_106_cast_fp16_to_fp16, y = var_1394_cast_fp16)[name = tensor("written_v_cast_fp16")]; + tensor k_cache5_to_fp16_dtype_0 = const()[name = tensor("k_cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache5_to_fp16 = cast(dtype = k_cache5_to_fp16_dtype_0, x = k_cache5)[name = tensor("cast_93")]; + tensor var_1405_cast_fp16 = mul(x = k_cache5_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_1405_cast_fp16")]; + tensor var_1410 = const()[name = tensor("op_1410"), val = tensor([1, 512, 16, 64])]; + tensor var_1411_cast_fp16 = reshape(shape = var_1410, x = written_k_cast_fp16)[name = tensor("op_1411_cast_fp16")]; + tensor new_k_cache_cast_fp16 = add(x = var_1405_cast_fp16, y = var_1411_cast_fp16)[name = tensor("new_k_cache_cast_fp16")]; + tensor new_k_cache_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache5_to_fp16_dtype_0 = const()[name = tensor("v_cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache5_to_fp16 = cast(dtype = v_cache5_to_fp16_dtype_0, x = v_cache5)[name = tensor("cast_91")]; + tensor var_1414_cast_fp16 = mul(x = v_cache5_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_1414_cast_fp16")]; + tensor var_1419 = const()[name = tensor("op_1419"), val = tensor([1, 512, 16, 64])]; + tensor var_1420_cast_fp16 = reshape(shape = var_1419, x = written_v_cast_fp16)[name = tensor("op_1420_cast_fp16")]; + tensor new_v_cache_cast_fp16 = add(x = var_1414_cast_fp16, y = var_1420_cast_fp16)[name = tensor("new_v_cache_cast_fp16")]; + tensor new_v_cache_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor position5_to_fp16_dtype_0 = const()[name = tensor("position5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position5_to_fp16 = cast(dtype = position5_to_fp16_dtype_0, x = position5)[name = tensor("cast_89")]; + tensor var_1427_cast_fp16 = add(x = position5_to_fp16, y = valid_len_to_fp16)[name = tensor("op_1427_cast_fp16")]; + tensor var_1427_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1427_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor new_position5 = cast(dtype = var_1427_cast_fp16_to_fp32_dtype_0, x = var_1427_cast_fp16)[name = tensor("cast_88")]; + tensor new_v_cache5 = cast(dtype = new_v_cache_cast_fp16_to_fp32_dtype_0, x = new_v_cache_cast_fp16)[name = tensor("cast_90")]; + tensor new_k_cache5 = cast(dtype = new_k_cache_cast_fp16_to_fp32_dtype_0, x = new_k_cache_cast_fp16)[name = tensor("cast_92")]; + tensor new_position4 = cast(dtype = var_1288_cast_fp16_to_fp32_dtype_0, x = var_1288_cast_fp16)[name = tensor("cast_94")]; + tensor new_v_cache4 = cast(dtype = new_v_cache_9_cast_fp16_to_fp32_dtype_0, x = new_v_cache_9_cast_fp16)[name = tensor("cast_96")]; + tensor new_k_cache4 = cast(dtype = new_k_cache_9_cast_fp16_to_fp32_dtype_0, x = new_k_cache_9_cast_fp16)[name = tensor("cast_98")]; + tensor new_position3 = cast(dtype = var_1057_cast_fp16_to_fp32_dtype_0, x = var_1057_cast_fp16)[name = tensor("cast_100")]; + tensor new_v_cache3 = cast(dtype = new_v_cache_7_cast_fp16_to_fp32_dtype_0, x = new_v_cache_7_cast_fp16)[name = tensor("cast_102")]; + tensor new_k_cache3 = cast(dtype = new_k_cache_7_cast_fp16_to_fp32_dtype_0, x = new_k_cache_7_cast_fp16)[name = tensor("cast_104")]; + tensor new_position2 = cast(dtype = var_826_cast_fp16_to_fp32_dtype_0, x = var_826_cast_fp16)[name = tensor("cast_106")]; + tensor new_v_cache2 = cast(dtype = new_v_cache_5_cast_fp16_to_fp32_dtype_0, x = new_v_cache_5_cast_fp16)[name = tensor("cast_108")]; + tensor new_k_cache2 = cast(dtype = new_k_cache_5_cast_fp16_to_fp32_dtype_0, x = new_k_cache_5_cast_fp16)[name = tensor("cast_110")]; + tensor new_position1 = cast(dtype = var_595_cast_fp16_to_fp32_dtype_0, x = var_595_cast_fp16)[name = tensor("cast_112")]; + tensor new_v_cache1 = cast(dtype = new_v_cache_3_cast_fp16_to_fp32_dtype_0, x = new_v_cache_3_cast_fp16)[name = tensor("cast_114")]; + tensor new_k_cache1 = cast(dtype = new_k_cache_3_cast_fp16_to_fp32_dtype_0, x = new_k_cache_3_cast_fp16)[name = tensor("cast_116")]; + tensor new_position0 = cast(dtype = var_364_cast_fp16_to_fp32_dtype_0, x = var_135_cast_fp16)[name = tensor("cast_118")]; + tensor new_v_cache0 = cast(dtype = new_v_cache_1_cast_fp16_to_fp32_dtype_0, x = new_v_cache_1_cast_fp16)[name = tensor("cast_119")]; + tensor new_k_cache0 = cast(dtype = new_k_cache_1_cast_fp16_to_fp32_dtype_0, x = new_k_cache_1_cast_fp16)[name = tensor("cast_121")]; + } -> (new_k_cache0, new_v_cache0, new_position0, new_k_cache1, new_v_cache1, new_position1, new_k_cache2, new_v_cache2, new_position2, new_k_cache3, new_v_cache3, new_position3, new_k_cache4, new_v_cache4, new_position4, new_k_cache5, new_v_cache5, new_position5); +} \ No newline at end of file diff --git a/v2.1/german/cond_prefill_ane.mlmodelc/weights/weight.bin b/v2.1/german/cond_prefill_ane.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..40694b5ea4758c30ef7e03596e370a74a84b8eb9 --- /dev/null +++ b/v2.1/german/cond_prefill_ane.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fe878d1ec2a8cf93e18df624b496550348e511f9928cdeec022a96977202e0ca +size 132187392 diff --git a/v2.1/german/cond_prefill_ane.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/german/cond_prefill_ane.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..4636088a17019120d03aae175cde21ab3c787baa --- /dev/null +++ b/v2.1/german/cond_prefill_ane.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2c906248cc4a46c35f294b69715ca98e926d16481f147611643eb7de61a57b57 +size 119086 diff --git a/v2.1/german/cond_prefill_ane.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/german/cond_prefill_ane.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..40694b5ea4758c30ef7e03596e370a74a84b8eb9 --- /dev/null +++ b/v2.1/german/cond_prefill_ane.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fe878d1ec2a8cf93e18df624b496550348e511f9928cdeec022a96977202e0ca +size 132187392 diff --git a/v2.1/german/cond_prefill_ane.mlpackage/Manifest.json b/v2.1/german/cond_prefill_ane.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..6830281361bf90b3b3b3819ab3489f187a110aa3 --- /dev/null +++ b/v2.1/german/cond_prefill_ane.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "53B49E7F-74E1-4363-A1C7-034818B33FE6": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "E1C4CAF0-9078-4918-A76A-C5C7B7DD70C2": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "E1C4CAF0-9078-4918-A76A-C5C7B7DD70C2" +} diff --git a/v2.1/german/flowlm_step_ane.mlmodelc/analytics/coremldata.bin b/v2.1/german/flowlm_step_ane.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..543423f90ef3b6941488ab6c21e768574fc6ddf0 --- /dev/null +++ b/v2.1/german/flowlm_step_ane.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b7312c96a30caf1a7281c3bf9737acc6a16ff187c0090f6f5d7550d8f95f29a7 +size 243 diff --git a/v2.1/german/flowlm_step_ane.mlmodelc/coremldata.bin b/v2.1/german/flowlm_step_ane.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..1dca3816763ca8530d04f6f6071b84a2f9315721 --- /dev/null +++ b/v2.1/german/flowlm_step_ane.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:401971692598e997822a77864ac719fdd547fae7f72b0cb85c5719dfdd9f9e02 +size 1414 diff --git a/v2.1/german/flowlm_step_ane.mlmodelc/model.mil b/v2.1/german/flowlm_step_ane.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..6e49299a2c6ef23d1d840b3b4ac19e3ef6c1d171 --- /dev/null +++ b/v2.1/german/flowlm_step_ane.mlmodelc/model.mil @@ -0,0 +1,945 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.9.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor k_cache0, tensor k_cache1, tensor k_cache2, tensor k_cache3, tensor k_cache4, tensor k_cache5, tensor position0, tensor position1, tensor position2, tensor position3, tensor position4, tensor position5, tensor sequence, tensor v_cache0, tensor v_cache1, tensor v_cache2, tensor v_cache3, tensor v_cache4, tensor v_cache5) { + tensor sequence_to_fp16_dtype_0 = const()[name = tensor("sequence_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor input_linear_weight_to_fp16 = const()[name = tensor("input_linear_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(65664)))]; + tensor sequence_to_fp16 = cast(dtype = sequence_to_fp16_dtype_0, x = sequence)[name = tensor("cast_146")]; + tensor linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = input_linear_weight_to_fp16, x = sequence_to_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; + tensor norm0_1_weight_to_fp16 = const()[name = tensor("norm0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67776)))]; + tensor norm0_1_bias_to_fp16 = const()[name = tensor("norm0_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(69888)))]; + tensor var_65_to_fp16 = const()[name = tensor("op_65_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_5_cast_fp16 = layer_norm(axes = input_5_axes_0, beta = norm0_1_bias_to_fp16, epsilon = var_65_to_fp16, gamma = norm0_1_weight_to_fp16, x = linear_0_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor attn0_in_proj_weight_to_fp16 = const()[name = tensor("attn0_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(72000)))]; + tensor linear_1_bias_0_to_fp16 = const()[name = tensor("linear_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6363520)))]; + tensor linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn0_in_proj_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("linear_1_cast_fp16")]; + tensor var_78 = const()[name = tensor("op_78"), val = tensor([1, 3, 16, 64])]; + tensor qkv_1_cast_fp16 = reshape(shape = var_78, x = linear_1_cast_fp16)[name = tensor("qkv_1_cast_fp16")]; + tensor x_1_begin_0 = const()[name = tensor("x_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_1_end_0 = const()[name = tensor("x_1_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_1_end_mask_0 = const()[name = tensor("x_1_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_1_squeeze_mask_0 = const()[name = tensor("x_1_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_1_cast_fp16 = slice_by_index(begin = x_1_begin_0, end = x_1_end_0, end_mask = x_1_end_mask_0, squeeze_mask = x_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("x_1_cast_fp16")]; + tensor x_3_begin_0 = const()[name = tensor("x_3_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_3_end_0 = const()[name = tensor("x_3_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_3_end_mask_0 = const()[name = tensor("x_3_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_3_squeeze_mask_0 = const()[name = tensor("x_3_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_3_cast_fp16 = slice_by_index(begin = x_3_begin_0, end = x_3_end_0, end_mask = x_3_end_mask_0, squeeze_mask = x_3_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("x_3_cast_fp16")]; + tensor v_1_begin_0 = const()[name = tensor("v_1_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_1_end_0 = const()[name = tensor("v_1_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_1_end_mask_0 = const()[name = tensor("v_1_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_1_squeeze_mask_0 = const()[name = tensor("v_1_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_1_cast_fp16 = slice_by_index(begin = v_1_begin_0, end = v_1_end_0, end_mask = v_1_end_mask_0, squeeze_mask = v_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("v_1_cast_fp16")]; + tensor var_107 = const()[name = tensor("op_107"), val = tensor([1, 1, 1])]; + tensor position0_to_fp16_dtype_0 = const()[name = tensor("position0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position0_to_fp16 = cast(dtype = position0_to_fp16_dtype_0, x = position0)[name = tensor("cast_145")]; + tensor var_108_cast_fp16 = reshape(shape = var_107, x = position0_to_fp16)[name = tensor("op_108_cast_fp16")]; + tensor rope_freqs_to_fp16 = const()[name = tensor("rope_freqs_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6369728)))]; + tensor angles_1_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_108_cast_fp16)[name = tensor("angles_1_cast_fp16")]; + tensor rotr_1_cast_fp16 = cos(x = angles_1_cast_fp16)[name = tensor("rotr_1_cast_fp16")]; + tensor roti_1_cast_fp16 = sin(x = angles_1_cast_fp16)[name = tensor("roti_1_cast_fp16")]; + tensor var_116 = const()[name = tensor("op_116"), val = tensor([1, 16, 32, 2])]; + tensor pairs_1_cast_fp16 = reshape(shape = var_116, x = x_1_cast_fp16)[name = tensor("pairs_1_cast_fp16")]; + tensor xr_1_begin_0 = const()[name = tensor("xr_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_1_end_0 = const()[name = tensor("xr_1_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_1_end_mask_0 = const()[name = tensor("xr_1_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_1_squeeze_mask_0 = const()[name = tensor("xr_1_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_1_cast_fp16 = slice_by_index(begin = xr_1_begin_0, end = xr_1_end_0, end_mask = xr_1_end_mask_0, squeeze_mask = xr_1_squeeze_mask_0, x = pairs_1_cast_fp16)[name = tensor("xr_1_cast_fp16")]; + tensor xi_1_begin_0 = const()[name = tensor("xi_1_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_1_end_0 = const()[name = tensor("xi_1_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_1_end_mask_0 = const()[name = tensor("xi_1_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_1_squeeze_mask_0 = const()[name = tensor("xi_1_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_1_cast_fp16 = slice_by_index(begin = xi_1_begin_0, end = xi_1_end_0, end_mask = xi_1_end_mask_0, squeeze_mask = xi_1_squeeze_mask_0, x = pairs_1_cast_fp16)[name = tensor("xi_1_cast_fp16")]; + tensor var_124_cast_fp16 = mul(x = xr_1_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_124_cast_fp16")]; + tensor var_125_cast_fp16 = mul(x = xi_1_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_125_cast_fp16")]; + tensor xor__1_cast_fp16 = sub(x = var_124_cast_fp16, y = var_125_cast_fp16)[name = tensor("xor__1_cast_fp16")]; + tensor var_128_cast_fp16 = mul(x = xr_1_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_128_cast_fp16")]; + tensor var_129_cast_fp16 = mul(x = xi_1_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_129_cast_fp16")]; + tensor xoi_1_cast_fp16 = add(x = var_128_cast_fp16, y = var_129_cast_fp16)[name = tensor("xoi_1_cast_fp16")]; + tensor out_1_axis_0 = const()[name = tensor("out_1_axis_0"), val = tensor(-1)]; + tensor out_1_cast_fp16 = stack(axis = out_1_axis_0, values = (xor__1_cast_fp16, xoi_1_cast_fp16))[name = tensor("out_1_cast_fp16")]; + tensor var_152 = const()[name = tensor("op_152"), val = tensor([1, 16, 32, 2])]; + tensor pairs_3_cast_fp16 = reshape(shape = var_152, x = x_3_cast_fp16)[name = tensor("pairs_3_cast_fp16")]; + tensor xr_3_begin_0 = const()[name = tensor("xr_3_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_3_end_0 = const()[name = tensor("xr_3_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_3_end_mask_0 = const()[name = tensor("xr_3_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_3_squeeze_mask_0 = const()[name = tensor("xr_3_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_3_cast_fp16 = slice_by_index(begin = xr_3_begin_0, end = xr_3_end_0, end_mask = xr_3_end_mask_0, squeeze_mask = xr_3_squeeze_mask_0, x = pairs_3_cast_fp16)[name = tensor("xr_3_cast_fp16")]; + tensor xi_3_begin_0 = const()[name = tensor("xi_3_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_3_end_0 = const()[name = tensor("xi_3_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_3_end_mask_0 = const()[name = tensor("xi_3_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_3_squeeze_mask_0 = const()[name = tensor("xi_3_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_3_cast_fp16 = slice_by_index(begin = xi_3_begin_0, end = xi_3_end_0, end_mask = xi_3_end_mask_0, squeeze_mask = xi_3_squeeze_mask_0, x = pairs_3_cast_fp16)[name = tensor("xi_3_cast_fp16")]; + tensor var_160_cast_fp16 = mul(x = xr_3_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_160_cast_fp16")]; + tensor var_161_cast_fp16 = mul(x = xi_3_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_161_cast_fp16")]; + tensor xor__3_cast_fp16 = sub(x = var_160_cast_fp16, y = var_161_cast_fp16)[name = tensor("xor__3_cast_fp16")]; + tensor var_164_cast_fp16 = mul(x = xr_3_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_164_cast_fp16")]; + tensor var_165_cast_fp16 = mul(x = xi_3_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_165_cast_fp16")]; + tensor xoi_3_cast_fp16 = add(x = var_164_cast_fp16, y = var_165_cast_fp16)[name = tensor("xoi_3_cast_fp16")]; + tensor out_3_axis_0 = const()[name = tensor("out_3_axis_0"), val = tensor(-1)]; + tensor out_3_cast_fp16 = stack(axis = out_3_axis_0, values = (xor__3_cast_fp16, xoi_3_cast_fp16))[name = tensor("out_3_cast_fp16")]; + tensor _inversed_177_y_0_to_fp16 = const()[name = tensor("_inversed_177_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_177_cast_fp16 = mul(x = position0_to_fp16, y = _inversed_177_y_0_to_fp16)[name = tensor("_inversed_177_cast_fp16")]; + tensor var_178_cast_fp16 = floor(x = _inversed_177_cast_fp16)[name = tensor("op_178_cast_fp16")]; + tensor var_179_to_fp16 = const()[name = tensor("op_179_to_fp16"), val = tensor(0x1p+9)]; + tensor var_180_cast_fp16 = mul(x = var_178_cast_fp16, y = var_179_to_fp16)[name = tensor("op_180_cast_fp16")]; + tensor write_idx_1_cast_fp16 = sub(x = position0_to_fp16, y = var_180_cast_fp16)[name = tensor("write_idx_1_cast_fp16")]; + tensor slot_indices_to_fp16 = const()[name = tensor("slot_indices_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6369856)))]; + tensor var_183_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_1_cast_fp16)[name = tensor("op_183_cast_fp16")]; + tensor var_193 = const()[name = tensor("op_193"), val = tensor([1, 512, 1, 1])]; + tensor onehot_1_to_fp16_dtype_0 = const()[name = tensor("onehot_1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_183_cast_fp16_to_fp16 = cast(dtype = onehot_1_to_fp16_dtype_0, x = var_183_cast_fp16)[name = tensor("cast_144")]; + tensor onehot_3_cast_fp16 = reshape(shape = var_193, x = var_183_cast_fp16_to_fp16)[name = tensor("onehot_3_cast_fp16")]; + tensor var_195_to_fp16 = const()[name = tensor("op_195_to_fp16"), val = tensor(0x1p+0)]; + tensor var_197_cast_fp16 = sub(x = var_195_to_fp16, y = onehot_3_cast_fp16)[name = tensor("op_197_cast_fp16")]; + tensor k_cache0_to_fp16_dtype_0 = const()[name = tensor("k_cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache0_to_fp16 = cast(dtype = k_cache0_to_fp16_dtype_0, x = k_cache0)[name = tensor("cast_143")]; + tensor var_198_cast_fp16 = mul(x = k_cache0_to_fp16, y = var_197_cast_fp16)[name = tensor("op_198_cast_fp16")]; + tensor var_203 = const()[name = tensor("op_203"), val = tensor([1, 1, 16, 64])]; + tensor var_204_cast_fp16 = reshape(shape = var_203, x = out_3_cast_fp16)[name = tensor("op_204_cast_fp16")]; + tensor var_205_cast_fp16 = mul(x = var_204_cast_fp16, y = onehot_3_cast_fp16)[name = tensor("op_205_cast_fp16")]; + tensor new_k_cache_1_cast_fp16 = add(x = var_198_cast_fp16, y = var_205_cast_fp16)[name = tensor("new_k_cache_1_cast_fp16")]; + tensor new_k_cache_1_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_1_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache0_to_fp16_dtype_0 = const()[name = tensor("v_cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache0_to_fp16 = cast(dtype = v_cache0_to_fp16_dtype_0, x = v_cache0)[name = tensor("cast_141")]; + tensor var_211_cast_fp16 = mul(x = v_cache0_to_fp16, y = var_197_cast_fp16)[name = tensor("op_211_cast_fp16")]; + tensor var_216 = const()[name = tensor("op_216"), val = tensor([1, 1, 16, 64])]; + tensor var_217_cast_fp16 = reshape(shape = var_216, x = v_1_cast_fp16)[name = tensor("op_217_cast_fp16")]; + tensor var_218_cast_fp16 = mul(x = var_217_cast_fp16, y = onehot_3_cast_fp16)[name = tensor("op_218_cast_fp16")]; + tensor new_v_cache_1_cast_fp16 = add(x = var_211_cast_fp16, y = var_218_cast_fp16)[name = tensor("new_v_cache_1_cast_fp16")]; + tensor new_v_cache_1_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_1_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_0_perm_0 = const()[name = tensor("transpose_0_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_231 = const()[name = tensor("op_231"), val = tensor([0, 2, 1, 3])]; + tensor var_237 = const()[name = tensor("op_237"), val = tensor([1, 16, 1, 64])]; + tensor q4_1_cast_fp16 = reshape(shape = var_237, x = out_1_cast_fp16)[name = tensor("q4_1_cast_fp16")]; + tensor var_242_transpose_x_0 = const()[name = tensor("op_242_transpose_x_0"), val = tensor(false)]; + tensor var_242_transpose_y_0 = const()[name = tensor("op_242_transpose_y_0"), val = tensor(false)]; + tensor transpose_0_cast_fp16 = transpose(perm = transpose_0_perm_0, x = new_k_cache_1_cast_fp16)[name = tensor("transpose_23")]; + tensor var_242_cast_fp16 = matmul(transpose_x = var_242_transpose_x_0, transpose_y = var_242_transpose_y_0, x = q4_1_cast_fp16, y = transpose_0_cast_fp16)[name = tensor("op_242_cast_fp16")]; + tensor var_243_to_fp16 = const()[name = tensor("op_243_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_1_cast_fp16 = mul(x = var_242_cast_fp16, y = var_243_to_fp16)[name = tensor("scores_1_cast_fp16")]; + tensor var_247 = const()[name = tensor("op_247"), val = tensor([1, 1])]; + tensor var_248_cast_fp16 = reshape(shape = var_247, x = position0_to_fp16)[name = tensor("op_248_cast_fp16")]; + tensor var_249_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_248_cast_fp16)[name = tensor("op_249_cast_fp16")]; + tensor mask_1_to_fp16_dtype_0 = const()[name = tensor("mask_1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_256_to_fp16 = const()[name = tensor("op_256_to_fp16"), val = tensor(0x1p+0)]; + tensor var_249_cast_fp16_to_fp16 = cast(dtype = mask_1_to_fp16_dtype_0, x = var_249_cast_fp16)[name = tensor("cast_139")]; + tensor var_257_cast_fp16 = sub(x = var_249_cast_fp16_to_fp16, y = var_256_to_fp16)[name = tensor("op_257_cast_fp16")]; + tensor var_262 = const()[name = tensor("op_262"), val = tensor([1, 1, 1, 512])]; + tensor var_263_cast_fp16 = reshape(shape = var_262, x = var_257_cast_fp16)[name = tensor("op_263_cast_fp16")]; + tensor var_264_to_fp16 = const()[name = tensor("op_264_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_265_cast_fp16 = mul(x = var_263_cast_fp16, y = var_264_to_fp16)[name = tensor("op_265_cast_fp16")]; + tensor scores_3_cast_fp16 = add(x = scores_1_cast_fp16, y = var_265_cast_fp16)[name = tensor("scores_3_cast_fp16")]; + tensor var_268 = const()[name = tensor("op_268"), val = tensor(-1)]; + tensor weights_1_cast_fp16 = softmax(axis = var_268, x = scores_3_cast_fp16)[name = tensor("weights_1_cast_fp16")]; + tensor attn_1_transpose_x_0 = const()[name = tensor("attn_1_transpose_x_0"), val = tensor(false)]; + tensor attn_1_transpose_y_0 = const()[name = tensor("attn_1_transpose_y_0"), val = tensor(false)]; + tensor values_1_cast_fp16 = transpose(perm = var_231, x = new_v_cache_1_cast_fp16)[name = tensor("transpose_22")]; + tensor attn_1_cast_fp16 = matmul(transpose_x = attn_1_transpose_x_0, transpose_y = attn_1_transpose_y_0, x = weights_1_cast_fp16, y = values_1_cast_fp16)[name = tensor("attn_1_cast_fp16")]; + tensor var_276 = const()[name = tensor("op_276"), val = tensor([0, 2, 1, 3])]; + tensor var_281 = const()[name = tensor("op_281"), val = tensor([1, 1, 1024])]; + tensor var_277_cast_fp16 = transpose(perm = var_276, x = attn_1_cast_fp16)[name = tensor("transpose_21")]; + tensor input_7_cast_fp16 = reshape(shape = var_281, x = var_277_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor attn0_out_proj_weight_to_fp16 = const()[name = tensor("attn0_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6370944)))]; + tensor linear_2_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn0_out_proj_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor var_287_to_fp16 = const()[name = tensor("op_287_to_fp16"), val = tensor(0x1p+0)]; + tensor var_288_cast_fp16 = add(x = position0_to_fp16, y = var_287_to_fp16)[name = tensor("op_288_cast_fp16")]; + tensor var_288_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_288_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_9_cast_fp16 = add(x = linear_0_cast_fp16, y = linear_2_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor input_11_axes_0 = const()[name = tensor("input_11_axes_0"), val = tensor([-1])]; + tensor norm0_2_weight_to_fp16 = const()[name = tensor("norm0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8468160)))]; + tensor norm0_2_bias_to_fp16 = const()[name = tensor("norm0_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8470272)))]; + tensor var_292_to_fp16 = const()[name = tensor("op_292_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_11_cast_fp16 = layer_norm(axes = input_11_axes_0, beta = norm0_2_bias_to_fp16, epsilon = var_292_to_fp16, gamma = norm0_2_weight_to_fp16, x = input_9_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor linear0_1_weight_to_fp16 = const()[name = tensor("linear0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8472384)))]; + tensor linear_3_bias_0_to_fp16 = const()[name = tensor("linear_3_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16861056)))]; + tensor linear_3_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear0_1_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("linear_3_cast_fp16")]; + tensor input_13_mode_0 = const()[name = tensor("input_13_mode_0"), val = tensor("EXACT")]; + tensor input_13_cast_fp16 = gelu(mode = input_13_mode_0, x = linear_3_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor linear0_2_weight_to_fp16 = const()[name = tensor("linear0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16869312)))]; + tensor linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear0_2_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor input_15_cast_fp16 = add(x = input_9_cast_fp16, y = linear_4_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor input_17_axes_0 = const()[name = tensor("input_17_axes_0"), val = tensor([-1])]; + tensor norm1_1_weight_to_fp16 = const()[name = tensor("norm1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25257984)))]; + tensor norm1_1_bias_to_fp16 = const()[name = tensor("norm1_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25260096)))]; + tensor var_309_to_fp16 = const()[name = tensor("op_309_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_17_cast_fp16 = layer_norm(axes = input_17_axes_0, beta = norm1_1_bias_to_fp16, epsilon = var_309_to_fp16, gamma = norm1_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor attn1_in_proj_weight_to_fp16 = const()[name = tensor("attn1_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25262208)))]; + tensor linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn1_in_proj_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor var_322 = const()[name = tensor("op_322"), val = tensor([1, 3, 16, 64])]; + tensor qkv_3_cast_fp16 = reshape(shape = var_322, x = linear_5_cast_fp16)[name = tensor("qkv_3_cast_fp16")]; + tensor x_5_begin_0 = const()[name = tensor("x_5_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_5_end_0 = const()[name = tensor("x_5_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_5_end_mask_0 = const()[name = tensor("x_5_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_5_squeeze_mask_0 = const()[name = tensor("x_5_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_5_cast_fp16 = slice_by_index(begin = x_5_begin_0, end = x_5_end_0, end_mask = x_5_end_mask_0, squeeze_mask = x_5_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("x_5_cast_fp16")]; + tensor x_7_begin_0 = const()[name = tensor("x_7_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_7_end_0 = const()[name = tensor("x_7_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_7_end_mask_0 = const()[name = tensor("x_7_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_7_squeeze_mask_0 = const()[name = tensor("x_7_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_7_cast_fp16 = slice_by_index(begin = x_7_begin_0, end = x_7_end_0, end_mask = x_7_end_mask_0, squeeze_mask = x_7_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("x_7_cast_fp16")]; + tensor v_3_begin_0 = const()[name = tensor("v_3_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_3_end_0 = const()[name = tensor("v_3_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_3_end_mask_0 = const()[name = tensor("v_3_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_3_squeeze_mask_0 = const()[name = tensor("v_3_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_3_cast_fp16 = slice_by_index(begin = v_3_begin_0, end = v_3_end_0, end_mask = v_3_end_mask_0, squeeze_mask = v_3_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("v_3_cast_fp16")]; + tensor var_351 = const()[name = tensor("op_351"), val = tensor([1, 1, 1])]; + tensor position1_to_fp16_dtype_0 = const()[name = tensor("position1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position1_to_fp16 = cast(dtype = position1_to_fp16_dtype_0, x = position1)[name = tensor("cast_137")]; + tensor var_352_cast_fp16 = reshape(shape = var_351, x = position1_to_fp16)[name = tensor("op_352_cast_fp16")]; + tensor angles_5_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_352_cast_fp16)[name = tensor("angles_5_cast_fp16")]; + tensor rotr_5_cast_fp16 = cos(x = angles_5_cast_fp16)[name = tensor("rotr_5_cast_fp16")]; + tensor roti_5_cast_fp16 = sin(x = angles_5_cast_fp16)[name = tensor("roti_5_cast_fp16")]; + tensor var_360 = const()[name = tensor("op_360"), val = tensor([1, 16, 32, 2])]; + tensor pairs_5_cast_fp16 = reshape(shape = var_360, x = x_5_cast_fp16)[name = tensor("pairs_5_cast_fp16")]; + tensor xr_5_begin_0 = const()[name = tensor("xr_5_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_5_end_0 = const()[name = tensor("xr_5_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_5_end_mask_0 = const()[name = tensor("xr_5_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_5_squeeze_mask_0 = const()[name = tensor("xr_5_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_5_cast_fp16 = slice_by_index(begin = xr_5_begin_0, end = xr_5_end_0, end_mask = xr_5_end_mask_0, squeeze_mask = xr_5_squeeze_mask_0, x = pairs_5_cast_fp16)[name = tensor("xr_5_cast_fp16")]; + tensor xi_5_begin_0 = const()[name = tensor("xi_5_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_5_end_0 = const()[name = tensor("xi_5_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_5_end_mask_0 = const()[name = tensor("xi_5_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_5_squeeze_mask_0 = const()[name = tensor("xi_5_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_5_cast_fp16 = slice_by_index(begin = xi_5_begin_0, end = xi_5_end_0, end_mask = xi_5_end_mask_0, squeeze_mask = xi_5_squeeze_mask_0, x = pairs_5_cast_fp16)[name = tensor("xi_5_cast_fp16")]; + tensor var_368_cast_fp16 = mul(x = xr_5_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_368_cast_fp16")]; + tensor var_369_cast_fp16 = mul(x = xi_5_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_369_cast_fp16")]; + tensor xor__5_cast_fp16 = sub(x = var_368_cast_fp16, y = var_369_cast_fp16)[name = tensor("xor__5_cast_fp16")]; + tensor var_372_cast_fp16 = mul(x = xr_5_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_372_cast_fp16")]; + tensor var_373_cast_fp16 = mul(x = xi_5_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_373_cast_fp16")]; + tensor xoi_5_cast_fp16 = add(x = var_372_cast_fp16, y = var_373_cast_fp16)[name = tensor("xoi_5_cast_fp16")]; + tensor out_5_axis_0 = const()[name = tensor("out_5_axis_0"), val = tensor(-1)]; + tensor out_5_cast_fp16 = stack(axis = out_5_axis_0, values = (xor__5_cast_fp16, xoi_5_cast_fp16))[name = tensor("out_5_cast_fp16")]; + tensor var_396 = const()[name = tensor("op_396"), val = tensor([1, 16, 32, 2])]; + tensor pairs_7_cast_fp16 = reshape(shape = var_396, x = x_7_cast_fp16)[name = tensor("pairs_7_cast_fp16")]; + tensor xr_7_begin_0 = const()[name = tensor("xr_7_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_7_end_0 = const()[name = tensor("xr_7_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_7_end_mask_0 = const()[name = tensor("xr_7_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_7_squeeze_mask_0 = const()[name = tensor("xr_7_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_7_cast_fp16 = slice_by_index(begin = xr_7_begin_0, end = xr_7_end_0, end_mask = xr_7_end_mask_0, squeeze_mask = xr_7_squeeze_mask_0, x = pairs_7_cast_fp16)[name = tensor("xr_7_cast_fp16")]; + tensor xi_7_begin_0 = const()[name = tensor("xi_7_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_7_end_0 = const()[name = tensor("xi_7_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_7_end_mask_0 = const()[name = tensor("xi_7_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_7_squeeze_mask_0 = const()[name = tensor("xi_7_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_7_cast_fp16 = slice_by_index(begin = xi_7_begin_0, end = xi_7_end_0, end_mask = xi_7_end_mask_0, squeeze_mask = xi_7_squeeze_mask_0, x = pairs_7_cast_fp16)[name = tensor("xi_7_cast_fp16")]; + tensor var_404_cast_fp16 = mul(x = xr_7_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_404_cast_fp16")]; + tensor var_405_cast_fp16 = mul(x = xi_7_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_405_cast_fp16")]; + tensor xor__7_cast_fp16 = sub(x = var_404_cast_fp16, y = var_405_cast_fp16)[name = tensor("xor__7_cast_fp16")]; + tensor var_408_cast_fp16 = mul(x = xr_7_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_408_cast_fp16")]; + tensor var_409_cast_fp16 = mul(x = xi_7_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_409_cast_fp16")]; + tensor xoi_7_cast_fp16 = add(x = var_408_cast_fp16, y = var_409_cast_fp16)[name = tensor("xoi_7_cast_fp16")]; + tensor out_7_axis_0 = const()[name = tensor("out_7_axis_0"), val = tensor(-1)]; + tensor out_7_cast_fp16 = stack(axis = out_7_axis_0, values = (xor__7_cast_fp16, xoi_7_cast_fp16))[name = tensor("out_7_cast_fp16")]; + tensor _inversed_421_y_0_to_fp16 = const()[name = tensor("_inversed_421_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_421_cast_fp16 = mul(x = position1_to_fp16, y = _inversed_421_y_0_to_fp16)[name = tensor("_inversed_421_cast_fp16")]; + tensor var_422_cast_fp16 = floor(x = _inversed_421_cast_fp16)[name = tensor("op_422_cast_fp16")]; + tensor var_423_to_fp16 = const()[name = tensor("op_423_to_fp16"), val = tensor(0x1p+9)]; + tensor var_424_cast_fp16 = mul(x = var_422_cast_fp16, y = var_423_to_fp16)[name = tensor("op_424_cast_fp16")]; + tensor write_idx_3_cast_fp16 = sub(x = position1_to_fp16, y = var_424_cast_fp16)[name = tensor("write_idx_3_cast_fp16")]; + tensor var_427_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_3_cast_fp16)[name = tensor("op_427_cast_fp16")]; + tensor var_437 = const()[name = tensor("op_437"), val = tensor([1, 512, 1, 1])]; + tensor onehot_5_to_fp16_dtype_0 = const()[name = tensor("onehot_5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_427_cast_fp16_to_fp16 = cast(dtype = onehot_5_to_fp16_dtype_0, x = var_427_cast_fp16)[name = tensor("cast_136")]; + tensor onehot_7_cast_fp16 = reshape(shape = var_437, x = var_427_cast_fp16_to_fp16)[name = tensor("onehot_7_cast_fp16")]; + tensor var_439_to_fp16 = const()[name = tensor("op_439_to_fp16"), val = tensor(0x1p+0)]; + tensor var_441_cast_fp16 = sub(x = var_439_to_fp16, y = onehot_7_cast_fp16)[name = tensor("op_441_cast_fp16")]; + tensor k_cache1_to_fp16_dtype_0 = const()[name = tensor("k_cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache1_to_fp16 = cast(dtype = k_cache1_to_fp16_dtype_0, x = k_cache1)[name = tensor("cast_135")]; + tensor var_442_cast_fp16 = mul(x = k_cache1_to_fp16, y = var_441_cast_fp16)[name = tensor("op_442_cast_fp16")]; + tensor var_447 = const()[name = tensor("op_447"), val = tensor([1, 1, 16, 64])]; + tensor var_448_cast_fp16 = reshape(shape = var_447, x = out_7_cast_fp16)[name = tensor("op_448_cast_fp16")]; + tensor var_449_cast_fp16 = mul(x = var_448_cast_fp16, y = onehot_7_cast_fp16)[name = tensor("op_449_cast_fp16")]; + tensor new_k_cache_3_cast_fp16 = add(x = var_442_cast_fp16, y = var_449_cast_fp16)[name = tensor("new_k_cache_3_cast_fp16")]; + tensor new_k_cache_3_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_3_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache1_to_fp16_dtype_0 = const()[name = tensor("v_cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache1_to_fp16 = cast(dtype = v_cache1_to_fp16_dtype_0, x = v_cache1)[name = tensor("cast_133")]; + tensor var_455_cast_fp16 = mul(x = v_cache1_to_fp16, y = var_441_cast_fp16)[name = tensor("op_455_cast_fp16")]; + tensor var_460 = const()[name = tensor("op_460"), val = tensor([1, 1, 16, 64])]; + tensor var_461_cast_fp16 = reshape(shape = var_460, x = v_3_cast_fp16)[name = tensor("op_461_cast_fp16")]; + tensor var_462_cast_fp16 = mul(x = var_461_cast_fp16, y = onehot_7_cast_fp16)[name = tensor("op_462_cast_fp16")]; + tensor new_v_cache_3_cast_fp16 = add(x = var_455_cast_fp16, y = var_462_cast_fp16)[name = tensor("new_v_cache_3_cast_fp16")]; + tensor new_v_cache_3_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_3_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_1_perm_0 = const()[name = tensor("transpose_1_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_475 = const()[name = tensor("op_475"), val = tensor([0, 2, 1, 3])]; + tensor var_481 = const()[name = tensor("op_481"), val = tensor([1, 16, 1, 64])]; + tensor q4_3_cast_fp16 = reshape(shape = var_481, x = out_5_cast_fp16)[name = tensor("q4_3_cast_fp16")]; + tensor var_486_transpose_x_0 = const()[name = tensor("op_486_transpose_x_0"), val = tensor(false)]; + tensor var_486_transpose_y_0 = const()[name = tensor("op_486_transpose_y_0"), val = tensor(false)]; + tensor transpose_1_cast_fp16 = transpose(perm = transpose_1_perm_0, x = new_k_cache_3_cast_fp16)[name = tensor("transpose_20")]; + tensor var_486_cast_fp16 = matmul(transpose_x = var_486_transpose_x_0, transpose_y = var_486_transpose_y_0, x = q4_3_cast_fp16, y = transpose_1_cast_fp16)[name = tensor("op_486_cast_fp16")]; + tensor var_487_to_fp16 = const()[name = tensor("op_487_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_5_cast_fp16 = mul(x = var_486_cast_fp16, y = var_487_to_fp16)[name = tensor("scores_5_cast_fp16")]; + tensor var_491 = const()[name = tensor("op_491"), val = tensor([1, 1])]; + tensor var_492_cast_fp16 = reshape(shape = var_491, x = position1_to_fp16)[name = tensor("op_492_cast_fp16")]; + tensor var_493_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_492_cast_fp16)[name = tensor("op_493_cast_fp16")]; + tensor mask_3_to_fp16_dtype_0 = const()[name = tensor("mask_3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_500_to_fp16 = const()[name = tensor("op_500_to_fp16"), val = tensor(0x1p+0)]; + tensor var_493_cast_fp16_to_fp16 = cast(dtype = mask_3_to_fp16_dtype_0, x = var_493_cast_fp16)[name = tensor("cast_131")]; + tensor var_501_cast_fp16 = sub(x = var_493_cast_fp16_to_fp16, y = var_500_to_fp16)[name = tensor("op_501_cast_fp16")]; + tensor var_506 = const()[name = tensor("op_506"), val = tensor([1, 1, 1, 512])]; + tensor var_507_cast_fp16 = reshape(shape = var_506, x = var_501_cast_fp16)[name = tensor("op_507_cast_fp16")]; + tensor var_508_to_fp16 = const()[name = tensor("op_508_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_509_cast_fp16 = mul(x = var_507_cast_fp16, y = var_508_to_fp16)[name = tensor("op_509_cast_fp16")]; + tensor scores_7_cast_fp16 = add(x = scores_5_cast_fp16, y = var_509_cast_fp16)[name = tensor("scores_7_cast_fp16")]; + tensor var_512 = const()[name = tensor("op_512"), val = tensor(-1)]; + tensor weights_3_cast_fp16 = softmax(axis = var_512, x = scores_7_cast_fp16)[name = tensor("weights_3_cast_fp16")]; + tensor attn_3_transpose_x_0 = const()[name = tensor("attn_3_transpose_x_0"), val = tensor(false)]; + tensor attn_3_transpose_y_0 = const()[name = tensor("attn_3_transpose_y_0"), val = tensor(false)]; + tensor values_3_cast_fp16 = transpose(perm = var_475, x = new_v_cache_3_cast_fp16)[name = tensor("transpose_19")]; + tensor attn_3_cast_fp16 = matmul(transpose_x = attn_3_transpose_x_0, transpose_y = attn_3_transpose_y_0, x = weights_3_cast_fp16, y = values_3_cast_fp16)[name = tensor("attn_3_cast_fp16")]; + tensor var_520 = const()[name = tensor("op_520"), val = tensor([0, 2, 1, 3])]; + tensor var_525 = const()[name = tensor("op_525"), val = tensor([1, 1, 1024])]; + tensor var_521_cast_fp16 = transpose(perm = var_520, x = attn_3_cast_fp16)[name = tensor("transpose_18")]; + tensor input_19_cast_fp16 = reshape(shape = var_525, x = var_521_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor attn1_out_proj_weight_to_fp16 = const()[name = tensor("attn1_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31553728)))]; + tensor linear_6_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn1_out_proj_weight_to_fp16, x = input_19_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor var_531_to_fp16 = const()[name = tensor("op_531_to_fp16"), val = tensor(0x1p+0)]; + tensor var_532_cast_fp16 = add(x = position1_to_fp16, y = var_531_to_fp16)[name = tensor("op_532_cast_fp16")]; + tensor var_532_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_532_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_21_cast_fp16 = add(x = input_15_cast_fp16, y = linear_6_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor input_23_axes_0 = const()[name = tensor("input_23_axes_0"), val = tensor([-1])]; + tensor norm1_2_weight_to_fp16 = const()[name = tensor("norm1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33650944)))]; + tensor norm1_2_bias_to_fp16 = const()[name = tensor("norm1_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33653056)))]; + tensor var_536_to_fp16 = const()[name = tensor("op_536_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_23_cast_fp16 = layer_norm(axes = input_23_axes_0, beta = norm1_2_bias_to_fp16, epsilon = var_536_to_fp16, gamma = norm1_2_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor linear1_1_weight_to_fp16 = const()[name = tensor("linear1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33655168)))]; + tensor linear_7_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear1_1_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_25_mode_0 = const()[name = tensor("input_25_mode_0"), val = tensor("EXACT")]; + tensor input_25_cast_fp16 = gelu(mode = input_25_mode_0, x = linear_7_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor linear1_2_weight_to_fp16 = const()[name = tensor("linear1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42043840)))]; + tensor linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear1_2_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor input_27_cast_fp16 = add(x = input_21_cast_fp16, y = linear_8_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor input_29_axes_0 = const()[name = tensor("input_29_axes_0"), val = tensor([-1])]; + tensor norm2_1_weight_to_fp16 = const()[name = tensor("norm2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50432512)))]; + tensor norm2_1_bias_to_fp16 = const()[name = tensor("norm2_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50434624)))]; + tensor var_553_to_fp16 = const()[name = tensor("op_553_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_29_cast_fp16 = layer_norm(axes = input_29_axes_0, beta = norm2_1_bias_to_fp16, epsilon = var_553_to_fp16, gamma = norm2_1_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor attn2_in_proj_weight_to_fp16 = const()[name = tensor("attn2_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50436736)))]; + tensor linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn2_in_proj_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor var_566 = const()[name = tensor("op_566"), val = tensor([1, 3, 16, 64])]; + tensor qkv_5_cast_fp16 = reshape(shape = var_566, x = linear_9_cast_fp16)[name = tensor("qkv_5_cast_fp16")]; + tensor x_9_begin_0 = const()[name = tensor("x_9_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_9_end_0 = const()[name = tensor("x_9_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_9_end_mask_0 = const()[name = tensor("x_9_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_9_squeeze_mask_0 = const()[name = tensor("x_9_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_9_cast_fp16 = slice_by_index(begin = x_9_begin_0, end = x_9_end_0, end_mask = x_9_end_mask_0, squeeze_mask = x_9_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("x_9_cast_fp16")]; + tensor x_11_begin_0 = const()[name = tensor("x_11_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_11_end_0 = const()[name = tensor("x_11_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_11_end_mask_0 = const()[name = tensor("x_11_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_11_squeeze_mask_0 = const()[name = tensor("x_11_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_11_cast_fp16 = slice_by_index(begin = x_11_begin_0, end = x_11_end_0, end_mask = x_11_end_mask_0, squeeze_mask = x_11_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("x_11_cast_fp16")]; + tensor v_5_begin_0 = const()[name = tensor("v_5_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_5_end_0 = const()[name = tensor("v_5_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_5_end_mask_0 = const()[name = tensor("v_5_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_5_squeeze_mask_0 = const()[name = tensor("v_5_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_5_cast_fp16 = slice_by_index(begin = v_5_begin_0, end = v_5_end_0, end_mask = v_5_end_mask_0, squeeze_mask = v_5_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("v_5_cast_fp16")]; + tensor var_595 = const()[name = tensor("op_595"), val = tensor([1, 1, 1])]; + tensor position2_to_fp16_dtype_0 = const()[name = tensor("position2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position2_to_fp16 = cast(dtype = position2_to_fp16_dtype_0, x = position2)[name = tensor("cast_129")]; + tensor var_596_cast_fp16 = reshape(shape = var_595, x = position2_to_fp16)[name = tensor("op_596_cast_fp16")]; + tensor angles_9_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_596_cast_fp16)[name = tensor("angles_9_cast_fp16")]; + tensor rotr_9_cast_fp16 = cos(x = angles_9_cast_fp16)[name = tensor("rotr_9_cast_fp16")]; + tensor roti_9_cast_fp16 = sin(x = angles_9_cast_fp16)[name = tensor("roti_9_cast_fp16")]; + tensor var_604 = const()[name = tensor("op_604"), val = tensor([1, 16, 32, 2])]; + tensor pairs_9_cast_fp16 = reshape(shape = var_604, x = x_9_cast_fp16)[name = tensor("pairs_9_cast_fp16")]; + tensor xr_9_begin_0 = const()[name = tensor("xr_9_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_9_end_0 = const()[name = tensor("xr_9_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_9_end_mask_0 = const()[name = tensor("xr_9_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_9_squeeze_mask_0 = const()[name = tensor("xr_9_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_9_cast_fp16 = slice_by_index(begin = xr_9_begin_0, end = xr_9_end_0, end_mask = xr_9_end_mask_0, squeeze_mask = xr_9_squeeze_mask_0, x = pairs_9_cast_fp16)[name = tensor("xr_9_cast_fp16")]; + tensor xi_9_begin_0 = const()[name = tensor("xi_9_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_9_end_0 = const()[name = tensor("xi_9_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_9_end_mask_0 = const()[name = tensor("xi_9_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_9_squeeze_mask_0 = const()[name = tensor("xi_9_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_9_cast_fp16 = slice_by_index(begin = xi_9_begin_0, end = xi_9_end_0, end_mask = xi_9_end_mask_0, squeeze_mask = xi_9_squeeze_mask_0, x = pairs_9_cast_fp16)[name = tensor("xi_9_cast_fp16")]; + tensor var_612_cast_fp16 = mul(x = xr_9_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_612_cast_fp16")]; + tensor var_613_cast_fp16 = mul(x = xi_9_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_613_cast_fp16")]; + tensor xor__9_cast_fp16 = sub(x = var_612_cast_fp16, y = var_613_cast_fp16)[name = tensor("xor__9_cast_fp16")]; + tensor var_616_cast_fp16 = mul(x = xr_9_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_616_cast_fp16")]; + tensor var_617_cast_fp16 = mul(x = xi_9_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_617_cast_fp16")]; + tensor xoi_9_cast_fp16 = add(x = var_616_cast_fp16, y = var_617_cast_fp16)[name = tensor("xoi_9_cast_fp16")]; + tensor out_9_axis_0 = const()[name = tensor("out_9_axis_0"), val = tensor(-1)]; + tensor out_9_cast_fp16 = stack(axis = out_9_axis_0, values = (xor__9_cast_fp16, xoi_9_cast_fp16))[name = tensor("out_9_cast_fp16")]; + tensor var_640 = const()[name = tensor("op_640"), val = tensor([1, 16, 32, 2])]; + tensor pairs_11_cast_fp16 = reshape(shape = var_640, x = x_11_cast_fp16)[name = tensor("pairs_11_cast_fp16")]; + tensor xr_11_begin_0 = const()[name = tensor("xr_11_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_11_end_0 = const()[name = tensor("xr_11_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_11_end_mask_0 = const()[name = tensor("xr_11_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_11_squeeze_mask_0 = const()[name = tensor("xr_11_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_11_cast_fp16 = slice_by_index(begin = xr_11_begin_0, end = xr_11_end_0, end_mask = xr_11_end_mask_0, squeeze_mask = xr_11_squeeze_mask_0, x = pairs_11_cast_fp16)[name = tensor("xr_11_cast_fp16")]; + tensor xi_11_begin_0 = const()[name = tensor("xi_11_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_11_end_0 = const()[name = tensor("xi_11_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_11_end_mask_0 = const()[name = tensor("xi_11_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_11_squeeze_mask_0 = const()[name = tensor("xi_11_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_11_cast_fp16 = slice_by_index(begin = xi_11_begin_0, end = xi_11_end_0, end_mask = xi_11_end_mask_0, squeeze_mask = xi_11_squeeze_mask_0, x = pairs_11_cast_fp16)[name = tensor("xi_11_cast_fp16")]; + tensor var_648_cast_fp16 = mul(x = xr_11_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_648_cast_fp16")]; + tensor var_649_cast_fp16 = mul(x = xi_11_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_649_cast_fp16")]; + tensor xor__11_cast_fp16 = sub(x = var_648_cast_fp16, y = var_649_cast_fp16)[name = tensor("xor__11_cast_fp16")]; + tensor var_652_cast_fp16 = mul(x = xr_11_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_652_cast_fp16")]; + tensor var_653_cast_fp16 = mul(x = xi_11_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_653_cast_fp16")]; + tensor xoi_11_cast_fp16 = add(x = var_652_cast_fp16, y = var_653_cast_fp16)[name = tensor("xoi_11_cast_fp16")]; + tensor out_11_axis_0 = const()[name = tensor("out_11_axis_0"), val = tensor(-1)]; + tensor out_11_cast_fp16 = stack(axis = out_11_axis_0, values = (xor__11_cast_fp16, xoi_11_cast_fp16))[name = tensor("out_11_cast_fp16")]; + tensor _inversed_665_y_0_to_fp16 = const()[name = tensor("_inversed_665_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_665_cast_fp16 = mul(x = position2_to_fp16, y = _inversed_665_y_0_to_fp16)[name = tensor("_inversed_665_cast_fp16")]; + tensor var_666_cast_fp16 = floor(x = _inversed_665_cast_fp16)[name = tensor("op_666_cast_fp16")]; + tensor var_667_to_fp16 = const()[name = tensor("op_667_to_fp16"), val = tensor(0x1p+9)]; + tensor var_668_cast_fp16 = mul(x = var_666_cast_fp16, y = var_667_to_fp16)[name = tensor("op_668_cast_fp16")]; + tensor write_idx_5_cast_fp16 = sub(x = position2_to_fp16, y = var_668_cast_fp16)[name = tensor("write_idx_5_cast_fp16")]; + tensor var_671_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_5_cast_fp16)[name = tensor("op_671_cast_fp16")]; + tensor var_681 = const()[name = tensor("op_681"), val = tensor([1, 512, 1, 1])]; + tensor onehot_9_to_fp16_dtype_0 = const()[name = tensor("onehot_9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_671_cast_fp16_to_fp16 = cast(dtype = onehot_9_to_fp16_dtype_0, x = var_671_cast_fp16)[name = tensor("cast_128")]; + tensor onehot_11_cast_fp16 = reshape(shape = var_681, x = var_671_cast_fp16_to_fp16)[name = tensor("onehot_11_cast_fp16")]; + tensor var_683_to_fp16 = const()[name = tensor("op_683_to_fp16"), val = tensor(0x1p+0)]; + tensor var_685_cast_fp16 = sub(x = var_683_to_fp16, y = onehot_11_cast_fp16)[name = tensor("op_685_cast_fp16")]; + tensor k_cache2_to_fp16_dtype_0 = const()[name = tensor("k_cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache2_to_fp16 = cast(dtype = k_cache2_to_fp16_dtype_0, x = k_cache2)[name = tensor("cast_127")]; + tensor var_686_cast_fp16 = mul(x = k_cache2_to_fp16, y = var_685_cast_fp16)[name = tensor("op_686_cast_fp16")]; + tensor var_691 = const()[name = tensor("op_691"), val = tensor([1, 1, 16, 64])]; + tensor var_692_cast_fp16 = reshape(shape = var_691, x = out_11_cast_fp16)[name = tensor("op_692_cast_fp16")]; + tensor var_693_cast_fp16 = mul(x = var_692_cast_fp16, y = onehot_11_cast_fp16)[name = tensor("op_693_cast_fp16")]; + tensor new_k_cache_5_cast_fp16 = add(x = var_686_cast_fp16, y = var_693_cast_fp16)[name = tensor("new_k_cache_5_cast_fp16")]; + tensor new_k_cache_5_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_5_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache2_to_fp16_dtype_0 = const()[name = tensor("v_cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache2_to_fp16 = cast(dtype = v_cache2_to_fp16_dtype_0, x = v_cache2)[name = tensor("cast_125")]; + tensor var_699_cast_fp16 = mul(x = v_cache2_to_fp16, y = var_685_cast_fp16)[name = tensor("op_699_cast_fp16")]; + tensor var_704 = const()[name = tensor("op_704"), val = tensor([1, 1, 16, 64])]; + tensor var_705_cast_fp16 = reshape(shape = var_704, x = v_5_cast_fp16)[name = tensor("op_705_cast_fp16")]; + tensor var_706_cast_fp16 = mul(x = var_705_cast_fp16, y = onehot_11_cast_fp16)[name = tensor("op_706_cast_fp16")]; + tensor new_v_cache_5_cast_fp16 = add(x = var_699_cast_fp16, y = var_706_cast_fp16)[name = tensor("new_v_cache_5_cast_fp16")]; + tensor new_v_cache_5_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_5_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_2_perm_0 = const()[name = tensor("transpose_2_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_719 = const()[name = tensor("op_719"), val = tensor([0, 2, 1, 3])]; + tensor var_725 = const()[name = tensor("op_725"), val = tensor([1, 16, 1, 64])]; + tensor q4_5_cast_fp16 = reshape(shape = var_725, x = out_9_cast_fp16)[name = tensor("q4_5_cast_fp16")]; + tensor var_730_transpose_x_0 = const()[name = tensor("op_730_transpose_x_0"), val = tensor(false)]; + tensor var_730_transpose_y_0 = const()[name = tensor("op_730_transpose_y_0"), val = tensor(false)]; + tensor transpose_2_cast_fp16 = transpose(perm = transpose_2_perm_0, x = new_k_cache_5_cast_fp16)[name = tensor("transpose_17")]; + tensor var_730_cast_fp16 = matmul(transpose_x = var_730_transpose_x_0, transpose_y = var_730_transpose_y_0, x = q4_5_cast_fp16, y = transpose_2_cast_fp16)[name = tensor("op_730_cast_fp16")]; + tensor var_731_to_fp16 = const()[name = tensor("op_731_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_9_cast_fp16 = mul(x = var_730_cast_fp16, y = var_731_to_fp16)[name = tensor("scores_9_cast_fp16")]; + tensor var_735 = const()[name = tensor("op_735"), val = tensor([1, 1])]; + tensor var_736_cast_fp16 = reshape(shape = var_735, x = position2_to_fp16)[name = tensor("op_736_cast_fp16")]; + tensor var_737_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_736_cast_fp16)[name = tensor("op_737_cast_fp16")]; + tensor mask_5_to_fp16_dtype_0 = const()[name = tensor("mask_5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_744_to_fp16 = const()[name = tensor("op_744_to_fp16"), val = tensor(0x1p+0)]; + tensor var_737_cast_fp16_to_fp16 = cast(dtype = mask_5_to_fp16_dtype_0, x = var_737_cast_fp16)[name = tensor("cast_123")]; + tensor var_745_cast_fp16 = sub(x = var_737_cast_fp16_to_fp16, y = var_744_to_fp16)[name = tensor("op_745_cast_fp16")]; + tensor var_750 = const()[name = tensor("op_750"), val = tensor([1, 1, 1, 512])]; + tensor var_751_cast_fp16 = reshape(shape = var_750, x = var_745_cast_fp16)[name = tensor("op_751_cast_fp16")]; + tensor var_752_to_fp16 = const()[name = tensor("op_752_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_753_cast_fp16 = mul(x = var_751_cast_fp16, y = var_752_to_fp16)[name = tensor("op_753_cast_fp16")]; + tensor scores_11_cast_fp16 = add(x = scores_9_cast_fp16, y = var_753_cast_fp16)[name = tensor("scores_11_cast_fp16")]; + tensor var_756 = const()[name = tensor("op_756"), val = tensor(-1)]; + tensor weights_5_cast_fp16 = softmax(axis = var_756, x = scores_11_cast_fp16)[name = tensor("weights_5_cast_fp16")]; + tensor attn_5_transpose_x_0 = const()[name = tensor("attn_5_transpose_x_0"), val = tensor(false)]; + tensor attn_5_transpose_y_0 = const()[name = tensor("attn_5_transpose_y_0"), val = tensor(false)]; + tensor values_5_cast_fp16 = transpose(perm = var_719, x = new_v_cache_5_cast_fp16)[name = tensor("transpose_16")]; + tensor attn_5_cast_fp16 = matmul(transpose_x = attn_5_transpose_x_0, transpose_y = attn_5_transpose_y_0, x = weights_5_cast_fp16, y = values_5_cast_fp16)[name = tensor("attn_5_cast_fp16")]; + tensor var_764 = const()[name = tensor("op_764"), val = tensor([0, 2, 1, 3])]; + tensor var_769 = const()[name = tensor("op_769"), val = tensor([1, 1, 1024])]; + tensor var_765_cast_fp16 = transpose(perm = var_764, x = attn_5_cast_fp16)[name = tensor("transpose_15")]; + tensor input_31_cast_fp16 = reshape(shape = var_769, x = var_765_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor attn2_out_proj_weight_to_fp16 = const()[name = tensor("attn2_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(56728256)))]; + tensor linear_10_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn2_out_proj_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor var_775_to_fp16 = const()[name = tensor("op_775_to_fp16"), val = tensor(0x1p+0)]; + tensor var_776_cast_fp16 = add(x = position2_to_fp16, y = var_775_to_fp16)[name = tensor("op_776_cast_fp16")]; + tensor var_776_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_776_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_33_cast_fp16 = add(x = input_27_cast_fp16, y = linear_10_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor input_35_axes_0 = const()[name = tensor("input_35_axes_0"), val = tensor([-1])]; + tensor norm2_2_weight_to_fp16 = const()[name = tensor("norm2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58825472)))]; + tensor norm2_2_bias_to_fp16 = const()[name = tensor("norm2_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58827584)))]; + tensor var_780_to_fp16 = const()[name = tensor("op_780_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_35_cast_fp16 = layer_norm(axes = input_35_axes_0, beta = norm2_2_bias_to_fp16, epsilon = var_780_to_fp16, gamma = norm2_2_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor linear2_1_weight_to_fp16 = const()[name = tensor("linear2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58829696)))]; + tensor linear_11_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear2_1_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor input_37_mode_0 = const()[name = tensor("input_37_mode_0"), val = tensor("EXACT")]; + tensor input_37_cast_fp16 = gelu(mode = input_37_mode_0, x = linear_11_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor linear2_2_weight_to_fp16 = const()[name = tensor("linear2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67218368)))]; + tensor linear_12_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear2_2_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor input_39_cast_fp16 = add(x = input_33_cast_fp16, y = linear_12_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor input_41_axes_0 = const()[name = tensor("input_41_axes_0"), val = tensor([-1])]; + tensor norm3_1_weight_to_fp16 = const()[name = tensor("norm3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75607040)))]; + tensor norm3_1_bias_to_fp16 = const()[name = tensor("norm3_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75609152)))]; + tensor var_797_to_fp16 = const()[name = tensor("op_797_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_41_cast_fp16 = layer_norm(axes = input_41_axes_0, beta = norm3_1_bias_to_fp16, epsilon = var_797_to_fp16, gamma = norm3_1_weight_to_fp16, x = input_39_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor attn3_in_proj_weight_to_fp16 = const()[name = tensor("attn3_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75611264)))]; + tensor linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn3_in_proj_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor var_810 = const()[name = tensor("op_810"), val = tensor([1, 3, 16, 64])]; + tensor qkv_7_cast_fp16 = reshape(shape = var_810, x = linear_13_cast_fp16)[name = tensor("qkv_7_cast_fp16")]; + tensor x_13_begin_0 = const()[name = tensor("x_13_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_13_end_0 = const()[name = tensor("x_13_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_13_end_mask_0 = const()[name = tensor("x_13_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_13_squeeze_mask_0 = const()[name = tensor("x_13_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_13_cast_fp16 = slice_by_index(begin = x_13_begin_0, end = x_13_end_0, end_mask = x_13_end_mask_0, squeeze_mask = x_13_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("x_13_cast_fp16")]; + tensor x_15_begin_0 = const()[name = tensor("x_15_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_15_end_0 = const()[name = tensor("x_15_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_15_end_mask_0 = const()[name = tensor("x_15_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_15_squeeze_mask_0 = const()[name = tensor("x_15_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_15_cast_fp16 = slice_by_index(begin = x_15_begin_0, end = x_15_end_0, end_mask = x_15_end_mask_0, squeeze_mask = x_15_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("x_15_cast_fp16")]; + tensor v_7_begin_0 = const()[name = tensor("v_7_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_7_end_0 = const()[name = tensor("v_7_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_7_end_mask_0 = const()[name = tensor("v_7_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_7_squeeze_mask_0 = const()[name = tensor("v_7_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_7_cast_fp16 = slice_by_index(begin = v_7_begin_0, end = v_7_end_0, end_mask = v_7_end_mask_0, squeeze_mask = v_7_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("v_7_cast_fp16")]; + tensor var_839 = const()[name = tensor("op_839"), val = tensor([1, 1, 1])]; + tensor position3_to_fp16_dtype_0 = const()[name = tensor("position3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position3_to_fp16 = cast(dtype = position3_to_fp16_dtype_0, x = position3)[name = tensor("cast_121")]; + tensor var_840_cast_fp16 = reshape(shape = var_839, x = position3_to_fp16)[name = tensor("op_840_cast_fp16")]; + tensor angles_13_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_840_cast_fp16)[name = tensor("angles_13_cast_fp16")]; + tensor rotr_13_cast_fp16 = cos(x = angles_13_cast_fp16)[name = tensor("rotr_13_cast_fp16")]; + tensor roti_13_cast_fp16 = sin(x = angles_13_cast_fp16)[name = tensor("roti_13_cast_fp16")]; + tensor var_848 = const()[name = tensor("op_848"), val = tensor([1, 16, 32, 2])]; + tensor pairs_13_cast_fp16 = reshape(shape = var_848, x = x_13_cast_fp16)[name = tensor("pairs_13_cast_fp16")]; + tensor xr_13_begin_0 = const()[name = tensor("xr_13_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_13_end_0 = const()[name = tensor("xr_13_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_13_end_mask_0 = const()[name = tensor("xr_13_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_13_squeeze_mask_0 = const()[name = tensor("xr_13_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_13_cast_fp16 = slice_by_index(begin = xr_13_begin_0, end = xr_13_end_0, end_mask = xr_13_end_mask_0, squeeze_mask = xr_13_squeeze_mask_0, x = pairs_13_cast_fp16)[name = tensor("xr_13_cast_fp16")]; + tensor xi_13_begin_0 = const()[name = tensor("xi_13_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_13_end_0 = const()[name = tensor("xi_13_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_13_end_mask_0 = const()[name = tensor("xi_13_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_13_squeeze_mask_0 = const()[name = tensor("xi_13_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_13_cast_fp16 = slice_by_index(begin = xi_13_begin_0, end = xi_13_end_0, end_mask = xi_13_end_mask_0, squeeze_mask = xi_13_squeeze_mask_0, x = pairs_13_cast_fp16)[name = tensor("xi_13_cast_fp16")]; + tensor var_856_cast_fp16 = mul(x = xr_13_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_856_cast_fp16")]; + tensor var_857_cast_fp16 = mul(x = xi_13_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_857_cast_fp16")]; + tensor xor__13_cast_fp16 = sub(x = var_856_cast_fp16, y = var_857_cast_fp16)[name = tensor("xor__13_cast_fp16")]; + tensor var_860_cast_fp16 = mul(x = xr_13_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_860_cast_fp16")]; + tensor var_861_cast_fp16 = mul(x = xi_13_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_861_cast_fp16")]; + tensor xoi_13_cast_fp16 = add(x = var_860_cast_fp16, y = var_861_cast_fp16)[name = tensor("xoi_13_cast_fp16")]; + tensor out_13_axis_0 = const()[name = tensor("out_13_axis_0"), val = tensor(-1)]; + tensor out_13_cast_fp16 = stack(axis = out_13_axis_0, values = (xor__13_cast_fp16, xoi_13_cast_fp16))[name = tensor("out_13_cast_fp16")]; + tensor var_884 = const()[name = tensor("op_884"), val = tensor([1, 16, 32, 2])]; + tensor pairs_15_cast_fp16 = reshape(shape = var_884, x = x_15_cast_fp16)[name = tensor("pairs_15_cast_fp16")]; + tensor xr_15_begin_0 = const()[name = tensor("xr_15_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_15_end_0 = const()[name = tensor("xr_15_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_15_end_mask_0 = const()[name = tensor("xr_15_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_15_squeeze_mask_0 = const()[name = tensor("xr_15_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_15_cast_fp16 = slice_by_index(begin = xr_15_begin_0, end = xr_15_end_0, end_mask = xr_15_end_mask_0, squeeze_mask = xr_15_squeeze_mask_0, x = pairs_15_cast_fp16)[name = tensor("xr_15_cast_fp16")]; + tensor xi_15_begin_0 = const()[name = tensor("xi_15_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_15_end_0 = const()[name = tensor("xi_15_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_15_end_mask_0 = const()[name = tensor("xi_15_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_15_squeeze_mask_0 = const()[name = tensor("xi_15_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_15_cast_fp16 = slice_by_index(begin = xi_15_begin_0, end = xi_15_end_0, end_mask = xi_15_end_mask_0, squeeze_mask = xi_15_squeeze_mask_0, x = pairs_15_cast_fp16)[name = tensor("xi_15_cast_fp16")]; + tensor var_892_cast_fp16 = mul(x = xr_15_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_892_cast_fp16")]; + tensor var_893_cast_fp16 = mul(x = xi_15_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_893_cast_fp16")]; + tensor xor__15_cast_fp16 = sub(x = var_892_cast_fp16, y = var_893_cast_fp16)[name = tensor("xor__15_cast_fp16")]; + tensor var_896_cast_fp16 = mul(x = xr_15_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_896_cast_fp16")]; + tensor var_897_cast_fp16 = mul(x = xi_15_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_897_cast_fp16")]; + tensor xoi_15_cast_fp16 = add(x = var_896_cast_fp16, y = var_897_cast_fp16)[name = tensor("xoi_15_cast_fp16")]; + tensor out_15_axis_0 = const()[name = tensor("out_15_axis_0"), val = tensor(-1)]; + tensor out_15_cast_fp16 = stack(axis = out_15_axis_0, values = (xor__15_cast_fp16, xoi_15_cast_fp16))[name = tensor("out_15_cast_fp16")]; + tensor _inversed_909_y_0_to_fp16 = const()[name = tensor("_inversed_909_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_909_cast_fp16 = mul(x = position3_to_fp16, y = _inversed_909_y_0_to_fp16)[name = tensor("_inversed_909_cast_fp16")]; + tensor var_910_cast_fp16 = floor(x = _inversed_909_cast_fp16)[name = tensor("op_910_cast_fp16")]; + tensor var_911_to_fp16 = const()[name = tensor("op_911_to_fp16"), val = tensor(0x1p+9)]; + tensor var_912_cast_fp16 = mul(x = var_910_cast_fp16, y = var_911_to_fp16)[name = tensor("op_912_cast_fp16")]; + tensor write_idx_7_cast_fp16 = sub(x = position3_to_fp16, y = var_912_cast_fp16)[name = tensor("write_idx_7_cast_fp16")]; + tensor var_915_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_7_cast_fp16)[name = tensor("op_915_cast_fp16")]; + tensor var_925 = const()[name = tensor("op_925"), val = tensor([1, 512, 1, 1])]; + tensor onehot_13_to_fp16_dtype_0 = const()[name = tensor("onehot_13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_915_cast_fp16_to_fp16 = cast(dtype = onehot_13_to_fp16_dtype_0, x = var_915_cast_fp16)[name = tensor("cast_120")]; + tensor onehot_15_cast_fp16 = reshape(shape = var_925, x = var_915_cast_fp16_to_fp16)[name = tensor("onehot_15_cast_fp16")]; + tensor var_927_to_fp16 = const()[name = tensor("op_927_to_fp16"), val = tensor(0x1p+0)]; + tensor var_929_cast_fp16 = sub(x = var_927_to_fp16, y = onehot_15_cast_fp16)[name = tensor("op_929_cast_fp16")]; + tensor k_cache3_to_fp16_dtype_0 = const()[name = tensor("k_cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache3_to_fp16 = cast(dtype = k_cache3_to_fp16_dtype_0, x = k_cache3)[name = tensor("cast_119")]; + tensor var_930_cast_fp16 = mul(x = k_cache3_to_fp16, y = var_929_cast_fp16)[name = tensor("op_930_cast_fp16")]; + tensor var_935 = const()[name = tensor("op_935"), val = tensor([1, 1, 16, 64])]; + tensor var_936_cast_fp16 = reshape(shape = var_935, x = out_15_cast_fp16)[name = tensor("op_936_cast_fp16")]; + tensor var_937_cast_fp16 = mul(x = var_936_cast_fp16, y = onehot_15_cast_fp16)[name = tensor("op_937_cast_fp16")]; + tensor new_k_cache_7_cast_fp16 = add(x = var_930_cast_fp16, y = var_937_cast_fp16)[name = tensor("new_k_cache_7_cast_fp16")]; + tensor new_k_cache_7_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_7_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache3_to_fp16_dtype_0 = const()[name = tensor("v_cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache3_to_fp16 = cast(dtype = v_cache3_to_fp16_dtype_0, x = v_cache3)[name = tensor("cast_117")]; + tensor var_943_cast_fp16 = mul(x = v_cache3_to_fp16, y = var_929_cast_fp16)[name = tensor("op_943_cast_fp16")]; + tensor var_948 = const()[name = tensor("op_948"), val = tensor([1, 1, 16, 64])]; + tensor var_949_cast_fp16 = reshape(shape = var_948, x = v_7_cast_fp16)[name = tensor("op_949_cast_fp16")]; + tensor var_950_cast_fp16 = mul(x = var_949_cast_fp16, y = onehot_15_cast_fp16)[name = tensor("op_950_cast_fp16")]; + tensor new_v_cache_7_cast_fp16 = add(x = var_943_cast_fp16, y = var_950_cast_fp16)[name = tensor("new_v_cache_7_cast_fp16")]; + tensor new_v_cache_7_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_7_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_3_perm_0 = const()[name = tensor("transpose_3_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_963 = const()[name = tensor("op_963"), val = tensor([0, 2, 1, 3])]; + tensor var_969 = const()[name = tensor("op_969"), val = tensor([1, 16, 1, 64])]; + tensor q4_7_cast_fp16 = reshape(shape = var_969, x = out_13_cast_fp16)[name = tensor("q4_7_cast_fp16")]; + tensor var_974_transpose_x_0 = const()[name = tensor("op_974_transpose_x_0"), val = tensor(false)]; + tensor var_974_transpose_y_0 = const()[name = tensor("op_974_transpose_y_0"), val = tensor(false)]; + tensor transpose_3_cast_fp16 = transpose(perm = transpose_3_perm_0, x = new_k_cache_7_cast_fp16)[name = tensor("transpose_14")]; + tensor var_974_cast_fp16 = matmul(transpose_x = var_974_transpose_x_0, transpose_y = var_974_transpose_y_0, x = q4_7_cast_fp16, y = transpose_3_cast_fp16)[name = tensor("op_974_cast_fp16")]; + tensor var_975_to_fp16 = const()[name = tensor("op_975_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_13_cast_fp16 = mul(x = var_974_cast_fp16, y = var_975_to_fp16)[name = tensor("scores_13_cast_fp16")]; + tensor var_979 = const()[name = tensor("op_979"), val = tensor([1, 1])]; + tensor var_980_cast_fp16 = reshape(shape = var_979, x = position3_to_fp16)[name = tensor("op_980_cast_fp16")]; + tensor var_981_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_980_cast_fp16)[name = tensor("op_981_cast_fp16")]; + tensor mask_7_to_fp16_dtype_0 = const()[name = tensor("mask_7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_988_to_fp16 = const()[name = tensor("op_988_to_fp16"), val = tensor(0x1p+0)]; + tensor var_981_cast_fp16_to_fp16 = cast(dtype = mask_7_to_fp16_dtype_0, x = var_981_cast_fp16)[name = tensor("cast_115")]; + tensor var_989_cast_fp16 = sub(x = var_981_cast_fp16_to_fp16, y = var_988_to_fp16)[name = tensor("op_989_cast_fp16")]; + tensor var_994 = const()[name = tensor("op_994"), val = tensor([1, 1, 1, 512])]; + tensor var_995_cast_fp16 = reshape(shape = var_994, x = var_989_cast_fp16)[name = tensor("op_995_cast_fp16")]; + tensor var_996_to_fp16 = const()[name = tensor("op_996_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_997_cast_fp16 = mul(x = var_995_cast_fp16, y = var_996_to_fp16)[name = tensor("op_997_cast_fp16")]; + tensor scores_15_cast_fp16 = add(x = scores_13_cast_fp16, y = var_997_cast_fp16)[name = tensor("scores_15_cast_fp16")]; + tensor var_1000 = const()[name = tensor("op_1000"), val = tensor(-1)]; + tensor weights_7_cast_fp16 = softmax(axis = var_1000, x = scores_15_cast_fp16)[name = tensor("weights_7_cast_fp16")]; + tensor attn_7_transpose_x_0 = const()[name = tensor("attn_7_transpose_x_0"), val = tensor(false)]; + tensor attn_7_transpose_y_0 = const()[name = tensor("attn_7_transpose_y_0"), val = tensor(false)]; + tensor values_7_cast_fp16 = transpose(perm = var_963, x = new_v_cache_7_cast_fp16)[name = tensor("transpose_13")]; + tensor attn_7_cast_fp16 = matmul(transpose_x = attn_7_transpose_x_0, transpose_y = attn_7_transpose_y_0, x = weights_7_cast_fp16, y = values_7_cast_fp16)[name = tensor("attn_7_cast_fp16")]; + tensor var_1008 = const()[name = tensor("op_1008"), val = tensor([0, 2, 1, 3])]; + tensor var_1013 = const()[name = tensor("op_1013"), val = tensor([1, 1, 1024])]; + tensor var_1009_cast_fp16 = transpose(perm = var_1008, x = attn_7_cast_fp16)[name = tensor("transpose_12")]; + tensor input_43_cast_fp16 = reshape(shape = var_1013, x = var_1009_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor attn3_out_proj_weight_to_fp16 = const()[name = tensor("attn3_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(81902784)))]; + tensor linear_14_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn3_out_proj_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor var_1019_to_fp16 = const()[name = tensor("op_1019_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1020_cast_fp16 = add(x = position3_to_fp16, y = var_1019_to_fp16)[name = tensor("op_1020_cast_fp16")]; + tensor var_1020_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1020_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_45_cast_fp16 = add(x = input_39_cast_fp16, y = linear_14_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor input_47_axes_0 = const()[name = tensor("input_47_axes_0"), val = tensor([-1])]; + tensor norm3_2_weight_to_fp16 = const()[name = tensor("norm3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84000000)))]; + tensor norm3_2_bias_to_fp16 = const()[name = tensor("norm3_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84002112)))]; + tensor var_1024_to_fp16 = const()[name = tensor("op_1024_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_47_cast_fp16 = layer_norm(axes = input_47_axes_0, beta = norm3_2_bias_to_fp16, epsilon = var_1024_to_fp16, gamma = norm3_2_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor linear3_1_weight_to_fp16 = const()[name = tensor("linear3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84004224)))]; + tensor linear_15_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear3_1_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor input_49_mode_0 = const()[name = tensor("input_49_mode_0"), val = tensor("EXACT")]; + tensor input_49_cast_fp16 = gelu(mode = input_49_mode_0, x = linear_15_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor linear3_2_weight_to_fp16 = const()[name = tensor("linear3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(92392896)))]; + tensor linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear3_2_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor input_51_cast_fp16 = add(x = input_45_cast_fp16, y = linear_16_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor input_53_axes_0 = const()[name = tensor("input_53_axes_0"), val = tensor([-1])]; + tensor norm4_1_weight_to_fp16 = const()[name = tensor("norm4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100781568)))]; + tensor norm4_1_bias_to_fp16 = const()[name = tensor("norm4_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100783680)))]; + tensor var_1041_to_fp16 = const()[name = tensor("op_1041_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_53_cast_fp16 = layer_norm(axes = input_53_axes_0, beta = norm4_1_bias_to_fp16, epsilon = var_1041_to_fp16, gamma = norm4_1_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor attn4_in_proj_weight_to_fp16 = const()[name = tensor("attn4_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100785792)))]; + tensor linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn4_in_proj_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor var_1054 = const()[name = tensor("op_1054"), val = tensor([1, 3, 16, 64])]; + tensor qkv_9_cast_fp16 = reshape(shape = var_1054, x = linear_17_cast_fp16)[name = tensor("qkv_9_cast_fp16")]; + tensor x_17_begin_0 = const()[name = tensor("x_17_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_17_end_0 = const()[name = tensor("x_17_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_17_end_mask_0 = const()[name = tensor("x_17_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_17_squeeze_mask_0 = const()[name = tensor("x_17_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_17_cast_fp16 = slice_by_index(begin = x_17_begin_0, end = x_17_end_0, end_mask = x_17_end_mask_0, squeeze_mask = x_17_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("x_17_cast_fp16")]; + tensor x_19_begin_0 = const()[name = tensor("x_19_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_19_end_0 = const()[name = tensor("x_19_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_19_end_mask_0 = const()[name = tensor("x_19_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_19_squeeze_mask_0 = const()[name = tensor("x_19_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_19_cast_fp16 = slice_by_index(begin = x_19_begin_0, end = x_19_end_0, end_mask = x_19_end_mask_0, squeeze_mask = x_19_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("x_19_cast_fp16")]; + tensor v_9_begin_0 = const()[name = tensor("v_9_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_9_end_0 = const()[name = tensor("v_9_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_9_end_mask_0 = const()[name = tensor("v_9_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_9_squeeze_mask_0 = const()[name = tensor("v_9_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_9_cast_fp16 = slice_by_index(begin = v_9_begin_0, end = v_9_end_0, end_mask = v_9_end_mask_0, squeeze_mask = v_9_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("v_9_cast_fp16")]; + tensor var_1083 = const()[name = tensor("op_1083"), val = tensor([1, 1, 1])]; + tensor position4_to_fp16_dtype_0 = const()[name = tensor("position4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position4_to_fp16 = cast(dtype = position4_to_fp16_dtype_0, x = position4)[name = tensor("cast_113")]; + tensor var_1084_cast_fp16 = reshape(shape = var_1083, x = position4_to_fp16)[name = tensor("op_1084_cast_fp16")]; + tensor angles_17_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_1084_cast_fp16)[name = tensor("angles_17_cast_fp16")]; + tensor rotr_17_cast_fp16 = cos(x = angles_17_cast_fp16)[name = tensor("rotr_17_cast_fp16")]; + tensor roti_17_cast_fp16 = sin(x = angles_17_cast_fp16)[name = tensor("roti_17_cast_fp16")]; + tensor var_1092 = const()[name = tensor("op_1092"), val = tensor([1, 16, 32, 2])]; + tensor pairs_17_cast_fp16 = reshape(shape = var_1092, x = x_17_cast_fp16)[name = tensor("pairs_17_cast_fp16")]; + tensor xr_17_begin_0 = const()[name = tensor("xr_17_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_17_end_0 = const()[name = tensor("xr_17_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_17_end_mask_0 = const()[name = tensor("xr_17_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_17_squeeze_mask_0 = const()[name = tensor("xr_17_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_17_cast_fp16 = slice_by_index(begin = xr_17_begin_0, end = xr_17_end_0, end_mask = xr_17_end_mask_0, squeeze_mask = xr_17_squeeze_mask_0, x = pairs_17_cast_fp16)[name = tensor("xr_17_cast_fp16")]; + tensor xi_17_begin_0 = const()[name = tensor("xi_17_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_17_end_0 = const()[name = tensor("xi_17_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_17_end_mask_0 = const()[name = tensor("xi_17_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_17_squeeze_mask_0 = const()[name = tensor("xi_17_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_17_cast_fp16 = slice_by_index(begin = xi_17_begin_0, end = xi_17_end_0, end_mask = xi_17_end_mask_0, squeeze_mask = xi_17_squeeze_mask_0, x = pairs_17_cast_fp16)[name = tensor("xi_17_cast_fp16")]; + tensor var_1100_cast_fp16 = mul(x = xr_17_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_1100_cast_fp16")]; + tensor var_1101_cast_fp16 = mul(x = xi_17_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_1101_cast_fp16")]; + tensor xor__17_cast_fp16 = sub(x = var_1100_cast_fp16, y = var_1101_cast_fp16)[name = tensor("xor__17_cast_fp16")]; + tensor var_1104_cast_fp16 = mul(x = xr_17_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_1104_cast_fp16")]; + tensor var_1105_cast_fp16 = mul(x = xi_17_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_1105_cast_fp16")]; + tensor xoi_17_cast_fp16 = add(x = var_1104_cast_fp16, y = var_1105_cast_fp16)[name = tensor("xoi_17_cast_fp16")]; + tensor out_17_axis_0 = const()[name = tensor("out_17_axis_0"), val = tensor(-1)]; + tensor out_17_cast_fp16 = stack(axis = out_17_axis_0, values = (xor__17_cast_fp16, xoi_17_cast_fp16))[name = tensor("out_17_cast_fp16")]; + tensor var_1128 = const()[name = tensor("op_1128"), val = tensor([1, 16, 32, 2])]; + tensor pairs_19_cast_fp16 = reshape(shape = var_1128, x = x_19_cast_fp16)[name = tensor("pairs_19_cast_fp16")]; + tensor xr_19_begin_0 = const()[name = tensor("xr_19_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_19_end_0 = const()[name = tensor("xr_19_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_19_end_mask_0 = const()[name = tensor("xr_19_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_19_squeeze_mask_0 = const()[name = tensor("xr_19_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_19_cast_fp16 = slice_by_index(begin = xr_19_begin_0, end = xr_19_end_0, end_mask = xr_19_end_mask_0, squeeze_mask = xr_19_squeeze_mask_0, x = pairs_19_cast_fp16)[name = tensor("xr_19_cast_fp16")]; + tensor xi_19_begin_0 = const()[name = tensor("xi_19_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_19_end_0 = const()[name = tensor("xi_19_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_19_end_mask_0 = const()[name = tensor("xi_19_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_19_squeeze_mask_0 = const()[name = tensor("xi_19_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_19_cast_fp16 = slice_by_index(begin = xi_19_begin_0, end = xi_19_end_0, end_mask = xi_19_end_mask_0, squeeze_mask = xi_19_squeeze_mask_0, x = pairs_19_cast_fp16)[name = tensor("xi_19_cast_fp16")]; + tensor var_1136_cast_fp16 = mul(x = xr_19_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_1136_cast_fp16")]; + tensor var_1137_cast_fp16 = mul(x = xi_19_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_1137_cast_fp16")]; + tensor xor__19_cast_fp16 = sub(x = var_1136_cast_fp16, y = var_1137_cast_fp16)[name = tensor("xor__19_cast_fp16")]; + tensor var_1140_cast_fp16 = mul(x = xr_19_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_1140_cast_fp16")]; + tensor var_1141_cast_fp16 = mul(x = xi_19_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_1141_cast_fp16")]; + tensor xoi_19_cast_fp16 = add(x = var_1140_cast_fp16, y = var_1141_cast_fp16)[name = tensor("xoi_19_cast_fp16")]; + tensor out_19_axis_0 = const()[name = tensor("out_19_axis_0"), val = tensor(-1)]; + tensor out_19_cast_fp16 = stack(axis = out_19_axis_0, values = (xor__19_cast_fp16, xoi_19_cast_fp16))[name = tensor("out_19_cast_fp16")]; + tensor _inversed_1153_y_0_to_fp16 = const()[name = tensor("_inversed_1153_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1153_cast_fp16 = mul(x = position4_to_fp16, y = _inversed_1153_y_0_to_fp16)[name = tensor("_inversed_1153_cast_fp16")]; + tensor var_1154_cast_fp16 = floor(x = _inversed_1153_cast_fp16)[name = tensor("op_1154_cast_fp16")]; + tensor var_1155_to_fp16 = const()[name = tensor("op_1155_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1156_cast_fp16 = mul(x = var_1154_cast_fp16, y = var_1155_to_fp16)[name = tensor("op_1156_cast_fp16")]; + tensor write_idx_9_cast_fp16 = sub(x = position4_to_fp16, y = var_1156_cast_fp16)[name = tensor("write_idx_9_cast_fp16")]; + tensor var_1159_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_9_cast_fp16)[name = tensor("op_1159_cast_fp16")]; + tensor var_1169 = const()[name = tensor("op_1169"), val = tensor([1, 512, 1, 1])]; + tensor onehot_17_to_fp16_dtype_0 = const()[name = tensor("onehot_17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_1159_cast_fp16_to_fp16 = cast(dtype = onehot_17_to_fp16_dtype_0, x = var_1159_cast_fp16)[name = tensor("cast_112")]; + tensor onehot_19_cast_fp16 = reshape(shape = var_1169, x = var_1159_cast_fp16_to_fp16)[name = tensor("onehot_19_cast_fp16")]; + tensor var_1171_to_fp16 = const()[name = tensor("op_1171_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1173_cast_fp16 = sub(x = var_1171_to_fp16, y = onehot_19_cast_fp16)[name = tensor("op_1173_cast_fp16")]; + tensor k_cache4_to_fp16_dtype_0 = const()[name = tensor("k_cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache4_to_fp16 = cast(dtype = k_cache4_to_fp16_dtype_0, x = k_cache4)[name = tensor("cast_111")]; + tensor var_1174_cast_fp16 = mul(x = k_cache4_to_fp16, y = var_1173_cast_fp16)[name = tensor("op_1174_cast_fp16")]; + tensor var_1179 = const()[name = tensor("op_1179"), val = tensor([1, 1, 16, 64])]; + tensor var_1180_cast_fp16 = reshape(shape = var_1179, x = out_19_cast_fp16)[name = tensor("op_1180_cast_fp16")]; + tensor var_1181_cast_fp16 = mul(x = var_1180_cast_fp16, y = onehot_19_cast_fp16)[name = tensor("op_1181_cast_fp16")]; + tensor new_k_cache_9_cast_fp16 = add(x = var_1174_cast_fp16, y = var_1181_cast_fp16)[name = tensor("new_k_cache_9_cast_fp16")]; + tensor new_k_cache_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache4_to_fp16_dtype_0 = const()[name = tensor("v_cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache4_to_fp16 = cast(dtype = v_cache4_to_fp16_dtype_0, x = v_cache4)[name = tensor("cast_109")]; + tensor var_1187_cast_fp16 = mul(x = v_cache4_to_fp16, y = var_1173_cast_fp16)[name = tensor("op_1187_cast_fp16")]; + tensor var_1192 = const()[name = tensor("op_1192"), val = tensor([1, 1, 16, 64])]; + tensor var_1193_cast_fp16 = reshape(shape = var_1192, x = v_9_cast_fp16)[name = tensor("op_1193_cast_fp16")]; + tensor var_1194_cast_fp16 = mul(x = var_1193_cast_fp16, y = onehot_19_cast_fp16)[name = tensor("op_1194_cast_fp16")]; + tensor new_v_cache_9_cast_fp16 = add(x = var_1187_cast_fp16, y = var_1194_cast_fp16)[name = tensor("new_v_cache_9_cast_fp16")]; + tensor new_v_cache_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_4_perm_0 = const()[name = tensor("transpose_4_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_1207 = const()[name = tensor("op_1207"), val = tensor([0, 2, 1, 3])]; + tensor var_1213 = const()[name = tensor("op_1213"), val = tensor([1, 16, 1, 64])]; + tensor q4_9_cast_fp16 = reshape(shape = var_1213, x = out_17_cast_fp16)[name = tensor("q4_9_cast_fp16")]; + tensor var_1218_transpose_x_0 = const()[name = tensor("op_1218_transpose_x_0"), val = tensor(false)]; + tensor var_1218_transpose_y_0 = const()[name = tensor("op_1218_transpose_y_0"), val = tensor(false)]; + tensor transpose_4_cast_fp16 = transpose(perm = transpose_4_perm_0, x = new_k_cache_9_cast_fp16)[name = tensor("transpose_11")]; + tensor var_1218_cast_fp16 = matmul(transpose_x = var_1218_transpose_x_0, transpose_y = var_1218_transpose_y_0, x = q4_9_cast_fp16, y = transpose_4_cast_fp16)[name = tensor("op_1218_cast_fp16")]; + tensor var_1219_to_fp16 = const()[name = tensor("op_1219_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_17_cast_fp16 = mul(x = var_1218_cast_fp16, y = var_1219_to_fp16)[name = tensor("scores_17_cast_fp16")]; + tensor var_1223 = const()[name = tensor("op_1223"), val = tensor([1, 1])]; + tensor var_1224_cast_fp16 = reshape(shape = var_1223, x = position4_to_fp16)[name = tensor("op_1224_cast_fp16")]; + tensor var_1225_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_1224_cast_fp16)[name = tensor("op_1225_cast_fp16")]; + tensor mask_9_to_fp16_dtype_0 = const()[name = tensor("mask_9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_1232_to_fp16 = const()[name = tensor("op_1232_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1225_cast_fp16_to_fp16 = cast(dtype = mask_9_to_fp16_dtype_0, x = var_1225_cast_fp16)[name = tensor("cast_107")]; + tensor var_1233_cast_fp16 = sub(x = var_1225_cast_fp16_to_fp16, y = var_1232_to_fp16)[name = tensor("op_1233_cast_fp16")]; + tensor var_1238 = const()[name = tensor("op_1238"), val = tensor([1, 1, 1, 512])]; + tensor var_1239_cast_fp16 = reshape(shape = var_1238, x = var_1233_cast_fp16)[name = tensor("op_1239_cast_fp16")]; + tensor var_1240_to_fp16 = const()[name = tensor("op_1240_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_1241_cast_fp16 = mul(x = var_1239_cast_fp16, y = var_1240_to_fp16)[name = tensor("op_1241_cast_fp16")]; + tensor scores_19_cast_fp16 = add(x = scores_17_cast_fp16, y = var_1241_cast_fp16)[name = tensor("scores_19_cast_fp16")]; + tensor var_1244 = const()[name = tensor("op_1244"), val = tensor(-1)]; + tensor weights_9_cast_fp16 = softmax(axis = var_1244, x = scores_19_cast_fp16)[name = tensor("weights_9_cast_fp16")]; + tensor attn_9_transpose_x_0 = const()[name = tensor("attn_9_transpose_x_0"), val = tensor(false)]; + tensor attn_9_transpose_y_0 = const()[name = tensor("attn_9_transpose_y_0"), val = tensor(false)]; + tensor values_9_cast_fp16 = transpose(perm = var_1207, x = new_v_cache_9_cast_fp16)[name = tensor("transpose_10")]; + tensor attn_9_cast_fp16 = matmul(transpose_x = attn_9_transpose_x_0, transpose_y = attn_9_transpose_y_0, x = weights_9_cast_fp16, y = values_9_cast_fp16)[name = tensor("attn_9_cast_fp16")]; + tensor var_1252 = const()[name = tensor("op_1252"), val = tensor([0, 2, 1, 3])]; + tensor var_1257 = const()[name = tensor("op_1257"), val = tensor([1, 1, 1024])]; + tensor var_1253_cast_fp16 = transpose(perm = var_1252, x = attn_9_cast_fp16)[name = tensor("transpose_9")]; + tensor input_55_cast_fp16 = reshape(shape = var_1257, x = var_1253_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor attn4_out_proj_weight_to_fp16 = const()[name = tensor("attn4_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(107077312)))]; + tensor linear_18_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn4_out_proj_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor var_1263_to_fp16 = const()[name = tensor("op_1263_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1264_cast_fp16 = add(x = position4_to_fp16, y = var_1263_to_fp16)[name = tensor("op_1264_cast_fp16")]; + tensor var_1264_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1264_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_57_cast_fp16 = add(x = input_51_cast_fp16, y = linear_18_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor input_59_axes_0 = const()[name = tensor("input_59_axes_0"), val = tensor([-1])]; + tensor norm4_2_weight_to_fp16 = const()[name = tensor("norm4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109174528)))]; + tensor norm4_2_bias_to_fp16 = const()[name = tensor("norm4_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109176640)))]; + tensor var_1268_to_fp16 = const()[name = tensor("op_1268_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_59_cast_fp16 = layer_norm(axes = input_59_axes_0, beta = norm4_2_bias_to_fp16, epsilon = var_1268_to_fp16, gamma = norm4_2_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor linear4_1_weight_to_fp16 = const()[name = tensor("linear4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109178752)))]; + tensor linear_19_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear4_1_weight_to_fp16, x = input_59_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_61_mode_0 = const()[name = tensor("input_61_mode_0"), val = tensor("EXACT")]; + tensor input_61_cast_fp16 = gelu(mode = input_61_mode_0, x = linear_19_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor linear4_2_weight_to_fp16 = const()[name = tensor("linear4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(117567424)))]; + tensor linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear4_2_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor input_63_cast_fp16 = add(x = input_57_cast_fp16, y = linear_20_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor input_65_axes_0 = const()[name = tensor("input_65_axes_0"), val = tensor([-1])]; + tensor norm5_1_weight_to_fp16 = const()[name = tensor("norm5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125956096)))]; + tensor norm5_1_bias_to_fp16 = const()[name = tensor("norm5_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125958208)))]; + tensor var_1285_to_fp16 = const()[name = tensor("op_1285_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_65_cast_fp16 = layer_norm(axes = input_65_axes_0, beta = norm5_1_bias_to_fp16, epsilon = var_1285_to_fp16, gamma = norm5_1_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("input_65_cast_fp16")]; + tensor attn5_in_proj_weight_to_fp16 = const()[name = tensor("attn5_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125960320)))]; + tensor linear_21_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn5_in_proj_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("linear_21_cast_fp16")]; + tensor var_1298 = const()[name = tensor("op_1298"), val = tensor([1, 3, 16, 64])]; + tensor qkv_cast_fp16 = reshape(shape = var_1298, x = linear_21_cast_fp16)[name = tensor("qkv_cast_fp16")]; + tensor x_21_begin_0 = const()[name = tensor("x_21_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_21_end_0 = const()[name = tensor("x_21_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_21_end_mask_0 = const()[name = tensor("x_21_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_21_squeeze_mask_0 = const()[name = tensor("x_21_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_21_cast_fp16 = slice_by_index(begin = x_21_begin_0, end = x_21_end_0, end_mask = x_21_end_mask_0, squeeze_mask = x_21_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("x_21_cast_fp16")]; + tensor x_begin_0 = const()[name = tensor("x_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_end_0 = const()[name = tensor("x_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_end_mask_0 = const()[name = tensor("x_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_squeeze_mask_0 = const()[name = tensor("x_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_cast_fp16 = slice_by_index(begin = x_begin_0, end = x_end_0, end_mask = x_end_mask_0, squeeze_mask = x_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("x_cast_fp16")]; + tensor v_begin_0 = const()[name = tensor("v_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_end_0 = const()[name = tensor("v_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_end_mask_0 = const()[name = tensor("v_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_squeeze_mask_0 = const()[name = tensor("v_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_cast_fp16 = slice_by_index(begin = v_begin_0, end = v_end_0, end_mask = v_end_mask_0, squeeze_mask = v_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("v_cast_fp16")]; + tensor var_1327 = const()[name = tensor("op_1327"), val = tensor([1, 1, 1])]; + tensor position5_to_fp16_dtype_0 = const()[name = tensor("position5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position5_to_fp16 = cast(dtype = position5_to_fp16_dtype_0, x = position5)[name = tensor("cast_105")]; + tensor var_1328_cast_fp16 = reshape(shape = var_1327, x = position5_to_fp16)[name = tensor("op_1328_cast_fp16")]; + tensor angles_21_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_1328_cast_fp16)[name = tensor("angles_21_cast_fp16")]; + tensor rotr_21_cast_fp16 = cos(x = angles_21_cast_fp16)[name = tensor("rotr_21_cast_fp16")]; + tensor roti_21_cast_fp16 = sin(x = angles_21_cast_fp16)[name = tensor("roti_21_cast_fp16")]; + tensor var_1336 = const()[name = tensor("op_1336"), val = tensor([1, 16, 32, 2])]; + tensor pairs_21_cast_fp16 = reshape(shape = var_1336, x = x_21_cast_fp16)[name = tensor("pairs_21_cast_fp16")]; + tensor xr_21_begin_0 = const()[name = tensor("xr_21_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_21_end_0 = const()[name = tensor("xr_21_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_21_end_mask_0 = const()[name = tensor("xr_21_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_21_squeeze_mask_0 = const()[name = tensor("xr_21_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_21_cast_fp16 = slice_by_index(begin = xr_21_begin_0, end = xr_21_end_0, end_mask = xr_21_end_mask_0, squeeze_mask = xr_21_squeeze_mask_0, x = pairs_21_cast_fp16)[name = tensor("xr_21_cast_fp16")]; + tensor xi_21_begin_0 = const()[name = tensor("xi_21_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_21_end_0 = const()[name = tensor("xi_21_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_21_end_mask_0 = const()[name = tensor("xi_21_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_21_squeeze_mask_0 = const()[name = tensor("xi_21_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_21_cast_fp16 = slice_by_index(begin = xi_21_begin_0, end = xi_21_end_0, end_mask = xi_21_end_mask_0, squeeze_mask = xi_21_squeeze_mask_0, x = pairs_21_cast_fp16)[name = tensor("xi_21_cast_fp16")]; + tensor var_1344_cast_fp16 = mul(x = xr_21_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_1344_cast_fp16")]; + tensor var_1345_cast_fp16 = mul(x = xi_21_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_1345_cast_fp16")]; + tensor xor__21_cast_fp16 = sub(x = var_1344_cast_fp16, y = var_1345_cast_fp16)[name = tensor("xor__21_cast_fp16")]; + tensor var_1348_cast_fp16 = mul(x = xr_21_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_1348_cast_fp16")]; + tensor var_1349_cast_fp16 = mul(x = xi_21_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_1349_cast_fp16")]; + tensor xoi_21_cast_fp16 = add(x = var_1348_cast_fp16, y = var_1349_cast_fp16)[name = tensor("xoi_21_cast_fp16")]; + tensor out_21_axis_0 = const()[name = tensor("out_21_axis_0"), val = tensor(-1)]; + tensor out_21_cast_fp16 = stack(axis = out_21_axis_0, values = (xor__21_cast_fp16, xoi_21_cast_fp16))[name = tensor("out_21_cast_fp16")]; + tensor var_1372 = const()[name = tensor("op_1372"), val = tensor([1, 16, 32, 2])]; + tensor pairs_cast_fp16 = reshape(shape = var_1372, x = x_cast_fp16)[name = tensor("pairs_cast_fp16")]; + tensor xr_begin_0 = const()[name = tensor("xr_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_end_0 = const()[name = tensor("xr_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_end_mask_0 = const()[name = tensor("xr_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_squeeze_mask_0 = const()[name = tensor("xr_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_cast_fp16 = slice_by_index(begin = xr_begin_0, end = xr_end_0, end_mask = xr_end_mask_0, squeeze_mask = xr_squeeze_mask_0, x = pairs_cast_fp16)[name = tensor("xr_cast_fp16")]; + tensor xi_begin_0 = const()[name = tensor("xi_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_end_0 = const()[name = tensor("xi_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_end_mask_0 = const()[name = tensor("xi_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_squeeze_mask_0 = const()[name = tensor("xi_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_cast_fp16 = slice_by_index(begin = xi_begin_0, end = xi_end_0, end_mask = xi_end_mask_0, squeeze_mask = xi_squeeze_mask_0, x = pairs_cast_fp16)[name = tensor("xi_cast_fp16")]; + tensor var_1380_cast_fp16 = mul(x = xr_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_1380_cast_fp16")]; + tensor var_1381_cast_fp16 = mul(x = xi_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_1381_cast_fp16")]; + tensor xor__cast_fp16 = sub(x = var_1380_cast_fp16, y = var_1381_cast_fp16)[name = tensor("xor__cast_fp16")]; + tensor var_1384_cast_fp16 = mul(x = xr_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_1384_cast_fp16")]; + tensor var_1385_cast_fp16 = mul(x = xi_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_1385_cast_fp16")]; + tensor xoi_cast_fp16 = add(x = var_1384_cast_fp16, y = var_1385_cast_fp16)[name = tensor("xoi_cast_fp16")]; + tensor out_axis_0 = const()[name = tensor("out_axis_0"), val = tensor(-1)]; + tensor out_cast_fp16 = stack(axis = out_axis_0, values = (xor__cast_fp16, xoi_cast_fp16))[name = tensor("out_cast_fp16")]; + tensor _inversed_1397_y_0_to_fp16 = const()[name = tensor("_inversed_1397_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1397_cast_fp16 = mul(x = position5_to_fp16, y = _inversed_1397_y_0_to_fp16)[name = tensor("_inversed_1397_cast_fp16")]; + tensor var_1398_cast_fp16 = floor(x = _inversed_1397_cast_fp16)[name = tensor("op_1398_cast_fp16")]; + tensor var_1399_to_fp16 = const()[name = tensor("op_1399_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1400_cast_fp16 = mul(x = var_1398_cast_fp16, y = var_1399_to_fp16)[name = tensor("op_1400_cast_fp16")]; + tensor write_idx_cast_fp16 = sub(x = position5_to_fp16, y = var_1400_cast_fp16)[name = tensor("write_idx_cast_fp16")]; + tensor var_1403_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_cast_fp16)[name = tensor("op_1403_cast_fp16")]; + tensor var_1413 = const()[name = tensor("op_1413"), val = tensor([1, 512, 1, 1])]; + tensor onehot_21_to_fp16_dtype_0 = const()[name = tensor("onehot_21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_1403_cast_fp16_to_fp16 = cast(dtype = onehot_21_to_fp16_dtype_0, x = var_1403_cast_fp16)[name = tensor("cast_104")]; + tensor onehot_cast_fp16 = reshape(shape = var_1413, x = var_1403_cast_fp16_to_fp16)[name = tensor("onehot_cast_fp16")]; + tensor var_1415_to_fp16 = const()[name = tensor("op_1415_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1417_cast_fp16 = sub(x = var_1415_to_fp16, y = onehot_cast_fp16)[name = tensor("op_1417_cast_fp16")]; + tensor k_cache5_to_fp16_dtype_0 = const()[name = tensor("k_cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache5_to_fp16 = cast(dtype = k_cache5_to_fp16_dtype_0, x = k_cache5)[name = tensor("cast_103")]; + tensor var_1418_cast_fp16 = mul(x = k_cache5_to_fp16, y = var_1417_cast_fp16)[name = tensor("op_1418_cast_fp16")]; + tensor var_1423 = const()[name = tensor("op_1423"), val = tensor([1, 1, 16, 64])]; + tensor var_1424_cast_fp16 = reshape(shape = var_1423, x = out_cast_fp16)[name = tensor("op_1424_cast_fp16")]; + tensor var_1425_cast_fp16 = mul(x = var_1424_cast_fp16, y = onehot_cast_fp16)[name = tensor("op_1425_cast_fp16")]; + tensor new_k_cache_cast_fp16 = add(x = var_1418_cast_fp16, y = var_1425_cast_fp16)[name = tensor("new_k_cache_cast_fp16")]; + tensor new_k_cache_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache5_to_fp16_dtype_0 = const()[name = tensor("v_cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache5_to_fp16 = cast(dtype = v_cache5_to_fp16_dtype_0, x = v_cache5)[name = tensor("cast_101")]; + tensor var_1431_cast_fp16 = mul(x = v_cache5_to_fp16, y = var_1417_cast_fp16)[name = tensor("op_1431_cast_fp16")]; + tensor var_1436 = const()[name = tensor("op_1436"), val = tensor([1, 1, 16, 64])]; + tensor var_1437_cast_fp16 = reshape(shape = var_1436, x = v_cast_fp16)[name = tensor("op_1437_cast_fp16")]; + tensor var_1438_cast_fp16 = mul(x = var_1437_cast_fp16, y = onehot_cast_fp16)[name = tensor("op_1438_cast_fp16")]; + tensor new_v_cache_cast_fp16 = add(x = var_1431_cast_fp16, y = var_1438_cast_fp16)[name = tensor("new_v_cache_cast_fp16")]; + tensor new_v_cache_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_5_perm_0 = const()[name = tensor("transpose_5_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_1451 = const()[name = tensor("op_1451"), val = tensor([0, 2, 1, 3])]; + tensor var_1457 = const()[name = tensor("op_1457"), val = tensor([1, 16, 1, 64])]; + tensor q4_cast_fp16 = reshape(shape = var_1457, x = out_21_cast_fp16)[name = tensor("q4_cast_fp16")]; + tensor var_1462_transpose_x_0 = const()[name = tensor("op_1462_transpose_x_0"), val = tensor(false)]; + tensor var_1462_transpose_y_0 = const()[name = tensor("op_1462_transpose_y_0"), val = tensor(false)]; + tensor transpose_5_cast_fp16 = transpose(perm = transpose_5_perm_0, x = new_k_cache_cast_fp16)[name = tensor("transpose_8")]; + tensor var_1462_cast_fp16 = matmul(transpose_x = var_1462_transpose_x_0, transpose_y = var_1462_transpose_y_0, x = q4_cast_fp16, y = transpose_5_cast_fp16)[name = tensor("op_1462_cast_fp16")]; + tensor var_1463_to_fp16 = const()[name = tensor("op_1463_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_21_cast_fp16 = mul(x = var_1462_cast_fp16, y = var_1463_to_fp16)[name = tensor("scores_21_cast_fp16")]; + tensor var_1467 = const()[name = tensor("op_1467"), val = tensor([1, 1])]; + tensor var_1468_cast_fp16 = reshape(shape = var_1467, x = position5_to_fp16)[name = tensor("op_1468_cast_fp16")]; + tensor var_1469_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_1468_cast_fp16)[name = tensor("op_1469_cast_fp16")]; + tensor mask_to_fp16_dtype_0 = const()[name = tensor("mask_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_1476_to_fp16 = const()[name = tensor("op_1476_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1469_cast_fp16_to_fp16 = cast(dtype = mask_to_fp16_dtype_0, x = var_1469_cast_fp16)[name = tensor("cast_99")]; + tensor var_1477_cast_fp16 = sub(x = var_1469_cast_fp16_to_fp16, y = var_1476_to_fp16)[name = tensor("op_1477_cast_fp16")]; + tensor var_1482 = const()[name = tensor("op_1482"), val = tensor([1, 1, 1, 512])]; + tensor var_1483_cast_fp16 = reshape(shape = var_1482, x = var_1477_cast_fp16)[name = tensor("op_1483_cast_fp16")]; + tensor var_1484_to_fp16 = const()[name = tensor("op_1484_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_1485_cast_fp16 = mul(x = var_1483_cast_fp16, y = var_1484_to_fp16)[name = tensor("op_1485_cast_fp16")]; + tensor scores_cast_fp16 = add(x = scores_21_cast_fp16, y = var_1485_cast_fp16)[name = tensor("scores_cast_fp16")]; + tensor var_1488 = const()[name = tensor("op_1488"), val = tensor(-1)]; + tensor weights_cast_fp16 = softmax(axis = var_1488, x = scores_cast_fp16)[name = tensor("weights_cast_fp16")]; + tensor attn_transpose_x_0 = const()[name = tensor("attn_transpose_x_0"), val = tensor(false)]; + tensor attn_transpose_y_0 = const()[name = tensor("attn_transpose_y_0"), val = tensor(false)]; + tensor values_cast_fp16 = transpose(perm = var_1451, x = new_v_cache_cast_fp16)[name = tensor("transpose_7")]; + tensor attn_cast_fp16 = matmul(transpose_x = attn_transpose_x_0, transpose_y = attn_transpose_y_0, x = weights_cast_fp16, y = values_cast_fp16)[name = tensor("attn_cast_fp16")]; + tensor var_1496 = const()[name = tensor("op_1496"), val = tensor([0, 2, 1, 3])]; + tensor var_1501 = const()[name = tensor("op_1501"), val = tensor([1, 1, 1024])]; + tensor var_1497_cast_fp16 = transpose(perm = var_1496, x = attn_cast_fp16)[name = tensor("transpose_6")]; + tensor input_67_cast_fp16 = reshape(shape = var_1501, x = var_1497_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor attn5_out_proj_weight_to_fp16 = const()[name = tensor("attn5_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132251840)))]; + tensor linear_22_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn5_out_proj_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("linear_22_cast_fp16")]; + tensor var_1507_to_fp16 = const()[name = tensor("op_1507_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1508_cast_fp16 = add(x = position5_to_fp16, y = var_1507_to_fp16)[name = tensor("op_1508_cast_fp16")]; + tensor var_1508_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1508_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_69_cast_fp16 = add(x = input_63_cast_fp16, y = linear_22_cast_fp16)[name = tensor("input_69_cast_fp16")]; + tensor input_71_axes_0 = const()[name = tensor("input_71_axes_0"), val = tensor([-1])]; + tensor norm5_2_weight_to_fp16 = const()[name = tensor("norm5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134349056)))]; + tensor norm5_2_bias_to_fp16 = const()[name = tensor("norm5_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134351168)))]; + tensor var_1512_to_fp16 = const()[name = tensor("op_1512_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_71_cast_fp16 = layer_norm(axes = input_71_axes_0, beta = norm5_2_bias_to_fp16, epsilon = var_1512_to_fp16, gamma = norm5_2_weight_to_fp16, x = input_69_cast_fp16)[name = tensor("input_71_cast_fp16")]; + tensor linear5_1_weight_to_fp16 = const()[name = tensor("linear5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134353280)))]; + tensor linear_23_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear5_1_weight_to_fp16, x = input_71_cast_fp16)[name = tensor("linear_23_cast_fp16")]; + tensor input_73_mode_0 = const()[name = tensor("input_73_mode_0"), val = tensor("EXACT")]; + tensor input_73_cast_fp16 = gelu(mode = input_73_mode_0, x = linear_23_cast_fp16)[name = tensor("input_73_cast_fp16")]; + tensor linear5_2_weight_to_fp16 = const()[name = tensor("linear5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142741952)))]; + tensor linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear5_2_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("linear_24_cast_fp16")]; + tensor input_75_cast_fp16 = add(x = input_69_cast_fp16, y = linear_24_cast_fp16)[name = tensor("input_75_cast_fp16")]; + tensor input_axes_0 = const()[name = tensor("input_axes_0"), val = tensor([-1])]; + tensor out_norm_weight_to_fp16 = const()[name = tensor("out_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151130624)))]; + tensor out_norm_bias_to_fp16 = const()[name = tensor("out_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151132736)))]; + tensor var_1529_to_fp16 = const()[name = tensor("op_1529_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_cast_fp16 = layer_norm(axes = input_axes_0, beta = out_norm_bias_to_fp16, epsilon = var_1529_to_fp16, gamma = out_norm_weight_to_fp16, x = input_75_cast_fp16)[name = tensor("input_cast_fp16")]; + tensor input_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("input_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor out_eos_weight_to_fp16 = const()[name = tensor("out_eos_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151134848)))]; + tensor out_eos_bias_to_fp16 = const()[name = tensor("out_eos_bias_to_fp16"), val = tensor([-0x1.3ep-3])]; + tensor linear_25_cast_fp16 = linear(bias = out_eos_bias_to_fp16, weight = out_eos_weight_to_fp16, x = input_cast_fp16)[name = tensor("linear_25_cast_fp16")]; + tensor linear_25_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("linear_25_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor is_eos = cast(dtype = linear_25_cast_fp16_to_fp32_dtype_0, x = linear_25_cast_fp16)[name = tensor("cast_96")]; + tensor transformer_out = cast(dtype = input_cast_fp16_to_fp32_dtype_0, x = input_cast_fp16)[name = tensor("cast_97")]; + tensor new_position5 = cast(dtype = var_1508_cast_fp16_to_fp32_dtype_0, x = var_1508_cast_fp16)[name = tensor("cast_98")]; + tensor new_v_cache5 = cast(dtype = new_v_cache_cast_fp16_to_fp32_dtype_0, x = new_v_cache_cast_fp16)[name = tensor("cast_100")]; + tensor new_k_cache5 = cast(dtype = new_k_cache_cast_fp16_to_fp32_dtype_0, x = new_k_cache_cast_fp16)[name = tensor("cast_102")]; + tensor new_position4 = cast(dtype = var_1264_cast_fp16_to_fp32_dtype_0, x = var_1264_cast_fp16)[name = tensor("cast_106")]; + tensor new_v_cache4 = cast(dtype = new_v_cache_9_cast_fp16_to_fp32_dtype_0, x = new_v_cache_9_cast_fp16)[name = tensor("cast_108")]; + tensor new_k_cache4 = cast(dtype = new_k_cache_9_cast_fp16_to_fp32_dtype_0, x = new_k_cache_9_cast_fp16)[name = tensor("cast_110")]; + tensor new_position3 = cast(dtype = var_1020_cast_fp16_to_fp32_dtype_0, x = var_1020_cast_fp16)[name = tensor("cast_114")]; + tensor new_v_cache3 = cast(dtype = new_v_cache_7_cast_fp16_to_fp32_dtype_0, x = new_v_cache_7_cast_fp16)[name = tensor("cast_116")]; + tensor new_k_cache3 = cast(dtype = new_k_cache_7_cast_fp16_to_fp32_dtype_0, x = new_k_cache_7_cast_fp16)[name = tensor("cast_118")]; + tensor new_position2 = cast(dtype = var_776_cast_fp16_to_fp32_dtype_0, x = var_776_cast_fp16)[name = tensor("cast_122")]; + tensor new_v_cache2 = cast(dtype = new_v_cache_5_cast_fp16_to_fp32_dtype_0, x = new_v_cache_5_cast_fp16)[name = tensor("cast_124")]; + tensor new_k_cache2 = cast(dtype = new_k_cache_5_cast_fp16_to_fp32_dtype_0, x = new_k_cache_5_cast_fp16)[name = tensor("cast_126")]; + tensor new_position1 = cast(dtype = var_532_cast_fp16_to_fp32_dtype_0, x = var_532_cast_fp16)[name = tensor("cast_130")]; + tensor new_v_cache1 = cast(dtype = new_v_cache_3_cast_fp16_to_fp32_dtype_0, x = new_v_cache_3_cast_fp16)[name = tensor("cast_132")]; + tensor new_k_cache1 = cast(dtype = new_k_cache_3_cast_fp16_to_fp32_dtype_0, x = new_k_cache_3_cast_fp16)[name = tensor("cast_134")]; + tensor new_position0 = cast(dtype = var_288_cast_fp16_to_fp32_dtype_0, x = var_288_cast_fp16)[name = tensor("cast_138")]; + tensor new_v_cache0 = cast(dtype = new_v_cache_1_cast_fp16_to_fp32_dtype_0, x = new_v_cache_1_cast_fp16)[name = tensor("cast_140")]; + tensor new_k_cache0 = cast(dtype = new_k_cache_1_cast_fp16_to_fp32_dtype_0, x = new_k_cache_1_cast_fp16)[name = tensor("cast_142")]; + } -> (transformer_out, is_eos, new_k_cache0, new_v_cache0, new_position0, new_k_cache1, new_v_cache1, new_position1, new_k_cache2, new_v_cache2, new_position2, new_k_cache3, new_v_cache3, new_position3, new_k_cache4, new_v_cache4, new_position4, new_k_cache5, new_v_cache5, new_position5); +} \ No newline at end of file diff --git a/v2.1/german/flowlm_step_ane.mlmodelc/weights/weight.bin b/v2.1/german/flowlm_step_ane.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..f083a95b73bd93036b964dc2dfbb25e950077369 --- /dev/null +++ b/v2.1/german/flowlm_step_ane.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9fb8fe67addcbe2f53775b5c98af5ea5c4fb75d18c6ca98f541222486ca970fc +size 151136960 diff --git a/v2.1/german/flowlm_step_ane.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/german/flowlm_step_ane.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..53423eda7de3288401d661bc0d2b130cbc1cf898 --- /dev/null +++ b/v2.1/german/flowlm_step_ane.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:79d7377491809bf467162356e72fe239499c194e7fe65b501c61552349742bf3 +size 137731 diff --git a/v2.1/german/flowlm_step_ane.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/german/flowlm_step_ane.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..f083a95b73bd93036b964dc2dfbb25e950077369 --- /dev/null +++ b/v2.1/german/flowlm_step_ane.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9fb8fe67addcbe2f53775b5c98af5ea5c4fb75d18c6ca98f541222486ca970fc +size 151136960 diff --git a/v2.1/german/flowlm_step_ane.mlpackage/Manifest.json b/v2.1/german/flowlm_step_ane.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..9593a0793895f6bdfe00abfffa12a244f45f53c7 --- /dev/null +++ b/v2.1/german/flowlm_step_ane.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "17B0B8F4-6486-4EA0-9903-517336F3163B": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "411DA685-305D-4C08-B68F-830A11580ACD": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "411DA685-305D-4C08-B68F-830A11580ACD" +} diff --git a/v2.1/german_24l/cond_prefill_ane.mlmodelc/analytics/coremldata.bin b/v2.1/german_24l/cond_prefill_ane.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..6dddd287ab4f46c99dd81f864a121eb8da858fee --- /dev/null +++ b/v2.1/german_24l/cond_prefill_ane.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:66777cc7d1dddabbd552bb65472beccd080601e92928b24a5167033ed0af4a62 +size 243 diff --git a/v2.1/german_24l/cond_prefill_ane.mlmodelc/coremldata.bin b/v2.1/german_24l/cond_prefill_ane.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..8361121093fcaa1c999a1507ebdadb587f65ed49 --- /dev/null +++ b/v2.1/german_24l/cond_prefill_ane.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3fdd745c516cc3d146416eee9aff93bf981070807ab6fa3510d742f4eb6b5e78 +size 4496 diff --git a/v2.1/german_24l/cond_prefill_ane.mlmodelc/model.mil b/v2.1/german_24l/cond_prefill_ane.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..f102b6e147b28d002e028bb659c273f9e9c7f4c9 --- /dev/null +++ b/v2.1/german_24l/cond_prefill_ane.mlmodelc/model.mil @@ -0,0 +1,3198 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.9.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor conditioning, tensor k_cache0, tensor k_cache1, tensor k_cache10, tensor k_cache11, tensor k_cache12, tensor k_cache13, tensor k_cache14, tensor k_cache15, tensor k_cache16, tensor k_cache17, tensor k_cache18, tensor k_cache19, tensor k_cache2, tensor k_cache20, tensor k_cache21, tensor k_cache22, tensor k_cache23, tensor k_cache3, tensor k_cache4, tensor k_cache5, tensor k_cache6, tensor k_cache7, tensor k_cache8, tensor k_cache9, tensor position0, tensor position1, tensor position10, tensor position11, tensor position12, tensor position13, tensor position14, tensor position15, tensor position16, tensor position17, tensor position18, tensor position19, tensor position2, tensor position20, tensor position21, tensor position22, tensor position23, tensor position3, tensor position4, tensor position5, tensor position6, tensor position7, tensor position8, tensor position9, tensor v_cache0, tensor v_cache1, tensor v_cache10, tensor v_cache11, tensor v_cache12, tensor v_cache13, tensor v_cache14, tensor v_cache15, tensor v_cache16, tensor v_cache17, tensor v_cache18, tensor v_cache19, tensor v_cache2, tensor v_cache20, tensor v_cache21, tensor v_cache22, tensor v_cache23, tensor v_cache3, tensor v_cache4, tensor v_cache5, tensor v_cache6, tensor v_cache7, tensor v_cache8, tensor v_cache9, tensor valid_len) { + tensor token_indices_to_fp16 = const()[name = tensor("token_indices_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor position0_to_fp16_dtype_0 = const()[name = tensor("position0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position0_to_fp16 = cast(dtype = position0_to_fp16_dtype_0, x = position0)[name = tensor("cast_470")]; + tensor ts_cast_fp16 = add(x = token_indices_to_fp16, y = position0_to_fp16)[name = tensor("ts_cast_fp16")]; + tensor var_231 = const()[name = tensor("op_231"), val = tensor([256, 1, 1])]; + tensor var_232_cast_fp16 = reshape(shape = var_231, x = ts_cast_fp16)[name = tensor("op_232_cast_fp16")]; + tensor var_227_to_fp16 = const()[name = tensor("op_227_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(640)))]; + tensor angles_cast_fp16 = mul(x = var_227_to_fp16, y = var_232_cast_fp16)[name = tensor("angles_cast_fp16")]; + tensor rotr_cast_fp16 = cos(x = angles_cast_fp16)[name = tensor("rotr_cast_fp16")]; + tensor roti_cast_fp16 = sin(x = angles_cast_fp16)[name = tensor("roti_cast_fp16")]; + tensor _inversed_237_y_0_to_fp16 = const()[name = tensor("_inversed_237_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_237_cast_fp16 = mul(x = ts_cast_fp16, y = _inversed_237_y_0_to_fp16)[name = tensor("_inversed_237_cast_fp16")]; + tensor var_238_cast_fp16 = floor(x = _inversed_237_cast_fp16)[name = tensor("op_238_cast_fp16")]; + tensor var_239_to_fp16 = const()[name = tensor("op_239_to_fp16"), val = tensor(0x1p+9)]; + tensor var_240_cast_fp16 = mul(x = var_238_cast_fp16, y = var_239_to_fp16)[name = tensor("op_240_cast_fp16")]; + tensor write_idx_1_cast_fp16 = sub(x = ts_cast_fp16, y = var_240_cast_fp16)[name = tensor("write_idx_1_cast_fp16")]; + tensor valid_len_to_fp16_dtype_0 = const()[name = tensor("valid_len_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor valid_len_to_fp16 = cast(dtype = valid_len_to_fp16_dtype_0, x = valid_len)[name = tensor("cast_469")]; + tensor var_246_cast_fp16 = less(x = token_indices_to_fp16, y = valid_len_to_fp16)[name = tensor("op_246_cast_fp16")]; + tensor token_valid_to_fp16_dtype_0 = const()[name = tensor("token_valid_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_246_cast_fp16_to_fp16 = cast(dtype = token_valid_to_fp16_dtype_0, x = var_246_cast_fp16)[name = tensor("cast_468")]; + tensor var_252_cast_fp16 = mul(x = write_idx_1_cast_fp16, y = var_246_cast_fp16_to_fp16)[name = tensor("op_252_cast_fp16")]; + tensor var_253_to_fp16 = const()[name = tensor("op_253_to_fp16"), val = tensor(0x1p+0)]; + tensor var_255_cast_fp16 = sub(x = var_253_to_fp16, y = var_246_cast_fp16_to_fp16)[name = tensor("op_255_cast_fp16")]; + tensor var_256_to_fp16 = const()[name = tensor("op_256_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_257_cast_fp16 = mul(x = var_255_cast_fp16, y = var_256_to_fp16)[name = tensor("op_257_cast_fp16")]; + tensor write_idx_cast_fp16 = add(x = var_252_cast_fp16, y = var_257_cast_fp16)[name = tensor("write_idx_cast_fp16")]; + tensor var_262 = const()[name = tensor("op_262"), val = tensor([256, 1])]; + tensor var_263_cast_fp16 = reshape(shape = var_262, x = write_idx_cast_fp16)[name = tensor("op_263_cast_fp16")]; + tensor var_267_to_fp16 = const()[name = tensor("op_267_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(768)))]; + tensor var_268_cast_fp16 = equal(x = var_263_cast_fp16, y = var_267_to_fp16)[name = tensor("op_268_cast_fp16")]; + tensor var_278_axes_0 = const()[name = tensor("op_278_axes_0"), val = tensor([0])]; + tensor var_278_keep_dims_0 = const()[name = tensor("op_278_keep_dims_0"), val = tensor(false)]; + tensor assign_to_fp16_dtype_0 = const()[name = tensor("assign_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_268_cast_fp16_to_fp16 = cast(dtype = assign_to_fp16_dtype_0, x = var_268_cast_fp16)[name = tensor("cast_467")]; + tensor var_278_cast_fp16 = reduce_sum(axes = var_278_axes_0, keep_dims = var_278_keep_dims_0, x = var_268_cast_fp16_to_fp16)[name = tensor("op_278_cast_fp16")]; + tensor const_0_to_fp16 = const()[name = tensor("const_0_to_fp16"), val = tensor(-inf)]; + tensor var_280_to_fp16 = const()[name = tensor("op_280_to_fp16"), val = tensor(0x1p+0)]; + tensor clip_0_cast_fp16 = clip(alpha = const_0_to_fp16, beta = var_280_to_fp16, x = var_278_cast_fp16)[name = tensor("clip_0_cast_fp16")]; + tensor var_284 = const()[name = tensor("op_284"), val = tensor([1, 1])]; + tensor var_285_cast_fp16 = reshape(shape = var_284, x = position0_to_fp16)[name = tensor("op_285_cast_fp16")]; + tensor var_289_to_fp16 = const()[name = tensor("op_289_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1856)))]; + tensor q_positions_cast_fp16 = add(x = var_285_cast_fp16, y = var_289_to_fp16)[name = tensor("q_positions_cast_fp16")]; + tensor var_297_cast_fp16 = add(x = position0_to_fp16, y = valid_len_to_fp16)[name = tensor("op_297_cast_fp16")]; + tensor var_300 = const()[name = tensor("op_300"), val = tensor([1, 1])]; + tensor valid_end_cast_fp16 = reshape(shape = var_300, x = var_297_cast_fp16)[name = tensor("valid_end_cast_fp16")]; + tensor var_302_cast_fp16 = less(x = var_267_to_fp16, y = valid_end_cast_fp16)[name = tensor("op_302_cast_fp16")]; + tensor var_303_cast_fp16 = less_equal(x = var_267_to_fp16, y = q_positions_cast_fp16)[name = tensor("op_303_cast_fp16")]; + tensor var_304 = logical_and(x = var_302_cast_fp16, y = var_303_cast_fp16)[name = tensor("op_304")]; + tensor mask_to_fp16_dtype_0 = const()[name = tensor("mask_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_311_to_fp16 = const()[name = tensor("op_311_to_fp16"), val = tensor(0x1p+0)]; + tensor var_304_to_fp16 = cast(dtype = mask_to_fp16_dtype_0, x = var_304)[name = tensor("cast_466")]; + tensor var_312_cast_fp16 = sub(x = var_304_to_fp16, y = var_311_to_fp16)[name = tensor("op_312_cast_fp16")]; + tensor var_313_to_fp16 = const()[name = tensor("op_313_to_fp16"), val = tensor(0x1.388p+13)]; + tensor mask_bias_cast_fp16 = mul(x = var_312_cast_fp16, y = var_313_to_fp16)[name = tensor("mask_bias_cast_fp16")]; + tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; + tensor conditioning_to_fp16_dtype_0 = const()[name = tensor("conditioning_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor norm0_1_weight_to_fp16 = const()[name = tensor("norm0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2432)))]; + tensor norm0_1_bias_to_fp16 = const()[name = tensor("norm0_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4544)))]; + tensor var_316_to_fp16 = const()[name = tensor("op_316_to_fp16"), val = tensor(0x1.5p-17)]; + tensor conditioning_to_fp16 = cast(dtype = conditioning_to_fp16_dtype_0, x = conditioning)[name = tensor("cast_465")]; + tensor input_3_cast_fp16 = layer_norm(axes = input_3_axes_0, beta = norm0_1_bias_to_fp16, epsilon = var_316_to_fp16, gamma = norm0_1_weight_to_fp16, x = conditioning_to_fp16)[name = tensor("input_3_cast_fp16")]; + tensor attn0_in_proj_weight_to_fp16 = const()[name = tensor("attn0_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6656)))]; + tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6298176)))]; + tensor linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn0_in_proj_weight_to_fp16, x = input_3_cast_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor var_330 = const()[name = tensor("op_330"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_1_cast_fp16 = reshape(shape = var_330, x = linear_0_cast_fp16)[name = tensor("qkv_1_cast_fp16")]; + tensor x_1_begin_0 = const()[name = tensor("x_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_1_end_0 = const()[name = tensor("x_1_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_1_end_mask_0 = const()[name = tensor("x_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_1_squeeze_mask_0 = const()[name = tensor("x_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_1_cast_fp16 = slice_by_index(begin = x_1_begin_0, end = x_1_end_0, end_mask = x_1_end_mask_0, squeeze_mask = x_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("x_1_cast_fp16")]; + tensor x_3_begin_0 = const()[name = tensor("x_3_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_3_end_0 = const()[name = tensor("x_3_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_3_end_mask_0 = const()[name = tensor("x_3_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_3_squeeze_mask_0 = const()[name = tensor("x_3_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_3_cast_fp16 = slice_by_index(begin = x_3_begin_0, end = x_3_end_0, end_mask = x_3_end_mask_0, squeeze_mask = x_3_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("x_3_cast_fp16")]; + tensor v_1_begin_0 = const()[name = tensor("v_1_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_1_end_0 = const()[name = tensor("v_1_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_1_end_mask_0 = const()[name = tensor("v_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_1_squeeze_mask_0 = const()[name = tensor("v_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_1_cast_fp16 = slice_by_index(begin = v_1_begin_0, end = v_1_end_0, end_mask = v_1_end_mask_0, squeeze_mask = v_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("v_1_cast_fp16")]; + tensor var_375 = const()[name = tensor("op_375"), val = tensor([256, 16, 32, 2])]; + tensor pairs_1_cast_fp16 = reshape(shape = var_375, x = x_1_cast_fp16)[name = tensor("pairs_1_cast_fp16")]; + tensor xr_1_begin_0 = const()[name = tensor("xr_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_1_end_0 = const()[name = tensor("xr_1_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_1_end_mask_0 = const()[name = tensor("xr_1_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_1_squeeze_mask_0 = const()[name = tensor("xr_1_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_1_cast_fp16 = slice_by_index(begin = xr_1_begin_0, end = xr_1_end_0, end_mask = xr_1_end_mask_0, squeeze_mask = xr_1_squeeze_mask_0, x = pairs_1_cast_fp16)[name = tensor("xr_1_cast_fp16")]; + tensor xi_1_begin_0 = const()[name = tensor("xi_1_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_1_end_0 = const()[name = tensor("xi_1_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_1_end_mask_0 = const()[name = tensor("xi_1_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_1_squeeze_mask_0 = const()[name = tensor("xi_1_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_1_cast_fp16 = slice_by_index(begin = xi_1_begin_0, end = xi_1_end_0, end_mask = xi_1_end_mask_0, squeeze_mask = xi_1_squeeze_mask_0, x = pairs_1_cast_fp16)[name = tensor("xi_1_cast_fp16")]; + tensor var_383_cast_fp16 = mul(x = xr_1_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_383_cast_fp16")]; + tensor var_384_cast_fp16 = mul(x = xi_1_cast_fp16, y = roti_cast_fp16)[name = tensor("op_384_cast_fp16")]; + tensor xor__1_cast_fp16 = sub(x = var_383_cast_fp16, y = var_384_cast_fp16)[name = tensor("xor__1_cast_fp16")]; + tensor var_387_cast_fp16 = mul(x = xr_1_cast_fp16, y = roti_cast_fp16)[name = tensor("op_387_cast_fp16")]; + tensor var_388_cast_fp16 = mul(x = xi_1_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_388_cast_fp16")]; + tensor xoi_1_cast_fp16 = add(x = var_387_cast_fp16, y = var_388_cast_fp16)[name = tensor("xoi_1_cast_fp16")]; + tensor out_1_axis_0 = const()[name = tensor("out_1_axis_0"), val = tensor(-1)]; + tensor out_1_cast_fp16 = stack(axis = out_1_axis_0, values = (xor__1_cast_fp16, xoi_1_cast_fp16))[name = tensor("out_1_cast_fp16")]; + tensor var_398 = const()[name = tensor("op_398"), val = tensor([1, 256, 16, 64])]; + tensor q_1_cast_fp16 = reshape(shape = var_398, x = out_1_cast_fp16)[name = tensor("q_1_cast_fp16")]; + tensor var_404 = const()[name = tensor("op_404"), val = tensor([256, 16, 32, 2])]; + tensor pairs_3_cast_fp16 = reshape(shape = var_404, x = x_3_cast_fp16)[name = tensor("pairs_3_cast_fp16")]; + tensor xr_3_begin_0 = const()[name = tensor("xr_3_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_3_end_0 = const()[name = tensor("xr_3_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_3_end_mask_0 = const()[name = tensor("xr_3_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_3_squeeze_mask_0 = const()[name = tensor("xr_3_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_3_cast_fp16 = slice_by_index(begin = xr_3_begin_0, end = xr_3_end_0, end_mask = xr_3_end_mask_0, squeeze_mask = xr_3_squeeze_mask_0, x = pairs_3_cast_fp16)[name = tensor("xr_3_cast_fp16")]; + tensor xi_3_begin_0 = const()[name = tensor("xi_3_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_3_end_0 = const()[name = tensor("xi_3_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_3_end_mask_0 = const()[name = tensor("xi_3_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_3_squeeze_mask_0 = const()[name = tensor("xi_3_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_3_cast_fp16 = slice_by_index(begin = xi_3_begin_0, end = xi_3_end_0, end_mask = xi_3_end_mask_0, squeeze_mask = xi_3_squeeze_mask_0, x = pairs_3_cast_fp16)[name = tensor("xi_3_cast_fp16")]; + tensor var_412_cast_fp16 = mul(x = xr_3_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_412_cast_fp16")]; + tensor var_413_cast_fp16 = mul(x = xi_3_cast_fp16, y = roti_cast_fp16)[name = tensor("op_413_cast_fp16")]; + tensor xor__3_cast_fp16 = sub(x = var_412_cast_fp16, y = var_413_cast_fp16)[name = tensor("xor__3_cast_fp16")]; + tensor var_416_cast_fp16 = mul(x = xr_3_cast_fp16, y = roti_cast_fp16)[name = tensor("op_416_cast_fp16")]; + tensor var_417_cast_fp16 = mul(x = xi_3_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_417_cast_fp16")]; + tensor xoi_3_cast_fp16 = add(x = var_416_cast_fp16, y = var_417_cast_fp16)[name = tensor("xoi_3_cast_fp16")]; + tensor out_3_axis_0 = const()[name = tensor("out_3_axis_0"), val = tensor(-1)]; + tensor out_3_cast_fp16 = stack(axis = out_3_axis_0, values = (xor__3_cast_fp16, xoi_3_cast_fp16))[name = tensor("out_3_cast_fp16")]; + tensor var_434 = const()[name = tensor("op_434"), val = tensor([256, 1024])]; + tensor var_435_cast_fp16 = reshape(shape = var_434, x = out_3_cast_fp16)[name = tensor("op_435_cast_fp16")]; + tensor written_k_1_transpose_x_1 = const()[name = tensor("written_k_1_transpose_x_1"), val = tensor(true)]; + tensor written_k_1_transpose_y_1 = const()[name = tensor("written_k_1_transpose_y_1"), val = tensor(false)]; + tensor written_k_1_cast_fp16 = matmul(transpose_x = written_k_1_transpose_x_1, transpose_y = written_k_1_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_435_cast_fp16)[name = tensor("written_k_1_cast_fp16")]; + tensor var_442 = const()[name = tensor("op_442"), val = tensor([256, 1024])]; + tensor var_443_cast_fp16 = reshape(shape = var_442, x = v_1_cast_fp16)[name = tensor("op_443_cast_fp16")]; + tensor written_v_1_transpose_x_1 = const()[name = tensor("written_v_1_transpose_x_1"), val = tensor(true)]; + tensor written_v_1_transpose_y_1 = const()[name = tensor("written_v_1_transpose_y_1"), val = tensor(false)]; + tensor written_v_1_cast_fp16 = matmul(transpose_x = written_v_1_transpose_x_1, transpose_y = written_v_1_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_443_cast_fp16)[name = tensor("written_v_1_cast_fp16")]; + tensor var_445_to_fp16 = const()[name = tensor("op_445_to_fp16"), val = tensor(0x1p+0)]; + tensor var_447_cast_fp16 = sub(x = var_445_to_fp16, y = clip_0_cast_fp16)[name = tensor("op_447_cast_fp16")]; + tensor var_452 = const()[name = tensor("op_452"), val = tensor([1, 512, 1, 1])]; + tensor keep_1_cast_fp16 = reshape(shape = var_452, x = var_447_cast_fp16)[name = tensor("keep_1_cast_fp16")]; + tensor k_cache0_to_fp16_dtype_0 = const()[name = tensor("k_cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache0_to_fp16 = cast(dtype = k_cache0_to_fp16_dtype_0, x = k_cache0)[name = tensor("cast_464")]; + tensor var_454_cast_fp16 = mul(x = k_cache0_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_454_cast_fp16")]; + tensor var_459 = const()[name = tensor("op_459"), val = tensor([1, 512, 16, 64])]; + tensor var_460_cast_fp16 = reshape(shape = var_459, x = written_k_1_cast_fp16)[name = tensor("op_460_cast_fp16")]; + tensor new_k_cache_1_cast_fp16 = add(x = var_454_cast_fp16, y = var_460_cast_fp16)[name = tensor("new_k_cache_1_cast_fp16")]; + tensor new_k_cache_1_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_1_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache0_to_fp16_dtype_0 = const()[name = tensor("v_cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache0_to_fp16 = cast(dtype = v_cache0_to_fp16_dtype_0, x = v_cache0)[name = tensor("cast_462")]; + tensor var_463_cast_fp16 = mul(x = v_cache0_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_463_cast_fp16")]; + tensor var_468 = const()[name = tensor("op_468"), val = tensor([1, 512, 16, 64])]; + tensor var_469_cast_fp16 = reshape(shape = var_468, x = written_v_1_cast_fp16)[name = tensor("op_469_cast_fp16")]; + tensor new_v_cache_1_cast_fp16 = add(x = var_463_cast_fp16, y = var_469_cast_fp16)[name = tensor("new_v_cache_1_cast_fp16")]; + tensor new_v_cache_1_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_1_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_482 = const()[name = tensor("op_482"), val = tensor([0, 2, 1, 3])]; + tensor var_493_transpose_x_0 = const()[name = tensor("op_493_transpose_x_0"), val = tensor(false)]; + tensor var_493_transpose_y_0 = const()[name = tensor("op_493_transpose_y_0"), val = tensor(false)]; + tensor transpose_69_perm_0 = const()[name = tensor("transpose_69_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_70_perm_0 = const()[name = tensor("transpose_70_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_70 = transpose(perm = transpose_70_perm_0, x = new_k_cache_1_cast_fp16)[name = tensor("transpose_204")]; + tensor transpose_69 = transpose(perm = transpose_69_perm_0, x = q_1_cast_fp16)[name = tensor("transpose_205")]; + tensor var_493_cast_fp16 = matmul(transpose_x = var_493_transpose_x_0, transpose_y = var_493_transpose_y_0, x = transpose_69, y = transpose_70)[name = tensor("op_493_cast_fp16")]; + tensor var_494_to_fp16 = const()[name = tensor("op_494_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_1_cast_fp16 = mul(x = var_493_cast_fp16, y = var_494_to_fp16)[name = tensor("scores_1_cast_fp16")]; + tensor var_500 = const()[name = tensor("op_500"), val = tensor([1, 1, 256, 512])]; + tensor var_501_cast_fp16 = reshape(shape = var_500, x = mask_bias_cast_fp16)[name = tensor("op_501_cast_fp16")]; + tensor scores_3_cast_fp16 = add(x = scores_1_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_3_cast_fp16")]; + tensor var_504 = const()[name = tensor("op_504"), val = tensor(-1)]; + tensor weights_1_cast_fp16 = softmax(axis = var_504, x = scores_3_cast_fp16)[name = tensor("weights_1_cast_fp16")]; + tensor attn_1_transpose_x_0 = const()[name = tensor("attn_1_transpose_x_0"), val = tensor(false)]; + tensor attn_1_transpose_y_0 = const()[name = tensor("attn_1_transpose_y_0"), val = tensor(false)]; + tensor values_1_cast_fp16 = transpose(perm = var_482, x = new_v_cache_1_cast_fp16)[name = tensor("transpose_206")]; + tensor attn_1_cast_fp16 = matmul(transpose_x = attn_1_transpose_x_0, transpose_y = attn_1_transpose_y_0, x = weights_1_cast_fp16, y = values_1_cast_fp16)[name = tensor("attn_1_cast_fp16")]; + tensor var_512 = const()[name = tensor("op_512"), val = tensor([0, 2, 1, 3])]; + tensor var_517 = const()[name = tensor("op_517"), val = tensor([1, 256, 1024])]; + tensor var_513_cast_fp16 = transpose(perm = var_512, x = attn_1_cast_fp16)[name = tensor("transpose_203")]; + tensor input_5_cast_fp16 = reshape(shape = var_517, x = var_513_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor attn0_out_proj_weight_to_fp16 = const()[name = tensor("attn0_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6304384)))]; + tensor linear_1_bias_0_to_fp16 = const()[name = tensor("linear_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401600)))]; + tensor linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn0_out_proj_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("linear_1_cast_fp16")]; + tensor var_526_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_526_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_7_cast_fp16 = add(x = conditioning_to_fp16, y = linear_1_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor input_9_axes_0 = const()[name = tensor("input_9_axes_0"), val = tensor([-1])]; + tensor norm0_2_weight_to_fp16 = const()[name = tensor("norm0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8403712)))]; + tensor norm0_2_bias_to_fp16 = const()[name = tensor("norm0_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8405824)))]; + tensor var_530_to_fp16 = const()[name = tensor("op_530_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_9_cast_fp16 = layer_norm(axes = input_9_axes_0, beta = norm0_2_bias_to_fp16, epsilon = var_530_to_fp16, gamma = norm0_2_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor linear0_1_weight_to_fp16 = const()[name = tensor("linear0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8407936)))]; + tensor linear_2_bias_0_to_fp16 = const()[name = tensor("linear_2_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16796608)))]; + tensor linear_2_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear0_1_weight_to_fp16, x = input_9_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor input_11_mode_0 = const()[name = tensor("input_11_mode_0"), val = tensor("EXACT")]; + tensor input_11_cast_fp16 = gelu(mode = input_11_mode_0, x = linear_2_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor linear0_2_weight_to_fp16 = const()[name = tensor("linear0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16804864)))]; + tensor linear_3_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear0_2_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("linear_3_cast_fp16")]; + tensor input_13_cast_fp16 = add(x = input_7_cast_fp16, y = linear_3_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor input_15_axes_0 = const()[name = tensor("input_15_axes_0"), val = tensor([-1])]; + tensor norm1_1_weight_to_fp16 = const()[name = tensor("norm1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25193536)))]; + tensor norm1_1_bias_to_fp16 = const()[name = tensor("norm1_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25195648)))]; + tensor var_547_to_fp16 = const()[name = tensor("op_547_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_15_cast_fp16 = layer_norm(axes = input_15_axes_0, beta = norm1_1_bias_to_fp16, epsilon = var_547_to_fp16, gamma = norm1_1_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor attn1_in_proj_weight_to_fp16 = const()[name = tensor("attn1_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25197760)))]; + tensor linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn1_in_proj_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor var_561 = const()[name = tensor("op_561"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_3_cast_fp16 = reshape(shape = var_561, x = linear_4_cast_fp16)[name = tensor("qkv_3_cast_fp16")]; + tensor x_5_begin_0 = const()[name = tensor("x_5_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_5_end_0 = const()[name = tensor("x_5_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_5_end_mask_0 = const()[name = tensor("x_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_5_squeeze_mask_0 = const()[name = tensor("x_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_5_cast_fp16 = slice_by_index(begin = x_5_begin_0, end = x_5_end_0, end_mask = x_5_end_mask_0, squeeze_mask = x_5_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("x_5_cast_fp16")]; + tensor x_7_begin_0 = const()[name = tensor("x_7_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_7_end_0 = const()[name = tensor("x_7_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_7_end_mask_0 = const()[name = tensor("x_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_7_squeeze_mask_0 = const()[name = tensor("x_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_7_cast_fp16 = slice_by_index(begin = x_7_begin_0, end = x_7_end_0, end_mask = x_7_end_mask_0, squeeze_mask = x_7_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("x_7_cast_fp16")]; + tensor v_3_begin_0 = const()[name = tensor("v_3_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_3_end_0 = const()[name = tensor("v_3_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_3_end_mask_0 = const()[name = tensor("v_3_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_3_squeeze_mask_0 = const()[name = tensor("v_3_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_3_cast_fp16 = slice_by_index(begin = v_3_begin_0, end = v_3_end_0, end_mask = v_3_end_mask_0, squeeze_mask = v_3_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("v_3_cast_fp16")]; + tensor var_606 = const()[name = tensor("op_606"), val = tensor([256, 16, 32, 2])]; + tensor pairs_5_cast_fp16 = reshape(shape = var_606, x = x_5_cast_fp16)[name = tensor("pairs_5_cast_fp16")]; + tensor xr_5_begin_0 = const()[name = tensor("xr_5_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_5_end_0 = const()[name = tensor("xr_5_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_5_end_mask_0 = const()[name = tensor("xr_5_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_5_squeeze_mask_0 = const()[name = tensor("xr_5_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_5_cast_fp16 = slice_by_index(begin = xr_5_begin_0, end = xr_5_end_0, end_mask = xr_5_end_mask_0, squeeze_mask = xr_5_squeeze_mask_0, x = pairs_5_cast_fp16)[name = tensor("xr_5_cast_fp16")]; + tensor xi_5_begin_0 = const()[name = tensor("xi_5_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_5_end_0 = const()[name = tensor("xi_5_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_5_end_mask_0 = const()[name = tensor("xi_5_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_5_squeeze_mask_0 = const()[name = tensor("xi_5_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_5_cast_fp16 = slice_by_index(begin = xi_5_begin_0, end = xi_5_end_0, end_mask = xi_5_end_mask_0, squeeze_mask = xi_5_squeeze_mask_0, x = pairs_5_cast_fp16)[name = tensor("xi_5_cast_fp16")]; + tensor var_614_cast_fp16 = mul(x = xr_5_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_614_cast_fp16")]; + tensor var_615_cast_fp16 = mul(x = xi_5_cast_fp16, y = roti_cast_fp16)[name = tensor("op_615_cast_fp16")]; + tensor xor__5_cast_fp16 = sub(x = var_614_cast_fp16, y = var_615_cast_fp16)[name = tensor("xor__5_cast_fp16")]; + tensor var_618_cast_fp16 = mul(x = xr_5_cast_fp16, y = roti_cast_fp16)[name = tensor("op_618_cast_fp16")]; + tensor var_619_cast_fp16 = mul(x = xi_5_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_619_cast_fp16")]; + tensor xoi_5_cast_fp16 = add(x = var_618_cast_fp16, y = var_619_cast_fp16)[name = tensor("xoi_5_cast_fp16")]; + tensor out_5_axis_0 = const()[name = tensor("out_5_axis_0"), val = tensor(-1)]; + tensor out_5_cast_fp16 = stack(axis = out_5_axis_0, values = (xor__5_cast_fp16, xoi_5_cast_fp16))[name = tensor("out_5_cast_fp16")]; + tensor var_629 = const()[name = tensor("op_629"), val = tensor([1, 256, 16, 64])]; + tensor q_3_cast_fp16 = reshape(shape = var_629, x = out_5_cast_fp16)[name = tensor("q_3_cast_fp16")]; + tensor var_635 = const()[name = tensor("op_635"), val = tensor([256, 16, 32, 2])]; + tensor pairs_7_cast_fp16 = reshape(shape = var_635, x = x_7_cast_fp16)[name = tensor("pairs_7_cast_fp16")]; + tensor xr_7_begin_0 = const()[name = tensor("xr_7_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_7_end_0 = const()[name = tensor("xr_7_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_7_end_mask_0 = const()[name = tensor("xr_7_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_7_squeeze_mask_0 = const()[name = tensor("xr_7_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_7_cast_fp16 = slice_by_index(begin = xr_7_begin_0, end = xr_7_end_0, end_mask = xr_7_end_mask_0, squeeze_mask = xr_7_squeeze_mask_0, x = pairs_7_cast_fp16)[name = tensor("xr_7_cast_fp16")]; + tensor xi_7_begin_0 = const()[name = tensor("xi_7_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_7_end_0 = const()[name = tensor("xi_7_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_7_end_mask_0 = const()[name = tensor("xi_7_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_7_squeeze_mask_0 = const()[name = tensor("xi_7_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_7_cast_fp16 = slice_by_index(begin = xi_7_begin_0, end = xi_7_end_0, end_mask = xi_7_end_mask_0, squeeze_mask = xi_7_squeeze_mask_0, x = pairs_7_cast_fp16)[name = tensor("xi_7_cast_fp16")]; + tensor var_643_cast_fp16 = mul(x = xr_7_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_643_cast_fp16")]; + tensor var_644_cast_fp16 = mul(x = xi_7_cast_fp16, y = roti_cast_fp16)[name = tensor("op_644_cast_fp16")]; + tensor xor__7_cast_fp16 = sub(x = var_643_cast_fp16, y = var_644_cast_fp16)[name = tensor("xor__7_cast_fp16")]; + tensor var_647_cast_fp16 = mul(x = xr_7_cast_fp16, y = roti_cast_fp16)[name = tensor("op_647_cast_fp16")]; + tensor var_648_cast_fp16 = mul(x = xi_7_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_648_cast_fp16")]; + tensor xoi_7_cast_fp16 = add(x = var_647_cast_fp16, y = var_648_cast_fp16)[name = tensor("xoi_7_cast_fp16")]; + tensor out_7_axis_0 = const()[name = tensor("out_7_axis_0"), val = tensor(-1)]; + tensor out_7_cast_fp16 = stack(axis = out_7_axis_0, values = (xor__7_cast_fp16, xoi_7_cast_fp16))[name = tensor("out_7_cast_fp16")]; + tensor var_665 = const()[name = tensor("op_665"), val = tensor([256, 1024])]; + tensor var_666_cast_fp16 = reshape(shape = var_665, x = out_7_cast_fp16)[name = tensor("op_666_cast_fp16")]; + tensor written_k_3_transpose_x_1 = const()[name = tensor("written_k_3_transpose_x_1"), val = tensor(true)]; + tensor written_k_3_transpose_y_1 = const()[name = tensor("written_k_3_transpose_y_1"), val = tensor(false)]; + tensor written_k_3_cast_fp16 = matmul(transpose_x = written_k_3_transpose_x_1, transpose_y = written_k_3_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_666_cast_fp16)[name = tensor("written_k_3_cast_fp16")]; + tensor var_673 = const()[name = tensor("op_673"), val = tensor([256, 1024])]; + tensor var_674_cast_fp16 = reshape(shape = var_673, x = v_3_cast_fp16)[name = tensor("op_674_cast_fp16")]; + tensor written_v_3_transpose_x_1 = const()[name = tensor("written_v_3_transpose_x_1"), val = tensor(true)]; + tensor written_v_3_transpose_y_1 = const()[name = tensor("written_v_3_transpose_y_1"), val = tensor(false)]; + tensor written_v_3_cast_fp16 = matmul(transpose_x = written_v_3_transpose_x_1, transpose_y = written_v_3_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_674_cast_fp16)[name = tensor("written_v_3_cast_fp16")]; + tensor k_cache1_to_fp16_dtype_0 = const()[name = tensor("k_cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache1_to_fp16 = cast(dtype = k_cache1_to_fp16_dtype_0, x = k_cache1)[name = tensor("cast_459")]; + tensor var_685_cast_fp16 = mul(x = k_cache1_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_685_cast_fp16")]; + tensor var_690 = const()[name = tensor("op_690"), val = tensor([1, 512, 16, 64])]; + tensor var_691_cast_fp16 = reshape(shape = var_690, x = written_k_3_cast_fp16)[name = tensor("op_691_cast_fp16")]; + tensor new_k_cache_3_cast_fp16 = add(x = var_685_cast_fp16, y = var_691_cast_fp16)[name = tensor("new_k_cache_3_cast_fp16")]; + tensor new_k_cache_3_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_3_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache1_to_fp16_dtype_0 = const()[name = tensor("v_cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache1_to_fp16 = cast(dtype = v_cache1_to_fp16_dtype_0, x = v_cache1)[name = tensor("cast_457")]; + tensor var_694_cast_fp16 = mul(x = v_cache1_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_694_cast_fp16")]; + tensor var_699 = const()[name = tensor("op_699"), val = tensor([1, 512, 16, 64])]; + tensor var_700_cast_fp16 = reshape(shape = var_699, x = written_v_3_cast_fp16)[name = tensor("op_700_cast_fp16")]; + tensor new_v_cache_3_cast_fp16 = add(x = var_694_cast_fp16, y = var_700_cast_fp16)[name = tensor("new_v_cache_3_cast_fp16")]; + tensor new_v_cache_3_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_3_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_713 = const()[name = tensor("op_713"), val = tensor([0, 2, 1, 3])]; + tensor var_724_transpose_x_0 = const()[name = tensor("op_724_transpose_x_0"), val = tensor(false)]; + tensor var_724_transpose_y_0 = const()[name = tensor("op_724_transpose_y_0"), val = tensor(false)]; + tensor transpose_71_perm_0 = const()[name = tensor("transpose_71_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_72_perm_0 = const()[name = tensor("transpose_72_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_72 = transpose(perm = transpose_72_perm_0, x = new_k_cache_3_cast_fp16)[name = tensor("transpose_200")]; + tensor transpose_71 = transpose(perm = transpose_71_perm_0, x = q_3_cast_fp16)[name = tensor("transpose_201")]; + tensor var_724_cast_fp16 = matmul(transpose_x = var_724_transpose_x_0, transpose_y = var_724_transpose_y_0, x = transpose_71, y = transpose_72)[name = tensor("op_724_cast_fp16")]; + tensor var_725_to_fp16 = const()[name = tensor("op_725_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_5_cast_fp16 = mul(x = var_724_cast_fp16, y = var_725_to_fp16)[name = tensor("scores_5_cast_fp16")]; + tensor scores_7_cast_fp16 = add(x = scores_5_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_7_cast_fp16")]; + tensor var_735 = const()[name = tensor("op_735"), val = tensor(-1)]; + tensor weights_3_cast_fp16 = softmax(axis = var_735, x = scores_7_cast_fp16)[name = tensor("weights_3_cast_fp16")]; + tensor attn_3_transpose_x_0 = const()[name = tensor("attn_3_transpose_x_0"), val = tensor(false)]; + tensor attn_3_transpose_y_0 = const()[name = tensor("attn_3_transpose_y_0"), val = tensor(false)]; + tensor values_3_cast_fp16 = transpose(perm = var_713, x = new_v_cache_3_cast_fp16)[name = tensor("transpose_202")]; + tensor attn_3_cast_fp16 = matmul(transpose_x = attn_3_transpose_x_0, transpose_y = attn_3_transpose_y_0, x = weights_3_cast_fp16, y = values_3_cast_fp16)[name = tensor("attn_3_cast_fp16")]; + tensor var_743 = const()[name = tensor("op_743"), val = tensor([0, 2, 1, 3])]; + tensor var_748 = const()[name = tensor("op_748"), val = tensor([1, 256, 1024])]; + tensor var_744_cast_fp16 = transpose(perm = var_743, x = attn_3_cast_fp16)[name = tensor("transpose_199")]; + tensor input_17_cast_fp16 = reshape(shape = var_748, x = var_744_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor attn1_out_proj_weight_to_fp16 = const()[name = tensor("attn1_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31489280)))]; + tensor linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn1_out_proj_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor position1_to_fp16_dtype_0 = const()[name = tensor("position1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position1_to_fp16 = cast(dtype = position1_to_fp16_dtype_0, x = position1)[name = tensor("cast_455")]; + tensor var_757_cast_fp16 = add(x = position1_to_fp16, y = valid_len_to_fp16)[name = tensor("op_757_cast_fp16")]; + tensor var_757_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_757_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_19_cast_fp16 = add(x = input_13_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor input_21_axes_0 = const()[name = tensor("input_21_axes_0"), val = tensor([-1])]; + tensor norm1_2_weight_to_fp16 = const()[name = tensor("norm1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33586496)))]; + tensor norm1_2_bias_to_fp16 = const()[name = tensor("norm1_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33588608)))]; + tensor var_761_to_fp16 = const()[name = tensor("op_761_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_21_cast_fp16 = layer_norm(axes = input_21_axes_0, beta = norm1_2_bias_to_fp16, epsilon = var_761_to_fp16, gamma = norm1_2_weight_to_fp16, x = input_19_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor linear1_1_weight_to_fp16 = const()[name = tensor("linear1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33590720)))]; + tensor linear_6_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear1_1_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor input_23_mode_0 = const()[name = tensor("input_23_mode_0"), val = tensor("EXACT")]; + tensor input_23_cast_fp16 = gelu(mode = input_23_mode_0, x = linear_6_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor linear1_2_weight_to_fp16 = const()[name = tensor("linear1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41979392)))]; + tensor linear_7_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear1_2_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_25_cast_fp16 = add(x = input_19_cast_fp16, y = linear_7_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor input_27_axes_0 = const()[name = tensor("input_27_axes_0"), val = tensor([-1])]; + tensor norm2_1_weight_to_fp16 = const()[name = tensor("norm2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50368064)))]; + tensor norm2_1_bias_to_fp16 = const()[name = tensor("norm2_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50370176)))]; + tensor var_778_to_fp16 = const()[name = tensor("op_778_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = norm2_1_bias_to_fp16, epsilon = var_778_to_fp16, gamma = norm2_1_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor attn2_in_proj_weight_to_fp16 = const()[name = tensor("attn2_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50372288)))]; + tensor linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn2_in_proj_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor var_792 = const()[name = tensor("op_792"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_5_cast_fp16 = reshape(shape = var_792, x = linear_8_cast_fp16)[name = tensor("qkv_5_cast_fp16")]; + tensor x_9_begin_0 = const()[name = tensor("x_9_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_9_end_0 = const()[name = tensor("x_9_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_9_end_mask_0 = const()[name = tensor("x_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_9_squeeze_mask_0 = const()[name = tensor("x_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_9_cast_fp16 = slice_by_index(begin = x_9_begin_0, end = x_9_end_0, end_mask = x_9_end_mask_0, squeeze_mask = x_9_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("x_9_cast_fp16")]; + tensor x_11_begin_0 = const()[name = tensor("x_11_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_11_end_0 = const()[name = tensor("x_11_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_11_end_mask_0 = const()[name = tensor("x_11_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_11_squeeze_mask_0 = const()[name = tensor("x_11_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_11_cast_fp16 = slice_by_index(begin = x_11_begin_0, end = x_11_end_0, end_mask = x_11_end_mask_0, squeeze_mask = x_11_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("x_11_cast_fp16")]; + tensor v_5_begin_0 = const()[name = tensor("v_5_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_5_end_0 = const()[name = tensor("v_5_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_5_end_mask_0 = const()[name = tensor("v_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_5_squeeze_mask_0 = const()[name = tensor("v_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_5_cast_fp16 = slice_by_index(begin = v_5_begin_0, end = v_5_end_0, end_mask = v_5_end_mask_0, squeeze_mask = v_5_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("v_5_cast_fp16")]; + tensor var_837 = const()[name = tensor("op_837"), val = tensor([256, 16, 32, 2])]; + tensor pairs_9_cast_fp16 = reshape(shape = var_837, x = x_9_cast_fp16)[name = tensor("pairs_9_cast_fp16")]; + tensor xr_9_begin_0 = const()[name = tensor("xr_9_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_9_end_0 = const()[name = tensor("xr_9_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_9_end_mask_0 = const()[name = tensor("xr_9_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_9_squeeze_mask_0 = const()[name = tensor("xr_9_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_9_cast_fp16 = slice_by_index(begin = xr_9_begin_0, end = xr_9_end_0, end_mask = xr_9_end_mask_0, squeeze_mask = xr_9_squeeze_mask_0, x = pairs_9_cast_fp16)[name = tensor("xr_9_cast_fp16")]; + tensor xi_9_begin_0 = const()[name = tensor("xi_9_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_9_end_0 = const()[name = tensor("xi_9_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_9_end_mask_0 = const()[name = tensor("xi_9_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_9_squeeze_mask_0 = const()[name = tensor("xi_9_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_9_cast_fp16 = slice_by_index(begin = xi_9_begin_0, end = xi_9_end_0, end_mask = xi_9_end_mask_0, squeeze_mask = xi_9_squeeze_mask_0, x = pairs_9_cast_fp16)[name = tensor("xi_9_cast_fp16")]; + tensor var_845_cast_fp16 = mul(x = xr_9_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_845_cast_fp16")]; + tensor var_846_cast_fp16 = mul(x = xi_9_cast_fp16, y = roti_cast_fp16)[name = tensor("op_846_cast_fp16")]; + tensor xor__9_cast_fp16 = sub(x = var_845_cast_fp16, y = var_846_cast_fp16)[name = tensor("xor__9_cast_fp16")]; + tensor var_849_cast_fp16 = mul(x = xr_9_cast_fp16, y = roti_cast_fp16)[name = tensor("op_849_cast_fp16")]; + tensor var_850_cast_fp16 = mul(x = xi_9_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_850_cast_fp16")]; + tensor xoi_9_cast_fp16 = add(x = var_849_cast_fp16, y = var_850_cast_fp16)[name = tensor("xoi_9_cast_fp16")]; + tensor out_9_axis_0 = const()[name = tensor("out_9_axis_0"), val = tensor(-1)]; + tensor out_9_cast_fp16 = stack(axis = out_9_axis_0, values = (xor__9_cast_fp16, xoi_9_cast_fp16))[name = tensor("out_9_cast_fp16")]; + tensor var_860 = const()[name = tensor("op_860"), val = tensor([1, 256, 16, 64])]; + tensor q_5_cast_fp16 = reshape(shape = var_860, x = out_9_cast_fp16)[name = tensor("q_5_cast_fp16")]; + tensor var_866 = const()[name = tensor("op_866"), val = tensor([256, 16, 32, 2])]; + tensor pairs_11_cast_fp16 = reshape(shape = var_866, x = x_11_cast_fp16)[name = tensor("pairs_11_cast_fp16")]; + tensor xr_11_begin_0 = const()[name = tensor("xr_11_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_11_end_0 = const()[name = tensor("xr_11_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_11_end_mask_0 = const()[name = tensor("xr_11_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_11_squeeze_mask_0 = const()[name = tensor("xr_11_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_11_cast_fp16 = slice_by_index(begin = xr_11_begin_0, end = xr_11_end_0, end_mask = xr_11_end_mask_0, squeeze_mask = xr_11_squeeze_mask_0, x = pairs_11_cast_fp16)[name = tensor("xr_11_cast_fp16")]; + tensor xi_11_begin_0 = const()[name = tensor("xi_11_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_11_end_0 = const()[name = tensor("xi_11_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_11_end_mask_0 = const()[name = tensor("xi_11_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_11_squeeze_mask_0 = const()[name = tensor("xi_11_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_11_cast_fp16 = slice_by_index(begin = xi_11_begin_0, end = xi_11_end_0, end_mask = xi_11_end_mask_0, squeeze_mask = xi_11_squeeze_mask_0, x = pairs_11_cast_fp16)[name = tensor("xi_11_cast_fp16")]; + tensor var_874_cast_fp16 = mul(x = xr_11_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_874_cast_fp16")]; + tensor var_875_cast_fp16 = mul(x = xi_11_cast_fp16, y = roti_cast_fp16)[name = tensor("op_875_cast_fp16")]; + tensor xor__11_cast_fp16 = sub(x = var_874_cast_fp16, y = var_875_cast_fp16)[name = tensor("xor__11_cast_fp16")]; + tensor var_878_cast_fp16 = mul(x = xr_11_cast_fp16, y = roti_cast_fp16)[name = tensor("op_878_cast_fp16")]; + tensor var_879_cast_fp16 = mul(x = xi_11_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_879_cast_fp16")]; + tensor xoi_11_cast_fp16 = add(x = var_878_cast_fp16, y = var_879_cast_fp16)[name = tensor("xoi_11_cast_fp16")]; + tensor out_11_axis_0 = const()[name = tensor("out_11_axis_0"), val = tensor(-1)]; + tensor out_11_cast_fp16 = stack(axis = out_11_axis_0, values = (xor__11_cast_fp16, xoi_11_cast_fp16))[name = tensor("out_11_cast_fp16")]; + tensor var_896 = const()[name = tensor("op_896"), val = tensor([256, 1024])]; + tensor var_897_cast_fp16 = reshape(shape = var_896, x = out_11_cast_fp16)[name = tensor("op_897_cast_fp16")]; + tensor written_k_5_transpose_x_1 = const()[name = tensor("written_k_5_transpose_x_1"), val = tensor(true)]; + tensor written_k_5_transpose_y_1 = const()[name = tensor("written_k_5_transpose_y_1"), val = tensor(false)]; + tensor written_k_5_cast_fp16 = matmul(transpose_x = written_k_5_transpose_x_1, transpose_y = written_k_5_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_897_cast_fp16)[name = tensor("written_k_5_cast_fp16")]; + tensor var_904 = const()[name = tensor("op_904"), val = tensor([256, 1024])]; + tensor var_905_cast_fp16 = reshape(shape = var_904, x = v_5_cast_fp16)[name = tensor("op_905_cast_fp16")]; + tensor written_v_5_transpose_x_1 = const()[name = tensor("written_v_5_transpose_x_1"), val = tensor(true)]; + tensor written_v_5_transpose_y_1 = const()[name = tensor("written_v_5_transpose_y_1"), val = tensor(false)]; + tensor written_v_5_cast_fp16 = matmul(transpose_x = written_v_5_transpose_x_1, transpose_y = written_v_5_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_905_cast_fp16)[name = tensor("written_v_5_cast_fp16")]; + tensor k_cache2_to_fp16_dtype_0 = const()[name = tensor("k_cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache2_to_fp16 = cast(dtype = k_cache2_to_fp16_dtype_0, x = k_cache2)[name = tensor("cast_453")]; + tensor var_916_cast_fp16 = mul(x = k_cache2_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_916_cast_fp16")]; + tensor var_921 = const()[name = tensor("op_921"), val = tensor([1, 512, 16, 64])]; + tensor var_922_cast_fp16 = reshape(shape = var_921, x = written_k_5_cast_fp16)[name = tensor("op_922_cast_fp16")]; + tensor new_k_cache_5_cast_fp16 = add(x = var_916_cast_fp16, y = var_922_cast_fp16)[name = tensor("new_k_cache_5_cast_fp16")]; + tensor new_k_cache_5_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_5_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache2_to_fp16_dtype_0 = const()[name = tensor("v_cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache2_to_fp16 = cast(dtype = v_cache2_to_fp16_dtype_0, x = v_cache2)[name = tensor("cast_451")]; + tensor var_925_cast_fp16 = mul(x = v_cache2_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_925_cast_fp16")]; + tensor var_930 = const()[name = tensor("op_930"), val = tensor([1, 512, 16, 64])]; + tensor var_931_cast_fp16 = reshape(shape = var_930, x = written_v_5_cast_fp16)[name = tensor("op_931_cast_fp16")]; + tensor new_v_cache_5_cast_fp16 = add(x = var_925_cast_fp16, y = var_931_cast_fp16)[name = tensor("new_v_cache_5_cast_fp16")]; + tensor new_v_cache_5_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_5_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_944 = const()[name = tensor("op_944"), val = tensor([0, 2, 1, 3])]; + tensor var_955_transpose_x_0 = const()[name = tensor("op_955_transpose_x_0"), val = tensor(false)]; + tensor var_955_transpose_y_0 = const()[name = tensor("op_955_transpose_y_0"), val = tensor(false)]; + tensor transpose_73_perm_0 = const()[name = tensor("transpose_73_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_74_perm_0 = const()[name = tensor("transpose_74_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_74 = transpose(perm = transpose_74_perm_0, x = new_k_cache_5_cast_fp16)[name = tensor("transpose_196")]; + tensor transpose_73 = transpose(perm = transpose_73_perm_0, x = q_5_cast_fp16)[name = tensor("transpose_197")]; + tensor var_955_cast_fp16 = matmul(transpose_x = var_955_transpose_x_0, transpose_y = var_955_transpose_y_0, x = transpose_73, y = transpose_74)[name = tensor("op_955_cast_fp16")]; + tensor var_956_to_fp16 = const()[name = tensor("op_956_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_9_cast_fp16 = mul(x = var_955_cast_fp16, y = var_956_to_fp16)[name = tensor("scores_9_cast_fp16")]; + tensor scores_11_cast_fp16 = add(x = scores_9_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_11_cast_fp16")]; + tensor var_966 = const()[name = tensor("op_966"), val = tensor(-1)]; + tensor weights_5_cast_fp16 = softmax(axis = var_966, x = scores_11_cast_fp16)[name = tensor("weights_5_cast_fp16")]; + tensor attn_5_transpose_x_0 = const()[name = tensor("attn_5_transpose_x_0"), val = tensor(false)]; + tensor attn_5_transpose_y_0 = const()[name = tensor("attn_5_transpose_y_0"), val = tensor(false)]; + tensor values_5_cast_fp16 = transpose(perm = var_944, x = new_v_cache_5_cast_fp16)[name = tensor("transpose_198")]; + tensor attn_5_cast_fp16 = matmul(transpose_x = attn_5_transpose_x_0, transpose_y = attn_5_transpose_y_0, x = weights_5_cast_fp16, y = values_5_cast_fp16)[name = tensor("attn_5_cast_fp16")]; + tensor var_974 = const()[name = tensor("op_974"), val = tensor([0, 2, 1, 3])]; + tensor var_979 = const()[name = tensor("op_979"), val = tensor([1, 256, 1024])]; + tensor var_975_cast_fp16 = transpose(perm = var_974, x = attn_5_cast_fp16)[name = tensor("transpose_195")]; + tensor input_29_cast_fp16 = reshape(shape = var_979, x = var_975_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor attn2_out_proj_weight_to_fp16 = const()[name = tensor("attn2_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(56663808)))]; + tensor linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn2_out_proj_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor position2_to_fp16_dtype_0 = const()[name = tensor("position2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position2_to_fp16 = cast(dtype = position2_to_fp16_dtype_0, x = position2)[name = tensor("cast_449")]; + tensor var_988_cast_fp16 = add(x = position2_to_fp16, y = valid_len_to_fp16)[name = tensor("op_988_cast_fp16")]; + tensor var_988_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_988_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_31_cast_fp16 = add(x = input_25_cast_fp16, y = linear_9_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; + tensor norm2_2_weight_to_fp16 = const()[name = tensor("norm2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58761024)))]; + tensor norm2_2_bias_to_fp16 = const()[name = tensor("norm2_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58763136)))]; + tensor var_992_to_fp16 = const()[name = tensor("op_992_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_33_cast_fp16 = layer_norm(axes = input_33_axes_0, beta = norm2_2_bias_to_fp16, epsilon = var_992_to_fp16, gamma = norm2_2_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor linear2_1_weight_to_fp16 = const()[name = tensor("linear2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58765248)))]; + tensor linear_10_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear2_1_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor input_35_mode_0 = const()[name = tensor("input_35_mode_0"), val = tensor("EXACT")]; + tensor input_35_cast_fp16 = gelu(mode = input_35_mode_0, x = linear_10_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor linear2_2_weight_to_fp16 = const()[name = tensor("linear2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67153920)))]; + tensor linear_11_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear2_2_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor input_37_cast_fp16 = add(x = input_31_cast_fp16, y = linear_11_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor input_39_axes_0 = const()[name = tensor("input_39_axes_0"), val = tensor([-1])]; + tensor norm3_1_weight_to_fp16 = const()[name = tensor("norm3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75542592)))]; + tensor norm3_1_bias_to_fp16 = const()[name = tensor("norm3_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75544704)))]; + tensor var_1009_to_fp16 = const()[name = tensor("op_1009_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_39_cast_fp16 = layer_norm(axes = input_39_axes_0, beta = norm3_1_bias_to_fp16, epsilon = var_1009_to_fp16, gamma = norm3_1_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor attn3_in_proj_weight_to_fp16 = const()[name = tensor("attn3_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75546816)))]; + tensor linear_12_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn3_in_proj_weight_to_fp16, x = input_39_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor var_1023 = const()[name = tensor("op_1023"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_7_cast_fp16 = reshape(shape = var_1023, x = linear_12_cast_fp16)[name = tensor("qkv_7_cast_fp16")]; + tensor x_13_begin_0 = const()[name = tensor("x_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_13_end_0 = const()[name = tensor("x_13_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_13_end_mask_0 = const()[name = tensor("x_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_13_squeeze_mask_0 = const()[name = tensor("x_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_13_cast_fp16 = slice_by_index(begin = x_13_begin_0, end = x_13_end_0, end_mask = x_13_end_mask_0, squeeze_mask = x_13_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("x_13_cast_fp16")]; + tensor x_15_begin_0 = const()[name = tensor("x_15_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_15_end_0 = const()[name = tensor("x_15_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_15_end_mask_0 = const()[name = tensor("x_15_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_15_squeeze_mask_0 = const()[name = tensor("x_15_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_15_cast_fp16 = slice_by_index(begin = x_15_begin_0, end = x_15_end_0, end_mask = x_15_end_mask_0, squeeze_mask = x_15_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("x_15_cast_fp16")]; + tensor v_7_begin_0 = const()[name = tensor("v_7_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_7_end_0 = const()[name = tensor("v_7_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_7_end_mask_0 = const()[name = tensor("v_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_7_squeeze_mask_0 = const()[name = tensor("v_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_7_cast_fp16 = slice_by_index(begin = v_7_begin_0, end = v_7_end_0, end_mask = v_7_end_mask_0, squeeze_mask = v_7_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("v_7_cast_fp16")]; + tensor var_1068 = const()[name = tensor("op_1068"), val = tensor([256, 16, 32, 2])]; + tensor pairs_13_cast_fp16 = reshape(shape = var_1068, x = x_13_cast_fp16)[name = tensor("pairs_13_cast_fp16")]; + tensor xr_13_begin_0 = const()[name = tensor("xr_13_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_13_end_0 = const()[name = tensor("xr_13_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_13_end_mask_0 = const()[name = tensor("xr_13_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_13_squeeze_mask_0 = const()[name = tensor("xr_13_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_13_cast_fp16 = slice_by_index(begin = xr_13_begin_0, end = xr_13_end_0, end_mask = xr_13_end_mask_0, squeeze_mask = xr_13_squeeze_mask_0, x = pairs_13_cast_fp16)[name = tensor("xr_13_cast_fp16")]; + tensor xi_13_begin_0 = const()[name = tensor("xi_13_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_13_end_0 = const()[name = tensor("xi_13_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_13_end_mask_0 = const()[name = tensor("xi_13_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_13_squeeze_mask_0 = const()[name = tensor("xi_13_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_13_cast_fp16 = slice_by_index(begin = xi_13_begin_0, end = xi_13_end_0, end_mask = xi_13_end_mask_0, squeeze_mask = xi_13_squeeze_mask_0, x = pairs_13_cast_fp16)[name = tensor("xi_13_cast_fp16")]; + tensor var_1076_cast_fp16 = mul(x = xr_13_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1076_cast_fp16")]; + tensor var_1077_cast_fp16 = mul(x = xi_13_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1077_cast_fp16")]; + tensor xor__13_cast_fp16 = sub(x = var_1076_cast_fp16, y = var_1077_cast_fp16)[name = tensor("xor__13_cast_fp16")]; + tensor var_1080_cast_fp16 = mul(x = xr_13_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1080_cast_fp16")]; + tensor var_1081_cast_fp16 = mul(x = xi_13_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1081_cast_fp16")]; + tensor xoi_13_cast_fp16 = add(x = var_1080_cast_fp16, y = var_1081_cast_fp16)[name = tensor("xoi_13_cast_fp16")]; + tensor out_13_axis_0 = const()[name = tensor("out_13_axis_0"), val = tensor(-1)]; + tensor out_13_cast_fp16 = stack(axis = out_13_axis_0, values = (xor__13_cast_fp16, xoi_13_cast_fp16))[name = tensor("out_13_cast_fp16")]; + tensor var_1091 = const()[name = tensor("op_1091"), val = tensor([1, 256, 16, 64])]; + tensor q_7_cast_fp16 = reshape(shape = var_1091, x = out_13_cast_fp16)[name = tensor("q_7_cast_fp16")]; + tensor var_1097 = const()[name = tensor("op_1097"), val = tensor([256, 16, 32, 2])]; + tensor pairs_15_cast_fp16 = reshape(shape = var_1097, x = x_15_cast_fp16)[name = tensor("pairs_15_cast_fp16")]; + tensor xr_15_begin_0 = const()[name = tensor("xr_15_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_15_end_0 = const()[name = tensor("xr_15_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_15_end_mask_0 = const()[name = tensor("xr_15_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_15_squeeze_mask_0 = const()[name = tensor("xr_15_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_15_cast_fp16 = slice_by_index(begin = xr_15_begin_0, end = xr_15_end_0, end_mask = xr_15_end_mask_0, squeeze_mask = xr_15_squeeze_mask_0, x = pairs_15_cast_fp16)[name = tensor("xr_15_cast_fp16")]; + tensor xi_15_begin_0 = const()[name = tensor("xi_15_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_15_end_0 = const()[name = tensor("xi_15_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_15_end_mask_0 = const()[name = tensor("xi_15_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_15_squeeze_mask_0 = const()[name = tensor("xi_15_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_15_cast_fp16 = slice_by_index(begin = xi_15_begin_0, end = xi_15_end_0, end_mask = xi_15_end_mask_0, squeeze_mask = xi_15_squeeze_mask_0, x = pairs_15_cast_fp16)[name = tensor("xi_15_cast_fp16")]; + tensor var_1105_cast_fp16 = mul(x = xr_15_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1105_cast_fp16")]; + tensor var_1106_cast_fp16 = mul(x = xi_15_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1106_cast_fp16")]; + tensor xor__15_cast_fp16 = sub(x = var_1105_cast_fp16, y = var_1106_cast_fp16)[name = tensor("xor__15_cast_fp16")]; + tensor var_1109_cast_fp16 = mul(x = xr_15_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1109_cast_fp16")]; + tensor var_1110_cast_fp16 = mul(x = xi_15_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1110_cast_fp16")]; + tensor xoi_15_cast_fp16 = add(x = var_1109_cast_fp16, y = var_1110_cast_fp16)[name = tensor("xoi_15_cast_fp16")]; + tensor out_15_axis_0 = const()[name = tensor("out_15_axis_0"), val = tensor(-1)]; + tensor out_15_cast_fp16 = stack(axis = out_15_axis_0, values = (xor__15_cast_fp16, xoi_15_cast_fp16))[name = tensor("out_15_cast_fp16")]; + tensor var_1127 = const()[name = tensor("op_1127"), val = tensor([256, 1024])]; + tensor var_1128_cast_fp16 = reshape(shape = var_1127, x = out_15_cast_fp16)[name = tensor("op_1128_cast_fp16")]; + tensor written_k_7_transpose_x_1 = const()[name = tensor("written_k_7_transpose_x_1"), val = tensor(true)]; + tensor written_k_7_transpose_y_1 = const()[name = tensor("written_k_7_transpose_y_1"), val = tensor(false)]; + tensor written_k_7_cast_fp16 = matmul(transpose_x = written_k_7_transpose_x_1, transpose_y = written_k_7_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_1128_cast_fp16)[name = tensor("written_k_7_cast_fp16")]; + tensor var_1135 = const()[name = tensor("op_1135"), val = tensor([256, 1024])]; + tensor var_1136_cast_fp16 = reshape(shape = var_1135, x = v_7_cast_fp16)[name = tensor("op_1136_cast_fp16")]; + tensor written_v_7_transpose_x_1 = const()[name = tensor("written_v_7_transpose_x_1"), val = tensor(true)]; + tensor written_v_7_transpose_y_1 = const()[name = tensor("written_v_7_transpose_y_1"), val = tensor(false)]; + tensor written_v_7_cast_fp16 = matmul(transpose_x = written_v_7_transpose_x_1, transpose_y = written_v_7_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_1136_cast_fp16)[name = tensor("written_v_7_cast_fp16")]; + tensor k_cache3_to_fp16_dtype_0 = const()[name = tensor("k_cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache3_to_fp16 = cast(dtype = k_cache3_to_fp16_dtype_0, x = k_cache3)[name = tensor("cast_447")]; + tensor var_1147_cast_fp16 = mul(x = k_cache3_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_1147_cast_fp16")]; + tensor var_1152 = const()[name = tensor("op_1152"), val = tensor([1, 512, 16, 64])]; + tensor var_1153_cast_fp16 = reshape(shape = var_1152, x = written_k_7_cast_fp16)[name = tensor("op_1153_cast_fp16")]; + tensor new_k_cache_7_cast_fp16 = add(x = var_1147_cast_fp16, y = var_1153_cast_fp16)[name = tensor("new_k_cache_7_cast_fp16")]; + tensor new_k_cache_7_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_7_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache3_to_fp16_dtype_0 = const()[name = tensor("v_cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache3_to_fp16 = cast(dtype = v_cache3_to_fp16_dtype_0, x = v_cache3)[name = tensor("cast_445")]; + tensor var_1156_cast_fp16 = mul(x = v_cache3_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_1156_cast_fp16")]; + tensor var_1161 = const()[name = tensor("op_1161"), val = tensor([1, 512, 16, 64])]; + tensor var_1162_cast_fp16 = reshape(shape = var_1161, x = written_v_7_cast_fp16)[name = tensor("op_1162_cast_fp16")]; + tensor new_v_cache_7_cast_fp16 = add(x = var_1156_cast_fp16, y = var_1162_cast_fp16)[name = tensor("new_v_cache_7_cast_fp16")]; + tensor new_v_cache_7_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_7_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_1175 = const()[name = tensor("op_1175"), val = tensor([0, 2, 1, 3])]; + tensor var_1186_transpose_x_0 = const()[name = tensor("op_1186_transpose_x_0"), val = tensor(false)]; + tensor var_1186_transpose_y_0 = const()[name = tensor("op_1186_transpose_y_0"), val = tensor(false)]; + tensor transpose_75_perm_0 = const()[name = tensor("transpose_75_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_76_perm_0 = const()[name = tensor("transpose_76_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_76 = transpose(perm = transpose_76_perm_0, x = new_k_cache_7_cast_fp16)[name = tensor("transpose_192")]; + tensor transpose_75 = transpose(perm = transpose_75_perm_0, x = q_7_cast_fp16)[name = tensor("transpose_193")]; + tensor var_1186_cast_fp16 = matmul(transpose_x = var_1186_transpose_x_0, transpose_y = var_1186_transpose_y_0, x = transpose_75, y = transpose_76)[name = tensor("op_1186_cast_fp16")]; + tensor var_1187_to_fp16 = const()[name = tensor("op_1187_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_13_cast_fp16 = mul(x = var_1186_cast_fp16, y = var_1187_to_fp16)[name = tensor("scores_13_cast_fp16")]; + tensor scores_15_cast_fp16 = add(x = scores_13_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_15_cast_fp16")]; + tensor var_1197 = const()[name = tensor("op_1197"), val = tensor(-1)]; + tensor weights_7_cast_fp16 = softmax(axis = var_1197, x = scores_15_cast_fp16)[name = tensor("weights_7_cast_fp16")]; + tensor attn_7_transpose_x_0 = const()[name = tensor("attn_7_transpose_x_0"), val = tensor(false)]; + tensor attn_7_transpose_y_0 = const()[name = tensor("attn_7_transpose_y_0"), val = tensor(false)]; + tensor values_7_cast_fp16 = transpose(perm = var_1175, x = new_v_cache_7_cast_fp16)[name = tensor("transpose_194")]; + tensor attn_7_cast_fp16 = matmul(transpose_x = attn_7_transpose_x_0, transpose_y = attn_7_transpose_y_0, x = weights_7_cast_fp16, y = values_7_cast_fp16)[name = tensor("attn_7_cast_fp16")]; + tensor var_1205 = const()[name = tensor("op_1205"), val = tensor([0, 2, 1, 3])]; + tensor var_1210 = const()[name = tensor("op_1210"), val = tensor([1, 256, 1024])]; + tensor var_1206_cast_fp16 = transpose(perm = var_1205, x = attn_7_cast_fp16)[name = tensor("transpose_191")]; + tensor input_41_cast_fp16 = reshape(shape = var_1210, x = var_1206_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor attn3_out_proj_weight_to_fp16 = const()[name = tensor("attn3_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(81838336)))]; + tensor linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn3_out_proj_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor position3_to_fp16_dtype_0 = const()[name = tensor("position3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position3_to_fp16 = cast(dtype = position3_to_fp16_dtype_0, x = position3)[name = tensor("cast_443")]; + tensor var_1219_cast_fp16 = add(x = position3_to_fp16, y = valid_len_to_fp16)[name = tensor("op_1219_cast_fp16")]; + tensor var_1219_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1219_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_43_cast_fp16 = add(x = input_37_cast_fp16, y = linear_13_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; + tensor norm3_2_weight_to_fp16 = const()[name = tensor("norm3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(83935552)))]; + tensor norm3_2_bias_to_fp16 = const()[name = tensor("norm3_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(83937664)))]; + tensor var_1223_to_fp16 = const()[name = tensor("op_1223_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_45_cast_fp16 = layer_norm(axes = input_45_axes_0, beta = norm3_2_bias_to_fp16, epsilon = var_1223_to_fp16, gamma = norm3_2_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor linear3_1_weight_to_fp16 = const()[name = tensor("linear3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(83939776)))]; + tensor linear_14_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear3_1_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor input_47_mode_0 = const()[name = tensor("input_47_mode_0"), val = tensor("EXACT")]; + tensor input_47_cast_fp16 = gelu(mode = input_47_mode_0, x = linear_14_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor linear3_2_weight_to_fp16 = const()[name = tensor("linear3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(92328448)))]; + tensor linear_15_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear3_2_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor input_49_cast_fp16 = add(x = input_43_cast_fp16, y = linear_15_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor input_51_axes_0 = const()[name = tensor("input_51_axes_0"), val = tensor([-1])]; + tensor norm4_1_weight_to_fp16 = const()[name = tensor("norm4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100717120)))]; + tensor norm4_1_bias_to_fp16 = const()[name = tensor("norm4_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100719232)))]; + tensor var_1240_to_fp16 = const()[name = tensor("op_1240_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_51_cast_fp16 = layer_norm(axes = input_51_axes_0, beta = norm4_1_bias_to_fp16, epsilon = var_1240_to_fp16, gamma = norm4_1_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor attn4_in_proj_weight_to_fp16 = const()[name = tensor("attn4_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100721344)))]; + tensor linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn4_in_proj_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor var_1254 = const()[name = tensor("op_1254"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_9_cast_fp16 = reshape(shape = var_1254, x = linear_16_cast_fp16)[name = tensor("qkv_9_cast_fp16")]; + tensor x_17_begin_0 = const()[name = tensor("x_17_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_17_end_0 = const()[name = tensor("x_17_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_17_end_mask_0 = const()[name = tensor("x_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_17_squeeze_mask_0 = const()[name = tensor("x_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_17_cast_fp16 = slice_by_index(begin = x_17_begin_0, end = x_17_end_0, end_mask = x_17_end_mask_0, squeeze_mask = x_17_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("x_17_cast_fp16")]; + tensor x_19_begin_0 = const()[name = tensor("x_19_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_19_end_0 = const()[name = tensor("x_19_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_19_end_mask_0 = const()[name = tensor("x_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_19_squeeze_mask_0 = const()[name = tensor("x_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_19_cast_fp16 = slice_by_index(begin = x_19_begin_0, end = x_19_end_0, end_mask = x_19_end_mask_0, squeeze_mask = x_19_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("x_19_cast_fp16")]; + tensor v_9_begin_0 = const()[name = tensor("v_9_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_9_end_0 = const()[name = tensor("v_9_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_9_end_mask_0 = const()[name = tensor("v_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_9_squeeze_mask_0 = const()[name = tensor("v_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_9_cast_fp16 = slice_by_index(begin = v_9_begin_0, end = v_9_end_0, end_mask = v_9_end_mask_0, squeeze_mask = v_9_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("v_9_cast_fp16")]; + tensor var_1299 = const()[name = tensor("op_1299"), val = tensor([256, 16, 32, 2])]; + tensor pairs_17_cast_fp16 = reshape(shape = var_1299, x = x_17_cast_fp16)[name = tensor("pairs_17_cast_fp16")]; + tensor xr_17_begin_0 = const()[name = tensor("xr_17_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_17_end_0 = const()[name = tensor("xr_17_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_17_end_mask_0 = const()[name = tensor("xr_17_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_17_squeeze_mask_0 = const()[name = tensor("xr_17_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_17_cast_fp16 = slice_by_index(begin = xr_17_begin_0, end = xr_17_end_0, end_mask = xr_17_end_mask_0, squeeze_mask = xr_17_squeeze_mask_0, x = pairs_17_cast_fp16)[name = tensor("xr_17_cast_fp16")]; + tensor xi_17_begin_0 = const()[name = tensor("xi_17_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_17_end_0 = const()[name = tensor("xi_17_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_17_end_mask_0 = const()[name = tensor("xi_17_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_17_squeeze_mask_0 = const()[name = tensor("xi_17_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_17_cast_fp16 = slice_by_index(begin = xi_17_begin_0, end = xi_17_end_0, end_mask = xi_17_end_mask_0, squeeze_mask = xi_17_squeeze_mask_0, x = pairs_17_cast_fp16)[name = tensor("xi_17_cast_fp16")]; + tensor var_1307_cast_fp16 = mul(x = xr_17_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1307_cast_fp16")]; + tensor var_1308_cast_fp16 = mul(x = xi_17_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1308_cast_fp16")]; + tensor xor__17_cast_fp16 = sub(x = var_1307_cast_fp16, y = var_1308_cast_fp16)[name = tensor("xor__17_cast_fp16")]; + tensor var_1311_cast_fp16 = mul(x = xr_17_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1311_cast_fp16")]; + tensor var_1312_cast_fp16 = mul(x = xi_17_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1312_cast_fp16")]; + tensor xoi_17_cast_fp16 = add(x = var_1311_cast_fp16, y = var_1312_cast_fp16)[name = tensor("xoi_17_cast_fp16")]; + tensor out_17_axis_0 = const()[name = tensor("out_17_axis_0"), val = tensor(-1)]; + tensor out_17_cast_fp16 = stack(axis = out_17_axis_0, values = (xor__17_cast_fp16, xoi_17_cast_fp16))[name = tensor("out_17_cast_fp16")]; + tensor var_1322 = const()[name = tensor("op_1322"), val = tensor([1, 256, 16, 64])]; + tensor q_9_cast_fp16 = reshape(shape = var_1322, x = out_17_cast_fp16)[name = tensor("q_9_cast_fp16")]; + tensor var_1328 = const()[name = tensor("op_1328"), val = tensor([256, 16, 32, 2])]; + tensor pairs_19_cast_fp16 = reshape(shape = var_1328, x = x_19_cast_fp16)[name = tensor("pairs_19_cast_fp16")]; + tensor xr_19_begin_0 = const()[name = tensor("xr_19_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_19_end_0 = const()[name = tensor("xr_19_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_19_end_mask_0 = const()[name = tensor("xr_19_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_19_squeeze_mask_0 = const()[name = tensor("xr_19_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_19_cast_fp16 = slice_by_index(begin = xr_19_begin_0, end = xr_19_end_0, end_mask = xr_19_end_mask_0, squeeze_mask = xr_19_squeeze_mask_0, x = pairs_19_cast_fp16)[name = tensor("xr_19_cast_fp16")]; + tensor xi_19_begin_0 = const()[name = tensor("xi_19_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_19_end_0 = const()[name = tensor("xi_19_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_19_end_mask_0 = const()[name = tensor("xi_19_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_19_squeeze_mask_0 = const()[name = tensor("xi_19_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_19_cast_fp16 = slice_by_index(begin = xi_19_begin_0, end = xi_19_end_0, end_mask = xi_19_end_mask_0, squeeze_mask = xi_19_squeeze_mask_0, x = pairs_19_cast_fp16)[name = tensor("xi_19_cast_fp16")]; + tensor var_1336_cast_fp16 = mul(x = xr_19_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1336_cast_fp16")]; + tensor var_1337_cast_fp16 = mul(x = xi_19_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1337_cast_fp16")]; + tensor xor__19_cast_fp16 = sub(x = var_1336_cast_fp16, y = var_1337_cast_fp16)[name = tensor("xor__19_cast_fp16")]; + tensor var_1340_cast_fp16 = mul(x = xr_19_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1340_cast_fp16")]; + tensor var_1341_cast_fp16 = mul(x = xi_19_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1341_cast_fp16")]; + tensor xoi_19_cast_fp16 = add(x = var_1340_cast_fp16, y = var_1341_cast_fp16)[name = tensor("xoi_19_cast_fp16")]; + tensor out_19_axis_0 = const()[name = tensor("out_19_axis_0"), val = tensor(-1)]; + tensor out_19_cast_fp16 = stack(axis = out_19_axis_0, values = (xor__19_cast_fp16, xoi_19_cast_fp16))[name = tensor("out_19_cast_fp16")]; + tensor var_1358 = const()[name = tensor("op_1358"), val = tensor([256, 1024])]; + tensor var_1359_cast_fp16 = reshape(shape = var_1358, x = out_19_cast_fp16)[name = tensor("op_1359_cast_fp16")]; + tensor written_k_9_transpose_x_1 = const()[name = tensor("written_k_9_transpose_x_1"), val = tensor(true)]; + tensor written_k_9_transpose_y_1 = const()[name = tensor("written_k_9_transpose_y_1"), val = tensor(false)]; + tensor written_k_9_cast_fp16 = matmul(transpose_x = written_k_9_transpose_x_1, transpose_y = written_k_9_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_1359_cast_fp16)[name = tensor("written_k_9_cast_fp16")]; + tensor var_1366 = const()[name = tensor("op_1366"), val = tensor([256, 1024])]; + tensor var_1367_cast_fp16 = reshape(shape = var_1366, x = v_9_cast_fp16)[name = tensor("op_1367_cast_fp16")]; + tensor written_v_9_transpose_x_1 = const()[name = tensor("written_v_9_transpose_x_1"), val = tensor(true)]; + tensor written_v_9_transpose_y_1 = const()[name = tensor("written_v_9_transpose_y_1"), val = tensor(false)]; + tensor written_v_9_cast_fp16 = matmul(transpose_x = written_v_9_transpose_x_1, transpose_y = written_v_9_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_1367_cast_fp16)[name = tensor("written_v_9_cast_fp16")]; + tensor k_cache4_to_fp16_dtype_0 = const()[name = tensor("k_cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache4_to_fp16 = cast(dtype = k_cache4_to_fp16_dtype_0, x = k_cache4)[name = tensor("cast_441")]; + tensor var_1378_cast_fp16 = mul(x = k_cache4_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_1378_cast_fp16")]; + tensor var_1383 = const()[name = tensor("op_1383"), val = tensor([1, 512, 16, 64])]; + tensor var_1384_cast_fp16 = reshape(shape = var_1383, x = written_k_9_cast_fp16)[name = tensor("op_1384_cast_fp16")]; + tensor new_k_cache_9_cast_fp16 = add(x = var_1378_cast_fp16, y = var_1384_cast_fp16)[name = tensor("new_k_cache_9_cast_fp16")]; + tensor new_k_cache_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache4_to_fp16_dtype_0 = const()[name = tensor("v_cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache4_to_fp16 = cast(dtype = v_cache4_to_fp16_dtype_0, x = v_cache4)[name = tensor("cast_439")]; + tensor var_1387_cast_fp16 = mul(x = v_cache4_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_1387_cast_fp16")]; + tensor var_1392 = const()[name = tensor("op_1392"), val = tensor([1, 512, 16, 64])]; + tensor var_1393_cast_fp16 = reshape(shape = var_1392, x = written_v_9_cast_fp16)[name = tensor("op_1393_cast_fp16")]; + tensor new_v_cache_9_cast_fp16 = add(x = var_1387_cast_fp16, y = var_1393_cast_fp16)[name = tensor("new_v_cache_9_cast_fp16")]; + tensor new_v_cache_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_1406 = const()[name = tensor("op_1406"), val = tensor([0, 2, 1, 3])]; + tensor var_1417_transpose_x_0 = const()[name = tensor("op_1417_transpose_x_0"), val = tensor(false)]; + tensor var_1417_transpose_y_0 = const()[name = tensor("op_1417_transpose_y_0"), val = tensor(false)]; + tensor transpose_77_perm_0 = const()[name = tensor("transpose_77_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_78_perm_0 = const()[name = tensor("transpose_78_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_78 = transpose(perm = transpose_78_perm_0, x = new_k_cache_9_cast_fp16)[name = tensor("transpose_188")]; + tensor transpose_77 = transpose(perm = transpose_77_perm_0, x = q_9_cast_fp16)[name = tensor("transpose_189")]; + tensor var_1417_cast_fp16 = matmul(transpose_x = var_1417_transpose_x_0, transpose_y = var_1417_transpose_y_0, x = transpose_77, y = transpose_78)[name = tensor("op_1417_cast_fp16")]; + tensor var_1418_to_fp16 = const()[name = tensor("op_1418_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_17_cast_fp16 = mul(x = var_1417_cast_fp16, y = var_1418_to_fp16)[name = tensor("scores_17_cast_fp16")]; + tensor scores_19_cast_fp16 = add(x = scores_17_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_19_cast_fp16")]; + tensor var_1428 = const()[name = tensor("op_1428"), val = tensor(-1)]; + tensor weights_9_cast_fp16 = softmax(axis = var_1428, x = scores_19_cast_fp16)[name = tensor("weights_9_cast_fp16")]; + tensor attn_9_transpose_x_0 = const()[name = tensor("attn_9_transpose_x_0"), val = tensor(false)]; + tensor attn_9_transpose_y_0 = const()[name = tensor("attn_9_transpose_y_0"), val = tensor(false)]; + tensor values_9_cast_fp16 = transpose(perm = var_1406, x = new_v_cache_9_cast_fp16)[name = tensor("transpose_190")]; + tensor attn_9_cast_fp16 = matmul(transpose_x = attn_9_transpose_x_0, transpose_y = attn_9_transpose_y_0, x = weights_9_cast_fp16, y = values_9_cast_fp16)[name = tensor("attn_9_cast_fp16")]; + tensor var_1436 = const()[name = tensor("op_1436"), val = tensor([0, 2, 1, 3])]; + tensor var_1441 = const()[name = tensor("op_1441"), val = tensor([1, 256, 1024])]; + tensor var_1437_cast_fp16 = transpose(perm = var_1436, x = attn_9_cast_fp16)[name = tensor("transpose_187")]; + tensor input_53_cast_fp16 = reshape(shape = var_1441, x = var_1437_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor attn4_out_proj_weight_to_fp16 = const()[name = tensor("attn4_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(107012864)))]; + tensor linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn4_out_proj_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor position4_to_fp16_dtype_0 = const()[name = tensor("position4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position4_to_fp16 = cast(dtype = position4_to_fp16_dtype_0, x = position4)[name = tensor("cast_437")]; + tensor var_1450_cast_fp16 = add(x = position4_to_fp16, y = valid_len_to_fp16)[name = tensor("op_1450_cast_fp16")]; + tensor var_1450_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1450_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_55_cast_fp16 = add(x = input_49_cast_fp16, y = linear_17_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor input_57_axes_0 = const()[name = tensor("input_57_axes_0"), val = tensor([-1])]; + tensor norm4_2_weight_to_fp16 = const()[name = tensor("norm4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109110080)))]; + tensor norm4_2_bias_to_fp16 = const()[name = tensor("norm4_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109112192)))]; + tensor var_1454_to_fp16 = const()[name = tensor("op_1454_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_57_cast_fp16 = layer_norm(axes = input_57_axes_0, beta = norm4_2_bias_to_fp16, epsilon = var_1454_to_fp16, gamma = norm4_2_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor linear4_1_weight_to_fp16 = const()[name = tensor("linear4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109114304)))]; + tensor linear_18_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear4_1_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor input_59_mode_0 = const()[name = tensor("input_59_mode_0"), val = tensor("EXACT")]; + tensor input_59_cast_fp16 = gelu(mode = input_59_mode_0, x = linear_18_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor linear4_2_weight_to_fp16 = const()[name = tensor("linear4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(117502976)))]; + tensor linear_19_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear4_2_weight_to_fp16, x = input_59_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_61_cast_fp16 = add(x = input_55_cast_fp16, y = linear_19_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor input_63_axes_0 = const()[name = tensor("input_63_axes_0"), val = tensor([-1])]; + tensor norm5_1_weight_to_fp16 = const()[name = tensor("norm5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125891648)))]; + tensor norm5_1_bias_to_fp16 = const()[name = tensor("norm5_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125893760)))]; + tensor var_1471_to_fp16 = const()[name = tensor("op_1471_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_63_cast_fp16 = layer_norm(axes = input_63_axes_0, beta = norm5_1_bias_to_fp16, epsilon = var_1471_to_fp16, gamma = norm5_1_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor attn5_in_proj_weight_to_fp16 = const()[name = tensor("attn5_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125895872)))]; + tensor linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn5_in_proj_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor var_1485 = const()[name = tensor("op_1485"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_11_cast_fp16 = reshape(shape = var_1485, x = linear_20_cast_fp16)[name = tensor("qkv_11_cast_fp16")]; + tensor x_21_begin_0 = const()[name = tensor("x_21_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_21_end_0 = const()[name = tensor("x_21_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_21_end_mask_0 = const()[name = tensor("x_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_21_squeeze_mask_0 = const()[name = tensor("x_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_21_cast_fp16 = slice_by_index(begin = x_21_begin_0, end = x_21_end_0, end_mask = x_21_end_mask_0, squeeze_mask = x_21_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("x_21_cast_fp16")]; + tensor x_23_begin_0 = const()[name = tensor("x_23_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_23_end_0 = const()[name = tensor("x_23_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_23_end_mask_0 = const()[name = tensor("x_23_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_23_squeeze_mask_0 = const()[name = tensor("x_23_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_23_cast_fp16 = slice_by_index(begin = x_23_begin_0, end = x_23_end_0, end_mask = x_23_end_mask_0, squeeze_mask = x_23_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("x_23_cast_fp16")]; + tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_11_cast_fp16 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("v_11_cast_fp16")]; + tensor var_1530 = const()[name = tensor("op_1530"), val = tensor([256, 16, 32, 2])]; + tensor pairs_21_cast_fp16 = reshape(shape = var_1530, x = x_21_cast_fp16)[name = tensor("pairs_21_cast_fp16")]; + tensor xr_21_begin_0 = const()[name = tensor("xr_21_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_21_end_0 = const()[name = tensor("xr_21_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_21_end_mask_0 = const()[name = tensor("xr_21_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_21_squeeze_mask_0 = const()[name = tensor("xr_21_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_21_cast_fp16 = slice_by_index(begin = xr_21_begin_0, end = xr_21_end_0, end_mask = xr_21_end_mask_0, squeeze_mask = xr_21_squeeze_mask_0, x = pairs_21_cast_fp16)[name = tensor("xr_21_cast_fp16")]; + tensor xi_21_begin_0 = const()[name = tensor("xi_21_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_21_end_0 = const()[name = tensor("xi_21_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_21_end_mask_0 = const()[name = tensor("xi_21_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_21_squeeze_mask_0 = const()[name = tensor("xi_21_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_21_cast_fp16 = slice_by_index(begin = xi_21_begin_0, end = xi_21_end_0, end_mask = xi_21_end_mask_0, squeeze_mask = xi_21_squeeze_mask_0, x = pairs_21_cast_fp16)[name = tensor("xi_21_cast_fp16")]; + tensor var_1538_cast_fp16 = mul(x = xr_21_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1538_cast_fp16")]; + tensor var_1539_cast_fp16 = mul(x = xi_21_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1539_cast_fp16")]; + tensor xor__21_cast_fp16 = sub(x = var_1538_cast_fp16, y = var_1539_cast_fp16)[name = tensor("xor__21_cast_fp16")]; + tensor var_1542_cast_fp16 = mul(x = xr_21_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1542_cast_fp16")]; + tensor var_1543_cast_fp16 = mul(x = xi_21_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1543_cast_fp16")]; + tensor xoi_21_cast_fp16 = add(x = var_1542_cast_fp16, y = var_1543_cast_fp16)[name = tensor("xoi_21_cast_fp16")]; + tensor out_21_axis_0 = const()[name = tensor("out_21_axis_0"), val = tensor(-1)]; + tensor out_21_cast_fp16 = stack(axis = out_21_axis_0, values = (xor__21_cast_fp16, xoi_21_cast_fp16))[name = tensor("out_21_cast_fp16")]; + tensor var_1553 = const()[name = tensor("op_1553"), val = tensor([1, 256, 16, 64])]; + tensor q_11_cast_fp16 = reshape(shape = var_1553, x = out_21_cast_fp16)[name = tensor("q_11_cast_fp16")]; + tensor var_1559 = const()[name = tensor("op_1559"), val = tensor([256, 16, 32, 2])]; + tensor pairs_23_cast_fp16 = reshape(shape = var_1559, x = x_23_cast_fp16)[name = tensor("pairs_23_cast_fp16")]; + tensor xr_23_begin_0 = const()[name = tensor("xr_23_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_23_end_0 = const()[name = tensor("xr_23_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_23_end_mask_0 = const()[name = tensor("xr_23_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_23_squeeze_mask_0 = const()[name = tensor("xr_23_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_23_cast_fp16 = slice_by_index(begin = xr_23_begin_0, end = xr_23_end_0, end_mask = xr_23_end_mask_0, squeeze_mask = xr_23_squeeze_mask_0, x = pairs_23_cast_fp16)[name = tensor("xr_23_cast_fp16")]; + tensor xi_23_begin_0 = const()[name = tensor("xi_23_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_23_end_0 = const()[name = tensor("xi_23_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_23_end_mask_0 = const()[name = tensor("xi_23_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_23_squeeze_mask_0 = const()[name = tensor("xi_23_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_23_cast_fp16 = slice_by_index(begin = xi_23_begin_0, end = xi_23_end_0, end_mask = xi_23_end_mask_0, squeeze_mask = xi_23_squeeze_mask_0, x = pairs_23_cast_fp16)[name = tensor("xi_23_cast_fp16")]; + tensor var_1567_cast_fp16 = mul(x = xr_23_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1567_cast_fp16")]; + tensor var_1568_cast_fp16 = mul(x = xi_23_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1568_cast_fp16")]; + tensor xor__23_cast_fp16 = sub(x = var_1567_cast_fp16, y = var_1568_cast_fp16)[name = tensor("xor__23_cast_fp16")]; + tensor var_1571_cast_fp16 = mul(x = xr_23_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1571_cast_fp16")]; + tensor var_1572_cast_fp16 = mul(x = xi_23_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1572_cast_fp16")]; + tensor xoi_23_cast_fp16 = add(x = var_1571_cast_fp16, y = var_1572_cast_fp16)[name = tensor("xoi_23_cast_fp16")]; + tensor out_23_axis_0 = const()[name = tensor("out_23_axis_0"), val = tensor(-1)]; + tensor out_23_cast_fp16 = stack(axis = out_23_axis_0, values = (xor__23_cast_fp16, xoi_23_cast_fp16))[name = tensor("out_23_cast_fp16")]; + tensor var_1589 = const()[name = tensor("op_1589"), val = tensor([256, 1024])]; + tensor var_1590_cast_fp16 = reshape(shape = var_1589, x = out_23_cast_fp16)[name = tensor("op_1590_cast_fp16")]; + tensor written_k_11_transpose_x_1 = const()[name = tensor("written_k_11_transpose_x_1"), val = tensor(true)]; + tensor written_k_11_transpose_y_1 = const()[name = tensor("written_k_11_transpose_y_1"), val = tensor(false)]; + tensor written_k_11_cast_fp16 = matmul(transpose_x = written_k_11_transpose_x_1, transpose_y = written_k_11_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_1590_cast_fp16)[name = tensor("written_k_11_cast_fp16")]; + tensor var_1597 = const()[name = tensor("op_1597"), val = tensor([256, 1024])]; + tensor var_1598_cast_fp16 = reshape(shape = var_1597, x = v_11_cast_fp16)[name = tensor("op_1598_cast_fp16")]; + tensor written_v_11_transpose_x_1 = const()[name = tensor("written_v_11_transpose_x_1"), val = tensor(true)]; + tensor written_v_11_transpose_y_1 = const()[name = tensor("written_v_11_transpose_y_1"), val = tensor(false)]; + tensor written_v_11_cast_fp16 = matmul(transpose_x = written_v_11_transpose_x_1, transpose_y = written_v_11_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_1598_cast_fp16)[name = tensor("written_v_11_cast_fp16")]; + tensor k_cache5_to_fp16_dtype_0 = const()[name = tensor("k_cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache5_to_fp16 = cast(dtype = k_cache5_to_fp16_dtype_0, x = k_cache5)[name = tensor("cast_435")]; + tensor var_1609_cast_fp16 = mul(x = k_cache5_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_1609_cast_fp16")]; + tensor var_1614 = const()[name = tensor("op_1614"), val = tensor([1, 512, 16, 64])]; + tensor var_1615_cast_fp16 = reshape(shape = var_1614, x = written_k_11_cast_fp16)[name = tensor("op_1615_cast_fp16")]; + tensor new_k_cache_11_cast_fp16 = add(x = var_1609_cast_fp16, y = var_1615_cast_fp16)[name = tensor("new_k_cache_11_cast_fp16")]; + tensor new_k_cache_11_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_11_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache5_to_fp16_dtype_0 = const()[name = tensor("v_cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache5_to_fp16 = cast(dtype = v_cache5_to_fp16_dtype_0, x = v_cache5)[name = tensor("cast_433")]; + tensor var_1618_cast_fp16 = mul(x = v_cache5_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_1618_cast_fp16")]; + tensor var_1623 = const()[name = tensor("op_1623"), val = tensor([1, 512, 16, 64])]; + tensor var_1624_cast_fp16 = reshape(shape = var_1623, x = written_v_11_cast_fp16)[name = tensor("op_1624_cast_fp16")]; + tensor new_v_cache_11_cast_fp16 = add(x = var_1618_cast_fp16, y = var_1624_cast_fp16)[name = tensor("new_v_cache_11_cast_fp16")]; + tensor new_v_cache_11_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_11_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_1637 = const()[name = tensor("op_1637"), val = tensor([0, 2, 1, 3])]; + tensor var_1648_transpose_x_0 = const()[name = tensor("op_1648_transpose_x_0"), val = tensor(false)]; + tensor var_1648_transpose_y_0 = const()[name = tensor("op_1648_transpose_y_0"), val = tensor(false)]; + tensor transpose_79_perm_0 = const()[name = tensor("transpose_79_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_80_perm_0 = const()[name = tensor("transpose_80_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_80 = transpose(perm = transpose_80_perm_0, x = new_k_cache_11_cast_fp16)[name = tensor("transpose_184")]; + tensor transpose_79 = transpose(perm = transpose_79_perm_0, x = q_11_cast_fp16)[name = tensor("transpose_185")]; + tensor var_1648_cast_fp16 = matmul(transpose_x = var_1648_transpose_x_0, transpose_y = var_1648_transpose_y_0, x = transpose_79, y = transpose_80)[name = tensor("op_1648_cast_fp16")]; + tensor var_1649_to_fp16 = const()[name = tensor("op_1649_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_21_cast_fp16 = mul(x = var_1648_cast_fp16, y = var_1649_to_fp16)[name = tensor("scores_21_cast_fp16")]; + tensor scores_23_cast_fp16 = add(x = scores_21_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_23_cast_fp16")]; + tensor var_1659 = const()[name = tensor("op_1659"), val = tensor(-1)]; + tensor weights_11_cast_fp16 = softmax(axis = var_1659, x = scores_23_cast_fp16)[name = tensor("weights_11_cast_fp16")]; + tensor attn_11_transpose_x_0 = const()[name = tensor("attn_11_transpose_x_0"), val = tensor(false)]; + tensor attn_11_transpose_y_0 = const()[name = tensor("attn_11_transpose_y_0"), val = tensor(false)]; + tensor values_11_cast_fp16 = transpose(perm = var_1637, x = new_v_cache_11_cast_fp16)[name = tensor("transpose_186")]; + tensor attn_11_cast_fp16 = matmul(transpose_x = attn_11_transpose_x_0, transpose_y = attn_11_transpose_y_0, x = weights_11_cast_fp16, y = values_11_cast_fp16)[name = tensor("attn_11_cast_fp16")]; + tensor var_1667 = const()[name = tensor("op_1667"), val = tensor([0, 2, 1, 3])]; + tensor var_1672 = const()[name = tensor("op_1672"), val = tensor([1, 256, 1024])]; + tensor var_1668_cast_fp16 = transpose(perm = var_1667, x = attn_11_cast_fp16)[name = tensor("transpose_183")]; + tensor input_65_cast_fp16 = reshape(shape = var_1672, x = var_1668_cast_fp16)[name = tensor("input_65_cast_fp16")]; + tensor attn5_out_proj_weight_to_fp16 = const()[name = tensor("attn5_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132187392)))]; + tensor linear_21_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn5_out_proj_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("linear_21_cast_fp16")]; + tensor position5_to_fp16_dtype_0 = const()[name = tensor("position5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position5_to_fp16 = cast(dtype = position5_to_fp16_dtype_0, x = position5)[name = tensor("cast_431")]; + tensor var_1681_cast_fp16 = add(x = position5_to_fp16, y = valid_len_to_fp16)[name = tensor("op_1681_cast_fp16")]; + tensor var_1681_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1681_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_67_cast_fp16 = add(x = input_61_cast_fp16, y = linear_21_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor input_69_axes_0 = const()[name = tensor("input_69_axes_0"), val = tensor([-1])]; + tensor norm5_2_weight_to_fp16 = const()[name = tensor("norm5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134284608)))]; + tensor norm5_2_bias_to_fp16 = const()[name = tensor("norm5_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134286720)))]; + tensor var_1685_to_fp16 = const()[name = tensor("op_1685_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_69_cast_fp16 = layer_norm(axes = input_69_axes_0, beta = norm5_2_bias_to_fp16, epsilon = var_1685_to_fp16, gamma = norm5_2_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("input_69_cast_fp16")]; + tensor linear5_1_weight_to_fp16 = const()[name = tensor("linear5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134288832)))]; + tensor linear_22_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear5_1_weight_to_fp16, x = input_69_cast_fp16)[name = tensor("linear_22_cast_fp16")]; + tensor input_71_mode_0 = const()[name = tensor("input_71_mode_0"), val = tensor("EXACT")]; + tensor input_71_cast_fp16 = gelu(mode = input_71_mode_0, x = linear_22_cast_fp16)[name = tensor("input_71_cast_fp16")]; + tensor linear5_2_weight_to_fp16 = const()[name = tensor("linear5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142677504)))]; + tensor linear_23_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear5_2_weight_to_fp16, x = input_71_cast_fp16)[name = tensor("linear_23_cast_fp16")]; + tensor input_73_cast_fp16 = add(x = input_67_cast_fp16, y = linear_23_cast_fp16)[name = tensor("input_73_cast_fp16")]; + tensor input_75_axes_0 = const()[name = tensor("input_75_axes_0"), val = tensor([-1])]; + tensor norm6_1_weight_to_fp16 = const()[name = tensor("norm6_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151066176)))]; + tensor norm6_1_bias_to_fp16 = const()[name = tensor("norm6_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151068288)))]; + tensor var_1702_to_fp16 = const()[name = tensor("op_1702_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_75_cast_fp16 = layer_norm(axes = input_75_axes_0, beta = norm6_1_bias_to_fp16, epsilon = var_1702_to_fp16, gamma = norm6_1_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("input_75_cast_fp16")]; + tensor attn6_in_proj_weight_to_fp16 = const()[name = tensor("attn6_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151070400)))]; + tensor linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn6_in_proj_weight_to_fp16, x = input_75_cast_fp16)[name = tensor("linear_24_cast_fp16")]; + tensor var_1716 = const()[name = tensor("op_1716"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_13_cast_fp16 = reshape(shape = var_1716, x = linear_24_cast_fp16)[name = tensor("qkv_13_cast_fp16")]; + tensor x_25_begin_0 = const()[name = tensor("x_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_25_end_0 = const()[name = tensor("x_25_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_25_end_mask_0 = const()[name = tensor("x_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_25_squeeze_mask_0 = const()[name = tensor("x_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_25_cast_fp16 = slice_by_index(begin = x_25_begin_0, end = x_25_end_0, end_mask = x_25_end_mask_0, squeeze_mask = x_25_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("x_25_cast_fp16")]; + tensor x_27_begin_0 = const()[name = tensor("x_27_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_27_end_0 = const()[name = tensor("x_27_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_27_end_mask_0 = const()[name = tensor("x_27_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_27_squeeze_mask_0 = const()[name = tensor("x_27_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_27_cast_fp16 = slice_by_index(begin = x_27_begin_0, end = x_27_end_0, end_mask = x_27_end_mask_0, squeeze_mask = x_27_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("x_27_cast_fp16")]; + tensor v_13_begin_0 = const()[name = tensor("v_13_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_13_end_0 = const()[name = tensor("v_13_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_13_end_mask_0 = const()[name = tensor("v_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_13_squeeze_mask_0 = const()[name = tensor("v_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_13_cast_fp16 = slice_by_index(begin = v_13_begin_0, end = v_13_end_0, end_mask = v_13_end_mask_0, squeeze_mask = v_13_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("v_13_cast_fp16")]; + tensor var_1761 = const()[name = tensor("op_1761"), val = tensor([256, 16, 32, 2])]; + tensor pairs_25_cast_fp16 = reshape(shape = var_1761, x = x_25_cast_fp16)[name = tensor("pairs_25_cast_fp16")]; + tensor xr_25_begin_0 = const()[name = tensor("xr_25_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_25_end_0 = const()[name = tensor("xr_25_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_25_end_mask_0 = const()[name = tensor("xr_25_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_25_squeeze_mask_0 = const()[name = tensor("xr_25_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_25_cast_fp16 = slice_by_index(begin = xr_25_begin_0, end = xr_25_end_0, end_mask = xr_25_end_mask_0, squeeze_mask = xr_25_squeeze_mask_0, x = pairs_25_cast_fp16)[name = tensor("xr_25_cast_fp16")]; + tensor xi_25_begin_0 = const()[name = tensor("xi_25_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_25_end_0 = const()[name = tensor("xi_25_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_25_end_mask_0 = const()[name = tensor("xi_25_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_25_squeeze_mask_0 = const()[name = tensor("xi_25_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_25_cast_fp16 = slice_by_index(begin = xi_25_begin_0, end = xi_25_end_0, end_mask = xi_25_end_mask_0, squeeze_mask = xi_25_squeeze_mask_0, x = pairs_25_cast_fp16)[name = tensor("xi_25_cast_fp16")]; + tensor var_1769_cast_fp16 = mul(x = xr_25_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1769_cast_fp16")]; + tensor var_1770_cast_fp16 = mul(x = xi_25_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1770_cast_fp16")]; + tensor xor__25_cast_fp16 = sub(x = var_1769_cast_fp16, y = var_1770_cast_fp16)[name = tensor("xor__25_cast_fp16")]; + tensor var_1773_cast_fp16 = mul(x = xr_25_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1773_cast_fp16")]; + tensor var_1774_cast_fp16 = mul(x = xi_25_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1774_cast_fp16")]; + tensor xoi_25_cast_fp16 = add(x = var_1773_cast_fp16, y = var_1774_cast_fp16)[name = tensor("xoi_25_cast_fp16")]; + tensor out_25_axis_0 = const()[name = tensor("out_25_axis_0"), val = tensor(-1)]; + tensor out_25_cast_fp16 = stack(axis = out_25_axis_0, values = (xor__25_cast_fp16, xoi_25_cast_fp16))[name = tensor("out_25_cast_fp16")]; + tensor var_1784 = const()[name = tensor("op_1784"), val = tensor([1, 256, 16, 64])]; + tensor q_13_cast_fp16 = reshape(shape = var_1784, x = out_25_cast_fp16)[name = tensor("q_13_cast_fp16")]; + tensor var_1790 = const()[name = tensor("op_1790"), val = tensor([256, 16, 32, 2])]; + tensor pairs_27_cast_fp16 = reshape(shape = var_1790, x = x_27_cast_fp16)[name = tensor("pairs_27_cast_fp16")]; + tensor xr_27_begin_0 = const()[name = tensor("xr_27_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_27_end_0 = const()[name = tensor("xr_27_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_27_end_mask_0 = const()[name = tensor("xr_27_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_27_squeeze_mask_0 = const()[name = tensor("xr_27_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_27_cast_fp16 = slice_by_index(begin = xr_27_begin_0, end = xr_27_end_0, end_mask = xr_27_end_mask_0, squeeze_mask = xr_27_squeeze_mask_0, x = pairs_27_cast_fp16)[name = tensor("xr_27_cast_fp16")]; + tensor xi_27_begin_0 = const()[name = tensor("xi_27_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_27_end_0 = const()[name = tensor("xi_27_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_27_end_mask_0 = const()[name = tensor("xi_27_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_27_squeeze_mask_0 = const()[name = tensor("xi_27_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_27_cast_fp16 = slice_by_index(begin = xi_27_begin_0, end = xi_27_end_0, end_mask = xi_27_end_mask_0, squeeze_mask = xi_27_squeeze_mask_0, x = pairs_27_cast_fp16)[name = tensor("xi_27_cast_fp16")]; + tensor var_1798_cast_fp16 = mul(x = xr_27_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1798_cast_fp16")]; + tensor var_1799_cast_fp16 = mul(x = xi_27_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1799_cast_fp16")]; + tensor xor__27_cast_fp16 = sub(x = var_1798_cast_fp16, y = var_1799_cast_fp16)[name = tensor("xor__27_cast_fp16")]; + tensor var_1802_cast_fp16 = mul(x = xr_27_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1802_cast_fp16")]; + tensor var_1803_cast_fp16 = mul(x = xi_27_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1803_cast_fp16")]; + tensor xoi_27_cast_fp16 = add(x = var_1802_cast_fp16, y = var_1803_cast_fp16)[name = tensor("xoi_27_cast_fp16")]; + tensor out_27_axis_0 = const()[name = tensor("out_27_axis_0"), val = tensor(-1)]; + tensor out_27_cast_fp16 = stack(axis = out_27_axis_0, values = (xor__27_cast_fp16, xoi_27_cast_fp16))[name = tensor("out_27_cast_fp16")]; + tensor var_1820 = const()[name = tensor("op_1820"), val = tensor([256, 1024])]; + tensor var_1821_cast_fp16 = reshape(shape = var_1820, x = out_27_cast_fp16)[name = tensor("op_1821_cast_fp16")]; + tensor written_k_13_transpose_x_1 = const()[name = tensor("written_k_13_transpose_x_1"), val = tensor(true)]; + tensor written_k_13_transpose_y_1 = const()[name = tensor("written_k_13_transpose_y_1"), val = tensor(false)]; + tensor written_k_13_cast_fp16 = matmul(transpose_x = written_k_13_transpose_x_1, transpose_y = written_k_13_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_1821_cast_fp16)[name = tensor("written_k_13_cast_fp16")]; + tensor var_1828 = const()[name = tensor("op_1828"), val = tensor([256, 1024])]; + tensor var_1829_cast_fp16 = reshape(shape = var_1828, x = v_13_cast_fp16)[name = tensor("op_1829_cast_fp16")]; + tensor written_v_13_transpose_x_1 = const()[name = tensor("written_v_13_transpose_x_1"), val = tensor(true)]; + tensor written_v_13_transpose_y_1 = const()[name = tensor("written_v_13_transpose_y_1"), val = tensor(false)]; + tensor written_v_13_cast_fp16 = matmul(transpose_x = written_v_13_transpose_x_1, transpose_y = written_v_13_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_1829_cast_fp16)[name = tensor("written_v_13_cast_fp16")]; + tensor k_cache6_to_fp16_dtype_0 = const()[name = tensor("k_cache6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache6_to_fp16 = cast(dtype = k_cache6_to_fp16_dtype_0, x = k_cache6)[name = tensor("cast_429")]; + tensor var_1840_cast_fp16 = mul(x = k_cache6_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_1840_cast_fp16")]; + tensor var_1845 = const()[name = tensor("op_1845"), val = tensor([1, 512, 16, 64])]; + tensor var_1846_cast_fp16 = reshape(shape = var_1845, x = written_k_13_cast_fp16)[name = tensor("op_1846_cast_fp16")]; + tensor new_k_cache_13_cast_fp16 = add(x = var_1840_cast_fp16, y = var_1846_cast_fp16)[name = tensor("new_k_cache_13_cast_fp16")]; + tensor new_k_cache_13_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_13_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache6_to_fp16_dtype_0 = const()[name = tensor("v_cache6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache6_to_fp16 = cast(dtype = v_cache6_to_fp16_dtype_0, x = v_cache6)[name = tensor("cast_427")]; + tensor var_1849_cast_fp16 = mul(x = v_cache6_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_1849_cast_fp16")]; + tensor var_1854 = const()[name = tensor("op_1854"), val = tensor([1, 512, 16, 64])]; + tensor var_1855_cast_fp16 = reshape(shape = var_1854, x = written_v_13_cast_fp16)[name = tensor("op_1855_cast_fp16")]; + tensor new_v_cache_13_cast_fp16 = add(x = var_1849_cast_fp16, y = var_1855_cast_fp16)[name = tensor("new_v_cache_13_cast_fp16")]; + tensor new_v_cache_13_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_13_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_1868 = const()[name = tensor("op_1868"), val = tensor([0, 2, 1, 3])]; + tensor var_1879_transpose_x_0 = const()[name = tensor("op_1879_transpose_x_0"), val = tensor(false)]; + tensor var_1879_transpose_y_0 = const()[name = tensor("op_1879_transpose_y_0"), val = tensor(false)]; + tensor transpose_81_perm_0 = const()[name = tensor("transpose_81_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_82_perm_0 = const()[name = tensor("transpose_82_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_82 = transpose(perm = transpose_82_perm_0, x = new_k_cache_13_cast_fp16)[name = tensor("transpose_180")]; + tensor transpose_81 = transpose(perm = transpose_81_perm_0, x = q_13_cast_fp16)[name = tensor("transpose_181")]; + tensor var_1879_cast_fp16 = matmul(transpose_x = var_1879_transpose_x_0, transpose_y = var_1879_transpose_y_0, x = transpose_81, y = transpose_82)[name = tensor("op_1879_cast_fp16")]; + tensor var_1880_to_fp16 = const()[name = tensor("op_1880_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_25_cast_fp16 = mul(x = var_1879_cast_fp16, y = var_1880_to_fp16)[name = tensor("scores_25_cast_fp16")]; + tensor scores_27_cast_fp16 = add(x = scores_25_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_27_cast_fp16")]; + tensor var_1890 = const()[name = tensor("op_1890"), val = tensor(-1)]; + tensor weights_13_cast_fp16 = softmax(axis = var_1890, x = scores_27_cast_fp16)[name = tensor("weights_13_cast_fp16")]; + tensor attn_13_transpose_x_0 = const()[name = tensor("attn_13_transpose_x_0"), val = tensor(false)]; + tensor attn_13_transpose_y_0 = const()[name = tensor("attn_13_transpose_y_0"), val = tensor(false)]; + tensor values_13_cast_fp16 = transpose(perm = var_1868, x = new_v_cache_13_cast_fp16)[name = tensor("transpose_182")]; + tensor attn_13_cast_fp16 = matmul(transpose_x = attn_13_transpose_x_0, transpose_y = attn_13_transpose_y_0, x = weights_13_cast_fp16, y = values_13_cast_fp16)[name = tensor("attn_13_cast_fp16")]; + tensor var_1898 = const()[name = tensor("op_1898"), val = tensor([0, 2, 1, 3])]; + tensor var_1903 = const()[name = tensor("op_1903"), val = tensor([1, 256, 1024])]; + tensor var_1899_cast_fp16 = transpose(perm = var_1898, x = attn_13_cast_fp16)[name = tensor("transpose_179")]; + tensor input_77_cast_fp16 = reshape(shape = var_1903, x = var_1899_cast_fp16)[name = tensor("input_77_cast_fp16")]; + tensor attn6_out_proj_weight_to_fp16 = const()[name = tensor("attn6_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(157361920)))]; + tensor linear_25_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn6_out_proj_weight_to_fp16, x = input_77_cast_fp16)[name = tensor("linear_25_cast_fp16")]; + tensor position6_to_fp16_dtype_0 = const()[name = tensor("position6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position6_to_fp16 = cast(dtype = position6_to_fp16_dtype_0, x = position6)[name = tensor("cast_425")]; + tensor var_1912_cast_fp16 = add(x = position6_to_fp16, y = valid_len_to_fp16)[name = tensor("op_1912_cast_fp16")]; + tensor var_1912_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1912_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_79_cast_fp16 = add(x = input_73_cast_fp16, y = linear_25_cast_fp16)[name = tensor("input_79_cast_fp16")]; + tensor input_81_axes_0 = const()[name = tensor("input_81_axes_0"), val = tensor([-1])]; + tensor norm6_2_weight_to_fp16 = const()[name = tensor("norm6_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159459136)))]; + tensor norm6_2_bias_to_fp16 = const()[name = tensor("norm6_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159461248)))]; + tensor var_1916_to_fp16 = const()[name = tensor("op_1916_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_81_cast_fp16 = layer_norm(axes = input_81_axes_0, beta = norm6_2_bias_to_fp16, epsilon = var_1916_to_fp16, gamma = norm6_2_weight_to_fp16, x = input_79_cast_fp16)[name = tensor("input_81_cast_fp16")]; + tensor linear6_1_weight_to_fp16 = const()[name = tensor("linear6_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159463360)))]; + tensor linear_26_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear6_1_weight_to_fp16, x = input_81_cast_fp16)[name = tensor("linear_26_cast_fp16")]; + tensor input_83_mode_0 = const()[name = tensor("input_83_mode_0"), val = tensor("EXACT")]; + tensor input_83_cast_fp16 = gelu(mode = input_83_mode_0, x = linear_26_cast_fp16)[name = tensor("input_83_cast_fp16")]; + tensor linear6_2_weight_to_fp16 = const()[name = tensor("linear6_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(167852032)))]; + tensor linear_27_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear6_2_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_27_cast_fp16")]; + tensor input_85_cast_fp16 = add(x = input_79_cast_fp16, y = linear_27_cast_fp16)[name = tensor("input_85_cast_fp16")]; + tensor input_87_axes_0 = const()[name = tensor("input_87_axes_0"), val = tensor([-1])]; + tensor norm7_1_weight_to_fp16 = const()[name = tensor("norm7_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176240704)))]; + tensor norm7_1_bias_to_fp16 = const()[name = tensor("norm7_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176242816)))]; + tensor var_1933_to_fp16 = const()[name = tensor("op_1933_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_87_cast_fp16 = layer_norm(axes = input_87_axes_0, beta = norm7_1_bias_to_fp16, epsilon = var_1933_to_fp16, gamma = norm7_1_weight_to_fp16, x = input_85_cast_fp16)[name = tensor("input_87_cast_fp16")]; + tensor attn7_in_proj_weight_to_fp16 = const()[name = tensor("attn7_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176244928)))]; + tensor linear_28_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn7_in_proj_weight_to_fp16, x = input_87_cast_fp16)[name = tensor("linear_28_cast_fp16")]; + tensor var_1947 = const()[name = tensor("op_1947"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_15_cast_fp16 = reshape(shape = var_1947, x = linear_28_cast_fp16)[name = tensor("qkv_15_cast_fp16")]; + tensor x_29_begin_0 = const()[name = tensor("x_29_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_29_end_0 = const()[name = tensor("x_29_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_29_end_mask_0 = const()[name = tensor("x_29_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_29_squeeze_mask_0 = const()[name = tensor("x_29_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_29_cast_fp16 = slice_by_index(begin = x_29_begin_0, end = x_29_end_0, end_mask = x_29_end_mask_0, squeeze_mask = x_29_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("x_29_cast_fp16")]; + tensor x_31_begin_0 = const()[name = tensor("x_31_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_31_end_0 = const()[name = tensor("x_31_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_31_end_mask_0 = const()[name = tensor("x_31_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_31_squeeze_mask_0 = const()[name = tensor("x_31_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_31_cast_fp16 = slice_by_index(begin = x_31_begin_0, end = x_31_end_0, end_mask = x_31_end_mask_0, squeeze_mask = x_31_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("x_31_cast_fp16")]; + tensor v_15_begin_0 = const()[name = tensor("v_15_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_15_end_0 = const()[name = tensor("v_15_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_15_end_mask_0 = const()[name = tensor("v_15_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_15_squeeze_mask_0 = const()[name = tensor("v_15_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_15_cast_fp16 = slice_by_index(begin = v_15_begin_0, end = v_15_end_0, end_mask = v_15_end_mask_0, squeeze_mask = v_15_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("v_15_cast_fp16")]; + tensor var_1992 = const()[name = tensor("op_1992"), val = tensor([256, 16, 32, 2])]; + tensor pairs_29_cast_fp16 = reshape(shape = var_1992, x = x_29_cast_fp16)[name = tensor("pairs_29_cast_fp16")]; + tensor xr_29_begin_0 = const()[name = tensor("xr_29_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_29_end_0 = const()[name = tensor("xr_29_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_29_end_mask_0 = const()[name = tensor("xr_29_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_29_squeeze_mask_0 = const()[name = tensor("xr_29_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_29_cast_fp16 = slice_by_index(begin = xr_29_begin_0, end = xr_29_end_0, end_mask = xr_29_end_mask_0, squeeze_mask = xr_29_squeeze_mask_0, x = pairs_29_cast_fp16)[name = tensor("xr_29_cast_fp16")]; + tensor xi_29_begin_0 = const()[name = tensor("xi_29_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_29_end_0 = const()[name = tensor("xi_29_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_29_end_mask_0 = const()[name = tensor("xi_29_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_29_squeeze_mask_0 = const()[name = tensor("xi_29_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_29_cast_fp16 = slice_by_index(begin = xi_29_begin_0, end = xi_29_end_0, end_mask = xi_29_end_mask_0, squeeze_mask = xi_29_squeeze_mask_0, x = pairs_29_cast_fp16)[name = tensor("xi_29_cast_fp16")]; + tensor var_2000_cast_fp16 = mul(x = xr_29_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2000_cast_fp16")]; + tensor var_2001_cast_fp16 = mul(x = xi_29_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2001_cast_fp16")]; + tensor xor__29_cast_fp16 = sub(x = var_2000_cast_fp16, y = var_2001_cast_fp16)[name = tensor("xor__29_cast_fp16")]; + tensor var_2004_cast_fp16 = mul(x = xr_29_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2004_cast_fp16")]; + tensor var_2005_cast_fp16 = mul(x = xi_29_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2005_cast_fp16")]; + tensor xoi_29_cast_fp16 = add(x = var_2004_cast_fp16, y = var_2005_cast_fp16)[name = tensor("xoi_29_cast_fp16")]; + tensor out_29_axis_0 = const()[name = tensor("out_29_axis_0"), val = tensor(-1)]; + tensor out_29_cast_fp16 = stack(axis = out_29_axis_0, values = (xor__29_cast_fp16, xoi_29_cast_fp16))[name = tensor("out_29_cast_fp16")]; + tensor var_2015 = const()[name = tensor("op_2015"), val = tensor([1, 256, 16, 64])]; + tensor q_15_cast_fp16 = reshape(shape = var_2015, x = out_29_cast_fp16)[name = tensor("q_15_cast_fp16")]; + tensor var_2021 = const()[name = tensor("op_2021"), val = tensor([256, 16, 32, 2])]; + tensor pairs_31_cast_fp16 = reshape(shape = var_2021, x = x_31_cast_fp16)[name = tensor("pairs_31_cast_fp16")]; + tensor xr_31_begin_0 = const()[name = tensor("xr_31_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_31_end_0 = const()[name = tensor("xr_31_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_31_end_mask_0 = const()[name = tensor("xr_31_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_31_squeeze_mask_0 = const()[name = tensor("xr_31_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_31_cast_fp16 = slice_by_index(begin = xr_31_begin_0, end = xr_31_end_0, end_mask = xr_31_end_mask_0, squeeze_mask = xr_31_squeeze_mask_0, x = pairs_31_cast_fp16)[name = tensor("xr_31_cast_fp16")]; + tensor xi_31_begin_0 = const()[name = tensor("xi_31_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_31_end_0 = const()[name = tensor("xi_31_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_31_end_mask_0 = const()[name = tensor("xi_31_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_31_squeeze_mask_0 = const()[name = tensor("xi_31_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_31_cast_fp16 = slice_by_index(begin = xi_31_begin_0, end = xi_31_end_0, end_mask = xi_31_end_mask_0, squeeze_mask = xi_31_squeeze_mask_0, x = pairs_31_cast_fp16)[name = tensor("xi_31_cast_fp16")]; + tensor var_2029_cast_fp16 = mul(x = xr_31_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2029_cast_fp16")]; + tensor var_2030_cast_fp16 = mul(x = xi_31_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2030_cast_fp16")]; + tensor xor__31_cast_fp16 = sub(x = var_2029_cast_fp16, y = var_2030_cast_fp16)[name = tensor("xor__31_cast_fp16")]; + tensor var_2033_cast_fp16 = mul(x = xr_31_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2033_cast_fp16")]; + tensor var_2034_cast_fp16 = mul(x = xi_31_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2034_cast_fp16")]; + tensor xoi_31_cast_fp16 = add(x = var_2033_cast_fp16, y = var_2034_cast_fp16)[name = tensor("xoi_31_cast_fp16")]; + tensor out_31_axis_0 = const()[name = tensor("out_31_axis_0"), val = tensor(-1)]; + tensor out_31_cast_fp16 = stack(axis = out_31_axis_0, values = (xor__31_cast_fp16, xoi_31_cast_fp16))[name = tensor("out_31_cast_fp16")]; + tensor var_2051 = const()[name = tensor("op_2051"), val = tensor([256, 1024])]; + tensor var_2052_cast_fp16 = reshape(shape = var_2051, x = out_31_cast_fp16)[name = tensor("op_2052_cast_fp16")]; + tensor written_k_15_transpose_x_1 = const()[name = tensor("written_k_15_transpose_x_1"), val = tensor(true)]; + tensor written_k_15_transpose_y_1 = const()[name = tensor("written_k_15_transpose_y_1"), val = tensor(false)]; + tensor written_k_15_cast_fp16 = matmul(transpose_x = written_k_15_transpose_x_1, transpose_y = written_k_15_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_2052_cast_fp16)[name = tensor("written_k_15_cast_fp16")]; + tensor var_2059 = const()[name = tensor("op_2059"), val = tensor([256, 1024])]; + tensor var_2060_cast_fp16 = reshape(shape = var_2059, x = v_15_cast_fp16)[name = tensor("op_2060_cast_fp16")]; + tensor written_v_15_transpose_x_1 = const()[name = tensor("written_v_15_transpose_x_1"), val = tensor(true)]; + tensor written_v_15_transpose_y_1 = const()[name = tensor("written_v_15_transpose_y_1"), val = tensor(false)]; + tensor written_v_15_cast_fp16 = matmul(transpose_x = written_v_15_transpose_x_1, transpose_y = written_v_15_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_2060_cast_fp16)[name = tensor("written_v_15_cast_fp16")]; + tensor k_cache7_to_fp16_dtype_0 = const()[name = tensor("k_cache7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache7_to_fp16 = cast(dtype = k_cache7_to_fp16_dtype_0, x = k_cache7)[name = tensor("cast_423")]; + tensor var_2071_cast_fp16 = mul(x = k_cache7_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_2071_cast_fp16")]; + tensor var_2076 = const()[name = tensor("op_2076"), val = tensor([1, 512, 16, 64])]; + tensor var_2077_cast_fp16 = reshape(shape = var_2076, x = written_k_15_cast_fp16)[name = tensor("op_2077_cast_fp16")]; + tensor new_k_cache_15_cast_fp16 = add(x = var_2071_cast_fp16, y = var_2077_cast_fp16)[name = tensor("new_k_cache_15_cast_fp16")]; + tensor new_k_cache_15_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_15_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache7_to_fp16_dtype_0 = const()[name = tensor("v_cache7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache7_to_fp16 = cast(dtype = v_cache7_to_fp16_dtype_0, x = v_cache7)[name = tensor("cast_421")]; + tensor var_2080_cast_fp16 = mul(x = v_cache7_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_2080_cast_fp16")]; + tensor var_2085 = const()[name = tensor("op_2085"), val = tensor([1, 512, 16, 64])]; + tensor var_2086_cast_fp16 = reshape(shape = var_2085, x = written_v_15_cast_fp16)[name = tensor("op_2086_cast_fp16")]; + tensor new_v_cache_15_cast_fp16 = add(x = var_2080_cast_fp16, y = var_2086_cast_fp16)[name = tensor("new_v_cache_15_cast_fp16")]; + tensor new_v_cache_15_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_15_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_2099 = const()[name = tensor("op_2099"), val = tensor([0, 2, 1, 3])]; + tensor var_2110_transpose_x_0 = const()[name = tensor("op_2110_transpose_x_0"), val = tensor(false)]; + tensor var_2110_transpose_y_0 = const()[name = tensor("op_2110_transpose_y_0"), val = tensor(false)]; + tensor transpose_83_perm_0 = const()[name = tensor("transpose_83_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_84_perm_0 = const()[name = tensor("transpose_84_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_84 = transpose(perm = transpose_84_perm_0, x = new_k_cache_15_cast_fp16)[name = tensor("transpose_176")]; + tensor transpose_83 = transpose(perm = transpose_83_perm_0, x = q_15_cast_fp16)[name = tensor("transpose_177")]; + tensor var_2110_cast_fp16 = matmul(transpose_x = var_2110_transpose_x_0, transpose_y = var_2110_transpose_y_0, x = transpose_83, y = transpose_84)[name = tensor("op_2110_cast_fp16")]; + tensor var_2111_to_fp16 = const()[name = tensor("op_2111_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_29_cast_fp16 = mul(x = var_2110_cast_fp16, y = var_2111_to_fp16)[name = tensor("scores_29_cast_fp16")]; + tensor scores_31_cast_fp16 = add(x = scores_29_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_31_cast_fp16")]; + tensor var_2121 = const()[name = tensor("op_2121"), val = tensor(-1)]; + tensor weights_15_cast_fp16 = softmax(axis = var_2121, x = scores_31_cast_fp16)[name = tensor("weights_15_cast_fp16")]; + tensor attn_15_transpose_x_0 = const()[name = tensor("attn_15_transpose_x_0"), val = tensor(false)]; + tensor attn_15_transpose_y_0 = const()[name = tensor("attn_15_transpose_y_0"), val = tensor(false)]; + tensor values_15_cast_fp16 = transpose(perm = var_2099, x = new_v_cache_15_cast_fp16)[name = tensor("transpose_178")]; + tensor attn_15_cast_fp16 = matmul(transpose_x = attn_15_transpose_x_0, transpose_y = attn_15_transpose_y_0, x = weights_15_cast_fp16, y = values_15_cast_fp16)[name = tensor("attn_15_cast_fp16")]; + tensor var_2129 = const()[name = tensor("op_2129"), val = tensor([0, 2, 1, 3])]; + tensor var_2134 = const()[name = tensor("op_2134"), val = tensor([1, 256, 1024])]; + tensor var_2130_cast_fp16 = transpose(perm = var_2129, x = attn_15_cast_fp16)[name = tensor("transpose_175")]; + tensor input_89_cast_fp16 = reshape(shape = var_2134, x = var_2130_cast_fp16)[name = tensor("input_89_cast_fp16")]; + tensor attn7_out_proj_weight_to_fp16 = const()[name = tensor("attn7_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(182536448)))]; + tensor linear_29_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn7_out_proj_weight_to_fp16, x = input_89_cast_fp16)[name = tensor("linear_29_cast_fp16")]; + tensor position7_to_fp16_dtype_0 = const()[name = tensor("position7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position7_to_fp16 = cast(dtype = position7_to_fp16_dtype_0, x = position7)[name = tensor("cast_419")]; + tensor var_2143_cast_fp16 = add(x = position7_to_fp16, y = valid_len_to_fp16)[name = tensor("op_2143_cast_fp16")]; + tensor var_2143_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2143_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_91_cast_fp16 = add(x = input_85_cast_fp16, y = linear_29_cast_fp16)[name = tensor("input_91_cast_fp16")]; + tensor input_93_axes_0 = const()[name = tensor("input_93_axes_0"), val = tensor([-1])]; + tensor norm7_2_weight_to_fp16 = const()[name = tensor("norm7_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184633664)))]; + tensor norm7_2_bias_to_fp16 = const()[name = tensor("norm7_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184635776)))]; + tensor var_2147_to_fp16 = const()[name = tensor("op_2147_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_93_cast_fp16 = layer_norm(axes = input_93_axes_0, beta = norm7_2_bias_to_fp16, epsilon = var_2147_to_fp16, gamma = norm7_2_weight_to_fp16, x = input_91_cast_fp16)[name = tensor("input_93_cast_fp16")]; + tensor linear7_1_weight_to_fp16 = const()[name = tensor("linear7_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184637888)))]; + tensor linear_30_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear7_1_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("linear_30_cast_fp16")]; + tensor input_95_mode_0 = const()[name = tensor("input_95_mode_0"), val = tensor("EXACT")]; + tensor input_95_cast_fp16 = gelu(mode = input_95_mode_0, x = linear_30_cast_fp16)[name = tensor("input_95_cast_fp16")]; + tensor linear7_2_weight_to_fp16 = const()[name = tensor("linear7_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(193026560)))]; + tensor linear_31_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear7_2_weight_to_fp16, x = input_95_cast_fp16)[name = tensor("linear_31_cast_fp16")]; + tensor input_97_cast_fp16 = add(x = input_91_cast_fp16, y = linear_31_cast_fp16)[name = tensor("input_97_cast_fp16")]; + tensor input_99_axes_0 = const()[name = tensor("input_99_axes_0"), val = tensor([-1])]; + tensor norm8_1_weight_to_fp16 = const()[name = tensor("norm8_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201415232)))]; + tensor norm8_1_bias_to_fp16 = const()[name = tensor("norm8_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201417344)))]; + tensor var_2164_to_fp16 = const()[name = tensor("op_2164_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_99_cast_fp16 = layer_norm(axes = input_99_axes_0, beta = norm8_1_bias_to_fp16, epsilon = var_2164_to_fp16, gamma = norm8_1_weight_to_fp16, x = input_97_cast_fp16)[name = tensor("input_99_cast_fp16")]; + tensor attn8_in_proj_weight_to_fp16 = const()[name = tensor("attn8_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201419456)))]; + tensor linear_32_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn8_in_proj_weight_to_fp16, x = input_99_cast_fp16)[name = tensor("linear_32_cast_fp16")]; + tensor var_2178 = const()[name = tensor("op_2178"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_17_cast_fp16 = reshape(shape = var_2178, x = linear_32_cast_fp16)[name = tensor("qkv_17_cast_fp16")]; + tensor x_33_begin_0 = const()[name = tensor("x_33_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_33_end_0 = const()[name = tensor("x_33_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_33_end_mask_0 = const()[name = tensor("x_33_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_33_squeeze_mask_0 = const()[name = tensor("x_33_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_33_cast_fp16 = slice_by_index(begin = x_33_begin_0, end = x_33_end_0, end_mask = x_33_end_mask_0, squeeze_mask = x_33_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("x_33_cast_fp16")]; + tensor x_35_begin_0 = const()[name = tensor("x_35_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_35_end_0 = const()[name = tensor("x_35_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_35_end_mask_0 = const()[name = tensor("x_35_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_35_squeeze_mask_0 = const()[name = tensor("x_35_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_35_cast_fp16 = slice_by_index(begin = x_35_begin_0, end = x_35_end_0, end_mask = x_35_end_mask_0, squeeze_mask = x_35_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("x_35_cast_fp16")]; + tensor v_17_begin_0 = const()[name = tensor("v_17_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_17_end_0 = const()[name = tensor("v_17_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_17_end_mask_0 = const()[name = tensor("v_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_17_squeeze_mask_0 = const()[name = tensor("v_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_17_cast_fp16 = slice_by_index(begin = v_17_begin_0, end = v_17_end_0, end_mask = v_17_end_mask_0, squeeze_mask = v_17_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("v_17_cast_fp16")]; + tensor var_2223 = const()[name = tensor("op_2223"), val = tensor([256, 16, 32, 2])]; + tensor pairs_33_cast_fp16 = reshape(shape = var_2223, x = x_33_cast_fp16)[name = tensor("pairs_33_cast_fp16")]; + tensor xr_33_begin_0 = const()[name = tensor("xr_33_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_33_end_0 = const()[name = tensor("xr_33_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_33_end_mask_0 = const()[name = tensor("xr_33_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_33_squeeze_mask_0 = const()[name = tensor("xr_33_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_33_cast_fp16 = slice_by_index(begin = xr_33_begin_0, end = xr_33_end_0, end_mask = xr_33_end_mask_0, squeeze_mask = xr_33_squeeze_mask_0, x = pairs_33_cast_fp16)[name = tensor("xr_33_cast_fp16")]; + tensor xi_33_begin_0 = const()[name = tensor("xi_33_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_33_end_0 = const()[name = tensor("xi_33_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_33_end_mask_0 = const()[name = tensor("xi_33_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_33_squeeze_mask_0 = const()[name = tensor("xi_33_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_33_cast_fp16 = slice_by_index(begin = xi_33_begin_0, end = xi_33_end_0, end_mask = xi_33_end_mask_0, squeeze_mask = xi_33_squeeze_mask_0, x = pairs_33_cast_fp16)[name = tensor("xi_33_cast_fp16")]; + tensor var_2231_cast_fp16 = mul(x = xr_33_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2231_cast_fp16")]; + tensor var_2232_cast_fp16 = mul(x = xi_33_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2232_cast_fp16")]; + tensor xor__33_cast_fp16 = sub(x = var_2231_cast_fp16, y = var_2232_cast_fp16)[name = tensor("xor__33_cast_fp16")]; + tensor var_2235_cast_fp16 = mul(x = xr_33_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2235_cast_fp16")]; + tensor var_2236_cast_fp16 = mul(x = xi_33_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2236_cast_fp16")]; + tensor xoi_33_cast_fp16 = add(x = var_2235_cast_fp16, y = var_2236_cast_fp16)[name = tensor("xoi_33_cast_fp16")]; + tensor out_33_axis_0 = const()[name = tensor("out_33_axis_0"), val = tensor(-1)]; + tensor out_33_cast_fp16 = stack(axis = out_33_axis_0, values = (xor__33_cast_fp16, xoi_33_cast_fp16))[name = tensor("out_33_cast_fp16")]; + tensor var_2246 = const()[name = tensor("op_2246"), val = tensor([1, 256, 16, 64])]; + tensor q_17_cast_fp16 = reshape(shape = var_2246, x = out_33_cast_fp16)[name = tensor("q_17_cast_fp16")]; + tensor var_2252 = const()[name = tensor("op_2252"), val = tensor([256, 16, 32, 2])]; + tensor pairs_35_cast_fp16 = reshape(shape = var_2252, x = x_35_cast_fp16)[name = tensor("pairs_35_cast_fp16")]; + tensor xr_35_begin_0 = const()[name = tensor("xr_35_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_35_end_0 = const()[name = tensor("xr_35_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_35_end_mask_0 = const()[name = tensor("xr_35_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_35_squeeze_mask_0 = const()[name = tensor("xr_35_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_35_cast_fp16 = slice_by_index(begin = xr_35_begin_0, end = xr_35_end_0, end_mask = xr_35_end_mask_0, squeeze_mask = xr_35_squeeze_mask_0, x = pairs_35_cast_fp16)[name = tensor("xr_35_cast_fp16")]; + tensor xi_35_begin_0 = const()[name = tensor("xi_35_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_35_end_0 = const()[name = tensor("xi_35_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_35_end_mask_0 = const()[name = tensor("xi_35_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_35_squeeze_mask_0 = const()[name = tensor("xi_35_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_35_cast_fp16 = slice_by_index(begin = xi_35_begin_0, end = xi_35_end_0, end_mask = xi_35_end_mask_0, squeeze_mask = xi_35_squeeze_mask_0, x = pairs_35_cast_fp16)[name = tensor("xi_35_cast_fp16")]; + tensor var_2260_cast_fp16 = mul(x = xr_35_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2260_cast_fp16")]; + tensor var_2261_cast_fp16 = mul(x = xi_35_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2261_cast_fp16")]; + tensor xor__35_cast_fp16 = sub(x = var_2260_cast_fp16, y = var_2261_cast_fp16)[name = tensor("xor__35_cast_fp16")]; + tensor var_2264_cast_fp16 = mul(x = xr_35_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2264_cast_fp16")]; + tensor var_2265_cast_fp16 = mul(x = xi_35_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2265_cast_fp16")]; + tensor xoi_35_cast_fp16 = add(x = var_2264_cast_fp16, y = var_2265_cast_fp16)[name = tensor("xoi_35_cast_fp16")]; + tensor out_35_axis_0 = const()[name = tensor("out_35_axis_0"), val = tensor(-1)]; + tensor out_35_cast_fp16 = stack(axis = out_35_axis_0, values = (xor__35_cast_fp16, xoi_35_cast_fp16))[name = tensor("out_35_cast_fp16")]; + tensor var_2282 = const()[name = tensor("op_2282"), val = tensor([256, 1024])]; + tensor var_2283_cast_fp16 = reshape(shape = var_2282, x = out_35_cast_fp16)[name = tensor("op_2283_cast_fp16")]; + tensor written_k_17_transpose_x_1 = const()[name = tensor("written_k_17_transpose_x_1"), val = tensor(true)]; + tensor written_k_17_transpose_y_1 = const()[name = tensor("written_k_17_transpose_y_1"), val = tensor(false)]; + tensor written_k_17_cast_fp16 = matmul(transpose_x = written_k_17_transpose_x_1, transpose_y = written_k_17_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_2283_cast_fp16)[name = tensor("written_k_17_cast_fp16")]; + tensor var_2290 = const()[name = tensor("op_2290"), val = tensor([256, 1024])]; + tensor var_2291_cast_fp16 = reshape(shape = var_2290, x = v_17_cast_fp16)[name = tensor("op_2291_cast_fp16")]; + tensor written_v_17_transpose_x_1 = const()[name = tensor("written_v_17_transpose_x_1"), val = tensor(true)]; + tensor written_v_17_transpose_y_1 = const()[name = tensor("written_v_17_transpose_y_1"), val = tensor(false)]; + tensor written_v_17_cast_fp16 = matmul(transpose_x = written_v_17_transpose_x_1, transpose_y = written_v_17_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_2291_cast_fp16)[name = tensor("written_v_17_cast_fp16")]; + tensor k_cache8_to_fp16_dtype_0 = const()[name = tensor("k_cache8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache8_to_fp16 = cast(dtype = k_cache8_to_fp16_dtype_0, x = k_cache8)[name = tensor("cast_417")]; + tensor var_2302_cast_fp16 = mul(x = k_cache8_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_2302_cast_fp16")]; + tensor var_2307 = const()[name = tensor("op_2307"), val = tensor([1, 512, 16, 64])]; + tensor var_2308_cast_fp16 = reshape(shape = var_2307, x = written_k_17_cast_fp16)[name = tensor("op_2308_cast_fp16")]; + tensor new_k_cache_17_cast_fp16 = add(x = var_2302_cast_fp16, y = var_2308_cast_fp16)[name = tensor("new_k_cache_17_cast_fp16")]; + tensor new_k_cache_17_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_17_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache8_to_fp16_dtype_0 = const()[name = tensor("v_cache8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache8_to_fp16 = cast(dtype = v_cache8_to_fp16_dtype_0, x = v_cache8)[name = tensor("cast_415")]; + tensor var_2311_cast_fp16 = mul(x = v_cache8_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_2311_cast_fp16")]; + tensor var_2316 = const()[name = tensor("op_2316"), val = tensor([1, 512, 16, 64])]; + tensor var_2317_cast_fp16 = reshape(shape = var_2316, x = written_v_17_cast_fp16)[name = tensor("op_2317_cast_fp16")]; + tensor new_v_cache_17_cast_fp16 = add(x = var_2311_cast_fp16, y = var_2317_cast_fp16)[name = tensor("new_v_cache_17_cast_fp16")]; + tensor new_v_cache_17_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_17_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_2330 = const()[name = tensor("op_2330"), val = tensor([0, 2, 1, 3])]; + tensor var_2341_transpose_x_0 = const()[name = tensor("op_2341_transpose_x_0"), val = tensor(false)]; + tensor var_2341_transpose_y_0 = const()[name = tensor("op_2341_transpose_y_0"), val = tensor(false)]; + tensor transpose_85_perm_0 = const()[name = tensor("transpose_85_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_86_perm_0 = const()[name = tensor("transpose_86_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_86 = transpose(perm = transpose_86_perm_0, x = new_k_cache_17_cast_fp16)[name = tensor("transpose_172")]; + tensor transpose_85 = transpose(perm = transpose_85_perm_0, x = q_17_cast_fp16)[name = tensor("transpose_173")]; + tensor var_2341_cast_fp16 = matmul(transpose_x = var_2341_transpose_x_0, transpose_y = var_2341_transpose_y_0, x = transpose_85, y = transpose_86)[name = tensor("op_2341_cast_fp16")]; + tensor var_2342_to_fp16 = const()[name = tensor("op_2342_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_33_cast_fp16 = mul(x = var_2341_cast_fp16, y = var_2342_to_fp16)[name = tensor("scores_33_cast_fp16")]; + tensor scores_35_cast_fp16 = add(x = scores_33_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_35_cast_fp16")]; + tensor var_2352 = const()[name = tensor("op_2352"), val = tensor(-1)]; + tensor weights_17_cast_fp16 = softmax(axis = var_2352, x = scores_35_cast_fp16)[name = tensor("weights_17_cast_fp16")]; + tensor attn_17_transpose_x_0 = const()[name = tensor("attn_17_transpose_x_0"), val = tensor(false)]; + tensor attn_17_transpose_y_0 = const()[name = tensor("attn_17_transpose_y_0"), val = tensor(false)]; + tensor values_17_cast_fp16 = transpose(perm = var_2330, x = new_v_cache_17_cast_fp16)[name = tensor("transpose_174")]; + tensor attn_17_cast_fp16 = matmul(transpose_x = attn_17_transpose_x_0, transpose_y = attn_17_transpose_y_0, x = weights_17_cast_fp16, y = values_17_cast_fp16)[name = tensor("attn_17_cast_fp16")]; + tensor var_2360 = const()[name = tensor("op_2360"), val = tensor([0, 2, 1, 3])]; + tensor var_2365 = const()[name = tensor("op_2365"), val = tensor([1, 256, 1024])]; + tensor var_2361_cast_fp16 = transpose(perm = var_2360, x = attn_17_cast_fp16)[name = tensor("transpose_171")]; + tensor input_101_cast_fp16 = reshape(shape = var_2365, x = var_2361_cast_fp16)[name = tensor("input_101_cast_fp16")]; + tensor attn8_out_proj_weight_to_fp16 = const()[name = tensor("attn8_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(207710976)))]; + tensor linear_33_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn8_out_proj_weight_to_fp16, x = input_101_cast_fp16)[name = tensor("linear_33_cast_fp16")]; + tensor position8_to_fp16_dtype_0 = const()[name = tensor("position8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position8_to_fp16 = cast(dtype = position8_to_fp16_dtype_0, x = position8)[name = tensor("cast_413")]; + tensor var_2374_cast_fp16 = add(x = position8_to_fp16, y = valid_len_to_fp16)[name = tensor("op_2374_cast_fp16")]; + tensor var_2374_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2374_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_103_cast_fp16 = add(x = input_97_cast_fp16, y = linear_33_cast_fp16)[name = tensor("input_103_cast_fp16")]; + tensor input_105_axes_0 = const()[name = tensor("input_105_axes_0"), val = tensor([-1])]; + tensor norm8_2_weight_to_fp16 = const()[name = tensor("norm8_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209808192)))]; + tensor norm8_2_bias_to_fp16 = const()[name = tensor("norm8_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209810304)))]; + tensor var_2378_to_fp16 = const()[name = tensor("op_2378_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_105_cast_fp16 = layer_norm(axes = input_105_axes_0, beta = norm8_2_bias_to_fp16, epsilon = var_2378_to_fp16, gamma = norm8_2_weight_to_fp16, x = input_103_cast_fp16)[name = tensor("input_105_cast_fp16")]; + tensor linear8_1_weight_to_fp16 = const()[name = tensor("linear8_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209812416)))]; + tensor linear_34_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear8_1_weight_to_fp16, x = input_105_cast_fp16)[name = tensor("linear_34_cast_fp16")]; + tensor input_107_mode_0 = const()[name = tensor("input_107_mode_0"), val = tensor("EXACT")]; + tensor input_107_cast_fp16 = gelu(mode = input_107_mode_0, x = linear_34_cast_fp16)[name = tensor("input_107_cast_fp16")]; + tensor linear8_2_weight_to_fp16 = const()[name = tensor("linear8_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(218201088)))]; + tensor linear_35_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear8_2_weight_to_fp16, x = input_107_cast_fp16)[name = tensor("linear_35_cast_fp16")]; + tensor input_109_cast_fp16 = add(x = input_103_cast_fp16, y = linear_35_cast_fp16)[name = tensor("input_109_cast_fp16")]; + tensor input_111_axes_0 = const()[name = tensor("input_111_axes_0"), val = tensor([-1])]; + tensor norm9_1_weight_to_fp16 = const()[name = tensor("norm9_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226589760)))]; + tensor norm9_1_bias_to_fp16 = const()[name = tensor("norm9_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226591872)))]; + tensor var_2395_to_fp16 = const()[name = tensor("op_2395_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_111_cast_fp16 = layer_norm(axes = input_111_axes_0, beta = norm9_1_bias_to_fp16, epsilon = var_2395_to_fp16, gamma = norm9_1_weight_to_fp16, x = input_109_cast_fp16)[name = tensor("input_111_cast_fp16")]; + tensor attn9_in_proj_weight_to_fp16 = const()[name = tensor("attn9_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226593984)))]; + tensor linear_36_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn9_in_proj_weight_to_fp16, x = input_111_cast_fp16)[name = tensor("linear_36_cast_fp16")]; + tensor var_2409 = const()[name = tensor("op_2409"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_19_cast_fp16 = reshape(shape = var_2409, x = linear_36_cast_fp16)[name = tensor("qkv_19_cast_fp16")]; + tensor x_37_begin_0 = const()[name = tensor("x_37_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_37_end_0 = const()[name = tensor("x_37_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_37_end_mask_0 = const()[name = tensor("x_37_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_37_squeeze_mask_0 = const()[name = tensor("x_37_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_37_cast_fp16 = slice_by_index(begin = x_37_begin_0, end = x_37_end_0, end_mask = x_37_end_mask_0, squeeze_mask = x_37_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("x_37_cast_fp16")]; + tensor x_39_begin_0 = const()[name = tensor("x_39_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_39_end_0 = const()[name = tensor("x_39_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_39_end_mask_0 = const()[name = tensor("x_39_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_39_squeeze_mask_0 = const()[name = tensor("x_39_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_39_cast_fp16 = slice_by_index(begin = x_39_begin_0, end = x_39_end_0, end_mask = x_39_end_mask_0, squeeze_mask = x_39_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("x_39_cast_fp16")]; + tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_19_cast_fp16 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("v_19_cast_fp16")]; + tensor var_2454 = const()[name = tensor("op_2454"), val = tensor([256, 16, 32, 2])]; + tensor pairs_37_cast_fp16 = reshape(shape = var_2454, x = x_37_cast_fp16)[name = tensor("pairs_37_cast_fp16")]; + tensor xr_37_begin_0 = const()[name = tensor("xr_37_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_37_end_0 = const()[name = tensor("xr_37_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_37_end_mask_0 = const()[name = tensor("xr_37_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_37_squeeze_mask_0 = const()[name = tensor("xr_37_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_37_cast_fp16 = slice_by_index(begin = xr_37_begin_0, end = xr_37_end_0, end_mask = xr_37_end_mask_0, squeeze_mask = xr_37_squeeze_mask_0, x = pairs_37_cast_fp16)[name = tensor("xr_37_cast_fp16")]; + tensor xi_37_begin_0 = const()[name = tensor("xi_37_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_37_end_0 = const()[name = tensor("xi_37_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_37_end_mask_0 = const()[name = tensor("xi_37_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_37_squeeze_mask_0 = const()[name = tensor("xi_37_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_37_cast_fp16 = slice_by_index(begin = xi_37_begin_0, end = xi_37_end_0, end_mask = xi_37_end_mask_0, squeeze_mask = xi_37_squeeze_mask_0, x = pairs_37_cast_fp16)[name = tensor("xi_37_cast_fp16")]; + tensor var_2462_cast_fp16 = mul(x = xr_37_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2462_cast_fp16")]; + tensor var_2463_cast_fp16 = mul(x = xi_37_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2463_cast_fp16")]; + tensor xor__37_cast_fp16 = sub(x = var_2462_cast_fp16, y = var_2463_cast_fp16)[name = tensor("xor__37_cast_fp16")]; + tensor var_2466_cast_fp16 = mul(x = xr_37_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2466_cast_fp16")]; + tensor var_2467_cast_fp16 = mul(x = xi_37_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2467_cast_fp16")]; + tensor xoi_37_cast_fp16 = add(x = var_2466_cast_fp16, y = var_2467_cast_fp16)[name = tensor("xoi_37_cast_fp16")]; + tensor out_37_axis_0 = const()[name = tensor("out_37_axis_0"), val = tensor(-1)]; + tensor out_37_cast_fp16 = stack(axis = out_37_axis_0, values = (xor__37_cast_fp16, xoi_37_cast_fp16))[name = tensor("out_37_cast_fp16")]; + tensor var_2477 = const()[name = tensor("op_2477"), val = tensor([1, 256, 16, 64])]; + tensor q_19_cast_fp16 = reshape(shape = var_2477, x = out_37_cast_fp16)[name = tensor("q_19_cast_fp16")]; + tensor var_2483 = const()[name = tensor("op_2483"), val = tensor([256, 16, 32, 2])]; + tensor pairs_39_cast_fp16 = reshape(shape = var_2483, x = x_39_cast_fp16)[name = tensor("pairs_39_cast_fp16")]; + tensor xr_39_begin_0 = const()[name = tensor("xr_39_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_39_end_0 = const()[name = tensor("xr_39_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_39_end_mask_0 = const()[name = tensor("xr_39_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_39_squeeze_mask_0 = const()[name = tensor("xr_39_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_39_cast_fp16 = slice_by_index(begin = xr_39_begin_0, end = xr_39_end_0, end_mask = xr_39_end_mask_0, squeeze_mask = xr_39_squeeze_mask_0, x = pairs_39_cast_fp16)[name = tensor("xr_39_cast_fp16")]; + tensor xi_39_begin_0 = const()[name = tensor("xi_39_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_39_end_0 = const()[name = tensor("xi_39_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_39_end_mask_0 = const()[name = tensor("xi_39_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_39_squeeze_mask_0 = const()[name = tensor("xi_39_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_39_cast_fp16 = slice_by_index(begin = xi_39_begin_0, end = xi_39_end_0, end_mask = xi_39_end_mask_0, squeeze_mask = xi_39_squeeze_mask_0, x = pairs_39_cast_fp16)[name = tensor("xi_39_cast_fp16")]; + tensor var_2491_cast_fp16 = mul(x = xr_39_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2491_cast_fp16")]; + tensor var_2492_cast_fp16 = mul(x = xi_39_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2492_cast_fp16")]; + tensor xor__39_cast_fp16 = sub(x = var_2491_cast_fp16, y = var_2492_cast_fp16)[name = tensor("xor__39_cast_fp16")]; + tensor var_2495_cast_fp16 = mul(x = xr_39_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2495_cast_fp16")]; + tensor var_2496_cast_fp16 = mul(x = xi_39_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2496_cast_fp16")]; + tensor xoi_39_cast_fp16 = add(x = var_2495_cast_fp16, y = var_2496_cast_fp16)[name = tensor("xoi_39_cast_fp16")]; + tensor out_39_axis_0 = const()[name = tensor("out_39_axis_0"), val = tensor(-1)]; + tensor out_39_cast_fp16 = stack(axis = out_39_axis_0, values = (xor__39_cast_fp16, xoi_39_cast_fp16))[name = tensor("out_39_cast_fp16")]; + tensor var_2513 = const()[name = tensor("op_2513"), val = tensor([256, 1024])]; + tensor var_2514_cast_fp16 = reshape(shape = var_2513, x = out_39_cast_fp16)[name = tensor("op_2514_cast_fp16")]; + tensor written_k_19_transpose_x_1 = const()[name = tensor("written_k_19_transpose_x_1"), val = tensor(true)]; + tensor written_k_19_transpose_y_1 = const()[name = tensor("written_k_19_transpose_y_1"), val = tensor(false)]; + tensor written_k_19_cast_fp16 = matmul(transpose_x = written_k_19_transpose_x_1, transpose_y = written_k_19_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_2514_cast_fp16)[name = tensor("written_k_19_cast_fp16")]; + tensor var_2521 = const()[name = tensor("op_2521"), val = tensor([256, 1024])]; + tensor var_2522_cast_fp16 = reshape(shape = var_2521, x = v_19_cast_fp16)[name = tensor("op_2522_cast_fp16")]; + tensor written_v_19_transpose_x_1 = const()[name = tensor("written_v_19_transpose_x_1"), val = tensor(true)]; + tensor written_v_19_transpose_y_1 = const()[name = tensor("written_v_19_transpose_y_1"), val = tensor(false)]; + tensor written_v_19_cast_fp16 = matmul(transpose_x = written_v_19_transpose_x_1, transpose_y = written_v_19_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_2522_cast_fp16)[name = tensor("written_v_19_cast_fp16")]; + tensor k_cache9_to_fp16_dtype_0 = const()[name = tensor("k_cache9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache9_to_fp16 = cast(dtype = k_cache9_to_fp16_dtype_0, x = k_cache9)[name = tensor("cast_411")]; + tensor var_2533_cast_fp16 = mul(x = k_cache9_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_2533_cast_fp16")]; + tensor var_2538 = const()[name = tensor("op_2538"), val = tensor([1, 512, 16, 64])]; + tensor var_2539_cast_fp16 = reshape(shape = var_2538, x = written_k_19_cast_fp16)[name = tensor("op_2539_cast_fp16")]; + tensor new_k_cache_19_cast_fp16 = add(x = var_2533_cast_fp16, y = var_2539_cast_fp16)[name = tensor("new_k_cache_19_cast_fp16")]; + tensor new_k_cache_19_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_19_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache9_to_fp16_dtype_0 = const()[name = tensor("v_cache9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache9_to_fp16 = cast(dtype = v_cache9_to_fp16_dtype_0, x = v_cache9)[name = tensor("cast_409")]; + tensor var_2542_cast_fp16 = mul(x = v_cache9_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_2542_cast_fp16")]; + tensor var_2547 = const()[name = tensor("op_2547"), val = tensor([1, 512, 16, 64])]; + tensor var_2548_cast_fp16 = reshape(shape = var_2547, x = written_v_19_cast_fp16)[name = tensor("op_2548_cast_fp16")]; + tensor new_v_cache_19_cast_fp16 = add(x = var_2542_cast_fp16, y = var_2548_cast_fp16)[name = tensor("new_v_cache_19_cast_fp16")]; + tensor new_v_cache_19_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_19_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_2561 = const()[name = tensor("op_2561"), val = tensor([0, 2, 1, 3])]; + tensor var_2572_transpose_x_0 = const()[name = tensor("op_2572_transpose_x_0"), val = tensor(false)]; + tensor var_2572_transpose_y_0 = const()[name = tensor("op_2572_transpose_y_0"), val = tensor(false)]; + tensor transpose_87_perm_0 = const()[name = tensor("transpose_87_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_88_perm_0 = const()[name = tensor("transpose_88_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_88 = transpose(perm = transpose_88_perm_0, x = new_k_cache_19_cast_fp16)[name = tensor("transpose_168")]; + tensor transpose_87 = transpose(perm = transpose_87_perm_0, x = q_19_cast_fp16)[name = tensor("transpose_169")]; + tensor var_2572_cast_fp16 = matmul(transpose_x = var_2572_transpose_x_0, transpose_y = var_2572_transpose_y_0, x = transpose_87, y = transpose_88)[name = tensor("op_2572_cast_fp16")]; + tensor var_2573_to_fp16 = const()[name = tensor("op_2573_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_37_cast_fp16 = mul(x = var_2572_cast_fp16, y = var_2573_to_fp16)[name = tensor("scores_37_cast_fp16")]; + tensor scores_39_cast_fp16 = add(x = scores_37_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_39_cast_fp16")]; + tensor var_2583 = const()[name = tensor("op_2583"), val = tensor(-1)]; + tensor weights_19_cast_fp16 = softmax(axis = var_2583, x = scores_39_cast_fp16)[name = tensor("weights_19_cast_fp16")]; + tensor attn_19_transpose_x_0 = const()[name = tensor("attn_19_transpose_x_0"), val = tensor(false)]; + tensor attn_19_transpose_y_0 = const()[name = tensor("attn_19_transpose_y_0"), val = tensor(false)]; + tensor values_19_cast_fp16 = transpose(perm = var_2561, x = new_v_cache_19_cast_fp16)[name = tensor("transpose_170")]; + tensor attn_19_cast_fp16 = matmul(transpose_x = attn_19_transpose_x_0, transpose_y = attn_19_transpose_y_0, x = weights_19_cast_fp16, y = values_19_cast_fp16)[name = tensor("attn_19_cast_fp16")]; + tensor var_2591 = const()[name = tensor("op_2591"), val = tensor([0, 2, 1, 3])]; + tensor var_2596 = const()[name = tensor("op_2596"), val = tensor([1, 256, 1024])]; + tensor var_2592_cast_fp16 = transpose(perm = var_2591, x = attn_19_cast_fp16)[name = tensor("transpose_167")]; + tensor input_113_cast_fp16 = reshape(shape = var_2596, x = var_2592_cast_fp16)[name = tensor("input_113_cast_fp16")]; + tensor attn9_out_proj_weight_to_fp16 = const()[name = tensor("attn9_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(232885504)))]; + tensor linear_37_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn9_out_proj_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("linear_37_cast_fp16")]; + tensor position9_to_fp16_dtype_0 = const()[name = tensor("position9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position9_to_fp16 = cast(dtype = position9_to_fp16_dtype_0, x = position9)[name = tensor("cast_407")]; + tensor var_2605_cast_fp16 = add(x = position9_to_fp16, y = valid_len_to_fp16)[name = tensor("op_2605_cast_fp16")]; + tensor var_2605_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2605_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_115_cast_fp16 = add(x = input_109_cast_fp16, y = linear_37_cast_fp16)[name = tensor("input_115_cast_fp16")]; + tensor input_117_axes_0 = const()[name = tensor("input_117_axes_0"), val = tensor([-1])]; + tensor norm9_2_weight_to_fp16 = const()[name = tensor("norm9_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(234982720)))]; + tensor norm9_2_bias_to_fp16 = const()[name = tensor("norm9_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(234984832)))]; + tensor var_2609_to_fp16 = const()[name = tensor("op_2609_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_117_cast_fp16 = layer_norm(axes = input_117_axes_0, beta = norm9_2_bias_to_fp16, epsilon = var_2609_to_fp16, gamma = norm9_2_weight_to_fp16, x = input_115_cast_fp16)[name = tensor("input_117_cast_fp16")]; + tensor linear9_1_weight_to_fp16 = const()[name = tensor("linear9_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(234986944)))]; + tensor linear_38_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear9_1_weight_to_fp16, x = input_117_cast_fp16)[name = tensor("linear_38_cast_fp16")]; + tensor input_119_mode_0 = const()[name = tensor("input_119_mode_0"), val = tensor("EXACT")]; + tensor input_119_cast_fp16 = gelu(mode = input_119_mode_0, x = linear_38_cast_fp16)[name = tensor("input_119_cast_fp16")]; + tensor linear9_2_weight_to_fp16 = const()[name = tensor("linear9_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(243375616)))]; + tensor linear_39_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear9_2_weight_to_fp16, x = input_119_cast_fp16)[name = tensor("linear_39_cast_fp16")]; + tensor input_121_cast_fp16 = add(x = input_115_cast_fp16, y = linear_39_cast_fp16)[name = tensor("input_121_cast_fp16")]; + tensor input_123_axes_0 = const()[name = tensor("input_123_axes_0"), val = tensor([-1])]; + tensor norm10_1_weight_to_fp16 = const()[name = tensor("norm10_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(251764288)))]; + tensor norm10_1_bias_to_fp16 = const()[name = tensor("norm10_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(251766400)))]; + tensor var_2626_to_fp16 = const()[name = tensor("op_2626_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_123_cast_fp16 = layer_norm(axes = input_123_axes_0, beta = norm10_1_bias_to_fp16, epsilon = var_2626_to_fp16, gamma = norm10_1_weight_to_fp16, x = input_121_cast_fp16)[name = tensor("input_123_cast_fp16")]; + tensor attn10_in_proj_weight_to_fp16 = const()[name = tensor("attn10_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(251768512)))]; + tensor linear_40_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn10_in_proj_weight_to_fp16, x = input_123_cast_fp16)[name = tensor("linear_40_cast_fp16")]; + tensor var_2640 = const()[name = tensor("op_2640"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_21_cast_fp16 = reshape(shape = var_2640, x = linear_40_cast_fp16)[name = tensor("qkv_21_cast_fp16")]; + tensor x_41_begin_0 = const()[name = tensor("x_41_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_41_end_0 = const()[name = tensor("x_41_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_41_end_mask_0 = const()[name = tensor("x_41_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_41_squeeze_mask_0 = const()[name = tensor("x_41_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_41_cast_fp16 = slice_by_index(begin = x_41_begin_0, end = x_41_end_0, end_mask = x_41_end_mask_0, squeeze_mask = x_41_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("x_41_cast_fp16")]; + tensor x_43_begin_0 = const()[name = tensor("x_43_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_43_end_0 = const()[name = tensor("x_43_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_43_end_mask_0 = const()[name = tensor("x_43_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_43_squeeze_mask_0 = const()[name = tensor("x_43_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_43_cast_fp16 = slice_by_index(begin = x_43_begin_0, end = x_43_end_0, end_mask = x_43_end_mask_0, squeeze_mask = x_43_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("x_43_cast_fp16")]; + tensor v_21_begin_0 = const()[name = tensor("v_21_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_21_end_0 = const()[name = tensor("v_21_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_21_end_mask_0 = const()[name = tensor("v_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_21_squeeze_mask_0 = const()[name = tensor("v_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_21_cast_fp16 = slice_by_index(begin = v_21_begin_0, end = v_21_end_0, end_mask = v_21_end_mask_0, squeeze_mask = v_21_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("v_21_cast_fp16")]; + tensor var_2685 = const()[name = tensor("op_2685"), val = tensor([256, 16, 32, 2])]; + tensor pairs_41_cast_fp16 = reshape(shape = var_2685, x = x_41_cast_fp16)[name = tensor("pairs_41_cast_fp16")]; + tensor xr_41_begin_0 = const()[name = tensor("xr_41_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_41_end_0 = const()[name = tensor("xr_41_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_41_end_mask_0 = const()[name = tensor("xr_41_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_41_squeeze_mask_0 = const()[name = tensor("xr_41_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_41_cast_fp16 = slice_by_index(begin = xr_41_begin_0, end = xr_41_end_0, end_mask = xr_41_end_mask_0, squeeze_mask = xr_41_squeeze_mask_0, x = pairs_41_cast_fp16)[name = tensor("xr_41_cast_fp16")]; + tensor xi_41_begin_0 = const()[name = tensor("xi_41_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_41_end_0 = const()[name = tensor("xi_41_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_41_end_mask_0 = const()[name = tensor("xi_41_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_41_squeeze_mask_0 = const()[name = tensor("xi_41_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_41_cast_fp16 = slice_by_index(begin = xi_41_begin_0, end = xi_41_end_0, end_mask = xi_41_end_mask_0, squeeze_mask = xi_41_squeeze_mask_0, x = pairs_41_cast_fp16)[name = tensor("xi_41_cast_fp16")]; + tensor var_2693_cast_fp16 = mul(x = xr_41_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2693_cast_fp16")]; + tensor var_2694_cast_fp16 = mul(x = xi_41_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2694_cast_fp16")]; + tensor xor__41_cast_fp16 = sub(x = var_2693_cast_fp16, y = var_2694_cast_fp16)[name = tensor("xor__41_cast_fp16")]; + tensor var_2697_cast_fp16 = mul(x = xr_41_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2697_cast_fp16")]; + tensor var_2698_cast_fp16 = mul(x = xi_41_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2698_cast_fp16")]; + tensor xoi_41_cast_fp16 = add(x = var_2697_cast_fp16, y = var_2698_cast_fp16)[name = tensor("xoi_41_cast_fp16")]; + tensor out_41_axis_0 = const()[name = tensor("out_41_axis_0"), val = tensor(-1)]; + tensor out_41_cast_fp16 = stack(axis = out_41_axis_0, values = (xor__41_cast_fp16, xoi_41_cast_fp16))[name = tensor("out_41_cast_fp16")]; + tensor var_2708 = const()[name = tensor("op_2708"), val = tensor([1, 256, 16, 64])]; + tensor q_21_cast_fp16 = reshape(shape = var_2708, x = out_41_cast_fp16)[name = tensor("q_21_cast_fp16")]; + tensor var_2714 = const()[name = tensor("op_2714"), val = tensor([256, 16, 32, 2])]; + tensor pairs_43_cast_fp16 = reshape(shape = var_2714, x = x_43_cast_fp16)[name = tensor("pairs_43_cast_fp16")]; + tensor xr_43_begin_0 = const()[name = tensor("xr_43_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_43_end_0 = const()[name = tensor("xr_43_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_43_end_mask_0 = const()[name = tensor("xr_43_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_43_squeeze_mask_0 = const()[name = tensor("xr_43_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_43_cast_fp16 = slice_by_index(begin = xr_43_begin_0, end = xr_43_end_0, end_mask = xr_43_end_mask_0, squeeze_mask = xr_43_squeeze_mask_0, x = pairs_43_cast_fp16)[name = tensor("xr_43_cast_fp16")]; + tensor xi_43_begin_0 = const()[name = tensor("xi_43_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_43_end_0 = const()[name = tensor("xi_43_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_43_end_mask_0 = const()[name = tensor("xi_43_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_43_squeeze_mask_0 = const()[name = tensor("xi_43_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_43_cast_fp16 = slice_by_index(begin = xi_43_begin_0, end = xi_43_end_0, end_mask = xi_43_end_mask_0, squeeze_mask = xi_43_squeeze_mask_0, x = pairs_43_cast_fp16)[name = tensor("xi_43_cast_fp16")]; + tensor var_2722_cast_fp16 = mul(x = xr_43_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2722_cast_fp16")]; + tensor var_2723_cast_fp16 = mul(x = xi_43_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2723_cast_fp16")]; + tensor xor__43_cast_fp16 = sub(x = var_2722_cast_fp16, y = var_2723_cast_fp16)[name = tensor("xor__43_cast_fp16")]; + tensor var_2726_cast_fp16 = mul(x = xr_43_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2726_cast_fp16")]; + tensor var_2727_cast_fp16 = mul(x = xi_43_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2727_cast_fp16")]; + tensor xoi_43_cast_fp16 = add(x = var_2726_cast_fp16, y = var_2727_cast_fp16)[name = tensor("xoi_43_cast_fp16")]; + tensor out_43_axis_0 = const()[name = tensor("out_43_axis_0"), val = tensor(-1)]; + tensor out_43_cast_fp16 = stack(axis = out_43_axis_0, values = (xor__43_cast_fp16, xoi_43_cast_fp16))[name = tensor("out_43_cast_fp16")]; + tensor var_2744 = const()[name = tensor("op_2744"), val = tensor([256, 1024])]; + tensor var_2745_cast_fp16 = reshape(shape = var_2744, x = out_43_cast_fp16)[name = tensor("op_2745_cast_fp16")]; + tensor written_k_21_transpose_x_1 = const()[name = tensor("written_k_21_transpose_x_1"), val = tensor(true)]; + tensor written_k_21_transpose_y_1 = const()[name = tensor("written_k_21_transpose_y_1"), val = tensor(false)]; + tensor written_k_21_cast_fp16 = matmul(transpose_x = written_k_21_transpose_x_1, transpose_y = written_k_21_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_2745_cast_fp16)[name = tensor("written_k_21_cast_fp16")]; + tensor var_2752 = const()[name = tensor("op_2752"), val = tensor([256, 1024])]; + tensor var_2753_cast_fp16 = reshape(shape = var_2752, x = v_21_cast_fp16)[name = tensor("op_2753_cast_fp16")]; + tensor written_v_21_transpose_x_1 = const()[name = tensor("written_v_21_transpose_x_1"), val = tensor(true)]; + tensor written_v_21_transpose_y_1 = const()[name = tensor("written_v_21_transpose_y_1"), val = tensor(false)]; + tensor written_v_21_cast_fp16 = matmul(transpose_x = written_v_21_transpose_x_1, transpose_y = written_v_21_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_2753_cast_fp16)[name = tensor("written_v_21_cast_fp16")]; + tensor k_cache10_to_fp16_dtype_0 = const()[name = tensor("k_cache10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache10_to_fp16 = cast(dtype = k_cache10_to_fp16_dtype_0, x = k_cache10)[name = tensor("cast_405")]; + tensor var_2764_cast_fp16 = mul(x = k_cache10_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_2764_cast_fp16")]; + tensor var_2769 = const()[name = tensor("op_2769"), val = tensor([1, 512, 16, 64])]; + tensor var_2770_cast_fp16 = reshape(shape = var_2769, x = written_k_21_cast_fp16)[name = tensor("op_2770_cast_fp16")]; + tensor new_k_cache_21_cast_fp16 = add(x = var_2764_cast_fp16, y = var_2770_cast_fp16)[name = tensor("new_k_cache_21_cast_fp16")]; + tensor new_k_cache_21_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_21_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache10_to_fp16_dtype_0 = const()[name = tensor("v_cache10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache10_to_fp16 = cast(dtype = v_cache10_to_fp16_dtype_0, x = v_cache10)[name = tensor("cast_403")]; + tensor var_2773_cast_fp16 = mul(x = v_cache10_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_2773_cast_fp16")]; + tensor var_2778 = const()[name = tensor("op_2778"), val = tensor([1, 512, 16, 64])]; + tensor var_2779_cast_fp16 = reshape(shape = var_2778, x = written_v_21_cast_fp16)[name = tensor("op_2779_cast_fp16")]; + tensor new_v_cache_21_cast_fp16 = add(x = var_2773_cast_fp16, y = var_2779_cast_fp16)[name = tensor("new_v_cache_21_cast_fp16")]; + tensor new_v_cache_21_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_21_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_2792 = const()[name = tensor("op_2792"), val = tensor([0, 2, 1, 3])]; + tensor var_2803_transpose_x_0 = const()[name = tensor("op_2803_transpose_x_0"), val = tensor(false)]; + tensor var_2803_transpose_y_0 = const()[name = tensor("op_2803_transpose_y_0"), val = tensor(false)]; + tensor transpose_89_perm_0 = const()[name = tensor("transpose_89_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_90_perm_0 = const()[name = tensor("transpose_90_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_90 = transpose(perm = transpose_90_perm_0, x = new_k_cache_21_cast_fp16)[name = tensor("transpose_164")]; + tensor transpose_89 = transpose(perm = transpose_89_perm_0, x = q_21_cast_fp16)[name = tensor("transpose_165")]; + tensor var_2803_cast_fp16 = matmul(transpose_x = var_2803_transpose_x_0, transpose_y = var_2803_transpose_y_0, x = transpose_89, y = transpose_90)[name = tensor("op_2803_cast_fp16")]; + tensor var_2804_to_fp16 = const()[name = tensor("op_2804_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_41_cast_fp16 = mul(x = var_2803_cast_fp16, y = var_2804_to_fp16)[name = tensor("scores_41_cast_fp16")]; + tensor scores_43_cast_fp16 = add(x = scores_41_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_43_cast_fp16")]; + tensor var_2814 = const()[name = tensor("op_2814"), val = tensor(-1)]; + tensor weights_21_cast_fp16 = softmax(axis = var_2814, x = scores_43_cast_fp16)[name = tensor("weights_21_cast_fp16")]; + tensor attn_21_transpose_x_0 = const()[name = tensor("attn_21_transpose_x_0"), val = tensor(false)]; + tensor attn_21_transpose_y_0 = const()[name = tensor("attn_21_transpose_y_0"), val = tensor(false)]; + tensor values_21_cast_fp16 = transpose(perm = var_2792, x = new_v_cache_21_cast_fp16)[name = tensor("transpose_166")]; + tensor attn_21_cast_fp16 = matmul(transpose_x = attn_21_transpose_x_0, transpose_y = attn_21_transpose_y_0, x = weights_21_cast_fp16, y = values_21_cast_fp16)[name = tensor("attn_21_cast_fp16")]; + tensor var_2822 = const()[name = tensor("op_2822"), val = tensor([0, 2, 1, 3])]; + tensor var_2827 = const()[name = tensor("op_2827"), val = tensor([1, 256, 1024])]; + tensor var_2823_cast_fp16 = transpose(perm = var_2822, x = attn_21_cast_fp16)[name = tensor("transpose_163")]; + tensor input_125_cast_fp16 = reshape(shape = var_2827, x = var_2823_cast_fp16)[name = tensor("input_125_cast_fp16")]; + tensor attn10_out_proj_weight_to_fp16 = const()[name = tensor("attn10_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(258060032)))]; + tensor linear_41_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn10_out_proj_weight_to_fp16, x = input_125_cast_fp16)[name = tensor("linear_41_cast_fp16")]; + tensor position10_to_fp16_dtype_0 = const()[name = tensor("position10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position10_to_fp16 = cast(dtype = position10_to_fp16_dtype_0, x = position10)[name = tensor("cast_401")]; + tensor var_2836_cast_fp16 = add(x = position10_to_fp16, y = valid_len_to_fp16)[name = tensor("op_2836_cast_fp16")]; + tensor var_2836_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2836_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_127_cast_fp16 = add(x = input_121_cast_fp16, y = linear_41_cast_fp16)[name = tensor("input_127_cast_fp16")]; + tensor input_129_axes_0 = const()[name = tensor("input_129_axes_0"), val = tensor([-1])]; + tensor norm10_2_weight_to_fp16 = const()[name = tensor("norm10_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260157248)))]; + tensor norm10_2_bias_to_fp16 = const()[name = tensor("norm10_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260159360)))]; + tensor var_2840_to_fp16 = const()[name = tensor("op_2840_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_129_cast_fp16 = layer_norm(axes = input_129_axes_0, beta = norm10_2_bias_to_fp16, epsilon = var_2840_to_fp16, gamma = norm10_2_weight_to_fp16, x = input_127_cast_fp16)[name = tensor("input_129_cast_fp16")]; + tensor linear10_1_weight_to_fp16 = const()[name = tensor("linear10_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260161472)))]; + tensor linear_42_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear10_1_weight_to_fp16, x = input_129_cast_fp16)[name = tensor("linear_42_cast_fp16")]; + tensor input_131_mode_0 = const()[name = tensor("input_131_mode_0"), val = tensor("EXACT")]; + tensor input_131_cast_fp16 = gelu(mode = input_131_mode_0, x = linear_42_cast_fp16)[name = tensor("input_131_cast_fp16")]; + tensor linear10_2_weight_to_fp16 = const()[name = tensor("linear10_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(268550144)))]; + tensor linear_43_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear10_2_weight_to_fp16, x = input_131_cast_fp16)[name = tensor("linear_43_cast_fp16")]; + tensor input_133_cast_fp16 = add(x = input_127_cast_fp16, y = linear_43_cast_fp16)[name = tensor("input_133_cast_fp16")]; + tensor input_135_axes_0 = const()[name = tensor("input_135_axes_0"), val = tensor([-1])]; + tensor norm11_1_weight_to_fp16 = const()[name = tensor("norm11_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(276938816)))]; + tensor norm11_1_bias_to_fp16 = const()[name = tensor("norm11_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(276940928)))]; + tensor var_2857_to_fp16 = const()[name = tensor("op_2857_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_135_cast_fp16 = layer_norm(axes = input_135_axes_0, beta = norm11_1_bias_to_fp16, epsilon = var_2857_to_fp16, gamma = norm11_1_weight_to_fp16, x = input_133_cast_fp16)[name = tensor("input_135_cast_fp16")]; + tensor attn11_in_proj_weight_to_fp16 = const()[name = tensor("attn11_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(276943040)))]; + tensor linear_44_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn11_in_proj_weight_to_fp16, x = input_135_cast_fp16)[name = tensor("linear_44_cast_fp16")]; + tensor var_2871 = const()[name = tensor("op_2871"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_23_cast_fp16 = reshape(shape = var_2871, x = linear_44_cast_fp16)[name = tensor("qkv_23_cast_fp16")]; + tensor x_45_begin_0 = const()[name = tensor("x_45_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_45_end_0 = const()[name = tensor("x_45_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_45_end_mask_0 = const()[name = tensor("x_45_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_45_squeeze_mask_0 = const()[name = tensor("x_45_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_45_cast_fp16 = slice_by_index(begin = x_45_begin_0, end = x_45_end_0, end_mask = x_45_end_mask_0, squeeze_mask = x_45_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("x_45_cast_fp16")]; + tensor x_47_begin_0 = const()[name = tensor("x_47_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_47_end_0 = const()[name = tensor("x_47_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_47_end_mask_0 = const()[name = tensor("x_47_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_47_squeeze_mask_0 = const()[name = tensor("x_47_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_47_cast_fp16 = slice_by_index(begin = x_47_begin_0, end = x_47_end_0, end_mask = x_47_end_mask_0, squeeze_mask = x_47_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("x_47_cast_fp16")]; + tensor v_23_begin_0 = const()[name = tensor("v_23_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_23_end_0 = const()[name = tensor("v_23_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_23_end_mask_0 = const()[name = tensor("v_23_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_23_squeeze_mask_0 = const()[name = tensor("v_23_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_23_cast_fp16 = slice_by_index(begin = v_23_begin_0, end = v_23_end_0, end_mask = v_23_end_mask_0, squeeze_mask = v_23_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("v_23_cast_fp16")]; + tensor var_2916 = const()[name = tensor("op_2916"), val = tensor([256, 16, 32, 2])]; + tensor pairs_45_cast_fp16 = reshape(shape = var_2916, x = x_45_cast_fp16)[name = tensor("pairs_45_cast_fp16")]; + tensor xr_45_begin_0 = const()[name = tensor("xr_45_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_45_end_0 = const()[name = tensor("xr_45_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_45_end_mask_0 = const()[name = tensor("xr_45_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_45_squeeze_mask_0 = const()[name = tensor("xr_45_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_45_cast_fp16 = slice_by_index(begin = xr_45_begin_0, end = xr_45_end_0, end_mask = xr_45_end_mask_0, squeeze_mask = xr_45_squeeze_mask_0, x = pairs_45_cast_fp16)[name = tensor("xr_45_cast_fp16")]; + tensor xi_45_begin_0 = const()[name = tensor("xi_45_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_45_end_0 = const()[name = tensor("xi_45_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_45_end_mask_0 = const()[name = tensor("xi_45_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_45_squeeze_mask_0 = const()[name = tensor("xi_45_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_45_cast_fp16 = slice_by_index(begin = xi_45_begin_0, end = xi_45_end_0, end_mask = xi_45_end_mask_0, squeeze_mask = xi_45_squeeze_mask_0, x = pairs_45_cast_fp16)[name = tensor("xi_45_cast_fp16")]; + tensor var_2924_cast_fp16 = mul(x = xr_45_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2924_cast_fp16")]; + tensor var_2925_cast_fp16 = mul(x = xi_45_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2925_cast_fp16")]; + tensor xor__45_cast_fp16 = sub(x = var_2924_cast_fp16, y = var_2925_cast_fp16)[name = tensor("xor__45_cast_fp16")]; + tensor var_2928_cast_fp16 = mul(x = xr_45_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2928_cast_fp16")]; + tensor var_2929_cast_fp16 = mul(x = xi_45_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2929_cast_fp16")]; + tensor xoi_45_cast_fp16 = add(x = var_2928_cast_fp16, y = var_2929_cast_fp16)[name = tensor("xoi_45_cast_fp16")]; + tensor out_45_axis_0 = const()[name = tensor("out_45_axis_0"), val = tensor(-1)]; + tensor out_45_cast_fp16 = stack(axis = out_45_axis_0, values = (xor__45_cast_fp16, xoi_45_cast_fp16))[name = tensor("out_45_cast_fp16")]; + tensor var_2939 = const()[name = tensor("op_2939"), val = tensor([1, 256, 16, 64])]; + tensor q_23_cast_fp16 = reshape(shape = var_2939, x = out_45_cast_fp16)[name = tensor("q_23_cast_fp16")]; + tensor var_2945 = const()[name = tensor("op_2945"), val = tensor([256, 16, 32, 2])]; + tensor pairs_47_cast_fp16 = reshape(shape = var_2945, x = x_47_cast_fp16)[name = tensor("pairs_47_cast_fp16")]; + tensor xr_47_begin_0 = const()[name = tensor("xr_47_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_47_end_0 = const()[name = tensor("xr_47_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_47_end_mask_0 = const()[name = tensor("xr_47_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_47_squeeze_mask_0 = const()[name = tensor("xr_47_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_47_cast_fp16 = slice_by_index(begin = xr_47_begin_0, end = xr_47_end_0, end_mask = xr_47_end_mask_0, squeeze_mask = xr_47_squeeze_mask_0, x = pairs_47_cast_fp16)[name = tensor("xr_47_cast_fp16")]; + tensor xi_47_begin_0 = const()[name = tensor("xi_47_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_47_end_0 = const()[name = tensor("xi_47_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_47_end_mask_0 = const()[name = tensor("xi_47_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_47_squeeze_mask_0 = const()[name = tensor("xi_47_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_47_cast_fp16 = slice_by_index(begin = xi_47_begin_0, end = xi_47_end_0, end_mask = xi_47_end_mask_0, squeeze_mask = xi_47_squeeze_mask_0, x = pairs_47_cast_fp16)[name = tensor("xi_47_cast_fp16")]; + tensor var_2953_cast_fp16 = mul(x = xr_47_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2953_cast_fp16")]; + tensor var_2954_cast_fp16 = mul(x = xi_47_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2954_cast_fp16")]; + tensor xor__47_cast_fp16 = sub(x = var_2953_cast_fp16, y = var_2954_cast_fp16)[name = tensor("xor__47_cast_fp16")]; + tensor var_2957_cast_fp16 = mul(x = xr_47_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2957_cast_fp16")]; + tensor var_2958_cast_fp16 = mul(x = xi_47_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2958_cast_fp16")]; + tensor xoi_47_cast_fp16 = add(x = var_2957_cast_fp16, y = var_2958_cast_fp16)[name = tensor("xoi_47_cast_fp16")]; + tensor out_47_axis_0 = const()[name = tensor("out_47_axis_0"), val = tensor(-1)]; + tensor out_47_cast_fp16 = stack(axis = out_47_axis_0, values = (xor__47_cast_fp16, xoi_47_cast_fp16))[name = tensor("out_47_cast_fp16")]; + tensor var_2975 = const()[name = tensor("op_2975"), val = tensor([256, 1024])]; + tensor var_2976_cast_fp16 = reshape(shape = var_2975, x = out_47_cast_fp16)[name = tensor("op_2976_cast_fp16")]; + tensor written_k_23_transpose_x_1 = const()[name = tensor("written_k_23_transpose_x_1"), val = tensor(true)]; + tensor written_k_23_transpose_y_1 = const()[name = tensor("written_k_23_transpose_y_1"), val = tensor(false)]; + tensor written_k_23_cast_fp16 = matmul(transpose_x = written_k_23_transpose_x_1, transpose_y = written_k_23_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_2976_cast_fp16)[name = tensor("written_k_23_cast_fp16")]; + tensor var_2983 = const()[name = tensor("op_2983"), val = tensor([256, 1024])]; + tensor var_2984_cast_fp16 = reshape(shape = var_2983, x = v_23_cast_fp16)[name = tensor("op_2984_cast_fp16")]; + tensor written_v_23_transpose_x_1 = const()[name = tensor("written_v_23_transpose_x_1"), val = tensor(true)]; + tensor written_v_23_transpose_y_1 = const()[name = tensor("written_v_23_transpose_y_1"), val = tensor(false)]; + tensor written_v_23_cast_fp16 = matmul(transpose_x = written_v_23_transpose_x_1, transpose_y = written_v_23_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_2984_cast_fp16)[name = tensor("written_v_23_cast_fp16")]; + tensor k_cache11_to_fp16_dtype_0 = const()[name = tensor("k_cache11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache11_to_fp16 = cast(dtype = k_cache11_to_fp16_dtype_0, x = k_cache11)[name = tensor("cast_399")]; + tensor var_2995_cast_fp16 = mul(x = k_cache11_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_2995_cast_fp16")]; + tensor var_3000 = const()[name = tensor("op_3000"), val = tensor([1, 512, 16, 64])]; + tensor var_3001_cast_fp16 = reshape(shape = var_3000, x = written_k_23_cast_fp16)[name = tensor("op_3001_cast_fp16")]; + tensor new_k_cache_23_cast_fp16 = add(x = var_2995_cast_fp16, y = var_3001_cast_fp16)[name = tensor("new_k_cache_23_cast_fp16")]; + tensor new_k_cache_23_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_23_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache11_to_fp16_dtype_0 = const()[name = tensor("v_cache11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache11_to_fp16 = cast(dtype = v_cache11_to_fp16_dtype_0, x = v_cache11)[name = tensor("cast_397")]; + tensor var_3004_cast_fp16 = mul(x = v_cache11_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_3004_cast_fp16")]; + tensor var_3009 = const()[name = tensor("op_3009"), val = tensor([1, 512, 16, 64])]; + tensor var_3010_cast_fp16 = reshape(shape = var_3009, x = written_v_23_cast_fp16)[name = tensor("op_3010_cast_fp16")]; + tensor new_v_cache_23_cast_fp16 = add(x = var_3004_cast_fp16, y = var_3010_cast_fp16)[name = tensor("new_v_cache_23_cast_fp16")]; + tensor new_v_cache_23_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_23_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_3023 = const()[name = tensor("op_3023"), val = tensor([0, 2, 1, 3])]; + tensor var_3034_transpose_x_0 = const()[name = tensor("op_3034_transpose_x_0"), val = tensor(false)]; + tensor var_3034_transpose_y_0 = const()[name = tensor("op_3034_transpose_y_0"), val = tensor(false)]; + tensor transpose_91_perm_0 = const()[name = tensor("transpose_91_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_92_perm_0 = const()[name = tensor("transpose_92_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_92 = transpose(perm = transpose_92_perm_0, x = new_k_cache_23_cast_fp16)[name = tensor("transpose_160")]; + tensor transpose_91 = transpose(perm = transpose_91_perm_0, x = q_23_cast_fp16)[name = tensor("transpose_161")]; + tensor var_3034_cast_fp16 = matmul(transpose_x = var_3034_transpose_x_0, transpose_y = var_3034_transpose_y_0, x = transpose_91, y = transpose_92)[name = tensor("op_3034_cast_fp16")]; + tensor var_3035_to_fp16 = const()[name = tensor("op_3035_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_45_cast_fp16 = mul(x = var_3034_cast_fp16, y = var_3035_to_fp16)[name = tensor("scores_45_cast_fp16")]; + tensor scores_47_cast_fp16 = add(x = scores_45_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_47_cast_fp16")]; + tensor var_3045 = const()[name = tensor("op_3045"), val = tensor(-1)]; + tensor weights_23_cast_fp16 = softmax(axis = var_3045, x = scores_47_cast_fp16)[name = tensor("weights_23_cast_fp16")]; + tensor attn_23_transpose_x_0 = const()[name = tensor("attn_23_transpose_x_0"), val = tensor(false)]; + tensor attn_23_transpose_y_0 = const()[name = tensor("attn_23_transpose_y_0"), val = tensor(false)]; + tensor values_23_cast_fp16 = transpose(perm = var_3023, x = new_v_cache_23_cast_fp16)[name = tensor("transpose_162")]; + tensor attn_23_cast_fp16 = matmul(transpose_x = attn_23_transpose_x_0, transpose_y = attn_23_transpose_y_0, x = weights_23_cast_fp16, y = values_23_cast_fp16)[name = tensor("attn_23_cast_fp16")]; + tensor var_3053 = const()[name = tensor("op_3053"), val = tensor([0, 2, 1, 3])]; + tensor var_3058 = const()[name = tensor("op_3058"), val = tensor([1, 256, 1024])]; + tensor var_3054_cast_fp16 = transpose(perm = var_3053, x = attn_23_cast_fp16)[name = tensor("transpose_159")]; + tensor input_137_cast_fp16 = reshape(shape = var_3058, x = var_3054_cast_fp16)[name = tensor("input_137_cast_fp16")]; + tensor attn11_out_proj_weight_to_fp16 = const()[name = tensor("attn11_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(283234560)))]; + tensor linear_45_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn11_out_proj_weight_to_fp16, x = input_137_cast_fp16)[name = tensor("linear_45_cast_fp16")]; + tensor position11_to_fp16_dtype_0 = const()[name = tensor("position11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position11_to_fp16 = cast(dtype = position11_to_fp16_dtype_0, x = position11)[name = tensor("cast_395")]; + tensor var_3067_cast_fp16 = add(x = position11_to_fp16, y = valid_len_to_fp16)[name = tensor("op_3067_cast_fp16")]; + tensor var_3067_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3067_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_139_cast_fp16 = add(x = input_133_cast_fp16, y = linear_45_cast_fp16)[name = tensor("input_139_cast_fp16")]; + tensor input_141_axes_0 = const()[name = tensor("input_141_axes_0"), val = tensor([-1])]; + tensor norm11_2_weight_to_fp16 = const()[name = tensor("norm11_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285331776)))]; + tensor norm11_2_bias_to_fp16 = const()[name = tensor("norm11_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285333888)))]; + tensor var_3071_to_fp16 = const()[name = tensor("op_3071_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_141_cast_fp16 = layer_norm(axes = input_141_axes_0, beta = norm11_2_bias_to_fp16, epsilon = var_3071_to_fp16, gamma = norm11_2_weight_to_fp16, x = input_139_cast_fp16)[name = tensor("input_141_cast_fp16")]; + tensor linear11_1_weight_to_fp16 = const()[name = tensor("linear11_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285336000)))]; + tensor linear_46_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear11_1_weight_to_fp16, x = input_141_cast_fp16)[name = tensor("linear_46_cast_fp16")]; + tensor input_143_mode_0 = const()[name = tensor("input_143_mode_0"), val = tensor("EXACT")]; + tensor input_143_cast_fp16 = gelu(mode = input_143_mode_0, x = linear_46_cast_fp16)[name = tensor("input_143_cast_fp16")]; + tensor linear11_2_weight_to_fp16 = const()[name = tensor("linear11_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293724672)))]; + tensor linear_47_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear11_2_weight_to_fp16, x = input_143_cast_fp16)[name = tensor("linear_47_cast_fp16")]; + tensor input_145_cast_fp16 = add(x = input_139_cast_fp16, y = linear_47_cast_fp16)[name = tensor("input_145_cast_fp16")]; + tensor input_147_axes_0 = const()[name = tensor("input_147_axes_0"), val = tensor([-1])]; + tensor norm12_1_weight_to_fp16 = const()[name = tensor("norm12_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302113344)))]; + tensor norm12_1_bias_to_fp16 = const()[name = tensor("norm12_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302115456)))]; + tensor var_3088_to_fp16 = const()[name = tensor("op_3088_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_147_cast_fp16 = layer_norm(axes = input_147_axes_0, beta = norm12_1_bias_to_fp16, epsilon = var_3088_to_fp16, gamma = norm12_1_weight_to_fp16, x = input_145_cast_fp16)[name = tensor("input_147_cast_fp16")]; + tensor attn12_in_proj_weight_to_fp16 = const()[name = tensor("attn12_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302117568)))]; + tensor linear_48_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn12_in_proj_weight_to_fp16, x = input_147_cast_fp16)[name = tensor("linear_48_cast_fp16")]; + tensor var_3102 = const()[name = tensor("op_3102"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_25_cast_fp16 = reshape(shape = var_3102, x = linear_48_cast_fp16)[name = tensor("qkv_25_cast_fp16")]; + tensor x_49_begin_0 = const()[name = tensor("x_49_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_49_end_0 = const()[name = tensor("x_49_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_49_end_mask_0 = const()[name = tensor("x_49_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_49_squeeze_mask_0 = const()[name = tensor("x_49_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_49_cast_fp16 = slice_by_index(begin = x_49_begin_0, end = x_49_end_0, end_mask = x_49_end_mask_0, squeeze_mask = x_49_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("x_49_cast_fp16")]; + tensor x_51_begin_0 = const()[name = tensor("x_51_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_51_end_0 = const()[name = tensor("x_51_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_51_end_mask_0 = const()[name = tensor("x_51_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_51_squeeze_mask_0 = const()[name = tensor("x_51_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_51_cast_fp16 = slice_by_index(begin = x_51_begin_0, end = x_51_end_0, end_mask = x_51_end_mask_0, squeeze_mask = x_51_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("x_51_cast_fp16")]; + tensor v_25_begin_0 = const()[name = tensor("v_25_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_25_end_0 = const()[name = tensor("v_25_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_25_end_mask_0 = const()[name = tensor("v_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_25_squeeze_mask_0 = const()[name = tensor("v_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_25_cast_fp16 = slice_by_index(begin = v_25_begin_0, end = v_25_end_0, end_mask = v_25_end_mask_0, squeeze_mask = v_25_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("v_25_cast_fp16")]; + tensor var_3147 = const()[name = tensor("op_3147"), val = tensor([256, 16, 32, 2])]; + tensor pairs_49_cast_fp16 = reshape(shape = var_3147, x = x_49_cast_fp16)[name = tensor("pairs_49_cast_fp16")]; + tensor xr_49_begin_0 = const()[name = tensor("xr_49_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_49_end_0 = const()[name = tensor("xr_49_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_49_end_mask_0 = const()[name = tensor("xr_49_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_49_squeeze_mask_0 = const()[name = tensor("xr_49_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_49_cast_fp16 = slice_by_index(begin = xr_49_begin_0, end = xr_49_end_0, end_mask = xr_49_end_mask_0, squeeze_mask = xr_49_squeeze_mask_0, x = pairs_49_cast_fp16)[name = tensor("xr_49_cast_fp16")]; + tensor xi_49_begin_0 = const()[name = tensor("xi_49_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_49_end_0 = const()[name = tensor("xi_49_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_49_end_mask_0 = const()[name = tensor("xi_49_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_49_squeeze_mask_0 = const()[name = tensor("xi_49_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_49_cast_fp16 = slice_by_index(begin = xi_49_begin_0, end = xi_49_end_0, end_mask = xi_49_end_mask_0, squeeze_mask = xi_49_squeeze_mask_0, x = pairs_49_cast_fp16)[name = tensor("xi_49_cast_fp16")]; + tensor var_3155_cast_fp16 = mul(x = xr_49_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3155_cast_fp16")]; + tensor var_3156_cast_fp16 = mul(x = xi_49_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3156_cast_fp16")]; + tensor xor__49_cast_fp16 = sub(x = var_3155_cast_fp16, y = var_3156_cast_fp16)[name = tensor("xor__49_cast_fp16")]; + tensor var_3159_cast_fp16 = mul(x = xr_49_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3159_cast_fp16")]; + tensor var_3160_cast_fp16 = mul(x = xi_49_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3160_cast_fp16")]; + tensor xoi_49_cast_fp16 = add(x = var_3159_cast_fp16, y = var_3160_cast_fp16)[name = tensor("xoi_49_cast_fp16")]; + tensor out_49_axis_0 = const()[name = tensor("out_49_axis_0"), val = tensor(-1)]; + tensor out_49_cast_fp16 = stack(axis = out_49_axis_0, values = (xor__49_cast_fp16, xoi_49_cast_fp16))[name = tensor("out_49_cast_fp16")]; + tensor var_3170 = const()[name = tensor("op_3170"), val = tensor([1, 256, 16, 64])]; + tensor q_25_cast_fp16 = reshape(shape = var_3170, x = out_49_cast_fp16)[name = tensor("q_25_cast_fp16")]; + tensor var_3176 = const()[name = tensor("op_3176"), val = tensor([256, 16, 32, 2])]; + tensor pairs_51_cast_fp16 = reshape(shape = var_3176, x = x_51_cast_fp16)[name = tensor("pairs_51_cast_fp16")]; + tensor xr_51_begin_0 = const()[name = tensor("xr_51_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_51_end_0 = const()[name = tensor("xr_51_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_51_end_mask_0 = const()[name = tensor("xr_51_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_51_squeeze_mask_0 = const()[name = tensor("xr_51_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_51_cast_fp16 = slice_by_index(begin = xr_51_begin_0, end = xr_51_end_0, end_mask = xr_51_end_mask_0, squeeze_mask = xr_51_squeeze_mask_0, x = pairs_51_cast_fp16)[name = tensor("xr_51_cast_fp16")]; + tensor xi_51_begin_0 = const()[name = tensor("xi_51_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_51_end_0 = const()[name = tensor("xi_51_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_51_end_mask_0 = const()[name = tensor("xi_51_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_51_squeeze_mask_0 = const()[name = tensor("xi_51_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_51_cast_fp16 = slice_by_index(begin = xi_51_begin_0, end = xi_51_end_0, end_mask = xi_51_end_mask_0, squeeze_mask = xi_51_squeeze_mask_0, x = pairs_51_cast_fp16)[name = tensor("xi_51_cast_fp16")]; + tensor var_3184_cast_fp16 = mul(x = xr_51_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3184_cast_fp16")]; + tensor var_3185_cast_fp16 = mul(x = xi_51_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3185_cast_fp16")]; + tensor xor__51_cast_fp16 = sub(x = var_3184_cast_fp16, y = var_3185_cast_fp16)[name = tensor("xor__51_cast_fp16")]; + tensor var_3188_cast_fp16 = mul(x = xr_51_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3188_cast_fp16")]; + tensor var_3189_cast_fp16 = mul(x = xi_51_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3189_cast_fp16")]; + tensor xoi_51_cast_fp16 = add(x = var_3188_cast_fp16, y = var_3189_cast_fp16)[name = tensor("xoi_51_cast_fp16")]; + tensor out_51_axis_0 = const()[name = tensor("out_51_axis_0"), val = tensor(-1)]; + tensor out_51_cast_fp16 = stack(axis = out_51_axis_0, values = (xor__51_cast_fp16, xoi_51_cast_fp16))[name = tensor("out_51_cast_fp16")]; + tensor var_3206 = const()[name = tensor("op_3206"), val = tensor([256, 1024])]; + tensor var_3207_cast_fp16 = reshape(shape = var_3206, x = out_51_cast_fp16)[name = tensor("op_3207_cast_fp16")]; + tensor written_k_25_transpose_x_1 = const()[name = tensor("written_k_25_transpose_x_1"), val = tensor(true)]; + tensor written_k_25_transpose_y_1 = const()[name = tensor("written_k_25_transpose_y_1"), val = tensor(false)]; + tensor written_k_25_cast_fp16 = matmul(transpose_x = written_k_25_transpose_x_1, transpose_y = written_k_25_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_3207_cast_fp16)[name = tensor("written_k_25_cast_fp16")]; + tensor var_3214 = const()[name = tensor("op_3214"), val = tensor([256, 1024])]; + tensor var_3215_cast_fp16 = reshape(shape = var_3214, x = v_25_cast_fp16)[name = tensor("op_3215_cast_fp16")]; + tensor written_v_25_transpose_x_1 = const()[name = tensor("written_v_25_transpose_x_1"), val = tensor(true)]; + tensor written_v_25_transpose_y_1 = const()[name = tensor("written_v_25_transpose_y_1"), val = tensor(false)]; + tensor written_v_25_cast_fp16 = matmul(transpose_x = written_v_25_transpose_x_1, transpose_y = written_v_25_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_3215_cast_fp16)[name = tensor("written_v_25_cast_fp16")]; + tensor k_cache12_to_fp16_dtype_0 = const()[name = tensor("k_cache12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache12_to_fp16 = cast(dtype = k_cache12_to_fp16_dtype_0, x = k_cache12)[name = tensor("cast_393")]; + tensor var_3226_cast_fp16 = mul(x = k_cache12_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_3226_cast_fp16")]; + tensor var_3231 = const()[name = tensor("op_3231"), val = tensor([1, 512, 16, 64])]; + tensor var_3232_cast_fp16 = reshape(shape = var_3231, x = written_k_25_cast_fp16)[name = tensor("op_3232_cast_fp16")]; + tensor new_k_cache_25_cast_fp16 = add(x = var_3226_cast_fp16, y = var_3232_cast_fp16)[name = tensor("new_k_cache_25_cast_fp16")]; + tensor new_k_cache_25_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_25_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache12_to_fp16_dtype_0 = const()[name = tensor("v_cache12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache12_to_fp16 = cast(dtype = v_cache12_to_fp16_dtype_0, x = v_cache12)[name = tensor("cast_391")]; + tensor var_3235_cast_fp16 = mul(x = v_cache12_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_3235_cast_fp16")]; + tensor var_3240 = const()[name = tensor("op_3240"), val = tensor([1, 512, 16, 64])]; + tensor var_3241_cast_fp16 = reshape(shape = var_3240, x = written_v_25_cast_fp16)[name = tensor("op_3241_cast_fp16")]; + tensor new_v_cache_25_cast_fp16 = add(x = var_3235_cast_fp16, y = var_3241_cast_fp16)[name = tensor("new_v_cache_25_cast_fp16")]; + tensor new_v_cache_25_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_25_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_3254 = const()[name = tensor("op_3254"), val = tensor([0, 2, 1, 3])]; + tensor var_3265_transpose_x_0 = const()[name = tensor("op_3265_transpose_x_0"), val = tensor(false)]; + tensor var_3265_transpose_y_0 = const()[name = tensor("op_3265_transpose_y_0"), val = tensor(false)]; + tensor transpose_93_perm_0 = const()[name = tensor("transpose_93_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_94_perm_0 = const()[name = tensor("transpose_94_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_94 = transpose(perm = transpose_94_perm_0, x = new_k_cache_25_cast_fp16)[name = tensor("transpose_156")]; + tensor transpose_93 = transpose(perm = transpose_93_perm_0, x = q_25_cast_fp16)[name = tensor("transpose_157")]; + tensor var_3265_cast_fp16 = matmul(transpose_x = var_3265_transpose_x_0, transpose_y = var_3265_transpose_y_0, x = transpose_93, y = transpose_94)[name = tensor("op_3265_cast_fp16")]; + tensor var_3266_to_fp16 = const()[name = tensor("op_3266_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_49_cast_fp16 = mul(x = var_3265_cast_fp16, y = var_3266_to_fp16)[name = tensor("scores_49_cast_fp16")]; + tensor scores_51_cast_fp16 = add(x = scores_49_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_51_cast_fp16")]; + tensor var_3276 = const()[name = tensor("op_3276"), val = tensor(-1)]; + tensor weights_25_cast_fp16 = softmax(axis = var_3276, x = scores_51_cast_fp16)[name = tensor("weights_25_cast_fp16")]; + tensor attn_25_transpose_x_0 = const()[name = tensor("attn_25_transpose_x_0"), val = tensor(false)]; + tensor attn_25_transpose_y_0 = const()[name = tensor("attn_25_transpose_y_0"), val = tensor(false)]; + tensor values_25_cast_fp16 = transpose(perm = var_3254, x = new_v_cache_25_cast_fp16)[name = tensor("transpose_158")]; + tensor attn_25_cast_fp16 = matmul(transpose_x = attn_25_transpose_x_0, transpose_y = attn_25_transpose_y_0, x = weights_25_cast_fp16, y = values_25_cast_fp16)[name = tensor("attn_25_cast_fp16")]; + tensor var_3284 = const()[name = tensor("op_3284"), val = tensor([0, 2, 1, 3])]; + tensor var_3289 = const()[name = tensor("op_3289"), val = tensor([1, 256, 1024])]; + tensor var_3285_cast_fp16 = transpose(perm = var_3284, x = attn_25_cast_fp16)[name = tensor("transpose_155")]; + tensor input_149_cast_fp16 = reshape(shape = var_3289, x = var_3285_cast_fp16)[name = tensor("input_149_cast_fp16")]; + tensor attn12_out_proj_weight_to_fp16 = const()[name = tensor("attn12_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(308409088)))]; + tensor linear_49_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn12_out_proj_weight_to_fp16, x = input_149_cast_fp16)[name = tensor("linear_49_cast_fp16")]; + tensor position12_to_fp16_dtype_0 = const()[name = tensor("position12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position12_to_fp16 = cast(dtype = position12_to_fp16_dtype_0, x = position12)[name = tensor("cast_389")]; + tensor var_3298_cast_fp16 = add(x = position12_to_fp16, y = valid_len_to_fp16)[name = tensor("op_3298_cast_fp16")]; + tensor var_3298_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3298_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_151_cast_fp16 = add(x = input_145_cast_fp16, y = linear_49_cast_fp16)[name = tensor("input_151_cast_fp16")]; + tensor input_153_axes_0 = const()[name = tensor("input_153_axes_0"), val = tensor([-1])]; + tensor norm12_2_weight_to_fp16 = const()[name = tensor("norm12_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(310506304)))]; + tensor norm12_2_bias_to_fp16 = const()[name = tensor("norm12_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(310508416)))]; + tensor var_3302_to_fp16 = const()[name = tensor("op_3302_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_153_cast_fp16 = layer_norm(axes = input_153_axes_0, beta = norm12_2_bias_to_fp16, epsilon = var_3302_to_fp16, gamma = norm12_2_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("input_153_cast_fp16")]; + tensor linear12_1_weight_to_fp16 = const()[name = tensor("linear12_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(310510528)))]; + tensor linear_50_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear12_1_weight_to_fp16, x = input_153_cast_fp16)[name = tensor("linear_50_cast_fp16")]; + tensor input_155_mode_0 = const()[name = tensor("input_155_mode_0"), val = tensor("EXACT")]; + tensor input_155_cast_fp16 = gelu(mode = input_155_mode_0, x = linear_50_cast_fp16)[name = tensor("input_155_cast_fp16")]; + tensor linear12_2_weight_to_fp16 = const()[name = tensor("linear12_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(318899200)))]; + tensor linear_51_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear12_2_weight_to_fp16, x = input_155_cast_fp16)[name = tensor("linear_51_cast_fp16")]; + tensor input_157_cast_fp16 = add(x = input_151_cast_fp16, y = linear_51_cast_fp16)[name = tensor("input_157_cast_fp16")]; + tensor input_159_axes_0 = const()[name = tensor("input_159_axes_0"), val = tensor([-1])]; + tensor norm13_1_weight_to_fp16 = const()[name = tensor("norm13_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(327287872)))]; + tensor norm13_1_bias_to_fp16 = const()[name = tensor("norm13_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(327289984)))]; + tensor var_3319_to_fp16 = const()[name = tensor("op_3319_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_159_cast_fp16 = layer_norm(axes = input_159_axes_0, beta = norm13_1_bias_to_fp16, epsilon = var_3319_to_fp16, gamma = norm13_1_weight_to_fp16, x = input_157_cast_fp16)[name = tensor("input_159_cast_fp16")]; + tensor attn13_in_proj_weight_to_fp16 = const()[name = tensor("attn13_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(327292096)))]; + tensor linear_52_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn13_in_proj_weight_to_fp16, x = input_159_cast_fp16)[name = tensor("linear_52_cast_fp16")]; + tensor var_3333 = const()[name = tensor("op_3333"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_27_cast_fp16 = reshape(shape = var_3333, x = linear_52_cast_fp16)[name = tensor("qkv_27_cast_fp16")]; + tensor x_53_begin_0 = const()[name = tensor("x_53_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_53_end_0 = const()[name = tensor("x_53_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_53_end_mask_0 = const()[name = tensor("x_53_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_53_squeeze_mask_0 = const()[name = tensor("x_53_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_53_cast_fp16 = slice_by_index(begin = x_53_begin_0, end = x_53_end_0, end_mask = x_53_end_mask_0, squeeze_mask = x_53_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("x_53_cast_fp16")]; + tensor x_55_begin_0 = const()[name = tensor("x_55_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_55_end_0 = const()[name = tensor("x_55_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_55_end_mask_0 = const()[name = tensor("x_55_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_55_squeeze_mask_0 = const()[name = tensor("x_55_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_55_cast_fp16 = slice_by_index(begin = x_55_begin_0, end = x_55_end_0, end_mask = x_55_end_mask_0, squeeze_mask = x_55_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("x_55_cast_fp16")]; + tensor v_27_begin_0 = const()[name = tensor("v_27_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_27_end_0 = const()[name = tensor("v_27_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_27_end_mask_0 = const()[name = tensor("v_27_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_27_squeeze_mask_0 = const()[name = tensor("v_27_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_27_cast_fp16 = slice_by_index(begin = v_27_begin_0, end = v_27_end_0, end_mask = v_27_end_mask_0, squeeze_mask = v_27_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("v_27_cast_fp16")]; + tensor var_3378 = const()[name = tensor("op_3378"), val = tensor([256, 16, 32, 2])]; + tensor pairs_53_cast_fp16 = reshape(shape = var_3378, x = x_53_cast_fp16)[name = tensor("pairs_53_cast_fp16")]; + tensor xr_53_begin_0 = const()[name = tensor("xr_53_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_53_end_0 = const()[name = tensor("xr_53_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_53_end_mask_0 = const()[name = tensor("xr_53_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_53_squeeze_mask_0 = const()[name = tensor("xr_53_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_53_cast_fp16 = slice_by_index(begin = xr_53_begin_0, end = xr_53_end_0, end_mask = xr_53_end_mask_0, squeeze_mask = xr_53_squeeze_mask_0, x = pairs_53_cast_fp16)[name = tensor("xr_53_cast_fp16")]; + tensor xi_53_begin_0 = const()[name = tensor("xi_53_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_53_end_0 = const()[name = tensor("xi_53_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_53_end_mask_0 = const()[name = tensor("xi_53_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_53_squeeze_mask_0 = const()[name = tensor("xi_53_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_53_cast_fp16 = slice_by_index(begin = xi_53_begin_0, end = xi_53_end_0, end_mask = xi_53_end_mask_0, squeeze_mask = xi_53_squeeze_mask_0, x = pairs_53_cast_fp16)[name = tensor("xi_53_cast_fp16")]; + tensor var_3386_cast_fp16 = mul(x = xr_53_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3386_cast_fp16")]; + tensor var_3387_cast_fp16 = mul(x = xi_53_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3387_cast_fp16")]; + tensor xor__53_cast_fp16 = sub(x = var_3386_cast_fp16, y = var_3387_cast_fp16)[name = tensor("xor__53_cast_fp16")]; + tensor var_3390_cast_fp16 = mul(x = xr_53_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3390_cast_fp16")]; + tensor var_3391_cast_fp16 = mul(x = xi_53_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3391_cast_fp16")]; + tensor xoi_53_cast_fp16 = add(x = var_3390_cast_fp16, y = var_3391_cast_fp16)[name = tensor("xoi_53_cast_fp16")]; + tensor out_53_axis_0 = const()[name = tensor("out_53_axis_0"), val = tensor(-1)]; + tensor out_53_cast_fp16 = stack(axis = out_53_axis_0, values = (xor__53_cast_fp16, xoi_53_cast_fp16))[name = tensor("out_53_cast_fp16")]; + tensor var_3401 = const()[name = tensor("op_3401"), val = tensor([1, 256, 16, 64])]; + tensor q_27_cast_fp16 = reshape(shape = var_3401, x = out_53_cast_fp16)[name = tensor("q_27_cast_fp16")]; + tensor var_3407 = const()[name = tensor("op_3407"), val = tensor([256, 16, 32, 2])]; + tensor pairs_55_cast_fp16 = reshape(shape = var_3407, x = x_55_cast_fp16)[name = tensor("pairs_55_cast_fp16")]; + tensor xr_55_begin_0 = const()[name = tensor("xr_55_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_55_end_0 = const()[name = tensor("xr_55_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_55_end_mask_0 = const()[name = tensor("xr_55_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_55_squeeze_mask_0 = const()[name = tensor("xr_55_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_55_cast_fp16 = slice_by_index(begin = xr_55_begin_0, end = xr_55_end_0, end_mask = xr_55_end_mask_0, squeeze_mask = xr_55_squeeze_mask_0, x = pairs_55_cast_fp16)[name = tensor("xr_55_cast_fp16")]; + tensor xi_55_begin_0 = const()[name = tensor("xi_55_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_55_end_0 = const()[name = tensor("xi_55_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_55_end_mask_0 = const()[name = tensor("xi_55_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_55_squeeze_mask_0 = const()[name = tensor("xi_55_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_55_cast_fp16 = slice_by_index(begin = xi_55_begin_0, end = xi_55_end_0, end_mask = xi_55_end_mask_0, squeeze_mask = xi_55_squeeze_mask_0, x = pairs_55_cast_fp16)[name = tensor("xi_55_cast_fp16")]; + tensor var_3415_cast_fp16 = mul(x = xr_55_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3415_cast_fp16")]; + tensor var_3416_cast_fp16 = mul(x = xi_55_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3416_cast_fp16")]; + tensor xor__55_cast_fp16 = sub(x = var_3415_cast_fp16, y = var_3416_cast_fp16)[name = tensor("xor__55_cast_fp16")]; + tensor var_3419_cast_fp16 = mul(x = xr_55_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3419_cast_fp16")]; + tensor var_3420_cast_fp16 = mul(x = xi_55_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3420_cast_fp16")]; + tensor xoi_55_cast_fp16 = add(x = var_3419_cast_fp16, y = var_3420_cast_fp16)[name = tensor("xoi_55_cast_fp16")]; + tensor out_55_axis_0 = const()[name = tensor("out_55_axis_0"), val = tensor(-1)]; + tensor out_55_cast_fp16 = stack(axis = out_55_axis_0, values = (xor__55_cast_fp16, xoi_55_cast_fp16))[name = tensor("out_55_cast_fp16")]; + tensor var_3437 = const()[name = tensor("op_3437"), val = tensor([256, 1024])]; + tensor var_3438_cast_fp16 = reshape(shape = var_3437, x = out_55_cast_fp16)[name = tensor("op_3438_cast_fp16")]; + tensor written_k_27_transpose_x_1 = const()[name = tensor("written_k_27_transpose_x_1"), val = tensor(true)]; + tensor written_k_27_transpose_y_1 = const()[name = tensor("written_k_27_transpose_y_1"), val = tensor(false)]; + tensor written_k_27_cast_fp16 = matmul(transpose_x = written_k_27_transpose_x_1, transpose_y = written_k_27_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_3438_cast_fp16)[name = tensor("written_k_27_cast_fp16")]; + tensor var_3445 = const()[name = tensor("op_3445"), val = tensor([256, 1024])]; + tensor var_3446_cast_fp16 = reshape(shape = var_3445, x = v_27_cast_fp16)[name = tensor("op_3446_cast_fp16")]; + tensor written_v_27_transpose_x_1 = const()[name = tensor("written_v_27_transpose_x_1"), val = tensor(true)]; + tensor written_v_27_transpose_y_1 = const()[name = tensor("written_v_27_transpose_y_1"), val = tensor(false)]; + tensor written_v_27_cast_fp16 = matmul(transpose_x = written_v_27_transpose_x_1, transpose_y = written_v_27_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_3446_cast_fp16)[name = tensor("written_v_27_cast_fp16")]; + tensor k_cache13_to_fp16_dtype_0 = const()[name = tensor("k_cache13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache13_to_fp16 = cast(dtype = k_cache13_to_fp16_dtype_0, x = k_cache13)[name = tensor("cast_387")]; + tensor var_3457_cast_fp16 = mul(x = k_cache13_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_3457_cast_fp16")]; + tensor var_3462 = const()[name = tensor("op_3462"), val = tensor([1, 512, 16, 64])]; + tensor var_3463_cast_fp16 = reshape(shape = var_3462, x = written_k_27_cast_fp16)[name = tensor("op_3463_cast_fp16")]; + tensor new_k_cache_27_cast_fp16 = add(x = var_3457_cast_fp16, y = var_3463_cast_fp16)[name = tensor("new_k_cache_27_cast_fp16")]; + tensor new_k_cache_27_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_27_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache13_to_fp16_dtype_0 = const()[name = tensor("v_cache13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache13_to_fp16 = cast(dtype = v_cache13_to_fp16_dtype_0, x = v_cache13)[name = tensor("cast_385")]; + tensor var_3466_cast_fp16 = mul(x = v_cache13_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_3466_cast_fp16")]; + tensor var_3471 = const()[name = tensor("op_3471"), val = tensor([1, 512, 16, 64])]; + tensor var_3472_cast_fp16 = reshape(shape = var_3471, x = written_v_27_cast_fp16)[name = tensor("op_3472_cast_fp16")]; + tensor new_v_cache_27_cast_fp16 = add(x = var_3466_cast_fp16, y = var_3472_cast_fp16)[name = tensor("new_v_cache_27_cast_fp16")]; + tensor new_v_cache_27_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_27_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_3485 = const()[name = tensor("op_3485"), val = tensor([0, 2, 1, 3])]; + tensor var_3496_transpose_x_0 = const()[name = tensor("op_3496_transpose_x_0"), val = tensor(false)]; + tensor var_3496_transpose_y_0 = const()[name = tensor("op_3496_transpose_y_0"), val = tensor(false)]; + tensor transpose_95_perm_0 = const()[name = tensor("transpose_95_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_96_perm_0 = const()[name = tensor("transpose_96_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_96 = transpose(perm = transpose_96_perm_0, x = new_k_cache_27_cast_fp16)[name = tensor("transpose_152")]; + tensor transpose_95 = transpose(perm = transpose_95_perm_0, x = q_27_cast_fp16)[name = tensor("transpose_153")]; + tensor var_3496_cast_fp16 = matmul(transpose_x = var_3496_transpose_x_0, transpose_y = var_3496_transpose_y_0, x = transpose_95, y = transpose_96)[name = tensor("op_3496_cast_fp16")]; + tensor var_3497_to_fp16 = const()[name = tensor("op_3497_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_53_cast_fp16 = mul(x = var_3496_cast_fp16, y = var_3497_to_fp16)[name = tensor("scores_53_cast_fp16")]; + tensor scores_55_cast_fp16 = add(x = scores_53_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_55_cast_fp16")]; + tensor var_3507 = const()[name = tensor("op_3507"), val = tensor(-1)]; + tensor weights_27_cast_fp16 = softmax(axis = var_3507, x = scores_55_cast_fp16)[name = tensor("weights_27_cast_fp16")]; + tensor attn_27_transpose_x_0 = const()[name = tensor("attn_27_transpose_x_0"), val = tensor(false)]; + tensor attn_27_transpose_y_0 = const()[name = tensor("attn_27_transpose_y_0"), val = tensor(false)]; + tensor values_27_cast_fp16 = transpose(perm = var_3485, x = new_v_cache_27_cast_fp16)[name = tensor("transpose_154")]; + tensor attn_27_cast_fp16 = matmul(transpose_x = attn_27_transpose_x_0, transpose_y = attn_27_transpose_y_0, x = weights_27_cast_fp16, y = values_27_cast_fp16)[name = tensor("attn_27_cast_fp16")]; + tensor var_3515 = const()[name = tensor("op_3515"), val = tensor([0, 2, 1, 3])]; + tensor var_3520 = const()[name = tensor("op_3520"), val = tensor([1, 256, 1024])]; + tensor var_3516_cast_fp16 = transpose(perm = var_3515, x = attn_27_cast_fp16)[name = tensor("transpose_151")]; + tensor input_161_cast_fp16 = reshape(shape = var_3520, x = var_3516_cast_fp16)[name = tensor("input_161_cast_fp16")]; + tensor attn13_out_proj_weight_to_fp16 = const()[name = tensor("attn13_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(333583616)))]; + tensor linear_53_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn13_out_proj_weight_to_fp16, x = input_161_cast_fp16)[name = tensor("linear_53_cast_fp16")]; + tensor position13_to_fp16_dtype_0 = const()[name = tensor("position13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position13_to_fp16 = cast(dtype = position13_to_fp16_dtype_0, x = position13)[name = tensor("cast_383")]; + tensor var_3529_cast_fp16 = add(x = position13_to_fp16, y = valid_len_to_fp16)[name = tensor("op_3529_cast_fp16")]; + tensor var_3529_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3529_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_163_cast_fp16 = add(x = input_157_cast_fp16, y = linear_53_cast_fp16)[name = tensor("input_163_cast_fp16")]; + tensor input_165_axes_0 = const()[name = tensor("input_165_axes_0"), val = tensor([-1])]; + tensor norm13_2_weight_to_fp16 = const()[name = tensor("norm13_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335680832)))]; + tensor norm13_2_bias_to_fp16 = const()[name = tensor("norm13_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335682944)))]; + tensor var_3533_to_fp16 = const()[name = tensor("op_3533_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_165_cast_fp16 = layer_norm(axes = input_165_axes_0, beta = norm13_2_bias_to_fp16, epsilon = var_3533_to_fp16, gamma = norm13_2_weight_to_fp16, x = input_163_cast_fp16)[name = tensor("input_165_cast_fp16")]; + tensor linear13_1_weight_to_fp16 = const()[name = tensor("linear13_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335685056)))]; + tensor linear_54_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear13_1_weight_to_fp16, x = input_165_cast_fp16)[name = tensor("linear_54_cast_fp16")]; + tensor input_167_mode_0 = const()[name = tensor("input_167_mode_0"), val = tensor("EXACT")]; + tensor input_167_cast_fp16 = gelu(mode = input_167_mode_0, x = linear_54_cast_fp16)[name = tensor("input_167_cast_fp16")]; + tensor linear13_2_weight_to_fp16 = const()[name = tensor("linear13_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(344073728)))]; + tensor linear_55_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear13_2_weight_to_fp16, x = input_167_cast_fp16)[name = tensor("linear_55_cast_fp16")]; + tensor input_169_cast_fp16 = add(x = input_163_cast_fp16, y = linear_55_cast_fp16)[name = tensor("input_169_cast_fp16")]; + tensor input_171_axes_0 = const()[name = tensor("input_171_axes_0"), val = tensor([-1])]; + tensor norm14_1_weight_to_fp16 = const()[name = tensor("norm14_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(352462400)))]; + tensor norm14_1_bias_to_fp16 = const()[name = tensor("norm14_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(352464512)))]; + tensor var_3550_to_fp16 = const()[name = tensor("op_3550_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_171_cast_fp16 = layer_norm(axes = input_171_axes_0, beta = norm14_1_bias_to_fp16, epsilon = var_3550_to_fp16, gamma = norm14_1_weight_to_fp16, x = input_169_cast_fp16)[name = tensor("input_171_cast_fp16")]; + tensor attn14_in_proj_weight_to_fp16 = const()[name = tensor("attn14_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(352466624)))]; + tensor linear_56_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn14_in_proj_weight_to_fp16, x = input_171_cast_fp16)[name = tensor("linear_56_cast_fp16")]; + tensor var_3564 = const()[name = tensor("op_3564"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_29_cast_fp16 = reshape(shape = var_3564, x = linear_56_cast_fp16)[name = tensor("qkv_29_cast_fp16")]; + tensor x_57_begin_0 = const()[name = tensor("x_57_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_57_end_0 = const()[name = tensor("x_57_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_57_end_mask_0 = const()[name = tensor("x_57_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_57_squeeze_mask_0 = const()[name = tensor("x_57_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_57_cast_fp16 = slice_by_index(begin = x_57_begin_0, end = x_57_end_0, end_mask = x_57_end_mask_0, squeeze_mask = x_57_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("x_57_cast_fp16")]; + tensor x_59_begin_0 = const()[name = tensor("x_59_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_59_end_0 = const()[name = tensor("x_59_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_59_end_mask_0 = const()[name = tensor("x_59_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_59_squeeze_mask_0 = const()[name = tensor("x_59_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_59_cast_fp16 = slice_by_index(begin = x_59_begin_0, end = x_59_end_0, end_mask = x_59_end_mask_0, squeeze_mask = x_59_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("x_59_cast_fp16")]; + tensor v_29_begin_0 = const()[name = tensor("v_29_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_29_end_0 = const()[name = tensor("v_29_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_29_end_mask_0 = const()[name = tensor("v_29_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_29_squeeze_mask_0 = const()[name = tensor("v_29_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_29_cast_fp16 = slice_by_index(begin = v_29_begin_0, end = v_29_end_0, end_mask = v_29_end_mask_0, squeeze_mask = v_29_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("v_29_cast_fp16")]; + tensor var_3609 = const()[name = tensor("op_3609"), val = tensor([256, 16, 32, 2])]; + tensor pairs_57_cast_fp16 = reshape(shape = var_3609, x = x_57_cast_fp16)[name = tensor("pairs_57_cast_fp16")]; + tensor xr_57_begin_0 = const()[name = tensor("xr_57_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_57_end_0 = const()[name = tensor("xr_57_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_57_end_mask_0 = const()[name = tensor("xr_57_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_57_squeeze_mask_0 = const()[name = tensor("xr_57_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_57_cast_fp16 = slice_by_index(begin = xr_57_begin_0, end = xr_57_end_0, end_mask = xr_57_end_mask_0, squeeze_mask = xr_57_squeeze_mask_0, x = pairs_57_cast_fp16)[name = tensor("xr_57_cast_fp16")]; + tensor xi_57_begin_0 = const()[name = tensor("xi_57_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_57_end_0 = const()[name = tensor("xi_57_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_57_end_mask_0 = const()[name = tensor("xi_57_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_57_squeeze_mask_0 = const()[name = tensor("xi_57_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_57_cast_fp16 = slice_by_index(begin = xi_57_begin_0, end = xi_57_end_0, end_mask = xi_57_end_mask_0, squeeze_mask = xi_57_squeeze_mask_0, x = pairs_57_cast_fp16)[name = tensor("xi_57_cast_fp16")]; + tensor var_3617_cast_fp16 = mul(x = xr_57_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3617_cast_fp16")]; + tensor var_3618_cast_fp16 = mul(x = xi_57_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3618_cast_fp16")]; + tensor xor__57_cast_fp16 = sub(x = var_3617_cast_fp16, y = var_3618_cast_fp16)[name = tensor("xor__57_cast_fp16")]; + tensor var_3621_cast_fp16 = mul(x = xr_57_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3621_cast_fp16")]; + tensor var_3622_cast_fp16 = mul(x = xi_57_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3622_cast_fp16")]; + tensor xoi_57_cast_fp16 = add(x = var_3621_cast_fp16, y = var_3622_cast_fp16)[name = tensor("xoi_57_cast_fp16")]; + tensor out_57_axis_0 = const()[name = tensor("out_57_axis_0"), val = tensor(-1)]; + tensor out_57_cast_fp16 = stack(axis = out_57_axis_0, values = (xor__57_cast_fp16, xoi_57_cast_fp16))[name = tensor("out_57_cast_fp16")]; + tensor var_3632 = const()[name = tensor("op_3632"), val = tensor([1, 256, 16, 64])]; + tensor q_29_cast_fp16 = reshape(shape = var_3632, x = out_57_cast_fp16)[name = tensor("q_29_cast_fp16")]; + tensor var_3638 = const()[name = tensor("op_3638"), val = tensor([256, 16, 32, 2])]; + tensor pairs_59_cast_fp16 = reshape(shape = var_3638, x = x_59_cast_fp16)[name = tensor("pairs_59_cast_fp16")]; + tensor xr_59_begin_0 = const()[name = tensor("xr_59_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_59_end_0 = const()[name = tensor("xr_59_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_59_end_mask_0 = const()[name = tensor("xr_59_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_59_squeeze_mask_0 = const()[name = tensor("xr_59_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_59_cast_fp16 = slice_by_index(begin = xr_59_begin_0, end = xr_59_end_0, end_mask = xr_59_end_mask_0, squeeze_mask = xr_59_squeeze_mask_0, x = pairs_59_cast_fp16)[name = tensor("xr_59_cast_fp16")]; + tensor xi_59_begin_0 = const()[name = tensor("xi_59_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_59_end_0 = const()[name = tensor("xi_59_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_59_end_mask_0 = const()[name = tensor("xi_59_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_59_squeeze_mask_0 = const()[name = tensor("xi_59_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_59_cast_fp16 = slice_by_index(begin = xi_59_begin_0, end = xi_59_end_0, end_mask = xi_59_end_mask_0, squeeze_mask = xi_59_squeeze_mask_0, x = pairs_59_cast_fp16)[name = tensor("xi_59_cast_fp16")]; + tensor var_3646_cast_fp16 = mul(x = xr_59_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3646_cast_fp16")]; + tensor var_3647_cast_fp16 = mul(x = xi_59_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3647_cast_fp16")]; + tensor xor__59_cast_fp16 = sub(x = var_3646_cast_fp16, y = var_3647_cast_fp16)[name = tensor("xor__59_cast_fp16")]; + tensor var_3650_cast_fp16 = mul(x = xr_59_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3650_cast_fp16")]; + tensor var_3651_cast_fp16 = mul(x = xi_59_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3651_cast_fp16")]; + tensor xoi_59_cast_fp16 = add(x = var_3650_cast_fp16, y = var_3651_cast_fp16)[name = tensor("xoi_59_cast_fp16")]; + tensor out_59_axis_0 = const()[name = tensor("out_59_axis_0"), val = tensor(-1)]; + tensor out_59_cast_fp16 = stack(axis = out_59_axis_0, values = (xor__59_cast_fp16, xoi_59_cast_fp16))[name = tensor("out_59_cast_fp16")]; + tensor var_3668 = const()[name = tensor("op_3668"), val = tensor([256, 1024])]; + tensor var_3669_cast_fp16 = reshape(shape = var_3668, x = out_59_cast_fp16)[name = tensor("op_3669_cast_fp16")]; + tensor written_k_29_transpose_x_1 = const()[name = tensor("written_k_29_transpose_x_1"), val = tensor(true)]; + tensor written_k_29_transpose_y_1 = const()[name = tensor("written_k_29_transpose_y_1"), val = tensor(false)]; + tensor written_k_29_cast_fp16 = matmul(transpose_x = written_k_29_transpose_x_1, transpose_y = written_k_29_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_3669_cast_fp16)[name = tensor("written_k_29_cast_fp16")]; + tensor var_3676 = const()[name = tensor("op_3676"), val = tensor([256, 1024])]; + tensor var_3677_cast_fp16 = reshape(shape = var_3676, x = v_29_cast_fp16)[name = tensor("op_3677_cast_fp16")]; + tensor written_v_29_transpose_x_1 = const()[name = tensor("written_v_29_transpose_x_1"), val = tensor(true)]; + tensor written_v_29_transpose_y_1 = const()[name = tensor("written_v_29_transpose_y_1"), val = tensor(false)]; + tensor written_v_29_cast_fp16 = matmul(transpose_x = written_v_29_transpose_x_1, transpose_y = written_v_29_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_3677_cast_fp16)[name = tensor("written_v_29_cast_fp16")]; + tensor k_cache14_to_fp16_dtype_0 = const()[name = tensor("k_cache14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache14_to_fp16 = cast(dtype = k_cache14_to_fp16_dtype_0, x = k_cache14)[name = tensor("cast_381")]; + tensor var_3688_cast_fp16 = mul(x = k_cache14_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_3688_cast_fp16")]; + tensor var_3693 = const()[name = tensor("op_3693"), val = tensor([1, 512, 16, 64])]; + tensor var_3694_cast_fp16 = reshape(shape = var_3693, x = written_k_29_cast_fp16)[name = tensor("op_3694_cast_fp16")]; + tensor new_k_cache_29_cast_fp16 = add(x = var_3688_cast_fp16, y = var_3694_cast_fp16)[name = tensor("new_k_cache_29_cast_fp16")]; + tensor new_k_cache_29_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_29_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache14_to_fp16_dtype_0 = const()[name = tensor("v_cache14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache14_to_fp16 = cast(dtype = v_cache14_to_fp16_dtype_0, x = v_cache14)[name = tensor("cast_379")]; + tensor var_3697_cast_fp16 = mul(x = v_cache14_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_3697_cast_fp16")]; + tensor var_3702 = const()[name = tensor("op_3702"), val = tensor([1, 512, 16, 64])]; + tensor var_3703_cast_fp16 = reshape(shape = var_3702, x = written_v_29_cast_fp16)[name = tensor("op_3703_cast_fp16")]; + tensor new_v_cache_29_cast_fp16 = add(x = var_3697_cast_fp16, y = var_3703_cast_fp16)[name = tensor("new_v_cache_29_cast_fp16")]; + tensor new_v_cache_29_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_29_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_3716 = const()[name = tensor("op_3716"), val = tensor([0, 2, 1, 3])]; + tensor var_3727_transpose_x_0 = const()[name = tensor("op_3727_transpose_x_0"), val = tensor(false)]; + tensor var_3727_transpose_y_0 = const()[name = tensor("op_3727_transpose_y_0"), val = tensor(false)]; + tensor transpose_97_perm_0 = const()[name = tensor("transpose_97_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_98_perm_0 = const()[name = tensor("transpose_98_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_98 = transpose(perm = transpose_98_perm_0, x = new_k_cache_29_cast_fp16)[name = tensor("transpose_148")]; + tensor transpose_97 = transpose(perm = transpose_97_perm_0, x = q_29_cast_fp16)[name = tensor("transpose_149")]; + tensor var_3727_cast_fp16 = matmul(transpose_x = var_3727_transpose_x_0, transpose_y = var_3727_transpose_y_0, x = transpose_97, y = transpose_98)[name = tensor("op_3727_cast_fp16")]; + tensor var_3728_to_fp16 = const()[name = tensor("op_3728_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_57_cast_fp16 = mul(x = var_3727_cast_fp16, y = var_3728_to_fp16)[name = tensor("scores_57_cast_fp16")]; + tensor scores_59_cast_fp16 = add(x = scores_57_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_59_cast_fp16")]; + tensor var_3738 = const()[name = tensor("op_3738"), val = tensor(-1)]; + tensor weights_29_cast_fp16 = softmax(axis = var_3738, x = scores_59_cast_fp16)[name = tensor("weights_29_cast_fp16")]; + tensor attn_29_transpose_x_0 = const()[name = tensor("attn_29_transpose_x_0"), val = tensor(false)]; + tensor attn_29_transpose_y_0 = const()[name = tensor("attn_29_transpose_y_0"), val = tensor(false)]; + tensor values_29_cast_fp16 = transpose(perm = var_3716, x = new_v_cache_29_cast_fp16)[name = tensor("transpose_150")]; + tensor attn_29_cast_fp16 = matmul(transpose_x = attn_29_transpose_x_0, transpose_y = attn_29_transpose_y_0, x = weights_29_cast_fp16, y = values_29_cast_fp16)[name = tensor("attn_29_cast_fp16")]; + tensor var_3746 = const()[name = tensor("op_3746"), val = tensor([0, 2, 1, 3])]; + tensor var_3751 = const()[name = tensor("op_3751"), val = tensor([1, 256, 1024])]; + tensor var_3747_cast_fp16 = transpose(perm = var_3746, x = attn_29_cast_fp16)[name = tensor("transpose_147")]; + tensor input_173_cast_fp16 = reshape(shape = var_3751, x = var_3747_cast_fp16)[name = tensor("input_173_cast_fp16")]; + tensor attn14_out_proj_weight_to_fp16 = const()[name = tensor("attn14_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(358758144)))]; + tensor linear_57_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn14_out_proj_weight_to_fp16, x = input_173_cast_fp16)[name = tensor("linear_57_cast_fp16")]; + tensor position14_to_fp16_dtype_0 = const()[name = tensor("position14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position14_to_fp16 = cast(dtype = position14_to_fp16_dtype_0, x = position14)[name = tensor("cast_377")]; + tensor var_3760_cast_fp16 = add(x = position14_to_fp16, y = valid_len_to_fp16)[name = tensor("op_3760_cast_fp16")]; + tensor var_3760_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3760_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_175_cast_fp16 = add(x = input_169_cast_fp16, y = linear_57_cast_fp16)[name = tensor("input_175_cast_fp16")]; + tensor input_177_axes_0 = const()[name = tensor("input_177_axes_0"), val = tensor([-1])]; + tensor norm14_2_weight_to_fp16 = const()[name = tensor("norm14_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(360855360)))]; + tensor norm14_2_bias_to_fp16 = const()[name = tensor("norm14_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(360857472)))]; + tensor var_3764_to_fp16 = const()[name = tensor("op_3764_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_177_cast_fp16 = layer_norm(axes = input_177_axes_0, beta = norm14_2_bias_to_fp16, epsilon = var_3764_to_fp16, gamma = norm14_2_weight_to_fp16, x = input_175_cast_fp16)[name = tensor("input_177_cast_fp16")]; + tensor linear14_1_weight_to_fp16 = const()[name = tensor("linear14_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(360859584)))]; + tensor linear_58_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear14_1_weight_to_fp16, x = input_177_cast_fp16)[name = tensor("linear_58_cast_fp16")]; + tensor input_179_mode_0 = const()[name = tensor("input_179_mode_0"), val = tensor("EXACT")]; + tensor input_179_cast_fp16 = gelu(mode = input_179_mode_0, x = linear_58_cast_fp16)[name = tensor("input_179_cast_fp16")]; + tensor linear14_2_weight_to_fp16 = const()[name = tensor("linear14_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(369248256)))]; + tensor linear_59_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear14_2_weight_to_fp16, x = input_179_cast_fp16)[name = tensor("linear_59_cast_fp16")]; + tensor input_181_cast_fp16 = add(x = input_175_cast_fp16, y = linear_59_cast_fp16)[name = tensor("input_181_cast_fp16")]; + tensor input_183_axes_0 = const()[name = tensor("input_183_axes_0"), val = tensor([-1])]; + tensor norm15_1_weight_to_fp16 = const()[name = tensor("norm15_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(377636928)))]; + tensor norm15_1_bias_to_fp16 = const()[name = tensor("norm15_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(377639040)))]; + tensor var_3781_to_fp16 = const()[name = tensor("op_3781_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_183_cast_fp16 = layer_norm(axes = input_183_axes_0, beta = norm15_1_bias_to_fp16, epsilon = var_3781_to_fp16, gamma = norm15_1_weight_to_fp16, x = input_181_cast_fp16)[name = tensor("input_183_cast_fp16")]; + tensor attn15_in_proj_weight_to_fp16 = const()[name = tensor("attn15_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(377641152)))]; + tensor linear_60_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn15_in_proj_weight_to_fp16, x = input_183_cast_fp16)[name = tensor("linear_60_cast_fp16")]; + tensor var_3795 = const()[name = tensor("op_3795"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_31_cast_fp16 = reshape(shape = var_3795, x = linear_60_cast_fp16)[name = tensor("qkv_31_cast_fp16")]; + tensor x_61_begin_0 = const()[name = tensor("x_61_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_61_end_0 = const()[name = tensor("x_61_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_61_end_mask_0 = const()[name = tensor("x_61_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_61_squeeze_mask_0 = const()[name = tensor("x_61_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_61_cast_fp16 = slice_by_index(begin = x_61_begin_0, end = x_61_end_0, end_mask = x_61_end_mask_0, squeeze_mask = x_61_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("x_61_cast_fp16")]; + tensor x_63_begin_0 = const()[name = tensor("x_63_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_63_end_0 = const()[name = tensor("x_63_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_63_end_mask_0 = const()[name = tensor("x_63_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_63_squeeze_mask_0 = const()[name = tensor("x_63_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_63_cast_fp16 = slice_by_index(begin = x_63_begin_0, end = x_63_end_0, end_mask = x_63_end_mask_0, squeeze_mask = x_63_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("x_63_cast_fp16")]; + tensor v_31_begin_0 = const()[name = tensor("v_31_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_31_end_0 = const()[name = tensor("v_31_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_31_end_mask_0 = const()[name = tensor("v_31_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_31_squeeze_mask_0 = const()[name = tensor("v_31_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_31_cast_fp16 = slice_by_index(begin = v_31_begin_0, end = v_31_end_0, end_mask = v_31_end_mask_0, squeeze_mask = v_31_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("v_31_cast_fp16")]; + tensor var_3840 = const()[name = tensor("op_3840"), val = tensor([256, 16, 32, 2])]; + tensor pairs_61_cast_fp16 = reshape(shape = var_3840, x = x_61_cast_fp16)[name = tensor("pairs_61_cast_fp16")]; + tensor xr_61_begin_0 = const()[name = tensor("xr_61_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_61_end_0 = const()[name = tensor("xr_61_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_61_end_mask_0 = const()[name = tensor("xr_61_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_61_squeeze_mask_0 = const()[name = tensor("xr_61_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_61_cast_fp16 = slice_by_index(begin = xr_61_begin_0, end = xr_61_end_0, end_mask = xr_61_end_mask_0, squeeze_mask = xr_61_squeeze_mask_0, x = pairs_61_cast_fp16)[name = tensor("xr_61_cast_fp16")]; + tensor xi_61_begin_0 = const()[name = tensor("xi_61_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_61_end_0 = const()[name = tensor("xi_61_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_61_end_mask_0 = const()[name = tensor("xi_61_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_61_squeeze_mask_0 = const()[name = tensor("xi_61_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_61_cast_fp16 = slice_by_index(begin = xi_61_begin_0, end = xi_61_end_0, end_mask = xi_61_end_mask_0, squeeze_mask = xi_61_squeeze_mask_0, x = pairs_61_cast_fp16)[name = tensor("xi_61_cast_fp16")]; + tensor var_3848_cast_fp16 = mul(x = xr_61_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3848_cast_fp16")]; + tensor var_3849_cast_fp16 = mul(x = xi_61_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3849_cast_fp16")]; + tensor xor__61_cast_fp16 = sub(x = var_3848_cast_fp16, y = var_3849_cast_fp16)[name = tensor("xor__61_cast_fp16")]; + tensor var_3852_cast_fp16 = mul(x = xr_61_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3852_cast_fp16")]; + tensor var_3853_cast_fp16 = mul(x = xi_61_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3853_cast_fp16")]; + tensor xoi_61_cast_fp16 = add(x = var_3852_cast_fp16, y = var_3853_cast_fp16)[name = tensor("xoi_61_cast_fp16")]; + tensor out_61_axis_0 = const()[name = tensor("out_61_axis_0"), val = tensor(-1)]; + tensor out_61_cast_fp16 = stack(axis = out_61_axis_0, values = (xor__61_cast_fp16, xoi_61_cast_fp16))[name = tensor("out_61_cast_fp16")]; + tensor var_3863 = const()[name = tensor("op_3863"), val = tensor([1, 256, 16, 64])]; + tensor q_31_cast_fp16 = reshape(shape = var_3863, x = out_61_cast_fp16)[name = tensor("q_31_cast_fp16")]; + tensor var_3869 = const()[name = tensor("op_3869"), val = tensor([256, 16, 32, 2])]; + tensor pairs_63_cast_fp16 = reshape(shape = var_3869, x = x_63_cast_fp16)[name = tensor("pairs_63_cast_fp16")]; + tensor xr_63_begin_0 = const()[name = tensor("xr_63_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_63_end_0 = const()[name = tensor("xr_63_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_63_end_mask_0 = const()[name = tensor("xr_63_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_63_squeeze_mask_0 = const()[name = tensor("xr_63_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_63_cast_fp16 = slice_by_index(begin = xr_63_begin_0, end = xr_63_end_0, end_mask = xr_63_end_mask_0, squeeze_mask = xr_63_squeeze_mask_0, x = pairs_63_cast_fp16)[name = tensor("xr_63_cast_fp16")]; + tensor xi_63_begin_0 = const()[name = tensor("xi_63_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_63_end_0 = const()[name = tensor("xi_63_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_63_end_mask_0 = const()[name = tensor("xi_63_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_63_squeeze_mask_0 = const()[name = tensor("xi_63_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_63_cast_fp16 = slice_by_index(begin = xi_63_begin_0, end = xi_63_end_0, end_mask = xi_63_end_mask_0, squeeze_mask = xi_63_squeeze_mask_0, x = pairs_63_cast_fp16)[name = tensor("xi_63_cast_fp16")]; + tensor var_3877_cast_fp16 = mul(x = xr_63_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3877_cast_fp16")]; + tensor var_3878_cast_fp16 = mul(x = xi_63_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3878_cast_fp16")]; + tensor xor__63_cast_fp16 = sub(x = var_3877_cast_fp16, y = var_3878_cast_fp16)[name = tensor("xor__63_cast_fp16")]; + tensor var_3881_cast_fp16 = mul(x = xr_63_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3881_cast_fp16")]; + tensor var_3882_cast_fp16 = mul(x = xi_63_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3882_cast_fp16")]; + tensor xoi_63_cast_fp16 = add(x = var_3881_cast_fp16, y = var_3882_cast_fp16)[name = tensor("xoi_63_cast_fp16")]; + tensor out_63_axis_0 = const()[name = tensor("out_63_axis_0"), val = tensor(-1)]; + tensor out_63_cast_fp16 = stack(axis = out_63_axis_0, values = (xor__63_cast_fp16, xoi_63_cast_fp16))[name = tensor("out_63_cast_fp16")]; + tensor var_3899 = const()[name = tensor("op_3899"), val = tensor([256, 1024])]; + tensor var_3900_cast_fp16 = reshape(shape = var_3899, x = out_63_cast_fp16)[name = tensor("op_3900_cast_fp16")]; + tensor written_k_31_transpose_x_1 = const()[name = tensor("written_k_31_transpose_x_1"), val = tensor(true)]; + tensor written_k_31_transpose_y_1 = const()[name = tensor("written_k_31_transpose_y_1"), val = tensor(false)]; + tensor written_k_31_cast_fp16 = matmul(transpose_x = written_k_31_transpose_x_1, transpose_y = written_k_31_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_3900_cast_fp16)[name = tensor("written_k_31_cast_fp16")]; + tensor var_3907 = const()[name = tensor("op_3907"), val = tensor([256, 1024])]; + tensor var_3908_cast_fp16 = reshape(shape = var_3907, x = v_31_cast_fp16)[name = tensor("op_3908_cast_fp16")]; + tensor written_v_31_transpose_x_1 = const()[name = tensor("written_v_31_transpose_x_1"), val = tensor(true)]; + tensor written_v_31_transpose_y_1 = const()[name = tensor("written_v_31_transpose_y_1"), val = tensor(false)]; + tensor written_v_31_cast_fp16 = matmul(transpose_x = written_v_31_transpose_x_1, transpose_y = written_v_31_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_3908_cast_fp16)[name = tensor("written_v_31_cast_fp16")]; + tensor k_cache15_to_fp16_dtype_0 = const()[name = tensor("k_cache15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache15_to_fp16 = cast(dtype = k_cache15_to_fp16_dtype_0, x = k_cache15)[name = tensor("cast_375")]; + tensor var_3919_cast_fp16 = mul(x = k_cache15_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_3919_cast_fp16")]; + tensor var_3924 = const()[name = tensor("op_3924"), val = tensor([1, 512, 16, 64])]; + tensor var_3925_cast_fp16 = reshape(shape = var_3924, x = written_k_31_cast_fp16)[name = tensor("op_3925_cast_fp16")]; + tensor new_k_cache_31_cast_fp16 = add(x = var_3919_cast_fp16, y = var_3925_cast_fp16)[name = tensor("new_k_cache_31_cast_fp16")]; + tensor new_k_cache_31_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_31_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache15_to_fp16_dtype_0 = const()[name = tensor("v_cache15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache15_to_fp16 = cast(dtype = v_cache15_to_fp16_dtype_0, x = v_cache15)[name = tensor("cast_373")]; + tensor var_3928_cast_fp16 = mul(x = v_cache15_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_3928_cast_fp16")]; + tensor var_3933 = const()[name = tensor("op_3933"), val = tensor([1, 512, 16, 64])]; + tensor var_3934_cast_fp16 = reshape(shape = var_3933, x = written_v_31_cast_fp16)[name = tensor("op_3934_cast_fp16")]; + tensor new_v_cache_31_cast_fp16 = add(x = var_3928_cast_fp16, y = var_3934_cast_fp16)[name = tensor("new_v_cache_31_cast_fp16")]; + tensor new_v_cache_31_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_31_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_3947 = const()[name = tensor("op_3947"), val = tensor([0, 2, 1, 3])]; + tensor var_3958_transpose_x_0 = const()[name = tensor("op_3958_transpose_x_0"), val = tensor(false)]; + tensor var_3958_transpose_y_0 = const()[name = tensor("op_3958_transpose_y_0"), val = tensor(false)]; + tensor transpose_99_perm_0 = const()[name = tensor("transpose_99_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_100_perm_0 = const()[name = tensor("transpose_100_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_100 = transpose(perm = transpose_100_perm_0, x = new_k_cache_31_cast_fp16)[name = tensor("transpose_144")]; + tensor transpose_99 = transpose(perm = transpose_99_perm_0, x = q_31_cast_fp16)[name = tensor("transpose_145")]; + tensor var_3958_cast_fp16 = matmul(transpose_x = var_3958_transpose_x_0, transpose_y = var_3958_transpose_y_0, x = transpose_99, y = transpose_100)[name = tensor("op_3958_cast_fp16")]; + tensor var_3959_to_fp16 = const()[name = tensor("op_3959_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_61_cast_fp16 = mul(x = var_3958_cast_fp16, y = var_3959_to_fp16)[name = tensor("scores_61_cast_fp16")]; + tensor scores_63_cast_fp16 = add(x = scores_61_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_63_cast_fp16")]; + tensor var_3969 = const()[name = tensor("op_3969"), val = tensor(-1)]; + tensor weights_31_cast_fp16 = softmax(axis = var_3969, x = scores_63_cast_fp16)[name = tensor("weights_31_cast_fp16")]; + tensor attn_31_transpose_x_0 = const()[name = tensor("attn_31_transpose_x_0"), val = tensor(false)]; + tensor attn_31_transpose_y_0 = const()[name = tensor("attn_31_transpose_y_0"), val = tensor(false)]; + tensor values_31_cast_fp16 = transpose(perm = var_3947, x = new_v_cache_31_cast_fp16)[name = tensor("transpose_146")]; + tensor attn_31_cast_fp16 = matmul(transpose_x = attn_31_transpose_x_0, transpose_y = attn_31_transpose_y_0, x = weights_31_cast_fp16, y = values_31_cast_fp16)[name = tensor("attn_31_cast_fp16")]; + tensor var_3977 = const()[name = tensor("op_3977"), val = tensor([0, 2, 1, 3])]; + tensor var_3982 = const()[name = tensor("op_3982"), val = tensor([1, 256, 1024])]; + tensor var_3978_cast_fp16 = transpose(perm = var_3977, x = attn_31_cast_fp16)[name = tensor("transpose_143")]; + tensor input_185_cast_fp16 = reshape(shape = var_3982, x = var_3978_cast_fp16)[name = tensor("input_185_cast_fp16")]; + tensor attn15_out_proj_weight_to_fp16 = const()[name = tensor("attn15_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(383932672)))]; + tensor linear_61_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn15_out_proj_weight_to_fp16, x = input_185_cast_fp16)[name = tensor("linear_61_cast_fp16")]; + tensor position15_to_fp16_dtype_0 = const()[name = tensor("position15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position15_to_fp16 = cast(dtype = position15_to_fp16_dtype_0, x = position15)[name = tensor("cast_371")]; + tensor var_3991_cast_fp16 = add(x = position15_to_fp16, y = valid_len_to_fp16)[name = tensor("op_3991_cast_fp16")]; + tensor var_3991_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3991_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_187_cast_fp16 = add(x = input_181_cast_fp16, y = linear_61_cast_fp16)[name = tensor("input_187_cast_fp16")]; + tensor input_189_axes_0 = const()[name = tensor("input_189_axes_0"), val = tensor([-1])]; + tensor norm15_2_weight_to_fp16 = const()[name = tensor("norm15_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(386029888)))]; + tensor norm15_2_bias_to_fp16 = const()[name = tensor("norm15_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(386032000)))]; + tensor var_3995_to_fp16 = const()[name = tensor("op_3995_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_189_cast_fp16 = layer_norm(axes = input_189_axes_0, beta = norm15_2_bias_to_fp16, epsilon = var_3995_to_fp16, gamma = norm15_2_weight_to_fp16, x = input_187_cast_fp16)[name = tensor("input_189_cast_fp16")]; + tensor linear15_1_weight_to_fp16 = const()[name = tensor("linear15_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(386034112)))]; + tensor linear_62_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear15_1_weight_to_fp16, x = input_189_cast_fp16)[name = tensor("linear_62_cast_fp16")]; + tensor input_191_mode_0 = const()[name = tensor("input_191_mode_0"), val = tensor("EXACT")]; + tensor input_191_cast_fp16 = gelu(mode = input_191_mode_0, x = linear_62_cast_fp16)[name = tensor("input_191_cast_fp16")]; + tensor linear15_2_weight_to_fp16 = const()[name = tensor("linear15_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(394422784)))]; + tensor linear_63_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear15_2_weight_to_fp16, x = input_191_cast_fp16)[name = tensor("linear_63_cast_fp16")]; + tensor input_193_cast_fp16 = add(x = input_187_cast_fp16, y = linear_63_cast_fp16)[name = tensor("input_193_cast_fp16")]; + tensor input_195_axes_0 = const()[name = tensor("input_195_axes_0"), val = tensor([-1])]; + tensor norm16_1_weight_to_fp16 = const()[name = tensor("norm16_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(402811456)))]; + tensor norm16_1_bias_to_fp16 = const()[name = tensor("norm16_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(402813568)))]; + tensor var_4012_to_fp16 = const()[name = tensor("op_4012_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_195_cast_fp16 = layer_norm(axes = input_195_axes_0, beta = norm16_1_bias_to_fp16, epsilon = var_4012_to_fp16, gamma = norm16_1_weight_to_fp16, x = input_193_cast_fp16)[name = tensor("input_195_cast_fp16")]; + tensor attn16_in_proj_weight_to_fp16 = const()[name = tensor("attn16_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(402815680)))]; + tensor linear_64_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn16_in_proj_weight_to_fp16, x = input_195_cast_fp16)[name = tensor("linear_64_cast_fp16")]; + tensor var_4026 = const()[name = tensor("op_4026"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_33_cast_fp16 = reshape(shape = var_4026, x = linear_64_cast_fp16)[name = tensor("qkv_33_cast_fp16")]; + tensor x_65_begin_0 = const()[name = tensor("x_65_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_65_end_0 = const()[name = tensor("x_65_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_65_end_mask_0 = const()[name = tensor("x_65_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_65_squeeze_mask_0 = const()[name = tensor("x_65_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_65_cast_fp16 = slice_by_index(begin = x_65_begin_0, end = x_65_end_0, end_mask = x_65_end_mask_0, squeeze_mask = x_65_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("x_65_cast_fp16")]; + tensor x_67_begin_0 = const()[name = tensor("x_67_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_67_end_0 = const()[name = tensor("x_67_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_67_end_mask_0 = const()[name = tensor("x_67_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_67_squeeze_mask_0 = const()[name = tensor("x_67_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_67_cast_fp16 = slice_by_index(begin = x_67_begin_0, end = x_67_end_0, end_mask = x_67_end_mask_0, squeeze_mask = x_67_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("x_67_cast_fp16")]; + tensor v_33_begin_0 = const()[name = tensor("v_33_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_33_end_0 = const()[name = tensor("v_33_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_33_end_mask_0 = const()[name = tensor("v_33_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_33_squeeze_mask_0 = const()[name = tensor("v_33_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_33_cast_fp16 = slice_by_index(begin = v_33_begin_0, end = v_33_end_0, end_mask = v_33_end_mask_0, squeeze_mask = v_33_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("v_33_cast_fp16")]; + tensor var_4071 = const()[name = tensor("op_4071"), val = tensor([256, 16, 32, 2])]; + tensor pairs_65_cast_fp16 = reshape(shape = var_4071, x = x_65_cast_fp16)[name = tensor("pairs_65_cast_fp16")]; + tensor xr_65_begin_0 = const()[name = tensor("xr_65_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_65_end_0 = const()[name = tensor("xr_65_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_65_end_mask_0 = const()[name = tensor("xr_65_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_65_squeeze_mask_0 = const()[name = tensor("xr_65_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_65_cast_fp16 = slice_by_index(begin = xr_65_begin_0, end = xr_65_end_0, end_mask = xr_65_end_mask_0, squeeze_mask = xr_65_squeeze_mask_0, x = pairs_65_cast_fp16)[name = tensor("xr_65_cast_fp16")]; + tensor xi_65_begin_0 = const()[name = tensor("xi_65_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_65_end_0 = const()[name = tensor("xi_65_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_65_end_mask_0 = const()[name = tensor("xi_65_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_65_squeeze_mask_0 = const()[name = tensor("xi_65_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_65_cast_fp16 = slice_by_index(begin = xi_65_begin_0, end = xi_65_end_0, end_mask = xi_65_end_mask_0, squeeze_mask = xi_65_squeeze_mask_0, x = pairs_65_cast_fp16)[name = tensor("xi_65_cast_fp16")]; + tensor var_4079_cast_fp16 = mul(x = xr_65_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4079_cast_fp16")]; + tensor var_4080_cast_fp16 = mul(x = xi_65_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4080_cast_fp16")]; + tensor xor__65_cast_fp16 = sub(x = var_4079_cast_fp16, y = var_4080_cast_fp16)[name = tensor("xor__65_cast_fp16")]; + tensor var_4083_cast_fp16 = mul(x = xr_65_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4083_cast_fp16")]; + tensor var_4084_cast_fp16 = mul(x = xi_65_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4084_cast_fp16")]; + tensor xoi_65_cast_fp16 = add(x = var_4083_cast_fp16, y = var_4084_cast_fp16)[name = tensor("xoi_65_cast_fp16")]; + tensor out_65_axis_0 = const()[name = tensor("out_65_axis_0"), val = tensor(-1)]; + tensor out_65_cast_fp16 = stack(axis = out_65_axis_0, values = (xor__65_cast_fp16, xoi_65_cast_fp16))[name = tensor("out_65_cast_fp16")]; + tensor var_4094 = const()[name = tensor("op_4094"), val = tensor([1, 256, 16, 64])]; + tensor q_33_cast_fp16 = reshape(shape = var_4094, x = out_65_cast_fp16)[name = tensor("q_33_cast_fp16")]; + tensor var_4100 = const()[name = tensor("op_4100"), val = tensor([256, 16, 32, 2])]; + tensor pairs_67_cast_fp16 = reshape(shape = var_4100, x = x_67_cast_fp16)[name = tensor("pairs_67_cast_fp16")]; + tensor xr_67_begin_0 = const()[name = tensor("xr_67_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_67_end_0 = const()[name = tensor("xr_67_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_67_end_mask_0 = const()[name = tensor("xr_67_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_67_squeeze_mask_0 = const()[name = tensor("xr_67_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_67_cast_fp16 = slice_by_index(begin = xr_67_begin_0, end = xr_67_end_0, end_mask = xr_67_end_mask_0, squeeze_mask = xr_67_squeeze_mask_0, x = pairs_67_cast_fp16)[name = tensor("xr_67_cast_fp16")]; + tensor xi_67_begin_0 = const()[name = tensor("xi_67_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_67_end_0 = const()[name = tensor("xi_67_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_67_end_mask_0 = const()[name = tensor("xi_67_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_67_squeeze_mask_0 = const()[name = tensor("xi_67_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_67_cast_fp16 = slice_by_index(begin = xi_67_begin_0, end = xi_67_end_0, end_mask = xi_67_end_mask_0, squeeze_mask = xi_67_squeeze_mask_0, x = pairs_67_cast_fp16)[name = tensor("xi_67_cast_fp16")]; + tensor var_4108_cast_fp16 = mul(x = xr_67_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4108_cast_fp16")]; + tensor var_4109_cast_fp16 = mul(x = xi_67_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4109_cast_fp16")]; + tensor xor__67_cast_fp16 = sub(x = var_4108_cast_fp16, y = var_4109_cast_fp16)[name = tensor("xor__67_cast_fp16")]; + tensor var_4112_cast_fp16 = mul(x = xr_67_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4112_cast_fp16")]; + tensor var_4113_cast_fp16 = mul(x = xi_67_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4113_cast_fp16")]; + tensor xoi_67_cast_fp16 = add(x = var_4112_cast_fp16, y = var_4113_cast_fp16)[name = tensor("xoi_67_cast_fp16")]; + tensor out_67_axis_0 = const()[name = tensor("out_67_axis_0"), val = tensor(-1)]; + tensor out_67_cast_fp16 = stack(axis = out_67_axis_0, values = (xor__67_cast_fp16, xoi_67_cast_fp16))[name = tensor("out_67_cast_fp16")]; + tensor var_4130 = const()[name = tensor("op_4130"), val = tensor([256, 1024])]; + tensor var_4131_cast_fp16 = reshape(shape = var_4130, x = out_67_cast_fp16)[name = tensor("op_4131_cast_fp16")]; + tensor written_k_33_transpose_x_1 = const()[name = tensor("written_k_33_transpose_x_1"), val = tensor(true)]; + tensor written_k_33_transpose_y_1 = const()[name = tensor("written_k_33_transpose_y_1"), val = tensor(false)]; + tensor written_k_33_cast_fp16 = matmul(transpose_x = written_k_33_transpose_x_1, transpose_y = written_k_33_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_4131_cast_fp16)[name = tensor("written_k_33_cast_fp16")]; + tensor var_4138 = const()[name = tensor("op_4138"), val = tensor([256, 1024])]; + tensor var_4139_cast_fp16 = reshape(shape = var_4138, x = v_33_cast_fp16)[name = tensor("op_4139_cast_fp16")]; + tensor written_v_33_transpose_x_1 = const()[name = tensor("written_v_33_transpose_x_1"), val = tensor(true)]; + tensor written_v_33_transpose_y_1 = const()[name = tensor("written_v_33_transpose_y_1"), val = tensor(false)]; + tensor written_v_33_cast_fp16 = matmul(transpose_x = written_v_33_transpose_x_1, transpose_y = written_v_33_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_4139_cast_fp16)[name = tensor("written_v_33_cast_fp16")]; + tensor k_cache16_to_fp16_dtype_0 = const()[name = tensor("k_cache16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache16_to_fp16 = cast(dtype = k_cache16_to_fp16_dtype_0, x = k_cache16)[name = tensor("cast_369")]; + tensor var_4150_cast_fp16 = mul(x = k_cache16_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_4150_cast_fp16")]; + tensor var_4155 = const()[name = tensor("op_4155"), val = tensor([1, 512, 16, 64])]; + tensor var_4156_cast_fp16 = reshape(shape = var_4155, x = written_k_33_cast_fp16)[name = tensor("op_4156_cast_fp16")]; + tensor new_k_cache_33_cast_fp16 = add(x = var_4150_cast_fp16, y = var_4156_cast_fp16)[name = tensor("new_k_cache_33_cast_fp16")]; + tensor new_k_cache_33_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_33_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache16_to_fp16_dtype_0 = const()[name = tensor("v_cache16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache16_to_fp16 = cast(dtype = v_cache16_to_fp16_dtype_0, x = v_cache16)[name = tensor("cast_367")]; + tensor var_4159_cast_fp16 = mul(x = v_cache16_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_4159_cast_fp16")]; + tensor var_4164 = const()[name = tensor("op_4164"), val = tensor([1, 512, 16, 64])]; + tensor var_4165_cast_fp16 = reshape(shape = var_4164, x = written_v_33_cast_fp16)[name = tensor("op_4165_cast_fp16")]; + tensor new_v_cache_33_cast_fp16 = add(x = var_4159_cast_fp16, y = var_4165_cast_fp16)[name = tensor("new_v_cache_33_cast_fp16")]; + tensor new_v_cache_33_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_33_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_4178 = const()[name = tensor("op_4178"), val = tensor([0, 2, 1, 3])]; + tensor var_4189_transpose_x_0 = const()[name = tensor("op_4189_transpose_x_0"), val = tensor(false)]; + tensor var_4189_transpose_y_0 = const()[name = tensor("op_4189_transpose_y_0"), val = tensor(false)]; + tensor transpose_101_perm_0 = const()[name = tensor("transpose_101_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_102_perm_0 = const()[name = tensor("transpose_102_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_102 = transpose(perm = transpose_102_perm_0, x = new_k_cache_33_cast_fp16)[name = tensor("transpose_140")]; + tensor transpose_101 = transpose(perm = transpose_101_perm_0, x = q_33_cast_fp16)[name = tensor("transpose_141")]; + tensor var_4189_cast_fp16 = matmul(transpose_x = var_4189_transpose_x_0, transpose_y = var_4189_transpose_y_0, x = transpose_101, y = transpose_102)[name = tensor("op_4189_cast_fp16")]; + tensor var_4190_to_fp16 = const()[name = tensor("op_4190_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_65_cast_fp16 = mul(x = var_4189_cast_fp16, y = var_4190_to_fp16)[name = tensor("scores_65_cast_fp16")]; + tensor scores_67_cast_fp16 = add(x = scores_65_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_67_cast_fp16")]; + tensor var_4200 = const()[name = tensor("op_4200"), val = tensor(-1)]; + tensor weights_33_cast_fp16 = softmax(axis = var_4200, x = scores_67_cast_fp16)[name = tensor("weights_33_cast_fp16")]; + tensor attn_33_transpose_x_0 = const()[name = tensor("attn_33_transpose_x_0"), val = tensor(false)]; + tensor attn_33_transpose_y_0 = const()[name = tensor("attn_33_transpose_y_0"), val = tensor(false)]; + tensor values_33_cast_fp16 = transpose(perm = var_4178, x = new_v_cache_33_cast_fp16)[name = tensor("transpose_142")]; + tensor attn_33_cast_fp16 = matmul(transpose_x = attn_33_transpose_x_0, transpose_y = attn_33_transpose_y_0, x = weights_33_cast_fp16, y = values_33_cast_fp16)[name = tensor("attn_33_cast_fp16")]; + tensor var_4208 = const()[name = tensor("op_4208"), val = tensor([0, 2, 1, 3])]; + tensor var_4213 = const()[name = tensor("op_4213"), val = tensor([1, 256, 1024])]; + tensor var_4209_cast_fp16 = transpose(perm = var_4208, x = attn_33_cast_fp16)[name = tensor("transpose_139")]; + tensor input_197_cast_fp16 = reshape(shape = var_4213, x = var_4209_cast_fp16)[name = tensor("input_197_cast_fp16")]; + tensor attn16_out_proj_weight_to_fp16 = const()[name = tensor("attn16_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(409107200)))]; + tensor linear_65_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn16_out_proj_weight_to_fp16, x = input_197_cast_fp16)[name = tensor("linear_65_cast_fp16")]; + tensor position16_to_fp16_dtype_0 = const()[name = tensor("position16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position16_to_fp16 = cast(dtype = position16_to_fp16_dtype_0, x = position16)[name = tensor("cast_365")]; + tensor var_4222_cast_fp16 = add(x = position16_to_fp16, y = valid_len_to_fp16)[name = tensor("op_4222_cast_fp16")]; + tensor var_4222_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4222_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_199_cast_fp16 = add(x = input_193_cast_fp16, y = linear_65_cast_fp16)[name = tensor("input_199_cast_fp16")]; + tensor input_201_axes_0 = const()[name = tensor("input_201_axes_0"), val = tensor([-1])]; + tensor norm16_2_weight_to_fp16 = const()[name = tensor("norm16_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411204416)))]; + tensor norm16_2_bias_to_fp16 = const()[name = tensor("norm16_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411206528)))]; + tensor var_4226_to_fp16 = const()[name = tensor("op_4226_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_201_cast_fp16 = layer_norm(axes = input_201_axes_0, beta = norm16_2_bias_to_fp16, epsilon = var_4226_to_fp16, gamma = norm16_2_weight_to_fp16, x = input_199_cast_fp16)[name = tensor("input_201_cast_fp16")]; + tensor linear16_1_weight_to_fp16 = const()[name = tensor("linear16_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411208640)))]; + tensor linear_66_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear16_1_weight_to_fp16, x = input_201_cast_fp16)[name = tensor("linear_66_cast_fp16")]; + tensor input_203_mode_0 = const()[name = tensor("input_203_mode_0"), val = tensor("EXACT")]; + tensor input_203_cast_fp16 = gelu(mode = input_203_mode_0, x = linear_66_cast_fp16)[name = tensor("input_203_cast_fp16")]; + tensor linear16_2_weight_to_fp16 = const()[name = tensor("linear16_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(419597312)))]; + tensor linear_67_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear16_2_weight_to_fp16, x = input_203_cast_fp16)[name = tensor("linear_67_cast_fp16")]; + tensor input_205_cast_fp16 = add(x = input_199_cast_fp16, y = linear_67_cast_fp16)[name = tensor("input_205_cast_fp16")]; + tensor input_207_axes_0 = const()[name = tensor("input_207_axes_0"), val = tensor([-1])]; + tensor norm17_1_weight_to_fp16 = const()[name = tensor("norm17_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(427985984)))]; + tensor norm17_1_bias_to_fp16 = const()[name = tensor("norm17_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(427988096)))]; + tensor var_4243_to_fp16 = const()[name = tensor("op_4243_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_207_cast_fp16 = layer_norm(axes = input_207_axes_0, beta = norm17_1_bias_to_fp16, epsilon = var_4243_to_fp16, gamma = norm17_1_weight_to_fp16, x = input_205_cast_fp16)[name = tensor("input_207_cast_fp16")]; + tensor attn17_in_proj_weight_to_fp16 = const()[name = tensor("attn17_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(427990208)))]; + tensor linear_68_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn17_in_proj_weight_to_fp16, x = input_207_cast_fp16)[name = tensor("linear_68_cast_fp16")]; + tensor var_4257 = const()[name = tensor("op_4257"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_35_cast_fp16 = reshape(shape = var_4257, x = linear_68_cast_fp16)[name = tensor("qkv_35_cast_fp16")]; + tensor x_69_begin_0 = const()[name = tensor("x_69_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_69_end_0 = const()[name = tensor("x_69_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_69_end_mask_0 = const()[name = tensor("x_69_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_69_squeeze_mask_0 = const()[name = tensor("x_69_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_69_cast_fp16 = slice_by_index(begin = x_69_begin_0, end = x_69_end_0, end_mask = x_69_end_mask_0, squeeze_mask = x_69_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("x_69_cast_fp16")]; + tensor x_71_begin_0 = const()[name = tensor("x_71_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_71_end_0 = const()[name = tensor("x_71_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_71_end_mask_0 = const()[name = tensor("x_71_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_71_squeeze_mask_0 = const()[name = tensor("x_71_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_71_cast_fp16 = slice_by_index(begin = x_71_begin_0, end = x_71_end_0, end_mask = x_71_end_mask_0, squeeze_mask = x_71_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("x_71_cast_fp16")]; + tensor v_35_begin_0 = const()[name = tensor("v_35_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_35_end_0 = const()[name = tensor("v_35_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_35_end_mask_0 = const()[name = tensor("v_35_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_35_squeeze_mask_0 = const()[name = tensor("v_35_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_35_cast_fp16 = slice_by_index(begin = v_35_begin_0, end = v_35_end_0, end_mask = v_35_end_mask_0, squeeze_mask = v_35_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("v_35_cast_fp16")]; + tensor var_4302 = const()[name = tensor("op_4302"), val = tensor([256, 16, 32, 2])]; + tensor pairs_69_cast_fp16 = reshape(shape = var_4302, x = x_69_cast_fp16)[name = tensor("pairs_69_cast_fp16")]; + tensor xr_69_begin_0 = const()[name = tensor("xr_69_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_69_end_0 = const()[name = tensor("xr_69_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_69_end_mask_0 = const()[name = tensor("xr_69_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_69_squeeze_mask_0 = const()[name = tensor("xr_69_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_69_cast_fp16 = slice_by_index(begin = xr_69_begin_0, end = xr_69_end_0, end_mask = xr_69_end_mask_0, squeeze_mask = xr_69_squeeze_mask_0, x = pairs_69_cast_fp16)[name = tensor("xr_69_cast_fp16")]; + tensor xi_69_begin_0 = const()[name = tensor("xi_69_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_69_end_0 = const()[name = tensor("xi_69_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_69_end_mask_0 = const()[name = tensor("xi_69_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_69_squeeze_mask_0 = const()[name = tensor("xi_69_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_69_cast_fp16 = slice_by_index(begin = xi_69_begin_0, end = xi_69_end_0, end_mask = xi_69_end_mask_0, squeeze_mask = xi_69_squeeze_mask_0, x = pairs_69_cast_fp16)[name = tensor("xi_69_cast_fp16")]; + tensor var_4310_cast_fp16 = mul(x = xr_69_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4310_cast_fp16")]; + tensor var_4311_cast_fp16 = mul(x = xi_69_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4311_cast_fp16")]; + tensor xor__69_cast_fp16 = sub(x = var_4310_cast_fp16, y = var_4311_cast_fp16)[name = tensor("xor__69_cast_fp16")]; + tensor var_4314_cast_fp16 = mul(x = xr_69_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4314_cast_fp16")]; + tensor var_4315_cast_fp16 = mul(x = xi_69_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4315_cast_fp16")]; + tensor xoi_69_cast_fp16 = add(x = var_4314_cast_fp16, y = var_4315_cast_fp16)[name = tensor("xoi_69_cast_fp16")]; + tensor out_69_axis_0 = const()[name = tensor("out_69_axis_0"), val = tensor(-1)]; + tensor out_69_cast_fp16 = stack(axis = out_69_axis_0, values = (xor__69_cast_fp16, xoi_69_cast_fp16))[name = tensor("out_69_cast_fp16")]; + tensor var_4325 = const()[name = tensor("op_4325"), val = tensor([1, 256, 16, 64])]; + tensor q_35_cast_fp16 = reshape(shape = var_4325, x = out_69_cast_fp16)[name = tensor("q_35_cast_fp16")]; + tensor var_4331 = const()[name = tensor("op_4331"), val = tensor([256, 16, 32, 2])]; + tensor pairs_71_cast_fp16 = reshape(shape = var_4331, x = x_71_cast_fp16)[name = tensor("pairs_71_cast_fp16")]; + tensor xr_71_begin_0 = const()[name = tensor("xr_71_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_71_end_0 = const()[name = tensor("xr_71_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_71_end_mask_0 = const()[name = tensor("xr_71_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_71_squeeze_mask_0 = const()[name = tensor("xr_71_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_71_cast_fp16 = slice_by_index(begin = xr_71_begin_0, end = xr_71_end_0, end_mask = xr_71_end_mask_0, squeeze_mask = xr_71_squeeze_mask_0, x = pairs_71_cast_fp16)[name = tensor("xr_71_cast_fp16")]; + tensor xi_71_begin_0 = const()[name = tensor("xi_71_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_71_end_0 = const()[name = tensor("xi_71_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_71_end_mask_0 = const()[name = tensor("xi_71_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_71_squeeze_mask_0 = const()[name = tensor("xi_71_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_71_cast_fp16 = slice_by_index(begin = xi_71_begin_0, end = xi_71_end_0, end_mask = xi_71_end_mask_0, squeeze_mask = xi_71_squeeze_mask_0, x = pairs_71_cast_fp16)[name = tensor("xi_71_cast_fp16")]; + tensor var_4339_cast_fp16 = mul(x = xr_71_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4339_cast_fp16")]; + tensor var_4340_cast_fp16 = mul(x = xi_71_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4340_cast_fp16")]; + tensor xor__71_cast_fp16 = sub(x = var_4339_cast_fp16, y = var_4340_cast_fp16)[name = tensor("xor__71_cast_fp16")]; + tensor var_4343_cast_fp16 = mul(x = xr_71_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4343_cast_fp16")]; + tensor var_4344_cast_fp16 = mul(x = xi_71_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4344_cast_fp16")]; + tensor xoi_71_cast_fp16 = add(x = var_4343_cast_fp16, y = var_4344_cast_fp16)[name = tensor("xoi_71_cast_fp16")]; + tensor out_71_axis_0 = const()[name = tensor("out_71_axis_0"), val = tensor(-1)]; + tensor out_71_cast_fp16 = stack(axis = out_71_axis_0, values = (xor__71_cast_fp16, xoi_71_cast_fp16))[name = tensor("out_71_cast_fp16")]; + tensor var_4361 = const()[name = tensor("op_4361"), val = tensor([256, 1024])]; + tensor var_4362_cast_fp16 = reshape(shape = var_4361, x = out_71_cast_fp16)[name = tensor("op_4362_cast_fp16")]; + tensor written_k_35_transpose_x_1 = const()[name = tensor("written_k_35_transpose_x_1"), val = tensor(true)]; + tensor written_k_35_transpose_y_1 = const()[name = tensor("written_k_35_transpose_y_1"), val = tensor(false)]; + tensor written_k_35_cast_fp16 = matmul(transpose_x = written_k_35_transpose_x_1, transpose_y = written_k_35_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_4362_cast_fp16)[name = tensor("written_k_35_cast_fp16")]; + tensor var_4369 = const()[name = tensor("op_4369"), val = tensor([256, 1024])]; + tensor var_4370_cast_fp16 = reshape(shape = var_4369, x = v_35_cast_fp16)[name = tensor("op_4370_cast_fp16")]; + tensor written_v_35_transpose_x_1 = const()[name = tensor("written_v_35_transpose_x_1"), val = tensor(true)]; + tensor written_v_35_transpose_y_1 = const()[name = tensor("written_v_35_transpose_y_1"), val = tensor(false)]; + tensor written_v_35_cast_fp16 = matmul(transpose_x = written_v_35_transpose_x_1, transpose_y = written_v_35_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_4370_cast_fp16)[name = tensor("written_v_35_cast_fp16")]; + tensor k_cache17_to_fp16_dtype_0 = const()[name = tensor("k_cache17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache17_to_fp16 = cast(dtype = k_cache17_to_fp16_dtype_0, x = k_cache17)[name = tensor("cast_363")]; + tensor var_4381_cast_fp16 = mul(x = k_cache17_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_4381_cast_fp16")]; + tensor var_4386 = const()[name = tensor("op_4386"), val = tensor([1, 512, 16, 64])]; + tensor var_4387_cast_fp16 = reshape(shape = var_4386, x = written_k_35_cast_fp16)[name = tensor("op_4387_cast_fp16")]; + tensor new_k_cache_35_cast_fp16 = add(x = var_4381_cast_fp16, y = var_4387_cast_fp16)[name = tensor("new_k_cache_35_cast_fp16")]; + tensor new_k_cache_35_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_35_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache17_to_fp16_dtype_0 = const()[name = tensor("v_cache17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache17_to_fp16 = cast(dtype = v_cache17_to_fp16_dtype_0, x = v_cache17)[name = tensor("cast_361")]; + tensor var_4390_cast_fp16 = mul(x = v_cache17_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_4390_cast_fp16")]; + tensor var_4395 = const()[name = tensor("op_4395"), val = tensor([1, 512, 16, 64])]; + tensor var_4396_cast_fp16 = reshape(shape = var_4395, x = written_v_35_cast_fp16)[name = tensor("op_4396_cast_fp16")]; + tensor new_v_cache_35_cast_fp16 = add(x = var_4390_cast_fp16, y = var_4396_cast_fp16)[name = tensor("new_v_cache_35_cast_fp16")]; + tensor new_v_cache_35_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_35_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_4409 = const()[name = tensor("op_4409"), val = tensor([0, 2, 1, 3])]; + tensor var_4420_transpose_x_0 = const()[name = tensor("op_4420_transpose_x_0"), val = tensor(false)]; + tensor var_4420_transpose_y_0 = const()[name = tensor("op_4420_transpose_y_0"), val = tensor(false)]; + tensor transpose_103_perm_0 = const()[name = tensor("transpose_103_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_104_perm_0 = const()[name = tensor("transpose_104_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_104 = transpose(perm = transpose_104_perm_0, x = new_k_cache_35_cast_fp16)[name = tensor("transpose_136")]; + tensor transpose_103 = transpose(perm = transpose_103_perm_0, x = q_35_cast_fp16)[name = tensor("transpose_137")]; + tensor var_4420_cast_fp16 = matmul(transpose_x = var_4420_transpose_x_0, transpose_y = var_4420_transpose_y_0, x = transpose_103, y = transpose_104)[name = tensor("op_4420_cast_fp16")]; + tensor var_4421_to_fp16 = const()[name = tensor("op_4421_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_69_cast_fp16 = mul(x = var_4420_cast_fp16, y = var_4421_to_fp16)[name = tensor("scores_69_cast_fp16")]; + tensor scores_71_cast_fp16 = add(x = scores_69_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_71_cast_fp16")]; + tensor var_4431 = const()[name = tensor("op_4431"), val = tensor(-1)]; + tensor weights_35_cast_fp16 = softmax(axis = var_4431, x = scores_71_cast_fp16)[name = tensor("weights_35_cast_fp16")]; + tensor attn_35_transpose_x_0 = const()[name = tensor("attn_35_transpose_x_0"), val = tensor(false)]; + tensor attn_35_transpose_y_0 = const()[name = tensor("attn_35_transpose_y_0"), val = tensor(false)]; + tensor values_35_cast_fp16 = transpose(perm = var_4409, x = new_v_cache_35_cast_fp16)[name = tensor("transpose_138")]; + tensor attn_35_cast_fp16 = matmul(transpose_x = attn_35_transpose_x_0, transpose_y = attn_35_transpose_y_0, x = weights_35_cast_fp16, y = values_35_cast_fp16)[name = tensor("attn_35_cast_fp16")]; + tensor var_4439 = const()[name = tensor("op_4439"), val = tensor([0, 2, 1, 3])]; + tensor var_4444 = const()[name = tensor("op_4444"), val = tensor([1, 256, 1024])]; + tensor var_4440_cast_fp16 = transpose(perm = var_4439, x = attn_35_cast_fp16)[name = tensor("transpose_135")]; + tensor input_209_cast_fp16 = reshape(shape = var_4444, x = var_4440_cast_fp16)[name = tensor("input_209_cast_fp16")]; + tensor attn17_out_proj_weight_to_fp16 = const()[name = tensor("attn17_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(434281728)))]; + tensor linear_69_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn17_out_proj_weight_to_fp16, x = input_209_cast_fp16)[name = tensor("linear_69_cast_fp16")]; + tensor position17_to_fp16_dtype_0 = const()[name = tensor("position17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position17_to_fp16 = cast(dtype = position17_to_fp16_dtype_0, x = position17)[name = tensor("cast_359")]; + tensor var_4453_cast_fp16 = add(x = position17_to_fp16, y = valid_len_to_fp16)[name = tensor("op_4453_cast_fp16")]; + tensor var_4453_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4453_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_211_cast_fp16 = add(x = input_205_cast_fp16, y = linear_69_cast_fp16)[name = tensor("input_211_cast_fp16")]; + tensor input_213_axes_0 = const()[name = tensor("input_213_axes_0"), val = tensor([-1])]; + tensor norm17_2_weight_to_fp16 = const()[name = tensor("norm17_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(436378944)))]; + tensor norm17_2_bias_to_fp16 = const()[name = tensor("norm17_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(436381056)))]; + tensor var_4457_to_fp16 = const()[name = tensor("op_4457_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_213_cast_fp16 = layer_norm(axes = input_213_axes_0, beta = norm17_2_bias_to_fp16, epsilon = var_4457_to_fp16, gamma = norm17_2_weight_to_fp16, x = input_211_cast_fp16)[name = tensor("input_213_cast_fp16")]; + tensor linear17_1_weight_to_fp16 = const()[name = tensor("linear17_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(436383168)))]; + tensor linear_70_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear17_1_weight_to_fp16, x = input_213_cast_fp16)[name = tensor("linear_70_cast_fp16")]; + tensor input_215_mode_0 = const()[name = tensor("input_215_mode_0"), val = tensor("EXACT")]; + tensor input_215_cast_fp16 = gelu(mode = input_215_mode_0, x = linear_70_cast_fp16)[name = tensor("input_215_cast_fp16")]; + tensor linear17_2_weight_to_fp16 = const()[name = tensor("linear17_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(444771840)))]; + tensor linear_71_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear17_2_weight_to_fp16, x = input_215_cast_fp16)[name = tensor("linear_71_cast_fp16")]; + tensor input_217_cast_fp16 = add(x = input_211_cast_fp16, y = linear_71_cast_fp16)[name = tensor("input_217_cast_fp16")]; + tensor input_219_axes_0 = const()[name = tensor("input_219_axes_0"), val = tensor([-1])]; + tensor norm18_1_weight_to_fp16 = const()[name = tensor("norm18_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(453160512)))]; + tensor norm18_1_bias_to_fp16 = const()[name = tensor("norm18_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(453162624)))]; + tensor var_4474_to_fp16 = const()[name = tensor("op_4474_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_219_cast_fp16 = layer_norm(axes = input_219_axes_0, beta = norm18_1_bias_to_fp16, epsilon = var_4474_to_fp16, gamma = norm18_1_weight_to_fp16, x = input_217_cast_fp16)[name = tensor("input_219_cast_fp16")]; + tensor attn18_in_proj_weight_to_fp16 = const()[name = tensor("attn18_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(453164736)))]; + tensor linear_72_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn18_in_proj_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_72_cast_fp16")]; + tensor var_4488 = const()[name = tensor("op_4488"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_37_cast_fp16 = reshape(shape = var_4488, x = linear_72_cast_fp16)[name = tensor("qkv_37_cast_fp16")]; + tensor x_73_begin_0 = const()[name = tensor("x_73_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_73_end_0 = const()[name = tensor("x_73_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_73_end_mask_0 = const()[name = tensor("x_73_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_73_squeeze_mask_0 = const()[name = tensor("x_73_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_73_cast_fp16 = slice_by_index(begin = x_73_begin_0, end = x_73_end_0, end_mask = x_73_end_mask_0, squeeze_mask = x_73_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("x_73_cast_fp16")]; + tensor x_75_begin_0 = const()[name = tensor("x_75_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_75_end_0 = const()[name = tensor("x_75_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_75_end_mask_0 = const()[name = tensor("x_75_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_75_squeeze_mask_0 = const()[name = tensor("x_75_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_75_cast_fp16 = slice_by_index(begin = x_75_begin_0, end = x_75_end_0, end_mask = x_75_end_mask_0, squeeze_mask = x_75_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("x_75_cast_fp16")]; + tensor v_37_begin_0 = const()[name = tensor("v_37_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_37_end_0 = const()[name = tensor("v_37_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_37_end_mask_0 = const()[name = tensor("v_37_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_37_squeeze_mask_0 = const()[name = tensor("v_37_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_37_cast_fp16 = slice_by_index(begin = v_37_begin_0, end = v_37_end_0, end_mask = v_37_end_mask_0, squeeze_mask = v_37_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("v_37_cast_fp16")]; + tensor var_4533 = const()[name = tensor("op_4533"), val = tensor([256, 16, 32, 2])]; + tensor pairs_73_cast_fp16 = reshape(shape = var_4533, x = x_73_cast_fp16)[name = tensor("pairs_73_cast_fp16")]; + tensor xr_73_begin_0 = const()[name = tensor("xr_73_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_73_end_0 = const()[name = tensor("xr_73_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_73_end_mask_0 = const()[name = tensor("xr_73_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_73_squeeze_mask_0 = const()[name = tensor("xr_73_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_73_cast_fp16 = slice_by_index(begin = xr_73_begin_0, end = xr_73_end_0, end_mask = xr_73_end_mask_0, squeeze_mask = xr_73_squeeze_mask_0, x = pairs_73_cast_fp16)[name = tensor("xr_73_cast_fp16")]; + tensor xi_73_begin_0 = const()[name = tensor("xi_73_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_73_end_0 = const()[name = tensor("xi_73_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_73_end_mask_0 = const()[name = tensor("xi_73_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_73_squeeze_mask_0 = const()[name = tensor("xi_73_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_73_cast_fp16 = slice_by_index(begin = xi_73_begin_0, end = xi_73_end_0, end_mask = xi_73_end_mask_0, squeeze_mask = xi_73_squeeze_mask_0, x = pairs_73_cast_fp16)[name = tensor("xi_73_cast_fp16")]; + tensor var_4541_cast_fp16 = mul(x = xr_73_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4541_cast_fp16")]; + tensor var_4542_cast_fp16 = mul(x = xi_73_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4542_cast_fp16")]; + tensor xor__73_cast_fp16 = sub(x = var_4541_cast_fp16, y = var_4542_cast_fp16)[name = tensor("xor__73_cast_fp16")]; + tensor var_4545_cast_fp16 = mul(x = xr_73_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4545_cast_fp16")]; + tensor var_4546_cast_fp16 = mul(x = xi_73_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4546_cast_fp16")]; + tensor xoi_73_cast_fp16 = add(x = var_4545_cast_fp16, y = var_4546_cast_fp16)[name = tensor("xoi_73_cast_fp16")]; + tensor out_73_axis_0 = const()[name = tensor("out_73_axis_0"), val = tensor(-1)]; + tensor out_73_cast_fp16 = stack(axis = out_73_axis_0, values = (xor__73_cast_fp16, xoi_73_cast_fp16))[name = tensor("out_73_cast_fp16")]; + tensor var_4556 = const()[name = tensor("op_4556"), val = tensor([1, 256, 16, 64])]; + tensor q_37_cast_fp16 = reshape(shape = var_4556, x = out_73_cast_fp16)[name = tensor("q_37_cast_fp16")]; + tensor var_4562 = const()[name = tensor("op_4562"), val = tensor([256, 16, 32, 2])]; + tensor pairs_75_cast_fp16 = reshape(shape = var_4562, x = x_75_cast_fp16)[name = tensor("pairs_75_cast_fp16")]; + tensor xr_75_begin_0 = const()[name = tensor("xr_75_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_75_end_0 = const()[name = tensor("xr_75_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_75_end_mask_0 = const()[name = tensor("xr_75_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_75_squeeze_mask_0 = const()[name = tensor("xr_75_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_75_cast_fp16 = slice_by_index(begin = xr_75_begin_0, end = xr_75_end_0, end_mask = xr_75_end_mask_0, squeeze_mask = xr_75_squeeze_mask_0, x = pairs_75_cast_fp16)[name = tensor("xr_75_cast_fp16")]; + tensor xi_75_begin_0 = const()[name = tensor("xi_75_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_75_end_0 = const()[name = tensor("xi_75_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_75_end_mask_0 = const()[name = tensor("xi_75_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_75_squeeze_mask_0 = const()[name = tensor("xi_75_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_75_cast_fp16 = slice_by_index(begin = xi_75_begin_0, end = xi_75_end_0, end_mask = xi_75_end_mask_0, squeeze_mask = xi_75_squeeze_mask_0, x = pairs_75_cast_fp16)[name = tensor("xi_75_cast_fp16")]; + tensor var_4570_cast_fp16 = mul(x = xr_75_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4570_cast_fp16")]; + tensor var_4571_cast_fp16 = mul(x = xi_75_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4571_cast_fp16")]; + tensor xor__75_cast_fp16 = sub(x = var_4570_cast_fp16, y = var_4571_cast_fp16)[name = tensor("xor__75_cast_fp16")]; + tensor var_4574_cast_fp16 = mul(x = xr_75_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4574_cast_fp16")]; + tensor var_4575_cast_fp16 = mul(x = xi_75_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4575_cast_fp16")]; + tensor xoi_75_cast_fp16 = add(x = var_4574_cast_fp16, y = var_4575_cast_fp16)[name = tensor("xoi_75_cast_fp16")]; + tensor out_75_axis_0 = const()[name = tensor("out_75_axis_0"), val = tensor(-1)]; + tensor out_75_cast_fp16 = stack(axis = out_75_axis_0, values = (xor__75_cast_fp16, xoi_75_cast_fp16))[name = tensor("out_75_cast_fp16")]; + tensor var_4592 = const()[name = tensor("op_4592"), val = tensor([256, 1024])]; + tensor var_4593_cast_fp16 = reshape(shape = var_4592, x = out_75_cast_fp16)[name = tensor("op_4593_cast_fp16")]; + tensor written_k_37_transpose_x_1 = const()[name = tensor("written_k_37_transpose_x_1"), val = tensor(true)]; + tensor written_k_37_transpose_y_1 = const()[name = tensor("written_k_37_transpose_y_1"), val = tensor(false)]; + tensor written_k_37_cast_fp16 = matmul(transpose_x = written_k_37_transpose_x_1, transpose_y = written_k_37_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_4593_cast_fp16)[name = tensor("written_k_37_cast_fp16")]; + tensor var_4600 = const()[name = tensor("op_4600"), val = tensor([256, 1024])]; + tensor var_4601_cast_fp16 = reshape(shape = var_4600, x = v_37_cast_fp16)[name = tensor("op_4601_cast_fp16")]; + tensor written_v_37_transpose_x_1 = const()[name = tensor("written_v_37_transpose_x_1"), val = tensor(true)]; + tensor written_v_37_transpose_y_1 = const()[name = tensor("written_v_37_transpose_y_1"), val = tensor(false)]; + tensor written_v_37_cast_fp16 = matmul(transpose_x = written_v_37_transpose_x_1, transpose_y = written_v_37_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_4601_cast_fp16)[name = tensor("written_v_37_cast_fp16")]; + tensor k_cache18_to_fp16_dtype_0 = const()[name = tensor("k_cache18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache18_to_fp16 = cast(dtype = k_cache18_to_fp16_dtype_0, x = k_cache18)[name = tensor("cast_357")]; + tensor var_4612_cast_fp16 = mul(x = k_cache18_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_4612_cast_fp16")]; + tensor var_4617 = const()[name = tensor("op_4617"), val = tensor([1, 512, 16, 64])]; + tensor var_4618_cast_fp16 = reshape(shape = var_4617, x = written_k_37_cast_fp16)[name = tensor("op_4618_cast_fp16")]; + tensor new_k_cache_37_cast_fp16 = add(x = var_4612_cast_fp16, y = var_4618_cast_fp16)[name = tensor("new_k_cache_37_cast_fp16")]; + tensor new_k_cache_37_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_37_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache18_to_fp16_dtype_0 = const()[name = tensor("v_cache18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache18_to_fp16 = cast(dtype = v_cache18_to_fp16_dtype_0, x = v_cache18)[name = tensor("cast_355")]; + tensor var_4621_cast_fp16 = mul(x = v_cache18_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_4621_cast_fp16")]; + tensor var_4626 = const()[name = tensor("op_4626"), val = tensor([1, 512, 16, 64])]; + tensor var_4627_cast_fp16 = reshape(shape = var_4626, x = written_v_37_cast_fp16)[name = tensor("op_4627_cast_fp16")]; + tensor new_v_cache_37_cast_fp16 = add(x = var_4621_cast_fp16, y = var_4627_cast_fp16)[name = tensor("new_v_cache_37_cast_fp16")]; + tensor new_v_cache_37_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_37_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_4640 = const()[name = tensor("op_4640"), val = tensor([0, 2, 1, 3])]; + tensor var_4651_transpose_x_0 = const()[name = tensor("op_4651_transpose_x_0"), val = tensor(false)]; + tensor var_4651_transpose_y_0 = const()[name = tensor("op_4651_transpose_y_0"), val = tensor(false)]; + tensor transpose_105_perm_0 = const()[name = tensor("transpose_105_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_106_perm_0 = const()[name = tensor("transpose_106_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_106 = transpose(perm = transpose_106_perm_0, x = new_k_cache_37_cast_fp16)[name = tensor("transpose_132")]; + tensor transpose_105 = transpose(perm = transpose_105_perm_0, x = q_37_cast_fp16)[name = tensor("transpose_133")]; + tensor var_4651_cast_fp16 = matmul(transpose_x = var_4651_transpose_x_0, transpose_y = var_4651_transpose_y_0, x = transpose_105, y = transpose_106)[name = tensor("op_4651_cast_fp16")]; + tensor var_4652_to_fp16 = const()[name = tensor("op_4652_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_73_cast_fp16 = mul(x = var_4651_cast_fp16, y = var_4652_to_fp16)[name = tensor("scores_73_cast_fp16")]; + tensor scores_75_cast_fp16 = add(x = scores_73_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_75_cast_fp16")]; + tensor var_4662 = const()[name = tensor("op_4662"), val = tensor(-1)]; + tensor weights_37_cast_fp16 = softmax(axis = var_4662, x = scores_75_cast_fp16)[name = tensor("weights_37_cast_fp16")]; + tensor attn_37_transpose_x_0 = const()[name = tensor("attn_37_transpose_x_0"), val = tensor(false)]; + tensor attn_37_transpose_y_0 = const()[name = tensor("attn_37_transpose_y_0"), val = tensor(false)]; + tensor values_37_cast_fp16 = transpose(perm = var_4640, x = new_v_cache_37_cast_fp16)[name = tensor("transpose_134")]; + tensor attn_37_cast_fp16 = matmul(transpose_x = attn_37_transpose_x_0, transpose_y = attn_37_transpose_y_0, x = weights_37_cast_fp16, y = values_37_cast_fp16)[name = tensor("attn_37_cast_fp16")]; + tensor var_4670 = const()[name = tensor("op_4670"), val = tensor([0, 2, 1, 3])]; + tensor var_4675 = const()[name = tensor("op_4675"), val = tensor([1, 256, 1024])]; + tensor var_4671_cast_fp16 = transpose(perm = var_4670, x = attn_37_cast_fp16)[name = tensor("transpose_131")]; + tensor input_221_cast_fp16 = reshape(shape = var_4675, x = var_4671_cast_fp16)[name = tensor("input_221_cast_fp16")]; + tensor attn18_out_proj_weight_to_fp16 = const()[name = tensor("attn18_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(459456256)))]; + tensor linear_73_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn18_out_proj_weight_to_fp16, x = input_221_cast_fp16)[name = tensor("linear_73_cast_fp16")]; + tensor position18_to_fp16_dtype_0 = const()[name = tensor("position18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position18_to_fp16 = cast(dtype = position18_to_fp16_dtype_0, x = position18)[name = tensor("cast_353")]; + tensor var_4684_cast_fp16 = add(x = position18_to_fp16, y = valid_len_to_fp16)[name = tensor("op_4684_cast_fp16")]; + tensor var_4684_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4684_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_223_cast_fp16 = add(x = input_217_cast_fp16, y = linear_73_cast_fp16)[name = tensor("input_223_cast_fp16")]; + tensor input_225_axes_0 = const()[name = tensor("input_225_axes_0"), val = tensor([-1])]; + tensor norm18_2_weight_to_fp16 = const()[name = tensor("norm18_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(461553472)))]; + tensor norm18_2_bias_to_fp16 = const()[name = tensor("norm18_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(461555584)))]; + tensor var_4688_to_fp16 = const()[name = tensor("op_4688_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_225_cast_fp16 = layer_norm(axes = input_225_axes_0, beta = norm18_2_bias_to_fp16, epsilon = var_4688_to_fp16, gamma = norm18_2_weight_to_fp16, x = input_223_cast_fp16)[name = tensor("input_225_cast_fp16")]; + tensor linear18_1_weight_to_fp16 = const()[name = tensor("linear18_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(461557696)))]; + tensor linear_74_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear18_1_weight_to_fp16, x = input_225_cast_fp16)[name = tensor("linear_74_cast_fp16")]; + tensor input_227_mode_0 = const()[name = tensor("input_227_mode_0"), val = tensor("EXACT")]; + tensor input_227_cast_fp16 = gelu(mode = input_227_mode_0, x = linear_74_cast_fp16)[name = tensor("input_227_cast_fp16")]; + tensor linear18_2_weight_to_fp16 = const()[name = tensor("linear18_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(469946368)))]; + tensor linear_75_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear18_2_weight_to_fp16, x = input_227_cast_fp16)[name = tensor("linear_75_cast_fp16")]; + tensor input_229_cast_fp16 = add(x = input_223_cast_fp16, y = linear_75_cast_fp16)[name = tensor("input_229_cast_fp16")]; + tensor input_231_axes_0 = const()[name = tensor("input_231_axes_0"), val = tensor([-1])]; + tensor norm19_1_weight_to_fp16 = const()[name = tensor("norm19_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(478335040)))]; + tensor norm19_1_bias_to_fp16 = const()[name = tensor("norm19_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(478337152)))]; + tensor var_4705_to_fp16 = const()[name = tensor("op_4705_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_231_cast_fp16 = layer_norm(axes = input_231_axes_0, beta = norm19_1_bias_to_fp16, epsilon = var_4705_to_fp16, gamma = norm19_1_weight_to_fp16, x = input_229_cast_fp16)[name = tensor("input_231_cast_fp16")]; + tensor attn19_in_proj_weight_to_fp16 = const()[name = tensor("attn19_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(478339264)))]; + tensor linear_76_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn19_in_proj_weight_to_fp16, x = input_231_cast_fp16)[name = tensor("linear_76_cast_fp16")]; + tensor var_4719 = const()[name = tensor("op_4719"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_39_cast_fp16 = reshape(shape = var_4719, x = linear_76_cast_fp16)[name = tensor("qkv_39_cast_fp16")]; + tensor x_77_begin_0 = const()[name = tensor("x_77_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_77_end_0 = const()[name = tensor("x_77_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_77_end_mask_0 = const()[name = tensor("x_77_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_77_squeeze_mask_0 = const()[name = tensor("x_77_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_77_cast_fp16 = slice_by_index(begin = x_77_begin_0, end = x_77_end_0, end_mask = x_77_end_mask_0, squeeze_mask = x_77_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("x_77_cast_fp16")]; + tensor x_79_begin_0 = const()[name = tensor("x_79_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_79_end_0 = const()[name = tensor("x_79_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_79_end_mask_0 = const()[name = tensor("x_79_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_79_squeeze_mask_0 = const()[name = tensor("x_79_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_79_cast_fp16 = slice_by_index(begin = x_79_begin_0, end = x_79_end_0, end_mask = x_79_end_mask_0, squeeze_mask = x_79_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("x_79_cast_fp16")]; + tensor v_39_begin_0 = const()[name = tensor("v_39_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_39_end_0 = const()[name = tensor("v_39_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_39_end_mask_0 = const()[name = tensor("v_39_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_39_squeeze_mask_0 = const()[name = tensor("v_39_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_39_cast_fp16 = slice_by_index(begin = v_39_begin_0, end = v_39_end_0, end_mask = v_39_end_mask_0, squeeze_mask = v_39_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("v_39_cast_fp16")]; + tensor var_4764 = const()[name = tensor("op_4764"), val = tensor([256, 16, 32, 2])]; + tensor pairs_77_cast_fp16 = reshape(shape = var_4764, x = x_77_cast_fp16)[name = tensor("pairs_77_cast_fp16")]; + tensor xr_77_begin_0 = const()[name = tensor("xr_77_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_77_end_0 = const()[name = tensor("xr_77_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_77_end_mask_0 = const()[name = tensor("xr_77_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_77_squeeze_mask_0 = const()[name = tensor("xr_77_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_77_cast_fp16 = slice_by_index(begin = xr_77_begin_0, end = xr_77_end_0, end_mask = xr_77_end_mask_0, squeeze_mask = xr_77_squeeze_mask_0, x = pairs_77_cast_fp16)[name = tensor("xr_77_cast_fp16")]; + tensor xi_77_begin_0 = const()[name = tensor("xi_77_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_77_end_0 = const()[name = tensor("xi_77_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_77_end_mask_0 = const()[name = tensor("xi_77_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_77_squeeze_mask_0 = const()[name = tensor("xi_77_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_77_cast_fp16 = slice_by_index(begin = xi_77_begin_0, end = xi_77_end_0, end_mask = xi_77_end_mask_0, squeeze_mask = xi_77_squeeze_mask_0, x = pairs_77_cast_fp16)[name = tensor("xi_77_cast_fp16")]; + tensor var_4772_cast_fp16 = mul(x = xr_77_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4772_cast_fp16")]; + tensor var_4773_cast_fp16 = mul(x = xi_77_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4773_cast_fp16")]; + tensor xor__77_cast_fp16 = sub(x = var_4772_cast_fp16, y = var_4773_cast_fp16)[name = tensor("xor__77_cast_fp16")]; + tensor var_4776_cast_fp16 = mul(x = xr_77_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4776_cast_fp16")]; + tensor var_4777_cast_fp16 = mul(x = xi_77_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4777_cast_fp16")]; + tensor xoi_77_cast_fp16 = add(x = var_4776_cast_fp16, y = var_4777_cast_fp16)[name = tensor("xoi_77_cast_fp16")]; + tensor out_77_axis_0 = const()[name = tensor("out_77_axis_0"), val = tensor(-1)]; + tensor out_77_cast_fp16 = stack(axis = out_77_axis_0, values = (xor__77_cast_fp16, xoi_77_cast_fp16))[name = tensor("out_77_cast_fp16")]; + tensor var_4787 = const()[name = tensor("op_4787"), val = tensor([1, 256, 16, 64])]; + tensor q_39_cast_fp16 = reshape(shape = var_4787, x = out_77_cast_fp16)[name = tensor("q_39_cast_fp16")]; + tensor var_4793 = const()[name = tensor("op_4793"), val = tensor([256, 16, 32, 2])]; + tensor pairs_79_cast_fp16 = reshape(shape = var_4793, x = x_79_cast_fp16)[name = tensor("pairs_79_cast_fp16")]; + tensor xr_79_begin_0 = const()[name = tensor("xr_79_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_79_end_0 = const()[name = tensor("xr_79_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_79_end_mask_0 = const()[name = tensor("xr_79_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_79_squeeze_mask_0 = const()[name = tensor("xr_79_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_79_cast_fp16 = slice_by_index(begin = xr_79_begin_0, end = xr_79_end_0, end_mask = xr_79_end_mask_0, squeeze_mask = xr_79_squeeze_mask_0, x = pairs_79_cast_fp16)[name = tensor("xr_79_cast_fp16")]; + tensor xi_79_begin_0 = const()[name = tensor("xi_79_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_79_end_0 = const()[name = tensor("xi_79_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_79_end_mask_0 = const()[name = tensor("xi_79_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_79_squeeze_mask_0 = const()[name = tensor("xi_79_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_79_cast_fp16 = slice_by_index(begin = xi_79_begin_0, end = xi_79_end_0, end_mask = xi_79_end_mask_0, squeeze_mask = xi_79_squeeze_mask_0, x = pairs_79_cast_fp16)[name = tensor("xi_79_cast_fp16")]; + tensor var_4801_cast_fp16 = mul(x = xr_79_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4801_cast_fp16")]; + tensor var_4802_cast_fp16 = mul(x = xi_79_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4802_cast_fp16")]; + tensor xor__79_cast_fp16 = sub(x = var_4801_cast_fp16, y = var_4802_cast_fp16)[name = tensor("xor__79_cast_fp16")]; + tensor var_4805_cast_fp16 = mul(x = xr_79_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4805_cast_fp16")]; + tensor var_4806_cast_fp16 = mul(x = xi_79_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4806_cast_fp16")]; + tensor xoi_79_cast_fp16 = add(x = var_4805_cast_fp16, y = var_4806_cast_fp16)[name = tensor("xoi_79_cast_fp16")]; + tensor out_79_axis_0 = const()[name = tensor("out_79_axis_0"), val = tensor(-1)]; + tensor out_79_cast_fp16 = stack(axis = out_79_axis_0, values = (xor__79_cast_fp16, xoi_79_cast_fp16))[name = tensor("out_79_cast_fp16")]; + tensor var_4823 = const()[name = tensor("op_4823"), val = tensor([256, 1024])]; + tensor var_4824_cast_fp16 = reshape(shape = var_4823, x = out_79_cast_fp16)[name = tensor("op_4824_cast_fp16")]; + tensor written_k_39_transpose_x_1 = const()[name = tensor("written_k_39_transpose_x_1"), val = tensor(true)]; + tensor written_k_39_transpose_y_1 = const()[name = tensor("written_k_39_transpose_y_1"), val = tensor(false)]; + tensor written_k_39_cast_fp16 = matmul(transpose_x = written_k_39_transpose_x_1, transpose_y = written_k_39_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_4824_cast_fp16)[name = tensor("written_k_39_cast_fp16")]; + tensor var_4831 = const()[name = tensor("op_4831"), val = tensor([256, 1024])]; + tensor var_4832_cast_fp16 = reshape(shape = var_4831, x = v_39_cast_fp16)[name = tensor("op_4832_cast_fp16")]; + tensor written_v_39_transpose_x_1 = const()[name = tensor("written_v_39_transpose_x_1"), val = tensor(true)]; + tensor written_v_39_transpose_y_1 = const()[name = tensor("written_v_39_transpose_y_1"), val = tensor(false)]; + tensor written_v_39_cast_fp16 = matmul(transpose_x = written_v_39_transpose_x_1, transpose_y = written_v_39_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_4832_cast_fp16)[name = tensor("written_v_39_cast_fp16")]; + tensor k_cache19_to_fp16_dtype_0 = const()[name = tensor("k_cache19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache19_to_fp16 = cast(dtype = k_cache19_to_fp16_dtype_0, x = k_cache19)[name = tensor("cast_351")]; + tensor var_4843_cast_fp16 = mul(x = k_cache19_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_4843_cast_fp16")]; + tensor var_4848 = const()[name = tensor("op_4848"), val = tensor([1, 512, 16, 64])]; + tensor var_4849_cast_fp16 = reshape(shape = var_4848, x = written_k_39_cast_fp16)[name = tensor("op_4849_cast_fp16")]; + tensor new_k_cache_39_cast_fp16 = add(x = var_4843_cast_fp16, y = var_4849_cast_fp16)[name = tensor("new_k_cache_39_cast_fp16")]; + tensor new_k_cache_39_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_39_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache19_to_fp16_dtype_0 = const()[name = tensor("v_cache19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache19_to_fp16 = cast(dtype = v_cache19_to_fp16_dtype_0, x = v_cache19)[name = tensor("cast_349")]; + tensor var_4852_cast_fp16 = mul(x = v_cache19_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_4852_cast_fp16")]; + tensor var_4857 = const()[name = tensor("op_4857"), val = tensor([1, 512, 16, 64])]; + tensor var_4858_cast_fp16 = reshape(shape = var_4857, x = written_v_39_cast_fp16)[name = tensor("op_4858_cast_fp16")]; + tensor new_v_cache_39_cast_fp16 = add(x = var_4852_cast_fp16, y = var_4858_cast_fp16)[name = tensor("new_v_cache_39_cast_fp16")]; + tensor new_v_cache_39_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_39_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_4871 = const()[name = tensor("op_4871"), val = tensor([0, 2, 1, 3])]; + tensor var_4882_transpose_x_0 = const()[name = tensor("op_4882_transpose_x_0"), val = tensor(false)]; + tensor var_4882_transpose_y_0 = const()[name = tensor("op_4882_transpose_y_0"), val = tensor(false)]; + tensor transpose_107_perm_0 = const()[name = tensor("transpose_107_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_108_perm_0 = const()[name = tensor("transpose_108_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_108 = transpose(perm = transpose_108_perm_0, x = new_k_cache_39_cast_fp16)[name = tensor("transpose_128")]; + tensor transpose_107 = transpose(perm = transpose_107_perm_0, x = q_39_cast_fp16)[name = tensor("transpose_129")]; + tensor var_4882_cast_fp16 = matmul(transpose_x = var_4882_transpose_x_0, transpose_y = var_4882_transpose_y_0, x = transpose_107, y = transpose_108)[name = tensor("op_4882_cast_fp16")]; + tensor var_4883_to_fp16 = const()[name = tensor("op_4883_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_77_cast_fp16 = mul(x = var_4882_cast_fp16, y = var_4883_to_fp16)[name = tensor("scores_77_cast_fp16")]; + tensor scores_79_cast_fp16 = add(x = scores_77_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_79_cast_fp16")]; + tensor var_4893 = const()[name = tensor("op_4893"), val = tensor(-1)]; + tensor weights_39_cast_fp16 = softmax(axis = var_4893, x = scores_79_cast_fp16)[name = tensor("weights_39_cast_fp16")]; + tensor attn_39_transpose_x_0 = const()[name = tensor("attn_39_transpose_x_0"), val = tensor(false)]; + tensor attn_39_transpose_y_0 = const()[name = tensor("attn_39_transpose_y_0"), val = tensor(false)]; + tensor values_39_cast_fp16 = transpose(perm = var_4871, x = new_v_cache_39_cast_fp16)[name = tensor("transpose_130")]; + tensor attn_39_cast_fp16 = matmul(transpose_x = attn_39_transpose_x_0, transpose_y = attn_39_transpose_y_0, x = weights_39_cast_fp16, y = values_39_cast_fp16)[name = tensor("attn_39_cast_fp16")]; + tensor var_4901 = const()[name = tensor("op_4901"), val = tensor([0, 2, 1, 3])]; + tensor var_4906 = const()[name = tensor("op_4906"), val = tensor([1, 256, 1024])]; + tensor var_4902_cast_fp16 = transpose(perm = var_4901, x = attn_39_cast_fp16)[name = tensor("transpose_127")]; + tensor input_233_cast_fp16 = reshape(shape = var_4906, x = var_4902_cast_fp16)[name = tensor("input_233_cast_fp16")]; + tensor attn19_out_proj_weight_to_fp16 = const()[name = tensor("attn19_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(484630784)))]; + tensor linear_77_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn19_out_proj_weight_to_fp16, x = input_233_cast_fp16)[name = tensor("linear_77_cast_fp16")]; + tensor position19_to_fp16_dtype_0 = const()[name = tensor("position19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position19_to_fp16 = cast(dtype = position19_to_fp16_dtype_0, x = position19)[name = tensor("cast_347")]; + tensor var_4915_cast_fp16 = add(x = position19_to_fp16, y = valid_len_to_fp16)[name = tensor("op_4915_cast_fp16")]; + tensor var_4915_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4915_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_235_cast_fp16 = add(x = input_229_cast_fp16, y = linear_77_cast_fp16)[name = tensor("input_235_cast_fp16")]; + tensor input_237_axes_0 = const()[name = tensor("input_237_axes_0"), val = tensor([-1])]; + tensor norm19_2_weight_to_fp16 = const()[name = tensor("norm19_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486728000)))]; + tensor norm19_2_bias_to_fp16 = const()[name = tensor("norm19_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486730112)))]; + tensor var_4919_to_fp16 = const()[name = tensor("op_4919_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_237_cast_fp16 = layer_norm(axes = input_237_axes_0, beta = norm19_2_bias_to_fp16, epsilon = var_4919_to_fp16, gamma = norm19_2_weight_to_fp16, x = input_235_cast_fp16)[name = tensor("input_237_cast_fp16")]; + tensor linear19_1_weight_to_fp16 = const()[name = tensor("linear19_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486732224)))]; + tensor linear_78_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear19_1_weight_to_fp16, x = input_237_cast_fp16)[name = tensor("linear_78_cast_fp16")]; + tensor input_239_mode_0 = const()[name = tensor("input_239_mode_0"), val = tensor("EXACT")]; + tensor input_239_cast_fp16 = gelu(mode = input_239_mode_0, x = linear_78_cast_fp16)[name = tensor("input_239_cast_fp16")]; + tensor linear19_2_weight_to_fp16 = const()[name = tensor("linear19_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(495120896)))]; + tensor linear_79_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear19_2_weight_to_fp16, x = input_239_cast_fp16)[name = tensor("linear_79_cast_fp16")]; + tensor input_241_cast_fp16 = add(x = input_235_cast_fp16, y = linear_79_cast_fp16)[name = tensor("input_241_cast_fp16")]; + tensor input_243_axes_0 = const()[name = tensor("input_243_axes_0"), val = tensor([-1])]; + tensor norm20_1_weight_to_fp16 = const()[name = tensor("norm20_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(503509568)))]; + tensor norm20_1_bias_to_fp16 = const()[name = tensor("norm20_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(503511680)))]; + tensor var_4936_to_fp16 = const()[name = tensor("op_4936_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_243_cast_fp16 = layer_norm(axes = input_243_axes_0, beta = norm20_1_bias_to_fp16, epsilon = var_4936_to_fp16, gamma = norm20_1_weight_to_fp16, x = input_241_cast_fp16)[name = tensor("input_243_cast_fp16")]; + tensor attn20_in_proj_weight_to_fp16 = const()[name = tensor("attn20_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(503513792)))]; + tensor linear_80_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn20_in_proj_weight_to_fp16, x = input_243_cast_fp16)[name = tensor("linear_80_cast_fp16")]; + tensor var_4950 = const()[name = tensor("op_4950"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_41_cast_fp16 = reshape(shape = var_4950, x = linear_80_cast_fp16)[name = tensor("qkv_41_cast_fp16")]; + tensor x_81_begin_0 = const()[name = tensor("x_81_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_81_end_0 = const()[name = tensor("x_81_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_81_end_mask_0 = const()[name = tensor("x_81_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_81_squeeze_mask_0 = const()[name = tensor("x_81_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_81_cast_fp16 = slice_by_index(begin = x_81_begin_0, end = x_81_end_0, end_mask = x_81_end_mask_0, squeeze_mask = x_81_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("x_81_cast_fp16")]; + tensor x_83_begin_0 = const()[name = tensor("x_83_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_83_end_0 = const()[name = tensor("x_83_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_83_end_mask_0 = const()[name = tensor("x_83_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_83_squeeze_mask_0 = const()[name = tensor("x_83_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_83_cast_fp16 = slice_by_index(begin = x_83_begin_0, end = x_83_end_0, end_mask = x_83_end_mask_0, squeeze_mask = x_83_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("x_83_cast_fp16")]; + tensor v_41_begin_0 = const()[name = tensor("v_41_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_41_end_0 = const()[name = tensor("v_41_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_41_end_mask_0 = const()[name = tensor("v_41_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_41_squeeze_mask_0 = const()[name = tensor("v_41_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_41_cast_fp16 = slice_by_index(begin = v_41_begin_0, end = v_41_end_0, end_mask = v_41_end_mask_0, squeeze_mask = v_41_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("v_41_cast_fp16")]; + tensor var_4995 = const()[name = tensor("op_4995"), val = tensor([256, 16, 32, 2])]; + tensor pairs_81_cast_fp16 = reshape(shape = var_4995, x = x_81_cast_fp16)[name = tensor("pairs_81_cast_fp16")]; + tensor xr_81_begin_0 = const()[name = tensor("xr_81_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_81_end_0 = const()[name = tensor("xr_81_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_81_end_mask_0 = const()[name = tensor("xr_81_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_81_squeeze_mask_0 = const()[name = tensor("xr_81_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_81_cast_fp16 = slice_by_index(begin = xr_81_begin_0, end = xr_81_end_0, end_mask = xr_81_end_mask_0, squeeze_mask = xr_81_squeeze_mask_0, x = pairs_81_cast_fp16)[name = tensor("xr_81_cast_fp16")]; + tensor xi_81_begin_0 = const()[name = tensor("xi_81_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_81_end_0 = const()[name = tensor("xi_81_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_81_end_mask_0 = const()[name = tensor("xi_81_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_81_squeeze_mask_0 = const()[name = tensor("xi_81_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_81_cast_fp16 = slice_by_index(begin = xi_81_begin_0, end = xi_81_end_0, end_mask = xi_81_end_mask_0, squeeze_mask = xi_81_squeeze_mask_0, x = pairs_81_cast_fp16)[name = tensor("xi_81_cast_fp16")]; + tensor var_5003_cast_fp16 = mul(x = xr_81_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5003_cast_fp16")]; + tensor var_5004_cast_fp16 = mul(x = xi_81_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5004_cast_fp16")]; + tensor xor__81_cast_fp16 = sub(x = var_5003_cast_fp16, y = var_5004_cast_fp16)[name = tensor("xor__81_cast_fp16")]; + tensor var_5007_cast_fp16 = mul(x = xr_81_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5007_cast_fp16")]; + tensor var_5008_cast_fp16 = mul(x = xi_81_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5008_cast_fp16")]; + tensor xoi_81_cast_fp16 = add(x = var_5007_cast_fp16, y = var_5008_cast_fp16)[name = tensor("xoi_81_cast_fp16")]; + tensor out_81_axis_0 = const()[name = tensor("out_81_axis_0"), val = tensor(-1)]; + tensor out_81_cast_fp16 = stack(axis = out_81_axis_0, values = (xor__81_cast_fp16, xoi_81_cast_fp16))[name = tensor("out_81_cast_fp16")]; + tensor var_5018 = const()[name = tensor("op_5018"), val = tensor([1, 256, 16, 64])]; + tensor q_41_cast_fp16 = reshape(shape = var_5018, x = out_81_cast_fp16)[name = tensor("q_41_cast_fp16")]; + tensor var_5024 = const()[name = tensor("op_5024"), val = tensor([256, 16, 32, 2])]; + tensor pairs_83_cast_fp16 = reshape(shape = var_5024, x = x_83_cast_fp16)[name = tensor("pairs_83_cast_fp16")]; + tensor xr_83_begin_0 = const()[name = tensor("xr_83_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_83_end_0 = const()[name = tensor("xr_83_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_83_end_mask_0 = const()[name = tensor("xr_83_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_83_squeeze_mask_0 = const()[name = tensor("xr_83_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_83_cast_fp16 = slice_by_index(begin = xr_83_begin_0, end = xr_83_end_0, end_mask = xr_83_end_mask_0, squeeze_mask = xr_83_squeeze_mask_0, x = pairs_83_cast_fp16)[name = tensor("xr_83_cast_fp16")]; + tensor xi_83_begin_0 = const()[name = tensor("xi_83_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_83_end_0 = const()[name = tensor("xi_83_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_83_end_mask_0 = const()[name = tensor("xi_83_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_83_squeeze_mask_0 = const()[name = tensor("xi_83_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_83_cast_fp16 = slice_by_index(begin = xi_83_begin_0, end = xi_83_end_0, end_mask = xi_83_end_mask_0, squeeze_mask = xi_83_squeeze_mask_0, x = pairs_83_cast_fp16)[name = tensor("xi_83_cast_fp16")]; + tensor var_5032_cast_fp16 = mul(x = xr_83_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5032_cast_fp16")]; + tensor var_5033_cast_fp16 = mul(x = xi_83_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5033_cast_fp16")]; + tensor xor__83_cast_fp16 = sub(x = var_5032_cast_fp16, y = var_5033_cast_fp16)[name = tensor("xor__83_cast_fp16")]; + tensor var_5036_cast_fp16 = mul(x = xr_83_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5036_cast_fp16")]; + tensor var_5037_cast_fp16 = mul(x = xi_83_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5037_cast_fp16")]; + tensor xoi_83_cast_fp16 = add(x = var_5036_cast_fp16, y = var_5037_cast_fp16)[name = tensor("xoi_83_cast_fp16")]; + tensor out_83_axis_0 = const()[name = tensor("out_83_axis_0"), val = tensor(-1)]; + tensor out_83_cast_fp16 = stack(axis = out_83_axis_0, values = (xor__83_cast_fp16, xoi_83_cast_fp16))[name = tensor("out_83_cast_fp16")]; + tensor var_5054 = const()[name = tensor("op_5054"), val = tensor([256, 1024])]; + tensor var_5055_cast_fp16 = reshape(shape = var_5054, x = out_83_cast_fp16)[name = tensor("op_5055_cast_fp16")]; + tensor written_k_41_transpose_x_1 = const()[name = tensor("written_k_41_transpose_x_1"), val = tensor(true)]; + tensor written_k_41_transpose_y_1 = const()[name = tensor("written_k_41_transpose_y_1"), val = tensor(false)]; + tensor written_k_41_cast_fp16 = matmul(transpose_x = written_k_41_transpose_x_1, transpose_y = written_k_41_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_5055_cast_fp16)[name = tensor("written_k_41_cast_fp16")]; + tensor var_5062 = const()[name = tensor("op_5062"), val = tensor([256, 1024])]; + tensor var_5063_cast_fp16 = reshape(shape = var_5062, x = v_41_cast_fp16)[name = tensor("op_5063_cast_fp16")]; + tensor written_v_41_transpose_x_1 = const()[name = tensor("written_v_41_transpose_x_1"), val = tensor(true)]; + tensor written_v_41_transpose_y_1 = const()[name = tensor("written_v_41_transpose_y_1"), val = tensor(false)]; + tensor written_v_41_cast_fp16 = matmul(transpose_x = written_v_41_transpose_x_1, transpose_y = written_v_41_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_5063_cast_fp16)[name = tensor("written_v_41_cast_fp16")]; + tensor k_cache20_to_fp16_dtype_0 = const()[name = tensor("k_cache20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache20_to_fp16 = cast(dtype = k_cache20_to_fp16_dtype_0, x = k_cache20)[name = tensor("cast_345")]; + tensor var_5074_cast_fp16 = mul(x = k_cache20_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_5074_cast_fp16")]; + tensor var_5079 = const()[name = tensor("op_5079"), val = tensor([1, 512, 16, 64])]; + tensor var_5080_cast_fp16 = reshape(shape = var_5079, x = written_k_41_cast_fp16)[name = tensor("op_5080_cast_fp16")]; + tensor new_k_cache_41_cast_fp16 = add(x = var_5074_cast_fp16, y = var_5080_cast_fp16)[name = tensor("new_k_cache_41_cast_fp16")]; + tensor new_k_cache_41_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_41_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache20_to_fp16_dtype_0 = const()[name = tensor("v_cache20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache20_to_fp16 = cast(dtype = v_cache20_to_fp16_dtype_0, x = v_cache20)[name = tensor("cast_343")]; + tensor var_5083_cast_fp16 = mul(x = v_cache20_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_5083_cast_fp16")]; + tensor var_5088 = const()[name = tensor("op_5088"), val = tensor([1, 512, 16, 64])]; + tensor var_5089_cast_fp16 = reshape(shape = var_5088, x = written_v_41_cast_fp16)[name = tensor("op_5089_cast_fp16")]; + tensor new_v_cache_41_cast_fp16 = add(x = var_5083_cast_fp16, y = var_5089_cast_fp16)[name = tensor("new_v_cache_41_cast_fp16")]; + tensor new_v_cache_41_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_41_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_5102 = const()[name = tensor("op_5102"), val = tensor([0, 2, 1, 3])]; + tensor var_5113_transpose_x_0 = const()[name = tensor("op_5113_transpose_x_0"), val = tensor(false)]; + tensor var_5113_transpose_y_0 = const()[name = tensor("op_5113_transpose_y_0"), val = tensor(false)]; + tensor transpose_109_perm_0 = const()[name = tensor("transpose_109_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_110_perm_0 = const()[name = tensor("transpose_110_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_110 = transpose(perm = transpose_110_perm_0, x = new_k_cache_41_cast_fp16)[name = tensor("transpose_124")]; + tensor transpose_109 = transpose(perm = transpose_109_perm_0, x = q_41_cast_fp16)[name = tensor("transpose_125")]; + tensor var_5113_cast_fp16 = matmul(transpose_x = var_5113_transpose_x_0, transpose_y = var_5113_transpose_y_0, x = transpose_109, y = transpose_110)[name = tensor("op_5113_cast_fp16")]; + tensor var_5114_to_fp16 = const()[name = tensor("op_5114_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_81_cast_fp16 = mul(x = var_5113_cast_fp16, y = var_5114_to_fp16)[name = tensor("scores_81_cast_fp16")]; + tensor scores_83_cast_fp16 = add(x = scores_81_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_83_cast_fp16")]; + tensor var_5124 = const()[name = tensor("op_5124"), val = tensor(-1)]; + tensor weights_41_cast_fp16 = softmax(axis = var_5124, x = scores_83_cast_fp16)[name = tensor("weights_41_cast_fp16")]; + tensor attn_41_transpose_x_0 = const()[name = tensor("attn_41_transpose_x_0"), val = tensor(false)]; + tensor attn_41_transpose_y_0 = const()[name = tensor("attn_41_transpose_y_0"), val = tensor(false)]; + tensor values_41_cast_fp16 = transpose(perm = var_5102, x = new_v_cache_41_cast_fp16)[name = tensor("transpose_126")]; + tensor attn_41_cast_fp16 = matmul(transpose_x = attn_41_transpose_x_0, transpose_y = attn_41_transpose_y_0, x = weights_41_cast_fp16, y = values_41_cast_fp16)[name = tensor("attn_41_cast_fp16")]; + tensor var_5132 = const()[name = tensor("op_5132"), val = tensor([0, 2, 1, 3])]; + tensor var_5137 = const()[name = tensor("op_5137"), val = tensor([1, 256, 1024])]; + tensor var_5133_cast_fp16 = transpose(perm = var_5132, x = attn_41_cast_fp16)[name = tensor("transpose_123")]; + tensor input_245_cast_fp16 = reshape(shape = var_5137, x = var_5133_cast_fp16)[name = tensor("input_245_cast_fp16")]; + tensor attn20_out_proj_weight_to_fp16 = const()[name = tensor("attn20_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(509805312)))]; + tensor linear_81_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn20_out_proj_weight_to_fp16, x = input_245_cast_fp16)[name = tensor("linear_81_cast_fp16")]; + tensor position20_to_fp16_dtype_0 = const()[name = tensor("position20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position20_to_fp16 = cast(dtype = position20_to_fp16_dtype_0, x = position20)[name = tensor("cast_341")]; + tensor var_5146_cast_fp16 = add(x = position20_to_fp16, y = valid_len_to_fp16)[name = tensor("op_5146_cast_fp16")]; + tensor var_5146_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5146_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_247_cast_fp16 = add(x = input_241_cast_fp16, y = linear_81_cast_fp16)[name = tensor("input_247_cast_fp16")]; + tensor input_249_axes_0 = const()[name = tensor("input_249_axes_0"), val = tensor([-1])]; + tensor norm20_2_weight_to_fp16 = const()[name = tensor("norm20_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(511902528)))]; + tensor norm20_2_bias_to_fp16 = const()[name = tensor("norm20_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(511904640)))]; + tensor var_5150_to_fp16 = const()[name = tensor("op_5150_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_249_cast_fp16 = layer_norm(axes = input_249_axes_0, beta = norm20_2_bias_to_fp16, epsilon = var_5150_to_fp16, gamma = norm20_2_weight_to_fp16, x = input_247_cast_fp16)[name = tensor("input_249_cast_fp16")]; + tensor linear20_1_weight_to_fp16 = const()[name = tensor("linear20_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(511906752)))]; + tensor linear_82_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear20_1_weight_to_fp16, x = input_249_cast_fp16)[name = tensor("linear_82_cast_fp16")]; + tensor input_251_mode_0 = const()[name = tensor("input_251_mode_0"), val = tensor("EXACT")]; + tensor input_251_cast_fp16 = gelu(mode = input_251_mode_0, x = linear_82_cast_fp16)[name = tensor("input_251_cast_fp16")]; + tensor linear20_2_weight_to_fp16 = const()[name = tensor("linear20_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(520295424)))]; + tensor linear_83_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear20_2_weight_to_fp16, x = input_251_cast_fp16)[name = tensor("linear_83_cast_fp16")]; + tensor input_253_cast_fp16 = add(x = input_247_cast_fp16, y = linear_83_cast_fp16)[name = tensor("input_253_cast_fp16")]; + tensor input_255_axes_0 = const()[name = tensor("input_255_axes_0"), val = tensor([-1])]; + tensor norm21_1_weight_to_fp16 = const()[name = tensor("norm21_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(528684096)))]; + tensor norm21_1_bias_to_fp16 = const()[name = tensor("norm21_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(528686208)))]; + tensor var_5167_to_fp16 = const()[name = tensor("op_5167_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_255_cast_fp16 = layer_norm(axes = input_255_axes_0, beta = norm21_1_bias_to_fp16, epsilon = var_5167_to_fp16, gamma = norm21_1_weight_to_fp16, x = input_253_cast_fp16)[name = tensor("input_255_cast_fp16")]; + tensor attn21_in_proj_weight_to_fp16 = const()[name = tensor("attn21_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(528688320)))]; + tensor linear_84_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn21_in_proj_weight_to_fp16, x = input_255_cast_fp16)[name = tensor("linear_84_cast_fp16")]; + tensor var_5181 = const()[name = tensor("op_5181"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_43_cast_fp16 = reshape(shape = var_5181, x = linear_84_cast_fp16)[name = tensor("qkv_43_cast_fp16")]; + tensor x_85_begin_0 = const()[name = tensor("x_85_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_85_end_0 = const()[name = tensor("x_85_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_85_end_mask_0 = const()[name = tensor("x_85_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_85_squeeze_mask_0 = const()[name = tensor("x_85_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_85_cast_fp16 = slice_by_index(begin = x_85_begin_0, end = x_85_end_0, end_mask = x_85_end_mask_0, squeeze_mask = x_85_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("x_85_cast_fp16")]; + tensor x_87_begin_0 = const()[name = tensor("x_87_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_87_end_0 = const()[name = tensor("x_87_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_87_end_mask_0 = const()[name = tensor("x_87_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_87_squeeze_mask_0 = const()[name = tensor("x_87_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_87_cast_fp16 = slice_by_index(begin = x_87_begin_0, end = x_87_end_0, end_mask = x_87_end_mask_0, squeeze_mask = x_87_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("x_87_cast_fp16")]; + tensor v_43_begin_0 = const()[name = tensor("v_43_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_43_end_0 = const()[name = tensor("v_43_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_43_end_mask_0 = const()[name = tensor("v_43_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_43_squeeze_mask_0 = const()[name = tensor("v_43_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_43_cast_fp16 = slice_by_index(begin = v_43_begin_0, end = v_43_end_0, end_mask = v_43_end_mask_0, squeeze_mask = v_43_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("v_43_cast_fp16")]; + tensor var_5226 = const()[name = tensor("op_5226"), val = tensor([256, 16, 32, 2])]; + tensor pairs_85_cast_fp16 = reshape(shape = var_5226, x = x_85_cast_fp16)[name = tensor("pairs_85_cast_fp16")]; + tensor xr_85_begin_0 = const()[name = tensor("xr_85_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_85_end_0 = const()[name = tensor("xr_85_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_85_end_mask_0 = const()[name = tensor("xr_85_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_85_squeeze_mask_0 = const()[name = tensor("xr_85_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_85_cast_fp16 = slice_by_index(begin = xr_85_begin_0, end = xr_85_end_0, end_mask = xr_85_end_mask_0, squeeze_mask = xr_85_squeeze_mask_0, x = pairs_85_cast_fp16)[name = tensor("xr_85_cast_fp16")]; + tensor xi_85_begin_0 = const()[name = tensor("xi_85_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_85_end_0 = const()[name = tensor("xi_85_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_85_end_mask_0 = const()[name = tensor("xi_85_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_85_squeeze_mask_0 = const()[name = tensor("xi_85_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_85_cast_fp16 = slice_by_index(begin = xi_85_begin_0, end = xi_85_end_0, end_mask = xi_85_end_mask_0, squeeze_mask = xi_85_squeeze_mask_0, x = pairs_85_cast_fp16)[name = tensor("xi_85_cast_fp16")]; + tensor var_5234_cast_fp16 = mul(x = xr_85_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5234_cast_fp16")]; + tensor var_5235_cast_fp16 = mul(x = xi_85_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5235_cast_fp16")]; + tensor xor__85_cast_fp16 = sub(x = var_5234_cast_fp16, y = var_5235_cast_fp16)[name = tensor("xor__85_cast_fp16")]; + tensor var_5238_cast_fp16 = mul(x = xr_85_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5238_cast_fp16")]; + tensor var_5239_cast_fp16 = mul(x = xi_85_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5239_cast_fp16")]; + tensor xoi_85_cast_fp16 = add(x = var_5238_cast_fp16, y = var_5239_cast_fp16)[name = tensor("xoi_85_cast_fp16")]; + tensor out_85_axis_0 = const()[name = tensor("out_85_axis_0"), val = tensor(-1)]; + tensor out_85_cast_fp16 = stack(axis = out_85_axis_0, values = (xor__85_cast_fp16, xoi_85_cast_fp16))[name = tensor("out_85_cast_fp16")]; + tensor var_5249 = const()[name = tensor("op_5249"), val = tensor([1, 256, 16, 64])]; + tensor q_43_cast_fp16 = reshape(shape = var_5249, x = out_85_cast_fp16)[name = tensor("q_43_cast_fp16")]; + tensor var_5255 = const()[name = tensor("op_5255"), val = tensor([256, 16, 32, 2])]; + tensor pairs_87_cast_fp16 = reshape(shape = var_5255, x = x_87_cast_fp16)[name = tensor("pairs_87_cast_fp16")]; + tensor xr_87_begin_0 = const()[name = tensor("xr_87_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_87_end_0 = const()[name = tensor("xr_87_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_87_end_mask_0 = const()[name = tensor("xr_87_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_87_squeeze_mask_0 = const()[name = tensor("xr_87_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_87_cast_fp16 = slice_by_index(begin = xr_87_begin_0, end = xr_87_end_0, end_mask = xr_87_end_mask_0, squeeze_mask = xr_87_squeeze_mask_0, x = pairs_87_cast_fp16)[name = tensor("xr_87_cast_fp16")]; + tensor xi_87_begin_0 = const()[name = tensor("xi_87_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_87_end_0 = const()[name = tensor("xi_87_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_87_end_mask_0 = const()[name = tensor("xi_87_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_87_squeeze_mask_0 = const()[name = tensor("xi_87_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_87_cast_fp16 = slice_by_index(begin = xi_87_begin_0, end = xi_87_end_0, end_mask = xi_87_end_mask_0, squeeze_mask = xi_87_squeeze_mask_0, x = pairs_87_cast_fp16)[name = tensor("xi_87_cast_fp16")]; + tensor var_5263_cast_fp16 = mul(x = xr_87_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5263_cast_fp16")]; + tensor var_5264_cast_fp16 = mul(x = xi_87_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5264_cast_fp16")]; + tensor xor__87_cast_fp16 = sub(x = var_5263_cast_fp16, y = var_5264_cast_fp16)[name = tensor("xor__87_cast_fp16")]; + tensor var_5267_cast_fp16 = mul(x = xr_87_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5267_cast_fp16")]; + tensor var_5268_cast_fp16 = mul(x = xi_87_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5268_cast_fp16")]; + tensor xoi_87_cast_fp16 = add(x = var_5267_cast_fp16, y = var_5268_cast_fp16)[name = tensor("xoi_87_cast_fp16")]; + tensor out_87_axis_0 = const()[name = tensor("out_87_axis_0"), val = tensor(-1)]; + tensor out_87_cast_fp16 = stack(axis = out_87_axis_0, values = (xor__87_cast_fp16, xoi_87_cast_fp16))[name = tensor("out_87_cast_fp16")]; + tensor var_5285 = const()[name = tensor("op_5285"), val = tensor([256, 1024])]; + tensor var_5286_cast_fp16 = reshape(shape = var_5285, x = out_87_cast_fp16)[name = tensor("op_5286_cast_fp16")]; + tensor written_k_43_transpose_x_1 = const()[name = tensor("written_k_43_transpose_x_1"), val = tensor(true)]; + tensor written_k_43_transpose_y_1 = const()[name = tensor("written_k_43_transpose_y_1"), val = tensor(false)]; + tensor written_k_43_cast_fp16 = matmul(transpose_x = written_k_43_transpose_x_1, transpose_y = written_k_43_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_5286_cast_fp16)[name = tensor("written_k_43_cast_fp16")]; + tensor var_5293 = const()[name = tensor("op_5293"), val = tensor([256, 1024])]; + tensor var_5294_cast_fp16 = reshape(shape = var_5293, x = v_43_cast_fp16)[name = tensor("op_5294_cast_fp16")]; + tensor written_v_43_transpose_x_1 = const()[name = tensor("written_v_43_transpose_x_1"), val = tensor(true)]; + tensor written_v_43_transpose_y_1 = const()[name = tensor("written_v_43_transpose_y_1"), val = tensor(false)]; + tensor written_v_43_cast_fp16 = matmul(transpose_x = written_v_43_transpose_x_1, transpose_y = written_v_43_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_5294_cast_fp16)[name = tensor("written_v_43_cast_fp16")]; + tensor k_cache21_to_fp16_dtype_0 = const()[name = tensor("k_cache21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache21_to_fp16 = cast(dtype = k_cache21_to_fp16_dtype_0, x = k_cache21)[name = tensor("cast_339")]; + tensor var_5305_cast_fp16 = mul(x = k_cache21_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_5305_cast_fp16")]; + tensor var_5310 = const()[name = tensor("op_5310"), val = tensor([1, 512, 16, 64])]; + tensor var_5311_cast_fp16 = reshape(shape = var_5310, x = written_k_43_cast_fp16)[name = tensor("op_5311_cast_fp16")]; + tensor new_k_cache_43_cast_fp16 = add(x = var_5305_cast_fp16, y = var_5311_cast_fp16)[name = tensor("new_k_cache_43_cast_fp16")]; + tensor new_k_cache_43_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_43_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache21_to_fp16_dtype_0 = const()[name = tensor("v_cache21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache21_to_fp16 = cast(dtype = v_cache21_to_fp16_dtype_0, x = v_cache21)[name = tensor("cast_337")]; + tensor var_5314_cast_fp16 = mul(x = v_cache21_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_5314_cast_fp16")]; + tensor var_5319 = const()[name = tensor("op_5319"), val = tensor([1, 512, 16, 64])]; + tensor var_5320_cast_fp16 = reshape(shape = var_5319, x = written_v_43_cast_fp16)[name = tensor("op_5320_cast_fp16")]; + tensor new_v_cache_43_cast_fp16 = add(x = var_5314_cast_fp16, y = var_5320_cast_fp16)[name = tensor("new_v_cache_43_cast_fp16")]; + tensor new_v_cache_43_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_43_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_5333 = const()[name = tensor("op_5333"), val = tensor([0, 2, 1, 3])]; + tensor var_5344_transpose_x_0 = const()[name = tensor("op_5344_transpose_x_0"), val = tensor(false)]; + tensor var_5344_transpose_y_0 = const()[name = tensor("op_5344_transpose_y_0"), val = tensor(false)]; + tensor transpose_111_perm_0 = const()[name = tensor("transpose_111_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_112_perm_0 = const()[name = tensor("transpose_112_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_112 = transpose(perm = transpose_112_perm_0, x = new_k_cache_43_cast_fp16)[name = tensor("transpose_120")]; + tensor transpose_111 = transpose(perm = transpose_111_perm_0, x = q_43_cast_fp16)[name = tensor("transpose_121")]; + tensor var_5344_cast_fp16 = matmul(transpose_x = var_5344_transpose_x_0, transpose_y = var_5344_transpose_y_0, x = transpose_111, y = transpose_112)[name = tensor("op_5344_cast_fp16")]; + tensor var_5345_to_fp16 = const()[name = tensor("op_5345_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_85_cast_fp16 = mul(x = var_5344_cast_fp16, y = var_5345_to_fp16)[name = tensor("scores_85_cast_fp16")]; + tensor scores_87_cast_fp16 = add(x = scores_85_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_87_cast_fp16")]; + tensor var_5355 = const()[name = tensor("op_5355"), val = tensor(-1)]; + tensor weights_43_cast_fp16 = softmax(axis = var_5355, x = scores_87_cast_fp16)[name = tensor("weights_43_cast_fp16")]; + tensor attn_43_transpose_x_0 = const()[name = tensor("attn_43_transpose_x_0"), val = tensor(false)]; + tensor attn_43_transpose_y_0 = const()[name = tensor("attn_43_transpose_y_0"), val = tensor(false)]; + tensor values_43_cast_fp16 = transpose(perm = var_5333, x = new_v_cache_43_cast_fp16)[name = tensor("transpose_122")]; + tensor attn_43_cast_fp16 = matmul(transpose_x = attn_43_transpose_x_0, transpose_y = attn_43_transpose_y_0, x = weights_43_cast_fp16, y = values_43_cast_fp16)[name = tensor("attn_43_cast_fp16")]; + tensor var_5363 = const()[name = tensor("op_5363"), val = tensor([0, 2, 1, 3])]; + tensor var_5368 = const()[name = tensor("op_5368"), val = tensor([1, 256, 1024])]; + tensor var_5364_cast_fp16 = transpose(perm = var_5363, x = attn_43_cast_fp16)[name = tensor("transpose_119")]; + tensor input_257_cast_fp16 = reshape(shape = var_5368, x = var_5364_cast_fp16)[name = tensor("input_257_cast_fp16")]; + tensor attn21_out_proj_weight_to_fp16 = const()[name = tensor("attn21_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(534979840)))]; + tensor linear_85_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn21_out_proj_weight_to_fp16, x = input_257_cast_fp16)[name = tensor("linear_85_cast_fp16")]; + tensor position21_to_fp16_dtype_0 = const()[name = tensor("position21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position21_to_fp16 = cast(dtype = position21_to_fp16_dtype_0, x = position21)[name = tensor("cast_335")]; + tensor var_5377_cast_fp16 = add(x = position21_to_fp16, y = valid_len_to_fp16)[name = tensor("op_5377_cast_fp16")]; + tensor var_5377_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5377_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_259_cast_fp16 = add(x = input_253_cast_fp16, y = linear_85_cast_fp16)[name = tensor("input_259_cast_fp16")]; + tensor input_261_axes_0 = const()[name = tensor("input_261_axes_0"), val = tensor([-1])]; + tensor norm21_2_weight_to_fp16 = const()[name = tensor("norm21_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(537077056)))]; + tensor norm21_2_bias_to_fp16 = const()[name = tensor("norm21_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(537079168)))]; + tensor var_5381_to_fp16 = const()[name = tensor("op_5381_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_261_cast_fp16 = layer_norm(axes = input_261_axes_0, beta = norm21_2_bias_to_fp16, epsilon = var_5381_to_fp16, gamma = norm21_2_weight_to_fp16, x = input_259_cast_fp16)[name = tensor("input_261_cast_fp16")]; + tensor linear21_1_weight_to_fp16 = const()[name = tensor("linear21_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(537081280)))]; + tensor linear_86_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear21_1_weight_to_fp16, x = input_261_cast_fp16)[name = tensor("linear_86_cast_fp16")]; + tensor input_263_mode_0 = const()[name = tensor("input_263_mode_0"), val = tensor("EXACT")]; + tensor input_263_cast_fp16 = gelu(mode = input_263_mode_0, x = linear_86_cast_fp16)[name = tensor("input_263_cast_fp16")]; + tensor linear21_2_weight_to_fp16 = const()[name = tensor("linear21_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(545469952)))]; + tensor linear_87_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear21_2_weight_to_fp16, x = input_263_cast_fp16)[name = tensor("linear_87_cast_fp16")]; + tensor input_265_cast_fp16 = add(x = input_259_cast_fp16, y = linear_87_cast_fp16)[name = tensor("input_265_cast_fp16")]; + tensor input_267_axes_0 = const()[name = tensor("input_267_axes_0"), val = tensor([-1])]; + tensor norm22_1_weight_to_fp16 = const()[name = tensor("norm22_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(553858624)))]; + tensor norm22_1_bias_to_fp16 = const()[name = tensor("norm22_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(553860736)))]; + tensor var_5398_to_fp16 = const()[name = tensor("op_5398_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_267_cast_fp16 = layer_norm(axes = input_267_axes_0, beta = norm22_1_bias_to_fp16, epsilon = var_5398_to_fp16, gamma = norm22_1_weight_to_fp16, x = input_265_cast_fp16)[name = tensor("input_267_cast_fp16")]; + tensor attn22_in_proj_weight_to_fp16 = const()[name = tensor("attn22_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(553862848)))]; + tensor linear_88_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn22_in_proj_weight_to_fp16, x = input_267_cast_fp16)[name = tensor("linear_88_cast_fp16")]; + tensor var_5412 = const()[name = tensor("op_5412"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_45_cast_fp16 = reshape(shape = var_5412, x = linear_88_cast_fp16)[name = tensor("qkv_45_cast_fp16")]; + tensor x_89_begin_0 = const()[name = tensor("x_89_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_89_end_0 = const()[name = tensor("x_89_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_89_end_mask_0 = const()[name = tensor("x_89_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_89_squeeze_mask_0 = const()[name = tensor("x_89_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_89_cast_fp16 = slice_by_index(begin = x_89_begin_0, end = x_89_end_0, end_mask = x_89_end_mask_0, squeeze_mask = x_89_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("x_89_cast_fp16")]; + tensor x_91_begin_0 = const()[name = tensor("x_91_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_91_end_0 = const()[name = tensor("x_91_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_91_end_mask_0 = const()[name = tensor("x_91_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_91_squeeze_mask_0 = const()[name = tensor("x_91_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_91_cast_fp16 = slice_by_index(begin = x_91_begin_0, end = x_91_end_0, end_mask = x_91_end_mask_0, squeeze_mask = x_91_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("x_91_cast_fp16")]; + tensor v_45_begin_0 = const()[name = tensor("v_45_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_45_end_0 = const()[name = tensor("v_45_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_45_end_mask_0 = const()[name = tensor("v_45_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_45_squeeze_mask_0 = const()[name = tensor("v_45_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_45_cast_fp16 = slice_by_index(begin = v_45_begin_0, end = v_45_end_0, end_mask = v_45_end_mask_0, squeeze_mask = v_45_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("v_45_cast_fp16")]; + tensor var_5457 = const()[name = tensor("op_5457"), val = tensor([256, 16, 32, 2])]; + tensor pairs_89_cast_fp16 = reshape(shape = var_5457, x = x_89_cast_fp16)[name = tensor("pairs_89_cast_fp16")]; + tensor xr_89_begin_0 = const()[name = tensor("xr_89_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_89_end_0 = const()[name = tensor("xr_89_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_89_end_mask_0 = const()[name = tensor("xr_89_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_89_squeeze_mask_0 = const()[name = tensor("xr_89_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_89_cast_fp16 = slice_by_index(begin = xr_89_begin_0, end = xr_89_end_0, end_mask = xr_89_end_mask_0, squeeze_mask = xr_89_squeeze_mask_0, x = pairs_89_cast_fp16)[name = tensor("xr_89_cast_fp16")]; + tensor xi_89_begin_0 = const()[name = tensor("xi_89_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_89_end_0 = const()[name = tensor("xi_89_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_89_end_mask_0 = const()[name = tensor("xi_89_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_89_squeeze_mask_0 = const()[name = tensor("xi_89_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_89_cast_fp16 = slice_by_index(begin = xi_89_begin_0, end = xi_89_end_0, end_mask = xi_89_end_mask_0, squeeze_mask = xi_89_squeeze_mask_0, x = pairs_89_cast_fp16)[name = tensor("xi_89_cast_fp16")]; + tensor var_5465_cast_fp16 = mul(x = xr_89_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5465_cast_fp16")]; + tensor var_5466_cast_fp16 = mul(x = xi_89_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5466_cast_fp16")]; + tensor xor__89_cast_fp16 = sub(x = var_5465_cast_fp16, y = var_5466_cast_fp16)[name = tensor("xor__89_cast_fp16")]; + tensor var_5469_cast_fp16 = mul(x = xr_89_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5469_cast_fp16")]; + tensor var_5470_cast_fp16 = mul(x = xi_89_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5470_cast_fp16")]; + tensor xoi_89_cast_fp16 = add(x = var_5469_cast_fp16, y = var_5470_cast_fp16)[name = tensor("xoi_89_cast_fp16")]; + tensor out_89_axis_0 = const()[name = tensor("out_89_axis_0"), val = tensor(-1)]; + tensor out_89_cast_fp16 = stack(axis = out_89_axis_0, values = (xor__89_cast_fp16, xoi_89_cast_fp16))[name = tensor("out_89_cast_fp16")]; + tensor var_5480 = const()[name = tensor("op_5480"), val = tensor([1, 256, 16, 64])]; + tensor q_45_cast_fp16 = reshape(shape = var_5480, x = out_89_cast_fp16)[name = tensor("q_45_cast_fp16")]; + tensor var_5486 = const()[name = tensor("op_5486"), val = tensor([256, 16, 32, 2])]; + tensor pairs_91_cast_fp16 = reshape(shape = var_5486, x = x_91_cast_fp16)[name = tensor("pairs_91_cast_fp16")]; + tensor xr_91_begin_0 = const()[name = tensor("xr_91_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_91_end_0 = const()[name = tensor("xr_91_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_91_end_mask_0 = const()[name = tensor("xr_91_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_91_squeeze_mask_0 = const()[name = tensor("xr_91_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_91_cast_fp16 = slice_by_index(begin = xr_91_begin_0, end = xr_91_end_0, end_mask = xr_91_end_mask_0, squeeze_mask = xr_91_squeeze_mask_0, x = pairs_91_cast_fp16)[name = tensor("xr_91_cast_fp16")]; + tensor xi_91_begin_0 = const()[name = tensor("xi_91_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_91_end_0 = const()[name = tensor("xi_91_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_91_end_mask_0 = const()[name = tensor("xi_91_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_91_squeeze_mask_0 = const()[name = tensor("xi_91_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_91_cast_fp16 = slice_by_index(begin = xi_91_begin_0, end = xi_91_end_0, end_mask = xi_91_end_mask_0, squeeze_mask = xi_91_squeeze_mask_0, x = pairs_91_cast_fp16)[name = tensor("xi_91_cast_fp16")]; + tensor var_5494_cast_fp16 = mul(x = xr_91_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5494_cast_fp16")]; + tensor var_5495_cast_fp16 = mul(x = xi_91_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5495_cast_fp16")]; + tensor xor__91_cast_fp16 = sub(x = var_5494_cast_fp16, y = var_5495_cast_fp16)[name = tensor("xor__91_cast_fp16")]; + tensor var_5498_cast_fp16 = mul(x = xr_91_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5498_cast_fp16")]; + tensor var_5499_cast_fp16 = mul(x = xi_91_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5499_cast_fp16")]; + tensor xoi_91_cast_fp16 = add(x = var_5498_cast_fp16, y = var_5499_cast_fp16)[name = tensor("xoi_91_cast_fp16")]; + tensor out_91_axis_0 = const()[name = tensor("out_91_axis_0"), val = tensor(-1)]; + tensor out_91_cast_fp16 = stack(axis = out_91_axis_0, values = (xor__91_cast_fp16, xoi_91_cast_fp16))[name = tensor("out_91_cast_fp16")]; + tensor var_5516 = const()[name = tensor("op_5516"), val = tensor([256, 1024])]; + tensor var_5517_cast_fp16 = reshape(shape = var_5516, x = out_91_cast_fp16)[name = tensor("op_5517_cast_fp16")]; + tensor written_k_45_transpose_x_1 = const()[name = tensor("written_k_45_transpose_x_1"), val = tensor(true)]; + tensor written_k_45_transpose_y_1 = const()[name = tensor("written_k_45_transpose_y_1"), val = tensor(false)]; + tensor written_k_45_cast_fp16 = matmul(transpose_x = written_k_45_transpose_x_1, transpose_y = written_k_45_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_5517_cast_fp16)[name = tensor("written_k_45_cast_fp16")]; + tensor var_5524 = const()[name = tensor("op_5524"), val = tensor([256, 1024])]; + tensor var_5525_cast_fp16 = reshape(shape = var_5524, x = v_45_cast_fp16)[name = tensor("op_5525_cast_fp16")]; + tensor written_v_45_transpose_x_1 = const()[name = tensor("written_v_45_transpose_x_1"), val = tensor(true)]; + tensor written_v_45_transpose_y_1 = const()[name = tensor("written_v_45_transpose_y_1"), val = tensor(false)]; + tensor written_v_45_cast_fp16 = matmul(transpose_x = written_v_45_transpose_x_1, transpose_y = written_v_45_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_5525_cast_fp16)[name = tensor("written_v_45_cast_fp16")]; + tensor k_cache22_to_fp16_dtype_0 = const()[name = tensor("k_cache22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache22_to_fp16 = cast(dtype = k_cache22_to_fp16_dtype_0, x = k_cache22)[name = tensor("cast_333")]; + tensor var_5536_cast_fp16 = mul(x = k_cache22_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_5536_cast_fp16")]; + tensor var_5541 = const()[name = tensor("op_5541"), val = tensor([1, 512, 16, 64])]; + tensor var_5542_cast_fp16 = reshape(shape = var_5541, x = written_k_45_cast_fp16)[name = tensor("op_5542_cast_fp16")]; + tensor new_k_cache_45_cast_fp16 = add(x = var_5536_cast_fp16, y = var_5542_cast_fp16)[name = tensor("new_k_cache_45_cast_fp16")]; + tensor new_k_cache_45_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_45_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache22_to_fp16_dtype_0 = const()[name = tensor("v_cache22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache22_to_fp16 = cast(dtype = v_cache22_to_fp16_dtype_0, x = v_cache22)[name = tensor("cast_331")]; + tensor var_5545_cast_fp16 = mul(x = v_cache22_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_5545_cast_fp16")]; + tensor var_5550 = const()[name = tensor("op_5550"), val = tensor([1, 512, 16, 64])]; + tensor var_5551_cast_fp16 = reshape(shape = var_5550, x = written_v_45_cast_fp16)[name = tensor("op_5551_cast_fp16")]; + tensor new_v_cache_45_cast_fp16 = add(x = var_5545_cast_fp16, y = var_5551_cast_fp16)[name = tensor("new_v_cache_45_cast_fp16")]; + tensor new_v_cache_45_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_45_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_5564 = const()[name = tensor("op_5564"), val = tensor([0, 2, 1, 3])]; + tensor var_5575_transpose_x_0 = const()[name = tensor("op_5575_transpose_x_0"), val = tensor(false)]; + tensor var_5575_transpose_y_0 = const()[name = tensor("op_5575_transpose_y_0"), val = tensor(false)]; + tensor transpose_113_perm_0 = const()[name = tensor("transpose_113_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_114_perm_0 = const()[name = tensor("transpose_114_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_114 = transpose(perm = transpose_114_perm_0, x = new_k_cache_45_cast_fp16)[name = tensor("transpose_116")]; + tensor transpose_113 = transpose(perm = transpose_113_perm_0, x = q_45_cast_fp16)[name = tensor("transpose_117")]; + tensor var_5575_cast_fp16 = matmul(transpose_x = var_5575_transpose_x_0, transpose_y = var_5575_transpose_y_0, x = transpose_113, y = transpose_114)[name = tensor("op_5575_cast_fp16")]; + tensor var_5576_to_fp16 = const()[name = tensor("op_5576_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_89_cast_fp16 = mul(x = var_5575_cast_fp16, y = var_5576_to_fp16)[name = tensor("scores_89_cast_fp16")]; + tensor scores_91_cast_fp16 = add(x = scores_89_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_91_cast_fp16")]; + tensor var_5586 = const()[name = tensor("op_5586"), val = tensor(-1)]; + tensor weights_45_cast_fp16 = softmax(axis = var_5586, x = scores_91_cast_fp16)[name = tensor("weights_45_cast_fp16")]; + tensor attn_45_transpose_x_0 = const()[name = tensor("attn_45_transpose_x_0"), val = tensor(false)]; + tensor attn_45_transpose_y_0 = const()[name = tensor("attn_45_transpose_y_0"), val = tensor(false)]; + tensor values_45_cast_fp16 = transpose(perm = var_5564, x = new_v_cache_45_cast_fp16)[name = tensor("transpose_118")]; + tensor attn_45_cast_fp16 = matmul(transpose_x = attn_45_transpose_x_0, transpose_y = attn_45_transpose_y_0, x = weights_45_cast_fp16, y = values_45_cast_fp16)[name = tensor("attn_45_cast_fp16")]; + tensor var_5594 = const()[name = tensor("op_5594"), val = tensor([0, 2, 1, 3])]; + tensor var_5599 = const()[name = tensor("op_5599"), val = tensor([1, 256, 1024])]; + tensor var_5595_cast_fp16 = transpose(perm = var_5594, x = attn_45_cast_fp16)[name = tensor("transpose_115")]; + tensor input_269_cast_fp16 = reshape(shape = var_5599, x = var_5595_cast_fp16)[name = tensor("input_269_cast_fp16")]; + tensor attn22_out_proj_weight_to_fp16 = const()[name = tensor("attn22_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(560154368)))]; + tensor linear_89_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn22_out_proj_weight_to_fp16, x = input_269_cast_fp16)[name = tensor("linear_89_cast_fp16")]; + tensor position22_to_fp16_dtype_0 = const()[name = tensor("position22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position22_to_fp16 = cast(dtype = position22_to_fp16_dtype_0, x = position22)[name = tensor("cast_329")]; + tensor var_5608_cast_fp16 = add(x = position22_to_fp16, y = valid_len_to_fp16)[name = tensor("op_5608_cast_fp16")]; + tensor var_5608_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5608_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_271_cast_fp16 = add(x = input_265_cast_fp16, y = linear_89_cast_fp16)[name = tensor("input_271_cast_fp16")]; + tensor input_273_axes_0 = const()[name = tensor("input_273_axes_0"), val = tensor([-1])]; + tensor norm22_2_weight_to_fp16 = const()[name = tensor("norm22_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(562251584)))]; + tensor norm22_2_bias_to_fp16 = const()[name = tensor("norm22_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(562253696)))]; + tensor var_5612_to_fp16 = const()[name = tensor("op_5612_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_273_cast_fp16 = layer_norm(axes = input_273_axes_0, beta = norm22_2_bias_to_fp16, epsilon = var_5612_to_fp16, gamma = norm22_2_weight_to_fp16, x = input_271_cast_fp16)[name = tensor("input_273_cast_fp16")]; + tensor linear22_1_weight_to_fp16 = const()[name = tensor("linear22_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(562255808)))]; + tensor linear_90_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear22_1_weight_to_fp16, x = input_273_cast_fp16)[name = tensor("linear_90_cast_fp16")]; + tensor input_275_mode_0 = const()[name = tensor("input_275_mode_0"), val = tensor("EXACT")]; + tensor input_275_cast_fp16 = gelu(mode = input_275_mode_0, x = linear_90_cast_fp16)[name = tensor("input_275_cast_fp16")]; + tensor linear22_2_weight_to_fp16 = const()[name = tensor("linear22_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(570644480)))]; + tensor linear_91_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear22_2_weight_to_fp16, x = input_275_cast_fp16)[name = tensor("linear_91_cast_fp16")]; + tensor input_277_cast_fp16 = add(x = input_271_cast_fp16, y = linear_91_cast_fp16)[name = tensor("input_277_cast_fp16")]; + tensor input_279_axes_0 = const()[name = tensor("input_279_axes_0"), val = tensor([-1])]; + tensor norm23_1_weight_to_fp16 = const()[name = tensor("norm23_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(579033152)))]; + tensor norm23_1_bias_to_fp16 = const()[name = tensor("norm23_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(579035264)))]; + tensor var_5629_to_fp16 = const()[name = tensor("op_5629_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_279_cast_fp16 = layer_norm(axes = input_279_axes_0, beta = norm23_1_bias_to_fp16, epsilon = var_5629_to_fp16, gamma = norm23_1_weight_to_fp16, x = input_277_cast_fp16)[name = tensor("input_279_cast_fp16")]; + tensor attn23_in_proj_weight_to_fp16 = const()[name = tensor("attn23_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(579037376)))]; + tensor linear_92_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn23_in_proj_weight_to_fp16, x = input_279_cast_fp16)[name = tensor("linear_92_cast_fp16")]; + tensor var_5643 = const()[name = tensor("op_5643"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_cast_fp16 = reshape(shape = var_5643, x = linear_92_cast_fp16)[name = tensor("qkv_cast_fp16")]; + tensor x_begin_0 = const()[name = tensor("x_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_end_0 = const()[name = tensor("x_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_end_mask_0 = const()[name = tensor("x_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_squeeze_mask_0 = const()[name = tensor("x_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_cast_fp16 = slice_by_index(begin = x_begin_0, end = x_end_0, end_mask = x_end_mask_0, squeeze_mask = x_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("x_cast_fp16")]; + tensor v_begin_0 = const()[name = tensor("v_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_end_0 = const()[name = tensor("v_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_end_mask_0 = const()[name = tensor("v_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_squeeze_mask_0 = const()[name = tensor("v_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_cast_fp16 = slice_by_index(begin = v_begin_0, end = v_end_0, end_mask = v_end_mask_0, squeeze_mask = v_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("v_cast_fp16")]; + tensor var_5675 = const()[name = tensor("op_5675"), val = tensor([256, 16, 32, 2])]; + tensor pairs_cast_fp16 = reshape(shape = var_5675, x = x_cast_fp16)[name = tensor("pairs_cast_fp16")]; + tensor xr_begin_0 = const()[name = tensor("xr_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_end_0 = const()[name = tensor("xr_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_end_mask_0 = const()[name = tensor("xr_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_squeeze_mask_0 = const()[name = tensor("xr_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_cast_fp16 = slice_by_index(begin = xr_begin_0, end = xr_end_0, end_mask = xr_end_mask_0, squeeze_mask = xr_squeeze_mask_0, x = pairs_cast_fp16)[name = tensor("xr_cast_fp16")]; + tensor xi_begin_0 = const()[name = tensor("xi_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_end_0 = const()[name = tensor("xi_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_end_mask_0 = const()[name = tensor("xi_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_squeeze_mask_0 = const()[name = tensor("xi_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_cast_fp16 = slice_by_index(begin = xi_begin_0, end = xi_end_0, end_mask = xi_end_mask_0, squeeze_mask = xi_squeeze_mask_0, x = pairs_cast_fp16)[name = tensor("xi_cast_fp16")]; + tensor var_5683_cast_fp16 = mul(x = xr_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5683_cast_fp16")]; + tensor var_5684_cast_fp16 = mul(x = xi_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5684_cast_fp16")]; + tensor xor__cast_fp16 = sub(x = var_5683_cast_fp16, y = var_5684_cast_fp16)[name = tensor("xor__cast_fp16")]; + tensor var_5687_cast_fp16 = mul(x = xr_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5687_cast_fp16")]; + tensor var_5688_cast_fp16 = mul(x = xi_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5688_cast_fp16")]; + tensor xoi_cast_fp16 = add(x = var_5687_cast_fp16, y = var_5688_cast_fp16)[name = tensor("xoi_cast_fp16")]; + tensor out_axis_0 = const()[name = tensor("out_axis_0"), val = tensor(-1)]; + tensor out_cast_fp16 = stack(axis = out_axis_0, values = (xor__cast_fp16, xoi_cast_fp16))[name = tensor("out_cast_fp16")]; + tensor var_5705 = const()[name = tensor("op_5705"), val = tensor([256, 1024])]; + tensor var_5706_cast_fp16 = reshape(shape = var_5705, x = out_cast_fp16)[name = tensor("op_5706_cast_fp16")]; + tensor written_k_transpose_x_1 = const()[name = tensor("written_k_transpose_x_1"), val = tensor(true)]; + tensor written_k_transpose_y_1 = const()[name = tensor("written_k_transpose_y_1"), val = tensor(false)]; + tensor written_k_cast_fp16 = matmul(transpose_x = written_k_transpose_x_1, transpose_y = written_k_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_5706_cast_fp16)[name = tensor("written_k_cast_fp16")]; + tensor var_5713 = const()[name = tensor("op_5713"), val = tensor([256, 1024])]; + tensor var_5714_cast_fp16 = reshape(shape = var_5713, x = v_cast_fp16)[name = tensor("op_5714_cast_fp16")]; + tensor written_v_transpose_x_1 = const()[name = tensor("written_v_transpose_x_1"), val = tensor(true)]; + tensor written_v_transpose_y_1 = const()[name = tensor("written_v_transpose_y_1"), val = tensor(false)]; + tensor written_v_cast_fp16 = matmul(transpose_x = written_v_transpose_x_1, transpose_y = written_v_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_5714_cast_fp16)[name = tensor("written_v_cast_fp16")]; + tensor k_cache23_to_fp16_dtype_0 = const()[name = tensor("k_cache23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache23_to_fp16 = cast(dtype = k_cache23_to_fp16_dtype_0, x = k_cache23)[name = tensor("cast_327")]; + tensor var_5725_cast_fp16 = mul(x = k_cache23_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_5725_cast_fp16")]; + tensor var_5730 = const()[name = tensor("op_5730"), val = tensor([1, 512, 16, 64])]; + tensor var_5731_cast_fp16 = reshape(shape = var_5730, x = written_k_cast_fp16)[name = tensor("op_5731_cast_fp16")]; + tensor new_k_cache_cast_fp16 = add(x = var_5725_cast_fp16, y = var_5731_cast_fp16)[name = tensor("new_k_cache_cast_fp16")]; + tensor new_k_cache_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache23_to_fp16_dtype_0 = const()[name = tensor("v_cache23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache23_to_fp16 = cast(dtype = v_cache23_to_fp16_dtype_0, x = v_cache23)[name = tensor("cast_325")]; + tensor var_5734_cast_fp16 = mul(x = v_cache23_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_5734_cast_fp16")]; + tensor var_5739 = const()[name = tensor("op_5739"), val = tensor([1, 512, 16, 64])]; + tensor var_5740_cast_fp16 = reshape(shape = var_5739, x = written_v_cast_fp16)[name = tensor("op_5740_cast_fp16")]; + tensor new_v_cache_cast_fp16 = add(x = var_5734_cast_fp16, y = var_5740_cast_fp16)[name = tensor("new_v_cache_cast_fp16")]; + tensor new_v_cache_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor position23_to_fp16_dtype_0 = const()[name = tensor("position23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position23_to_fp16 = cast(dtype = position23_to_fp16_dtype_0, x = position23)[name = tensor("cast_323")]; + tensor var_5747_cast_fp16 = add(x = position23_to_fp16, y = valid_len_to_fp16)[name = tensor("op_5747_cast_fp16")]; + tensor var_5747_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5747_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor new_position23 = cast(dtype = var_5747_cast_fp16_to_fp32_dtype_0, x = var_5747_cast_fp16)[name = tensor("cast_322")]; + tensor new_v_cache23 = cast(dtype = new_v_cache_cast_fp16_to_fp32_dtype_0, x = new_v_cache_cast_fp16)[name = tensor("cast_324")]; + tensor new_k_cache23 = cast(dtype = new_k_cache_cast_fp16_to_fp32_dtype_0, x = new_k_cache_cast_fp16)[name = tensor("cast_326")]; + tensor new_position22 = cast(dtype = var_5608_cast_fp16_to_fp32_dtype_0, x = var_5608_cast_fp16)[name = tensor("cast_328")]; + tensor new_v_cache22 = cast(dtype = new_v_cache_45_cast_fp16_to_fp32_dtype_0, x = new_v_cache_45_cast_fp16)[name = tensor("cast_330")]; + tensor new_k_cache22 = cast(dtype = new_k_cache_45_cast_fp16_to_fp32_dtype_0, x = new_k_cache_45_cast_fp16)[name = tensor("cast_332")]; + tensor new_position21 = cast(dtype = var_5377_cast_fp16_to_fp32_dtype_0, x = var_5377_cast_fp16)[name = tensor("cast_334")]; + tensor new_v_cache21 = cast(dtype = new_v_cache_43_cast_fp16_to_fp32_dtype_0, x = new_v_cache_43_cast_fp16)[name = tensor("cast_336")]; + tensor new_k_cache21 = cast(dtype = new_k_cache_43_cast_fp16_to_fp32_dtype_0, x = new_k_cache_43_cast_fp16)[name = tensor("cast_338")]; + tensor new_position20 = cast(dtype = var_5146_cast_fp16_to_fp32_dtype_0, x = var_5146_cast_fp16)[name = tensor("cast_340")]; + tensor new_v_cache20 = cast(dtype = new_v_cache_41_cast_fp16_to_fp32_dtype_0, x = new_v_cache_41_cast_fp16)[name = tensor("cast_342")]; + tensor new_k_cache20 = cast(dtype = new_k_cache_41_cast_fp16_to_fp32_dtype_0, x = new_k_cache_41_cast_fp16)[name = tensor("cast_344")]; + tensor new_position19 = cast(dtype = var_4915_cast_fp16_to_fp32_dtype_0, x = var_4915_cast_fp16)[name = tensor("cast_346")]; + tensor new_v_cache19 = cast(dtype = new_v_cache_39_cast_fp16_to_fp32_dtype_0, x = new_v_cache_39_cast_fp16)[name = tensor("cast_348")]; + tensor new_k_cache19 = cast(dtype = new_k_cache_39_cast_fp16_to_fp32_dtype_0, x = new_k_cache_39_cast_fp16)[name = tensor("cast_350")]; + tensor new_position18 = cast(dtype = var_4684_cast_fp16_to_fp32_dtype_0, x = var_4684_cast_fp16)[name = tensor("cast_352")]; + tensor new_v_cache18 = cast(dtype = new_v_cache_37_cast_fp16_to_fp32_dtype_0, x = new_v_cache_37_cast_fp16)[name = tensor("cast_354")]; + tensor new_k_cache18 = cast(dtype = new_k_cache_37_cast_fp16_to_fp32_dtype_0, x = new_k_cache_37_cast_fp16)[name = tensor("cast_356")]; + tensor new_position17 = cast(dtype = var_4453_cast_fp16_to_fp32_dtype_0, x = var_4453_cast_fp16)[name = tensor("cast_358")]; + tensor new_v_cache17 = cast(dtype = new_v_cache_35_cast_fp16_to_fp32_dtype_0, x = new_v_cache_35_cast_fp16)[name = tensor("cast_360")]; + tensor new_k_cache17 = cast(dtype = new_k_cache_35_cast_fp16_to_fp32_dtype_0, x = new_k_cache_35_cast_fp16)[name = tensor("cast_362")]; + tensor new_position16 = cast(dtype = var_4222_cast_fp16_to_fp32_dtype_0, x = var_4222_cast_fp16)[name = tensor("cast_364")]; + tensor new_v_cache16 = cast(dtype = new_v_cache_33_cast_fp16_to_fp32_dtype_0, x = new_v_cache_33_cast_fp16)[name = tensor("cast_366")]; + tensor new_k_cache16 = cast(dtype = new_k_cache_33_cast_fp16_to_fp32_dtype_0, x = new_k_cache_33_cast_fp16)[name = tensor("cast_368")]; + tensor new_position15 = cast(dtype = var_3991_cast_fp16_to_fp32_dtype_0, x = var_3991_cast_fp16)[name = tensor("cast_370")]; + tensor new_v_cache15 = cast(dtype = new_v_cache_31_cast_fp16_to_fp32_dtype_0, x = new_v_cache_31_cast_fp16)[name = tensor("cast_372")]; + tensor new_k_cache15 = cast(dtype = new_k_cache_31_cast_fp16_to_fp32_dtype_0, x = new_k_cache_31_cast_fp16)[name = tensor("cast_374")]; + tensor new_position14 = cast(dtype = var_3760_cast_fp16_to_fp32_dtype_0, x = var_3760_cast_fp16)[name = tensor("cast_376")]; + tensor new_v_cache14 = cast(dtype = new_v_cache_29_cast_fp16_to_fp32_dtype_0, x = new_v_cache_29_cast_fp16)[name = tensor("cast_378")]; + tensor new_k_cache14 = cast(dtype = new_k_cache_29_cast_fp16_to_fp32_dtype_0, x = new_k_cache_29_cast_fp16)[name = tensor("cast_380")]; + tensor new_position13 = cast(dtype = var_3529_cast_fp16_to_fp32_dtype_0, x = var_3529_cast_fp16)[name = tensor("cast_382")]; + tensor new_v_cache13 = cast(dtype = new_v_cache_27_cast_fp16_to_fp32_dtype_0, x = new_v_cache_27_cast_fp16)[name = tensor("cast_384")]; + tensor new_k_cache13 = cast(dtype = new_k_cache_27_cast_fp16_to_fp32_dtype_0, x = new_k_cache_27_cast_fp16)[name = tensor("cast_386")]; + tensor new_position12 = cast(dtype = var_3298_cast_fp16_to_fp32_dtype_0, x = var_3298_cast_fp16)[name = tensor("cast_388")]; + tensor new_v_cache12 = cast(dtype = new_v_cache_25_cast_fp16_to_fp32_dtype_0, x = new_v_cache_25_cast_fp16)[name = tensor("cast_390")]; + tensor new_k_cache12 = cast(dtype = new_k_cache_25_cast_fp16_to_fp32_dtype_0, x = new_k_cache_25_cast_fp16)[name = tensor("cast_392")]; + tensor new_position11 = cast(dtype = var_3067_cast_fp16_to_fp32_dtype_0, x = var_3067_cast_fp16)[name = tensor("cast_394")]; + tensor new_v_cache11 = cast(dtype = new_v_cache_23_cast_fp16_to_fp32_dtype_0, x = new_v_cache_23_cast_fp16)[name = tensor("cast_396")]; + tensor new_k_cache11 = cast(dtype = new_k_cache_23_cast_fp16_to_fp32_dtype_0, x = new_k_cache_23_cast_fp16)[name = tensor("cast_398")]; + tensor new_position10 = cast(dtype = var_2836_cast_fp16_to_fp32_dtype_0, x = var_2836_cast_fp16)[name = tensor("cast_400")]; + tensor new_v_cache10 = cast(dtype = new_v_cache_21_cast_fp16_to_fp32_dtype_0, x = new_v_cache_21_cast_fp16)[name = tensor("cast_402")]; + tensor new_k_cache10 = cast(dtype = new_k_cache_21_cast_fp16_to_fp32_dtype_0, x = new_k_cache_21_cast_fp16)[name = tensor("cast_404")]; + tensor new_position9 = cast(dtype = var_2605_cast_fp16_to_fp32_dtype_0, x = var_2605_cast_fp16)[name = tensor("cast_406")]; + tensor new_v_cache9 = cast(dtype = new_v_cache_19_cast_fp16_to_fp32_dtype_0, x = new_v_cache_19_cast_fp16)[name = tensor("cast_408")]; + tensor new_k_cache9 = cast(dtype = new_k_cache_19_cast_fp16_to_fp32_dtype_0, x = new_k_cache_19_cast_fp16)[name = tensor("cast_410")]; + tensor new_position8 = cast(dtype = var_2374_cast_fp16_to_fp32_dtype_0, x = var_2374_cast_fp16)[name = tensor("cast_412")]; + tensor new_v_cache8 = cast(dtype = new_v_cache_17_cast_fp16_to_fp32_dtype_0, x = new_v_cache_17_cast_fp16)[name = tensor("cast_414")]; + tensor new_k_cache8 = cast(dtype = new_k_cache_17_cast_fp16_to_fp32_dtype_0, x = new_k_cache_17_cast_fp16)[name = tensor("cast_416")]; + tensor new_position7 = cast(dtype = var_2143_cast_fp16_to_fp32_dtype_0, x = var_2143_cast_fp16)[name = tensor("cast_418")]; + tensor new_v_cache7 = cast(dtype = new_v_cache_15_cast_fp16_to_fp32_dtype_0, x = new_v_cache_15_cast_fp16)[name = tensor("cast_420")]; + tensor new_k_cache7 = cast(dtype = new_k_cache_15_cast_fp16_to_fp32_dtype_0, x = new_k_cache_15_cast_fp16)[name = tensor("cast_422")]; + tensor new_position6 = cast(dtype = var_1912_cast_fp16_to_fp32_dtype_0, x = var_1912_cast_fp16)[name = tensor("cast_424")]; + tensor new_v_cache6 = cast(dtype = new_v_cache_13_cast_fp16_to_fp32_dtype_0, x = new_v_cache_13_cast_fp16)[name = tensor("cast_426")]; + tensor new_k_cache6 = cast(dtype = new_k_cache_13_cast_fp16_to_fp32_dtype_0, x = new_k_cache_13_cast_fp16)[name = tensor("cast_428")]; + tensor new_position5 = cast(dtype = var_1681_cast_fp16_to_fp32_dtype_0, x = var_1681_cast_fp16)[name = tensor("cast_430")]; + tensor new_v_cache5 = cast(dtype = new_v_cache_11_cast_fp16_to_fp32_dtype_0, x = new_v_cache_11_cast_fp16)[name = tensor("cast_432")]; + tensor new_k_cache5 = cast(dtype = new_k_cache_11_cast_fp16_to_fp32_dtype_0, x = new_k_cache_11_cast_fp16)[name = tensor("cast_434")]; + tensor new_position4 = cast(dtype = var_1450_cast_fp16_to_fp32_dtype_0, x = var_1450_cast_fp16)[name = tensor("cast_436")]; + tensor new_v_cache4 = cast(dtype = new_v_cache_9_cast_fp16_to_fp32_dtype_0, x = new_v_cache_9_cast_fp16)[name = tensor("cast_438")]; + tensor new_k_cache4 = cast(dtype = new_k_cache_9_cast_fp16_to_fp32_dtype_0, x = new_k_cache_9_cast_fp16)[name = tensor("cast_440")]; + tensor new_position3 = cast(dtype = var_1219_cast_fp16_to_fp32_dtype_0, x = var_1219_cast_fp16)[name = tensor("cast_442")]; + tensor new_v_cache3 = cast(dtype = new_v_cache_7_cast_fp16_to_fp32_dtype_0, x = new_v_cache_7_cast_fp16)[name = tensor("cast_444")]; + tensor new_k_cache3 = cast(dtype = new_k_cache_7_cast_fp16_to_fp32_dtype_0, x = new_k_cache_7_cast_fp16)[name = tensor("cast_446")]; + tensor new_position2 = cast(dtype = var_988_cast_fp16_to_fp32_dtype_0, x = var_988_cast_fp16)[name = tensor("cast_448")]; + tensor new_v_cache2 = cast(dtype = new_v_cache_5_cast_fp16_to_fp32_dtype_0, x = new_v_cache_5_cast_fp16)[name = tensor("cast_450")]; + tensor new_k_cache2 = cast(dtype = new_k_cache_5_cast_fp16_to_fp32_dtype_0, x = new_k_cache_5_cast_fp16)[name = tensor("cast_452")]; + tensor new_position1 = cast(dtype = var_757_cast_fp16_to_fp32_dtype_0, x = var_757_cast_fp16)[name = tensor("cast_454")]; + tensor new_v_cache1 = cast(dtype = new_v_cache_3_cast_fp16_to_fp32_dtype_0, x = new_v_cache_3_cast_fp16)[name = tensor("cast_456")]; + tensor new_k_cache1 = cast(dtype = new_k_cache_3_cast_fp16_to_fp32_dtype_0, x = new_k_cache_3_cast_fp16)[name = tensor("cast_458")]; + tensor new_position0 = cast(dtype = var_526_cast_fp16_to_fp32_dtype_0, x = var_297_cast_fp16)[name = tensor("cast_460")]; + tensor new_v_cache0 = cast(dtype = new_v_cache_1_cast_fp16_to_fp32_dtype_0, x = new_v_cache_1_cast_fp16)[name = tensor("cast_461")]; + tensor new_k_cache0 = cast(dtype = new_k_cache_1_cast_fp16_to_fp32_dtype_0, x = new_k_cache_1_cast_fp16)[name = tensor("cast_463")]; + } -> (new_k_cache0, new_v_cache0, new_position0, new_k_cache1, new_v_cache1, new_position1, new_k_cache2, new_v_cache2, new_position2, new_k_cache3, new_v_cache3, new_position3, new_k_cache4, new_v_cache4, new_position4, new_k_cache5, new_v_cache5, new_position5, new_k_cache6, new_v_cache6, new_position6, new_k_cache7, new_v_cache7, new_position7, new_k_cache8, new_v_cache8, new_position8, new_k_cache9, new_v_cache9, new_position9, new_k_cache10, new_v_cache10, new_position10, new_k_cache11, new_v_cache11, new_position11, new_k_cache12, new_v_cache12, new_position12, new_k_cache13, new_v_cache13, new_position13, new_k_cache14, new_v_cache14, new_position14, new_k_cache15, new_v_cache15, new_position15, new_k_cache16, new_v_cache16, new_position16, new_k_cache17, new_v_cache17, new_position17, new_k_cache18, new_v_cache18, new_position18, new_k_cache19, new_v_cache19, new_position19, new_k_cache20, new_v_cache20, new_position20, new_k_cache21, new_v_cache21, new_position21, new_k_cache22, new_v_cache22, new_position22, new_k_cache23, new_v_cache23, new_position23); +} \ No newline at end of file diff --git a/v2.1/german_24l/cond_prefill_ane.mlmodelc/weights/weight.bin b/v2.1/german_24l/cond_prefill_ane.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..49244c9e9c47eb13388277041932db080bdd9e91 --- /dev/null +++ b/v2.1/german_24l/cond_prefill_ane.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9e1fb0e46454fa037faeb15db2c47490fb346ad7d76245d527a061173a2f35f0 +size 585328896 diff --git a/v2.1/german_24l/cond_prefill_ane.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/german_24l/cond_prefill_ane.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..c7714655d84add4eaa746c38708af4a64cc29091 --- /dev/null +++ b/v2.1/german_24l/cond_prefill_ane.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:32f672d36f7cb3b8005ef9cedd67a672f031c22ac7586e0bd866aa2b13340c87 +size 483717 diff --git a/v2.1/german_24l/cond_prefill_ane.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/german_24l/cond_prefill_ane.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..49244c9e9c47eb13388277041932db080bdd9e91 --- /dev/null +++ b/v2.1/german_24l/cond_prefill_ane.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9e1fb0e46454fa037faeb15db2c47490fb346ad7d76245d527a061173a2f35f0 +size 585328896 diff --git a/v2.1/german_24l/cond_prefill_ane.mlpackage/Manifest.json b/v2.1/german_24l/cond_prefill_ane.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..1425c2e3be53cfdde2709981589bc4e664f48e93 --- /dev/null +++ b/v2.1/german_24l/cond_prefill_ane.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "339E5A7A-5554-4D6B-9A29-214BF2196C61": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "D22485B6-12A5-4A60-A06C-530D707502F0": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "D22485B6-12A5-4A60-A06C-530D707502F0" +} diff --git a/v2.1/german_24l/flowlm_step_ane.mlmodelc/analytics/coremldata.bin b/v2.1/german_24l/flowlm_step_ane.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..ce10686755614241437d688f84920b7fd5412cc6 --- /dev/null +++ b/v2.1/german_24l/flowlm_step_ane.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab071917524574d315739d312561ece6db88bdda107831abba69d82f42d7e225 +size 243 diff --git a/v2.1/german_24l/flowlm_step_ane.mlmodelc/coremldata.bin b/v2.1/german_24l/flowlm_step_ane.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..b80720dfe92891df5958f0fc42cac3bed3d09049 --- /dev/null +++ b/v2.1/german_24l/flowlm_step_ane.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7f681eb4e279a0ccdfc8045971a56d51b76f37550bab4567beb42e8ae1d93526 +size 4522 diff --git a/v2.1/german_24l/flowlm_step_ane.mlmodelc/model.mil b/v2.1/german_24l/flowlm_step_ane.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..9d4ce22e4ce7a6f9f3990282229571eb431d0297 --- /dev/null +++ b/v2.1/german_24l/flowlm_step_ane.mlmodelc/model.mil @@ -0,0 +1,3699 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.9.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor k_cache0, tensor k_cache1, tensor k_cache10, tensor k_cache11, tensor k_cache12, tensor k_cache13, tensor k_cache14, tensor k_cache15, tensor k_cache16, tensor k_cache17, tensor k_cache18, tensor k_cache19, tensor k_cache2, tensor k_cache20, tensor k_cache21, tensor k_cache22, tensor k_cache23, tensor k_cache3, tensor k_cache4, tensor k_cache5, tensor k_cache6, tensor k_cache7, tensor k_cache8, tensor k_cache9, tensor position0, tensor position1, tensor position10, tensor position11, tensor position12, tensor position13, tensor position14, tensor position15, tensor position16, tensor position17, tensor position18, tensor position19, tensor position2, tensor position20, tensor position21, tensor position22, tensor position23, tensor position3, tensor position4, tensor position5, tensor position6, tensor position7, tensor position8, tensor position9, tensor sequence, tensor v_cache0, tensor v_cache1, tensor v_cache10, tensor v_cache11, tensor v_cache12, tensor v_cache13, tensor v_cache14, tensor v_cache15, tensor v_cache16, tensor v_cache17, tensor v_cache18, tensor v_cache19, tensor v_cache2, tensor v_cache20, tensor v_cache21, tensor v_cache22, tensor v_cache23, tensor v_cache3, tensor v_cache4, tensor v_cache5, tensor v_cache6, tensor v_cache7, tensor v_cache8, tensor v_cache9) { + tensor sequence_to_fp16_dtype_0 = const()[name = tensor("sequence_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor input_linear_weight_to_fp16 = const()[name = tensor("input_linear_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(65664)))]; + tensor sequence_to_fp16 = cast(dtype = sequence_to_fp16_dtype_0, x = sequence)[name = tensor("cast_578")]; + tensor linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = input_linear_weight_to_fp16, x = sequence_to_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; + tensor norm0_1_weight_to_fp16 = const()[name = tensor("norm0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67776)))]; + tensor norm0_1_bias_to_fp16 = const()[name = tensor("norm0_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(69888)))]; + tensor var_227_to_fp16 = const()[name = tensor("op_227_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_5_cast_fp16 = layer_norm(axes = input_5_axes_0, beta = norm0_1_bias_to_fp16, epsilon = var_227_to_fp16, gamma = norm0_1_weight_to_fp16, x = linear_0_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor attn0_in_proj_weight_to_fp16 = const()[name = tensor("attn0_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(72000)))]; + tensor linear_1_bias_0_to_fp16 = const()[name = tensor("linear_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6363520)))]; + tensor linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn0_in_proj_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("linear_1_cast_fp16")]; + tensor var_240 = const()[name = tensor("op_240"), val = tensor([1, 3, 16, 64])]; + tensor qkv_1_cast_fp16 = reshape(shape = var_240, x = linear_1_cast_fp16)[name = tensor("qkv_1_cast_fp16")]; + tensor x_1_begin_0 = const()[name = tensor("x_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_1_end_0 = const()[name = tensor("x_1_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_1_end_mask_0 = const()[name = tensor("x_1_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_1_squeeze_mask_0 = const()[name = tensor("x_1_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_1_cast_fp16 = slice_by_index(begin = x_1_begin_0, end = x_1_end_0, end_mask = x_1_end_mask_0, squeeze_mask = x_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("x_1_cast_fp16")]; + tensor x_3_begin_0 = const()[name = tensor("x_3_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_3_end_0 = const()[name = tensor("x_3_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_3_end_mask_0 = const()[name = tensor("x_3_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_3_squeeze_mask_0 = const()[name = tensor("x_3_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_3_cast_fp16 = slice_by_index(begin = x_3_begin_0, end = x_3_end_0, end_mask = x_3_end_mask_0, squeeze_mask = x_3_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("x_3_cast_fp16")]; + tensor v_1_begin_0 = const()[name = tensor("v_1_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_1_end_0 = const()[name = tensor("v_1_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_1_end_mask_0 = const()[name = tensor("v_1_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_1_squeeze_mask_0 = const()[name = tensor("v_1_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_1_cast_fp16 = slice_by_index(begin = v_1_begin_0, end = v_1_end_0, end_mask = v_1_end_mask_0, squeeze_mask = v_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("v_1_cast_fp16")]; + tensor var_269 = const()[name = tensor("op_269"), val = tensor([1, 1, 1])]; + tensor position0_to_fp16_dtype_0 = const()[name = tensor("position0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position0_to_fp16 = cast(dtype = position0_to_fp16_dtype_0, x = position0)[name = tensor("cast_577")]; + tensor var_270_cast_fp16 = reshape(shape = var_269, x = position0_to_fp16)[name = tensor("op_270_cast_fp16")]; + tensor rope_freqs_to_fp16 = const()[name = tensor("rope_freqs_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6369728)))]; + tensor angles_1_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_270_cast_fp16)[name = tensor("angles_1_cast_fp16")]; + tensor rotr_1_cast_fp16 = cos(x = angles_1_cast_fp16)[name = tensor("rotr_1_cast_fp16")]; + tensor roti_1_cast_fp16 = sin(x = angles_1_cast_fp16)[name = tensor("roti_1_cast_fp16")]; + tensor var_278 = const()[name = tensor("op_278"), val = tensor([1, 16, 32, 2])]; + tensor pairs_1_cast_fp16 = reshape(shape = var_278, x = x_1_cast_fp16)[name = tensor("pairs_1_cast_fp16")]; + tensor xr_1_begin_0 = const()[name = tensor("xr_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_1_end_0 = const()[name = tensor("xr_1_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_1_end_mask_0 = const()[name = tensor("xr_1_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_1_squeeze_mask_0 = const()[name = tensor("xr_1_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_1_cast_fp16 = slice_by_index(begin = xr_1_begin_0, end = xr_1_end_0, end_mask = xr_1_end_mask_0, squeeze_mask = xr_1_squeeze_mask_0, x = pairs_1_cast_fp16)[name = tensor("xr_1_cast_fp16")]; + tensor xi_1_begin_0 = const()[name = tensor("xi_1_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_1_end_0 = const()[name = tensor("xi_1_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_1_end_mask_0 = const()[name = tensor("xi_1_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_1_squeeze_mask_0 = const()[name = tensor("xi_1_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_1_cast_fp16 = slice_by_index(begin = xi_1_begin_0, end = xi_1_end_0, end_mask = xi_1_end_mask_0, squeeze_mask = xi_1_squeeze_mask_0, x = pairs_1_cast_fp16)[name = tensor("xi_1_cast_fp16")]; + tensor var_286_cast_fp16 = mul(x = xr_1_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_286_cast_fp16")]; + tensor var_287_cast_fp16 = mul(x = xi_1_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_287_cast_fp16")]; + tensor xor__1_cast_fp16 = sub(x = var_286_cast_fp16, y = var_287_cast_fp16)[name = tensor("xor__1_cast_fp16")]; + tensor var_290_cast_fp16 = mul(x = xr_1_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_290_cast_fp16")]; + tensor var_291_cast_fp16 = mul(x = xi_1_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_291_cast_fp16")]; + tensor xoi_1_cast_fp16 = add(x = var_290_cast_fp16, y = var_291_cast_fp16)[name = tensor("xoi_1_cast_fp16")]; + tensor out_1_axis_0 = const()[name = tensor("out_1_axis_0"), val = tensor(-1)]; + tensor out_1_cast_fp16 = stack(axis = out_1_axis_0, values = (xor__1_cast_fp16, xoi_1_cast_fp16))[name = tensor("out_1_cast_fp16")]; + tensor var_314 = const()[name = tensor("op_314"), val = tensor([1, 16, 32, 2])]; + tensor pairs_3_cast_fp16 = reshape(shape = var_314, x = x_3_cast_fp16)[name = tensor("pairs_3_cast_fp16")]; + tensor xr_3_begin_0 = const()[name = tensor("xr_3_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_3_end_0 = const()[name = tensor("xr_3_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_3_end_mask_0 = const()[name = tensor("xr_3_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_3_squeeze_mask_0 = const()[name = tensor("xr_3_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_3_cast_fp16 = slice_by_index(begin = xr_3_begin_0, end = xr_3_end_0, end_mask = xr_3_end_mask_0, squeeze_mask = xr_3_squeeze_mask_0, x = pairs_3_cast_fp16)[name = tensor("xr_3_cast_fp16")]; + tensor xi_3_begin_0 = const()[name = tensor("xi_3_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_3_end_0 = const()[name = tensor("xi_3_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_3_end_mask_0 = const()[name = tensor("xi_3_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_3_squeeze_mask_0 = const()[name = tensor("xi_3_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_3_cast_fp16 = slice_by_index(begin = xi_3_begin_0, end = xi_3_end_0, end_mask = xi_3_end_mask_0, squeeze_mask = xi_3_squeeze_mask_0, x = pairs_3_cast_fp16)[name = tensor("xi_3_cast_fp16")]; + tensor var_322_cast_fp16 = mul(x = xr_3_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_322_cast_fp16")]; + tensor var_323_cast_fp16 = mul(x = xi_3_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_323_cast_fp16")]; + tensor xor__3_cast_fp16 = sub(x = var_322_cast_fp16, y = var_323_cast_fp16)[name = tensor("xor__3_cast_fp16")]; + tensor var_326_cast_fp16 = mul(x = xr_3_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_326_cast_fp16")]; + tensor var_327_cast_fp16 = mul(x = xi_3_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_327_cast_fp16")]; + tensor xoi_3_cast_fp16 = add(x = var_326_cast_fp16, y = var_327_cast_fp16)[name = tensor("xoi_3_cast_fp16")]; + tensor out_3_axis_0 = const()[name = tensor("out_3_axis_0"), val = tensor(-1)]; + tensor out_3_cast_fp16 = stack(axis = out_3_axis_0, values = (xor__3_cast_fp16, xoi_3_cast_fp16))[name = tensor("out_3_cast_fp16")]; + tensor _inversed_339_y_0_to_fp16 = const()[name = tensor("_inversed_339_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_339_cast_fp16 = mul(x = position0_to_fp16, y = _inversed_339_y_0_to_fp16)[name = tensor("_inversed_339_cast_fp16")]; + tensor var_340_cast_fp16 = floor(x = _inversed_339_cast_fp16)[name = tensor("op_340_cast_fp16")]; + tensor var_341_to_fp16 = const()[name = tensor("op_341_to_fp16"), val = tensor(0x1p+9)]; + tensor var_342_cast_fp16 = mul(x = var_340_cast_fp16, y = var_341_to_fp16)[name = tensor("op_342_cast_fp16")]; + tensor write_idx_1_cast_fp16 = sub(x = position0_to_fp16, y = var_342_cast_fp16)[name = tensor("write_idx_1_cast_fp16")]; + tensor slot_indices_to_fp16 = const()[name = tensor("slot_indices_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6369856)))]; + tensor var_345_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_1_cast_fp16)[name = tensor("op_345_cast_fp16")]; + tensor var_355 = const()[name = tensor("op_355"), val = tensor([1, 512, 1, 1])]; + tensor onehot_1_to_fp16_dtype_0 = const()[name = tensor("onehot_1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_345_cast_fp16_to_fp16 = cast(dtype = onehot_1_to_fp16_dtype_0, x = var_345_cast_fp16)[name = tensor("cast_576")]; + tensor onehot_3_cast_fp16 = reshape(shape = var_355, x = var_345_cast_fp16_to_fp16)[name = tensor("onehot_3_cast_fp16")]; + tensor var_357_to_fp16 = const()[name = tensor("op_357_to_fp16"), val = tensor(0x1p+0)]; + tensor var_359_cast_fp16 = sub(x = var_357_to_fp16, y = onehot_3_cast_fp16)[name = tensor("op_359_cast_fp16")]; + tensor k_cache0_to_fp16_dtype_0 = const()[name = tensor("k_cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache0_to_fp16 = cast(dtype = k_cache0_to_fp16_dtype_0, x = k_cache0)[name = tensor("cast_575")]; + tensor var_360_cast_fp16 = mul(x = k_cache0_to_fp16, y = var_359_cast_fp16)[name = tensor("op_360_cast_fp16")]; + tensor var_365 = const()[name = tensor("op_365"), val = tensor([1, 1, 16, 64])]; + tensor var_366_cast_fp16 = reshape(shape = var_365, x = out_3_cast_fp16)[name = tensor("op_366_cast_fp16")]; + tensor var_367_cast_fp16 = mul(x = var_366_cast_fp16, y = onehot_3_cast_fp16)[name = tensor("op_367_cast_fp16")]; + tensor new_k_cache_1_cast_fp16 = add(x = var_360_cast_fp16, y = var_367_cast_fp16)[name = tensor("new_k_cache_1_cast_fp16")]; + tensor new_k_cache_1_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_1_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache0_to_fp16_dtype_0 = const()[name = tensor("v_cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache0_to_fp16 = cast(dtype = v_cache0_to_fp16_dtype_0, x = v_cache0)[name = tensor("cast_573")]; + tensor var_373_cast_fp16 = mul(x = v_cache0_to_fp16, y = var_359_cast_fp16)[name = tensor("op_373_cast_fp16")]; + tensor var_378 = const()[name = tensor("op_378"), val = tensor([1, 1, 16, 64])]; + tensor var_379_cast_fp16 = reshape(shape = var_378, x = v_1_cast_fp16)[name = tensor("op_379_cast_fp16")]; + tensor var_380_cast_fp16 = mul(x = var_379_cast_fp16, y = onehot_3_cast_fp16)[name = tensor("op_380_cast_fp16")]; + tensor new_v_cache_1_cast_fp16 = add(x = var_373_cast_fp16, y = var_380_cast_fp16)[name = tensor("new_v_cache_1_cast_fp16")]; + tensor new_v_cache_1_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_1_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_0_perm_0 = const()[name = tensor("transpose_0_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_393 = const()[name = tensor("op_393"), val = tensor([0, 2, 1, 3])]; + tensor var_399 = const()[name = tensor("op_399"), val = tensor([1, 16, 1, 64])]; + tensor q4_1_cast_fp16 = reshape(shape = var_399, x = out_1_cast_fp16)[name = tensor("q4_1_cast_fp16")]; + tensor var_404_transpose_x_0 = const()[name = tensor("op_404_transpose_x_0"), val = tensor(false)]; + tensor var_404_transpose_y_0 = const()[name = tensor("op_404_transpose_y_0"), val = tensor(false)]; + tensor transpose_0_cast_fp16 = transpose(perm = transpose_0_perm_0, x = new_k_cache_1_cast_fp16)[name = tensor("transpose_95")]; + tensor var_404_cast_fp16 = matmul(transpose_x = var_404_transpose_x_0, transpose_y = var_404_transpose_y_0, x = q4_1_cast_fp16, y = transpose_0_cast_fp16)[name = tensor("op_404_cast_fp16")]; + tensor var_405_to_fp16 = const()[name = tensor("op_405_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_1_cast_fp16 = mul(x = var_404_cast_fp16, y = var_405_to_fp16)[name = tensor("scores_1_cast_fp16")]; + tensor var_409 = const()[name = tensor("op_409"), val = tensor([1, 1])]; + tensor var_410_cast_fp16 = reshape(shape = var_409, x = position0_to_fp16)[name = tensor("op_410_cast_fp16")]; + tensor var_411_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_410_cast_fp16)[name = tensor("op_411_cast_fp16")]; + tensor mask_1_to_fp16_dtype_0 = const()[name = tensor("mask_1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_418_to_fp16 = const()[name = tensor("op_418_to_fp16"), val = tensor(0x1p+0)]; + tensor var_411_cast_fp16_to_fp16 = cast(dtype = mask_1_to_fp16_dtype_0, x = var_411_cast_fp16)[name = tensor("cast_571")]; + tensor var_419_cast_fp16 = sub(x = var_411_cast_fp16_to_fp16, y = var_418_to_fp16)[name = tensor("op_419_cast_fp16")]; + tensor var_424 = const()[name = tensor("op_424"), val = tensor([1, 1, 1, 512])]; + tensor var_425_cast_fp16 = reshape(shape = var_424, x = var_419_cast_fp16)[name = tensor("op_425_cast_fp16")]; + tensor var_426_to_fp16 = const()[name = tensor("op_426_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_427_cast_fp16 = mul(x = var_425_cast_fp16, y = var_426_to_fp16)[name = tensor("op_427_cast_fp16")]; + tensor scores_3_cast_fp16 = add(x = scores_1_cast_fp16, y = var_427_cast_fp16)[name = tensor("scores_3_cast_fp16")]; + tensor var_430 = const()[name = tensor("op_430"), val = tensor(-1)]; + tensor weights_1_cast_fp16 = softmax(axis = var_430, x = scores_3_cast_fp16)[name = tensor("weights_1_cast_fp16")]; + tensor attn_1_transpose_x_0 = const()[name = tensor("attn_1_transpose_x_0"), val = tensor(false)]; + tensor attn_1_transpose_y_0 = const()[name = tensor("attn_1_transpose_y_0"), val = tensor(false)]; + tensor values_1_cast_fp16 = transpose(perm = var_393, x = new_v_cache_1_cast_fp16)[name = tensor("transpose_94")]; + tensor attn_1_cast_fp16 = matmul(transpose_x = attn_1_transpose_x_0, transpose_y = attn_1_transpose_y_0, x = weights_1_cast_fp16, y = values_1_cast_fp16)[name = tensor("attn_1_cast_fp16")]; + tensor var_438 = const()[name = tensor("op_438"), val = tensor([0, 2, 1, 3])]; + tensor var_443 = const()[name = tensor("op_443"), val = tensor([1, 1, 1024])]; + tensor var_439_cast_fp16 = transpose(perm = var_438, x = attn_1_cast_fp16)[name = tensor("transpose_93")]; + tensor input_7_cast_fp16 = reshape(shape = var_443, x = var_439_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor attn0_out_proj_weight_to_fp16 = const()[name = tensor("attn0_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6370944)))]; + tensor linear_2_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn0_out_proj_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor var_449_to_fp16 = const()[name = tensor("op_449_to_fp16"), val = tensor(0x1p+0)]; + tensor var_450_cast_fp16 = add(x = position0_to_fp16, y = var_449_to_fp16)[name = tensor("op_450_cast_fp16")]; + tensor var_450_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_450_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_9_cast_fp16 = add(x = linear_0_cast_fp16, y = linear_2_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor input_11_axes_0 = const()[name = tensor("input_11_axes_0"), val = tensor([-1])]; + tensor norm0_2_weight_to_fp16 = const()[name = tensor("norm0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8468160)))]; + tensor norm0_2_bias_to_fp16 = const()[name = tensor("norm0_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8470272)))]; + tensor var_454_to_fp16 = const()[name = tensor("op_454_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_11_cast_fp16 = layer_norm(axes = input_11_axes_0, beta = norm0_2_bias_to_fp16, epsilon = var_454_to_fp16, gamma = norm0_2_weight_to_fp16, x = input_9_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor linear0_1_weight_to_fp16 = const()[name = tensor("linear0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8472384)))]; + tensor linear_3_bias_0_to_fp16 = const()[name = tensor("linear_3_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16861056)))]; + tensor linear_3_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear0_1_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("linear_3_cast_fp16")]; + tensor input_13_mode_0 = const()[name = tensor("input_13_mode_0"), val = tensor("EXACT")]; + tensor input_13_cast_fp16 = gelu(mode = input_13_mode_0, x = linear_3_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor linear0_2_weight_to_fp16 = const()[name = tensor("linear0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16869312)))]; + tensor linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear0_2_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor input_15_cast_fp16 = add(x = input_9_cast_fp16, y = linear_4_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor input_17_axes_0 = const()[name = tensor("input_17_axes_0"), val = tensor([-1])]; + tensor norm1_1_weight_to_fp16 = const()[name = tensor("norm1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25257984)))]; + tensor norm1_1_bias_to_fp16 = const()[name = tensor("norm1_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25260096)))]; + tensor var_471_to_fp16 = const()[name = tensor("op_471_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_17_cast_fp16 = layer_norm(axes = input_17_axes_0, beta = norm1_1_bias_to_fp16, epsilon = var_471_to_fp16, gamma = norm1_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor attn1_in_proj_weight_to_fp16 = const()[name = tensor("attn1_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25262208)))]; + tensor linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn1_in_proj_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor var_484 = const()[name = tensor("op_484"), val = tensor([1, 3, 16, 64])]; + tensor qkv_3_cast_fp16 = reshape(shape = var_484, x = linear_5_cast_fp16)[name = tensor("qkv_3_cast_fp16")]; + tensor x_5_begin_0 = const()[name = tensor("x_5_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_5_end_0 = const()[name = tensor("x_5_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_5_end_mask_0 = const()[name = tensor("x_5_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_5_squeeze_mask_0 = const()[name = tensor("x_5_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_5_cast_fp16 = slice_by_index(begin = x_5_begin_0, end = x_5_end_0, end_mask = x_5_end_mask_0, squeeze_mask = x_5_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("x_5_cast_fp16")]; + tensor x_7_begin_0 = const()[name = tensor("x_7_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_7_end_0 = const()[name = tensor("x_7_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_7_end_mask_0 = const()[name = tensor("x_7_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_7_squeeze_mask_0 = const()[name = tensor("x_7_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_7_cast_fp16 = slice_by_index(begin = x_7_begin_0, end = x_7_end_0, end_mask = x_7_end_mask_0, squeeze_mask = x_7_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("x_7_cast_fp16")]; + tensor v_3_begin_0 = const()[name = tensor("v_3_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_3_end_0 = const()[name = tensor("v_3_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_3_end_mask_0 = const()[name = tensor("v_3_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_3_squeeze_mask_0 = const()[name = tensor("v_3_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_3_cast_fp16 = slice_by_index(begin = v_3_begin_0, end = v_3_end_0, end_mask = v_3_end_mask_0, squeeze_mask = v_3_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("v_3_cast_fp16")]; + tensor var_513 = const()[name = tensor("op_513"), val = tensor([1, 1, 1])]; + tensor position1_to_fp16_dtype_0 = const()[name = tensor("position1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position1_to_fp16 = cast(dtype = position1_to_fp16_dtype_0, x = position1)[name = tensor("cast_569")]; + tensor var_514_cast_fp16 = reshape(shape = var_513, x = position1_to_fp16)[name = tensor("op_514_cast_fp16")]; + tensor angles_5_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_514_cast_fp16)[name = tensor("angles_5_cast_fp16")]; + tensor rotr_5_cast_fp16 = cos(x = angles_5_cast_fp16)[name = tensor("rotr_5_cast_fp16")]; + tensor roti_5_cast_fp16 = sin(x = angles_5_cast_fp16)[name = tensor("roti_5_cast_fp16")]; + tensor var_522 = const()[name = tensor("op_522"), val = tensor([1, 16, 32, 2])]; + tensor pairs_5_cast_fp16 = reshape(shape = var_522, x = x_5_cast_fp16)[name = tensor("pairs_5_cast_fp16")]; + tensor xr_5_begin_0 = const()[name = tensor("xr_5_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_5_end_0 = const()[name = tensor("xr_5_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_5_end_mask_0 = const()[name = tensor("xr_5_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_5_squeeze_mask_0 = const()[name = tensor("xr_5_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_5_cast_fp16 = slice_by_index(begin = xr_5_begin_0, end = xr_5_end_0, end_mask = xr_5_end_mask_0, squeeze_mask = xr_5_squeeze_mask_0, x = pairs_5_cast_fp16)[name = tensor("xr_5_cast_fp16")]; + tensor xi_5_begin_0 = const()[name = tensor("xi_5_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_5_end_0 = const()[name = tensor("xi_5_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_5_end_mask_0 = const()[name = tensor("xi_5_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_5_squeeze_mask_0 = const()[name = tensor("xi_5_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_5_cast_fp16 = slice_by_index(begin = xi_5_begin_0, end = xi_5_end_0, end_mask = xi_5_end_mask_0, squeeze_mask = xi_5_squeeze_mask_0, x = pairs_5_cast_fp16)[name = tensor("xi_5_cast_fp16")]; + tensor var_530_cast_fp16 = mul(x = xr_5_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_530_cast_fp16")]; + tensor var_531_cast_fp16 = mul(x = xi_5_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_531_cast_fp16")]; + tensor xor__5_cast_fp16 = sub(x = var_530_cast_fp16, y = var_531_cast_fp16)[name = tensor("xor__5_cast_fp16")]; + tensor var_534_cast_fp16 = mul(x = xr_5_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_534_cast_fp16")]; + tensor var_535_cast_fp16 = mul(x = xi_5_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_535_cast_fp16")]; + tensor xoi_5_cast_fp16 = add(x = var_534_cast_fp16, y = var_535_cast_fp16)[name = tensor("xoi_5_cast_fp16")]; + tensor out_5_axis_0 = const()[name = tensor("out_5_axis_0"), val = tensor(-1)]; + tensor out_5_cast_fp16 = stack(axis = out_5_axis_0, values = (xor__5_cast_fp16, xoi_5_cast_fp16))[name = tensor("out_5_cast_fp16")]; + tensor var_558 = const()[name = tensor("op_558"), val = tensor([1, 16, 32, 2])]; + tensor pairs_7_cast_fp16 = reshape(shape = var_558, x = x_7_cast_fp16)[name = tensor("pairs_7_cast_fp16")]; + tensor xr_7_begin_0 = const()[name = tensor("xr_7_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_7_end_0 = const()[name = tensor("xr_7_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_7_end_mask_0 = const()[name = tensor("xr_7_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_7_squeeze_mask_0 = const()[name = tensor("xr_7_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_7_cast_fp16 = slice_by_index(begin = xr_7_begin_0, end = xr_7_end_0, end_mask = xr_7_end_mask_0, squeeze_mask = xr_7_squeeze_mask_0, x = pairs_7_cast_fp16)[name = tensor("xr_7_cast_fp16")]; + tensor xi_7_begin_0 = const()[name = tensor("xi_7_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_7_end_0 = const()[name = tensor("xi_7_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_7_end_mask_0 = const()[name = tensor("xi_7_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_7_squeeze_mask_0 = const()[name = tensor("xi_7_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_7_cast_fp16 = slice_by_index(begin = xi_7_begin_0, end = xi_7_end_0, end_mask = xi_7_end_mask_0, squeeze_mask = xi_7_squeeze_mask_0, x = pairs_7_cast_fp16)[name = tensor("xi_7_cast_fp16")]; + tensor var_566_cast_fp16 = mul(x = xr_7_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_566_cast_fp16")]; + tensor var_567_cast_fp16 = mul(x = xi_7_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_567_cast_fp16")]; + tensor xor__7_cast_fp16 = sub(x = var_566_cast_fp16, y = var_567_cast_fp16)[name = tensor("xor__7_cast_fp16")]; + tensor var_570_cast_fp16 = mul(x = xr_7_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_570_cast_fp16")]; + tensor var_571_cast_fp16 = mul(x = xi_7_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_571_cast_fp16")]; + tensor xoi_7_cast_fp16 = add(x = var_570_cast_fp16, y = var_571_cast_fp16)[name = tensor("xoi_7_cast_fp16")]; + tensor out_7_axis_0 = const()[name = tensor("out_7_axis_0"), val = tensor(-1)]; + tensor out_7_cast_fp16 = stack(axis = out_7_axis_0, values = (xor__7_cast_fp16, xoi_7_cast_fp16))[name = tensor("out_7_cast_fp16")]; + tensor _inversed_583_y_0_to_fp16 = const()[name = tensor("_inversed_583_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_583_cast_fp16 = mul(x = position1_to_fp16, y = _inversed_583_y_0_to_fp16)[name = tensor("_inversed_583_cast_fp16")]; + tensor var_584_cast_fp16 = floor(x = _inversed_583_cast_fp16)[name = tensor("op_584_cast_fp16")]; + tensor var_585_to_fp16 = const()[name = tensor("op_585_to_fp16"), val = tensor(0x1p+9)]; + tensor var_586_cast_fp16 = mul(x = var_584_cast_fp16, y = var_585_to_fp16)[name = tensor("op_586_cast_fp16")]; + tensor write_idx_3_cast_fp16 = sub(x = position1_to_fp16, y = var_586_cast_fp16)[name = tensor("write_idx_3_cast_fp16")]; + tensor var_589_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_3_cast_fp16)[name = tensor("op_589_cast_fp16")]; + tensor var_599 = const()[name = tensor("op_599"), val = tensor([1, 512, 1, 1])]; + tensor onehot_5_to_fp16_dtype_0 = const()[name = tensor("onehot_5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_589_cast_fp16_to_fp16 = cast(dtype = onehot_5_to_fp16_dtype_0, x = var_589_cast_fp16)[name = tensor("cast_568")]; + tensor onehot_7_cast_fp16 = reshape(shape = var_599, x = var_589_cast_fp16_to_fp16)[name = tensor("onehot_7_cast_fp16")]; + tensor var_601_to_fp16 = const()[name = tensor("op_601_to_fp16"), val = tensor(0x1p+0)]; + tensor var_603_cast_fp16 = sub(x = var_601_to_fp16, y = onehot_7_cast_fp16)[name = tensor("op_603_cast_fp16")]; + tensor k_cache1_to_fp16_dtype_0 = const()[name = tensor("k_cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache1_to_fp16 = cast(dtype = k_cache1_to_fp16_dtype_0, x = k_cache1)[name = tensor("cast_567")]; + tensor var_604_cast_fp16 = mul(x = k_cache1_to_fp16, y = var_603_cast_fp16)[name = tensor("op_604_cast_fp16")]; + tensor var_609 = const()[name = tensor("op_609"), val = tensor([1, 1, 16, 64])]; + tensor var_610_cast_fp16 = reshape(shape = var_609, x = out_7_cast_fp16)[name = tensor("op_610_cast_fp16")]; + tensor var_611_cast_fp16 = mul(x = var_610_cast_fp16, y = onehot_7_cast_fp16)[name = tensor("op_611_cast_fp16")]; + tensor new_k_cache_3_cast_fp16 = add(x = var_604_cast_fp16, y = var_611_cast_fp16)[name = tensor("new_k_cache_3_cast_fp16")]; + tensor new_k_cache_3_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_3_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache1_to_fp16_dtype_0 = const()[name = tensor("v_cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache1_to_fp16 = cast(dtype = v_cache1_to_fp16_dtype_0, x = v_cache1)[name = tensor("cast_565")]; + tensor var_617_cast_fp16 = mul(x = v_cache1_to_fp16, y = var_603_cast_fp16)[name = tensor("op_617_cast_fp16")]; + tensor var_622 = const()[name = tensor("op_622"), val = tensor([1, 1, 16, 64])]; + tensor var_623_cast_fp16 = reshape(shape = var_622, x = v_3_cast_fp16)[name = tensor("op_623_cast_fp16")]; + tensor var_624_cast_fp16 = mul(x = var_623_cast_fp16, y = onehot_7_cast_fp16)[name = tensor("op_624_cast_fp16")]; + tensor new_v_cache_3_cast_fp16 = add(x = var_617_cast_fp16, y = var_624_cast_fp16)[name = tensor("new_v_cache_3_cast_fp16")]; + tensor new_v_cache_3_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_3_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_1_perm_0 = const()[name = tensor("transpose_1_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_637 = const()[name = tensor("op_637"), val = tensor([0, 2, 1, 3])]; + tensor var_643 = const()[name = tensor("op_643"), val = tensor([1, 16, 1, 64])]; + tensor q4_3_cast_fp16 = reshape(shape = var_643, x = out_5_cast_fp16)[name = tensor("q4_3_cast_fp16")]; + tensor var_648_transpose_x_0 = const()[name = tensor("op_648_transpose_x_0"), val = tensor(false)]; + tensor var_648_transpose_y_0 = const()[name = tensor("op_648_transpose_y_0"), val = tensor(false)]; + tensor transpose_1_cast_fp16 = transpose(perm = transpose_1_perm_0, x = new_k_cache_3_cast_fp16)[name = tensor("transpose_92")]; + tensor var_648_cast_fp16 = matmul(transpose_x = var_648_transpose_x_0, transpose_y = var_648_transpose_y_0, x = q4_3_cast_fp16, y = transpose_1_cast_fp16)[name = tensor("op_648_cast_fp16")]; + tensor var_649_to_fp16 = const()[name = tensor("op_649_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_5_cast_fp16 = mul(x = var_648_cast_fp16, y = var_649_to_fp16)[name = tensor("scores_5_cast_fp16")]; + tensor var_653 = const()[name = tensor("op_653"), val = tensor([1, 1])]; + tensor var_654_cast_fp16 = reshape(shape = var_653, x = position1_to_fp16)[name = tensor("op_654_cast_fp16")]; + tensor var_655_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_654_cast_fp16)[name = tensor("op_655_cast_fp16")]; + tensor mask_3_to_fp16_dtype_0 = const()[name = tensor("mask_3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_662_to_fp16 = const()[name = tensor("op_662_to_fp16"), val = tensor(0x1p+0)]; + tensor var_655_cast_fp16_to_fp16 = cast(dtype = mask_3_to_fp16_dtype_0, x = var_655_cast_fp16)[name = tensor("cast_563")]; + tensor var_663_cast_fp16 = sub(x = var_655_cast_fp16_to_fp16, y = var_662_to_fp16)[name = tensor("op_663_cast_fp16")]; + tensor var_668 = const()[name = tensor("op_668"), val = tensor([1, 1, 1, 512])]; + tensor var_669_cast_fp16 = reshape(shape = var_668, x = var_663_cast_fp16)[name = tensor("op_669_cast_fp16")]; + tensor var_670_to_fp16 = const()[name = tensor("op_670_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_671_cast_fp16 = mul(x = var_669_cast_fp16, y = var_670_to_fp16)[name = tensor("op_671_cast_fp16")]; + tensor scores_7_cast_fp16 = add(x = scores_5_cast_fp16, y = var_671_cast_fp16)[name = tensor("scores_7_cast_fp16")]; + tensor var_674 = const()[name = tensor("op_674"), val = tensor(-1)]; + tensor weights_3_cast_fp16 = softmax(axis = var_674, x = scores_7_cast_fp16)[name = tensor("weights_3_cast_fp16")]; + tensor attn_3_transpose_x_0 = const()[name = tensor("attn_3_transpose_x_0"), val = tensor(false)]; + tensor attn_3_transpose_y_0 = const()[name = tensor("attn_3_transpose_y_0"), val = tensor(false)]; + tensor values_3_cast_fp16 = transpose(perm = var_637, x = new_v_cache_3_cast_fp16)[name = tensor("transpose_91")]; + tensor attn_3_cast_fp16 = matmul(transpose_x = attn_3_transpose_x_0, transpose_y = attn_3_transpose_y_0, x = weights_3_cast_fp16, y = values_3_cast_fp16)[name = tensor("attn_3_cast_fp16")]; + tensor var_682 = const()[name = tensor("op_682"), val = tensor([0, 2, 1, 3])]; + tensor var_687 = const()[name = tensor("op_687"), val = tensor([1, 1, 1024])]; + tensor var_683_cast_fp16 = transpose(perm = var_682, x = attn_3_cast_fp16)[name = tensor("transpose_90")]; + tensor input_19_cast_fp16 = reshape(shape = var_687, x = var_683_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor attn1_out_proj_weight_to_fp16 = const()[name = tensor("attn1_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31553728)))]; + tensor linear_6_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn1_out_proj_weight_to_fp16, x = input_19_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor var_693_to_fp16 = const()[name = tensor("op_693_to_fp16"), val = tensor(0x1p+0)]; + tensor var_694_cast_fp16 = add(x = position1_to_fp16, y = var_693_to_fp16)[name = tensor("op_694_cast_fp16")]; + tensor var_694_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_694_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_21_cast_fp16 = add(x = input_15_cast_fp16, y = linear_6_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor input_23_axes_0 = const()[name = tensor("input_23_axes_0"), val = tensor([-1])]; + tensor norm1_2_weight_to_fp16 = const()[name = tensor("norm1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33650944)))]; + tensor norm1_2_bias_to_fp16 = const()[name = tensor("norm1_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33653056)))]; + tensor var_698_to_fp16 = const()[name = tensor("op_698_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_23_cast_fp16 = layer_norm(axes = input_23_axes_0, beta = norm1_2_bias_to_fp16, epsilon = var_698_to_fp16, gamma = norm1_2_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor linear1_1_weight_to_fp16 = const()[name = tensor("linear1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33655168)))]; + tensor linear_7_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear1_1_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_25_mode_0 = const()[name = tensor("input_25_mode_0"), val = tensor("EXACT")]; + tensor input_25_cast_fp16 = gelu(mode = input_25_mode_0, x = linear_7_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor linear1_2_weight_to_fp16 = const()[name = tensor("linear1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42043840)))]; + tensor linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear1_2_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor input_27_cast_fp16 = add(x = input_21_cast_fp16, y = linear_8_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor input_29_axes_0 = const()[name = tensor("input_29_axes_0"), val = tensor([-1])]; + tensor norm2_1_weight_to_fp16 = const()[name = tensor("norm2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50432512)))]; + tensor norm2_1_bias_to_fp16 = const()[name = tensor("norm2_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50434624)))]; + tensor var_715_to_fp16 = const()[name = tensor("op_715_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_29_cast_fp16 = layer_norm(axes = input_29_axes_0, beta = norm2_1_bias_to_fp16, epsilon = var_715_to_fp16, gamma = norm2_1_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor attn2_in_proj_weight_to_fp16 = const()[name = tensor("attn2_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50436736)))]; + tensor linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn2_in_proj_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor var_728 = const()[name = tensor("op_728"), val = tensor([1, 3, 16, 64])]; + tensor qkv_5_cast_fp16 = reshape(shape = var_728, x = linear_9_cast_fp16)[name = tensor("qkv_5_cast_fp16")]; + tensor x_9_begin_0 = const()[name = tensor("x_9_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_9_end_0 = const()[name = tensor("x_9_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_9_end_mask_0 = const()[name = tensor("x_9_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_9_squeeze_mask_0 = const()[name = tensor("x_9_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_9_cast_fp16 = slice_by_index(begin = x_9_begin_0, end = x_9_end_0, end_mask = x_9_end_mask_0, squeeze_mask = x_9_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("x_9_cast_fp16")]; + tensor x_11_begin_0 = const()[name = tensor("x_11_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_11_end_0 = const()[name = tensor("x_11_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_11_end_mask_0 = const()[name = tensor("x_11_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_11_squeeze_mask_0 = const()[name = tensor("x_11_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_11_cast_fp16 = slice_by_index(begin = x_11_begin_0, end = x_11_end_0, end_mask = x_11_end_mask_0, squeeze_mask = x_11_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("x_11_cast_fp16")]; + tensor v_5_begin_0 = const()[name = tensor("v_5_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_5_end_0 = const()[name = tensor("v_5_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_5_end_mask_0 = const()[name = tensor("v_5_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_5_squeeze_mask_0 = const()[name = tensor("v_5_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_5_cast_fp16 = slice_by_index(begin = v_5_begin_0, end = v_5_end_0, end_mask = v_5_end_mask_0, squeeze_mask = v_5_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("v_5_cast_fp16")]; + tensor var_757 = const()[name = tensor("op_757"), val = tensor([1, 1, 1])]; + tensor position2_to_fp16_dtype_0 = const()[name = tensor("position2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position2_to_fp16 = cast(dtype = position2_to_fp16_dtype_0, x = position2)[name = tensor("cast_561")]; + tensor var_758_cast_fp16 = reshape(shape = var_757, x = position2_to_fp16)[name = tensor("op_758_cast_fp16")]; + tensor angles_9_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_758_cast_fp16)[name = tensor("angles_9_cast_fp16")]; + tensor rotr_9_cast_fp16 = cos(x = angles_9_cast_fp16)[name = tensor("rotr_9_cast_fp16")]; + tensor roti_9_cast_fp16 = sin(x = angles_9_cast_fp16)[name = tensor("roti_9_cast_fp16")]; + tensor var_766 = const()[name = tensor("op_766"), val = tensor([1, 16, 32, 2])]; + tensor pairs_9_cast_fp16 = reshape(shape = var_766, x = x_9_cast_fp16)[name = tensor("pairs_9_cast_fp16")]; + tensor xr_9_begin_0 = const()[name = tensor("xr_9_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_9_end_0 = const()[name = tensor("xr_9_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_9_end_mask_0 = const()[name = tensor("xr_9_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_9_squeeze_mask_0 = const()[name = tensor("xr_9_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_9_cast_fp16 = slice_by_index(begin = xr_9_begin_0, end = xr_9_end_0, end_mask = xr_9_end_mask_0, squeeze_mask = xr_9_squeeze_mask_0, x = pairs_9_cast_fp16)[name = tensor("xr_9_cast_fp16")]; + tensor xi_9_begin_0 = const()[name = tensor("xi_9_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_9_end_0 = const()[name = tensor("xi_9_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_9_end_mask_0 = const()[name = tensor("xi_9_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_9_squeeze_mask_0 = const()[name = tensor("xi_9_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_9_cast_fp16 = slice_by_index(begin = xi_9_begin_0, end = xi_9_end_0, end_mask = xi_9_end_mask_0, squeeze_mask = xi_9_squeeze_mask_0, x = pairs_9_cast_fp16)[name = tensor("xi_9_cast_fp16")]; + tensor var_774_cast_fp16 = mul(x = xr_9_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_774_cast_fp16")]; + tensor var_775_cast_fp16 = mul(x = xi_9_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_775_cast_fp16")]; + tensor xor__9_cast_fp16 = sub(x = var_774_cast_fp16, y = var_775_cast_fp16)[name = tensor("xor__9_cast_fp16")]; + tensor var_778_cast_fp16 = mul(x = xr_9_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_778_cast_fp16")]; + tensor var_779_cast_fp16 = mul(x = xi_9_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_779_cast_fp16")]; + tensor xoi_9_cast_fp16 = add(x = var_778_cast_fp16, y = var_779_cast_fp16)[name = tensor("xoi_9_cast_fp16")]; + tensor out_9_axis_0 = const()[name = tensor("out_9_axis_0"), val = tensor(-1)]; + tensor out_9_cast_fp16 = stack(axis = out_9_axis_0, values = (xor__9_cast_fp16, xoi_9_cast_fp16))[name = tensor("out_9_cast_fp16")]; + tensor var_802 = const()[name = tensor("op_802"), val = tensor([1, 16, 32, 2])]; + tensor pairs_11_cast_fp16 = reshape(shape = var_802, x = x_11_cast_fp16)[name = tensor("pairs_11_cast_fp16")]; + tensor xr_11_begin_0 = const()[name = tensor("xr_11_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_11_end_0 = const()[name = tensor("xr_11_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_11_end_mask_0 = const()[name = tensor("xr_11_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_11_squeeze_mask_0 = const()[name = tensor("xr_11_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_11_cast_fp16 = slice_by_index(begin = xr_11_begin_0, end = xr_11_end_0, end_mask = xr_11_end_mask_0, squeeze_mask = xr_11_squeeze_mask_0, x = pairs_11_cast_fp16)[name = tensor("xr_11_cast_fp16")]; + tensor xi_11_begin_0 = const()[name = tensor("xi_11_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_11_end_0 = const()[name = tensor("xi_11_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_11_end_mask_0 = const()[name = tensor("xi_11_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_11_squeeze_mask_0 = const()[name = tensor("xi_11_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_11_cast_fp16 = slice_by_index(begin = xi_11_begin_0, end = xi_11_end_0, end_mask = xi_11_end_mask_0, squeeze_mask = xi_11_squeeze_mask_0, x = pairs_11_cast_fp16)[name = tensor("xi_11_cast_fp16")]; + tensor var_810_cast_fp16 = mul(x = xr_11_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_810_cast_fp16")]; + tensor var_811_cast_fp16 = mul(x = xi_11_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_811_cast_fp16")]; + tensor xor__11_cast_fp16 = sub(x = var_810_cast_fp16, y = var_811_cast_fp16)[name = tensor("xor__11_cast_fp16")]; + tensor var_814_cast_fp16 = mul(x = xr_11_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_814_cast_fp16")]; + tensor var_815_cast_fp16 = mul(x = xi_11_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_815_cast_fp16")]; + tensor xoi_11_cast_fp16 = add(x = var_814_cast_fp16, y = var_815_cast_fp16)[name = tensor("xoi_11_cast_fp16")]; + tensor out_11_axis_0 = const()[name = tensor("out_11_axis_0"), val = tensor(-1)]; + tensor out_11_cast_fp16 = stack(axis = out_11_axis_0, values = (xor__11_cast_fp16, xoi_11_cast_fp16))[name = tensor("out_11_cast_fp16")]; + tensor _inversed_827_y_0_to_fp16 = const()[name = tensor("_inversed_827_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_827_cast_fp16 = mul(x = position2_to_fp16, y = _inversed_827_y_0_to_fp16)[name = tensor("_inversed_827_cast_fp16")]; + tensor var_828_cast_fp16 = floor(x = _inversed_827_cast_fp16)[name = tensor("op_828_cast_fp16")]; + tensor var_829_to_fp16 = const()[name = tensor("op_829_to_fp16"), val = tensor(0x1p+9)]; + tensor var_830_cast_fp16 = mul(x = var_828_cast_fp16, y = var_829_to_fp16)[name = tensor("op_830_cast_fp16")]; + tensor write_idx_5_cast_fp16 = sub(x = position2_to_fp16, y = var_830_cast_fp16)[name = tensor("write_idx_5_cast_fp16")]; + tensor var_833_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_5_cast_fp16)[name = tensor("op_833_cast_fp16")]; + tensor var_843 = const()[name = tensor("op_843"), val = tensor([1, 512, 1, 1])]; + tensor onehot_9_to_fp16_dtype_0 = const()[name = tensor("onehot_9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_833_cast_fp16_to_fp16 = cast(dtype = onehot_9_to_fp16_dtype_0, x = var_833_cast_fp16)[name = tensor("cast_560")]; + tensor onehot_11_cast_fp16 = reshape(shape = var_843, x = var_833_cast_fp16_to_fp16)[name = tensor("onehot_11_cast_fp16")]; + tensor var_845_to_fp16 = const()[name = tensor("op_845_to_fp16"), val = tensor(0x1p+0)]; + tensor var_847_cast_fp16 = sub(x = var_845_to_fp16, y = onehot_11_cast_fp16)[name = tensor("op_847_cast_fp16")]; + tensor k_cache2_to_fp16_dtype_0 = const()[name = tensor("k_cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache2_to_fp16 = cast(dtype = k_cache2_to_fp16_dtype_0, x = k_cache2)[name = tensor("cast_559")]; + tensor var_848_cast_fp16 = mul(x = k_cache2_to_fp16, y = var_847_cast_fp16)[name = tensor("op_848_cast_fp16")]; + tensor var_853 = const()[name = tensor("op_853"), val = tensor([1, 1, 16, 64])]; + tensor var_854_cast_fp16 = reshape(shape = var_853, x = out_11_cast_fp16)[name = tensor("op_854_cast_fp16")]; + tensor var_855_cast_fp16 = mul(x = var_854_cast_fp16, y = onehot_11_cast_fp16)[name = tensor("op_855_cast_fp16")]; + tensor new_k_cache_5_cast_fp16 = add(x = var_848_cast_fp16, y = var_855_cast_fp16)[name = tensor("new_k_cache_5_cast_fp16")]; + tensor new_k_cache_5_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_5_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache2_to_fp16_dtype_0 = const()[name = tensor("v_cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache2_to_fp16 = cast(dtype = v_cache2_to_fp16_dtype_0, x = v_cache2)[name = tensor("cast_557")]; + tensor var_861_cast_fp16 = mul(x = v_cache2_to_fp16, y = var_847_cast_fp16)[name = tensor("op_861_cast_fp16")]; + tensor var_866 = const()[name = tensor("op_866"), val = tensor([1, 1, 16, 64])]; + tensor var_867_cast_fp16 = reshape(shape = var_866, x = v_5_cast_fp16)[name = tensor("op_867_cast_fp16")]; + tensor var_868_cast_fp16 = mul(x = var_867_cast_fp16, y = onehot_11_cast_fp16)[name = tensor("op_868_cast_fp16")]; + tensor new_v_cache_5_cast_fp16 = add(x = var_861_cast_fp16, y = var_868_cast_fp16)[name = tensor("new_v_cache_5_cast_fp16")]; + tensor new_v_cache_5_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_5_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_2_perm_0 = const()[name = tensor("transpose_2_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_881 = const()[name = tensor("op_881"), val = tensor([0, 2, 1, 3])]; + tensor var_887 = const()[name = tensor("op_887"), val = tensor([1, 16, 1, 64])]; + tensor q4_5_cast_fp16 = reshape(shape = var_887, x = out_9_cast_fp16)[name = tensor("q4_5_cast_fp16")]; + tensor var_892_transpose_x_0 = const()[name = tensor("op_892_transpose_x_0"), val = tensor(false)]; + tensor var_892_transpose_y_0 = const()[name = tensor("op_892_transpose_y_0"), val = tensor(false)]; + tensor transpose_2_cast_fp16 = transpose(perm = transpose_2_perm_0, x = new_k_cache_5_cast_fp16)[name = tensor("transpose_89")]; + tensor var_892_cast_fp16 = matmul(transpose_x = var_892_transpose_x_0, transpose_y = var_892_transpose_y_0, x = q4_5_cast_fp16, y = transpose_2_cast_fp16)[name = tensor("op_892_cast_fp16")]; + tensor var_893_to_fp16 = const()[name = tensor("op_893_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_9_cast_fp16 = mul(x = var_892_cast_fp16, y = var_893_to_fp16)[name = tensor("scores_9_cast_fp16")]; + tensor var_897 = const()[name = tensor("op_897"), val = tensor([1, 1])]; + tensor var_898_cast_fp16 = reshape(shape = var_897, x = position2_to_fp16)[name = tensor("op_898_cast_fp16")]; + tensor var_899_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_898_cast_fp16)[name = tensor("op_899_cast_fp16")]; + tensor mask_5_to_fp16_dtype_0 = const()[name = tensor("mask_5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_906_to_fp16 = const()[name = tensor("op_906_to_fp16"), val = tensor(0x1p+0)]; + tensor var_899_cast_fp16_to_fp16 = cast(dtype = mask_5_to_fp16_dtype_0, x = var_899_cast_fp16)[name = tensor("cast_555")]; + tensor var_907_cast_fp16 = sub(x = var_899_cast_fp16_to_fp16, y = var_906_to_fp16)[name = tensor("op_907_cast_fp16")]; + tensor var_912 = const()[name = tensor("op_912"), val = tensor([1, 1, 1, 512])]; + tensor var_913_cast_fp16 = reshape(shape = var_912, x = var_907_cast_fp16)[name = tensor("op_913_cast_fp16")]; + tensor var_914_to_fp16 = const()[name = tensor("op_914_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_915_cast_fp16 = mul(x = var_913_cast_fp16, y = var_914_to_fp16)[name = tensor("op_915_cast_fp16")]; + tensor scores_11_cast_fp16 = add(x = scores_9_cast_fp16, y = var_915_cast_fp16)[name = tensor("scores_11_cast_fp16")]; + tensor var_918 = const()[name = tensor("op_918"), val = tensor(-1)]; + tensor weights_5_cast_fp16 = softmax(axis = var_918, x = scores_11_cast_fp16)[name = tensor("weights_5_cast_fp16")]; + tensor attn_5_transpose_x_0 = const()[name = tensor("attn_5_transpose_x_0"), val = tensor(false)]; + tensor attn_5_transpose_y_0 = const()[name = tensor("attn_5_transpose_y_0"), val = tensor(false)]; + tensor values_5_cast_fp16 = transpose(perm = var_881, x = new_v_cache_5_cast_fp16)[name = tensor("transpose_88")]; + tensor attn_5_cast_fp16 = matmul(transpose_x = attn_5_transpose_x_0, transpose_y = attn_5_transpose_y_0, x = weights_5_cast_fp16, y = values_5_cast_fp16)[name = tensor("attn_5_cast_fp16")]; + tensor var_926 = const()[name = tensor("op_926"), val = tensor([0, 2, 1, 3])]; + tensor var_931 = const()[name = tensor("op_931"), val = tensor([1, 1, 1024])]; + tensor var_927_cast_fp16 = transpose(perm = var_926, x = attn_5_cast_fp16)[name = tensor("transpose_87")]; + tensor input_31_cast_fp16 = reshape(shape = var_931, x = var_927_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor attn2_out_proj_weight_to_fp16 = const()[name = tensor("attn2_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(56728256)))]; + tensor linear_10_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn2_out_proj_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor var_937_to_fp16 = const()[name = tensor("op_937_to_fp16"), val = tensor(0x1p+0)]; + tensor var_938_cast_fp16 = add(x = position2_to_fp16, y = var_937_to_fp16)[name = tensor("op_938_cast_fp16")]; + tensor var_938_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_938_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_33_cast_fp16 = add(x = input_27_cast_fp16, y = linear_10_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor input_35_axes_0 = const()[name = tensor("input_35_axes_0"), val = tensor([-1])]; + tensor norm2_2_weight_to_fp16 = const()[name = tensor("norm2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58825472)))]; + tensor norm2_2_bias_to_fp16 = const()[name = tensor("norm2_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58827584)))]; + tensor var_942_to_fp16 = const()[name = tensor("op_942_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_35_cast_fp16 = layer_norm(axes = input_35_axes_0, beta = norm2_2_bias_to_fp16, epsilon = var_942_to_fp16, gamma = norm2_2_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor linear2_1_weight_to_fp16 = const()[name = tensor("linear2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58829696)))]; + tensor linear_11_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear2_1_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor input_37_mode_0 = const()[name = tensor("input_37_mode_0"), val = tensor("EXACT")]; + tensor input_37_cast_fp16 = gelu(mode = input_37_mode_0, x = linear_11_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor linear2_2_weight_to_fp16 = const()[name = tensor("linear2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67218368)))]; + tensor linear_12_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear2_2_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor input_39_cast_fp16 = add(x = input_33_cast_fp16, y = linear_12_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor input_41_axes_0 = const()[name = tensor("input_41_axes_0"), val = tensor([-1])]; + tensor norm3_1_weight_to_fp16 = const()[name = tensor("norm3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75607040)))]; + tensor norm3_1_bias_to_fp16 = const()[name = tensor("norm3_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75609152)))]; + tensor var_959_to_fp16 = const()[name = tensor("op_959_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_41_cast_fp16 = layer_norm(axes = input_41_axes_0, beta = norm3_1_bias_to_fp16, epsilon = var_959_to_fp16, gamma = norm3_1_weight_to_fp16, x = input_39_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor attn3_in_proj_weight_to_fp16 = const()[name = tensor("attn3_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75611264)))]; + tensor linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn3_in_proj_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor var_972 = const()[name = tensor("op_972"), val = tensor([1, 3, 16, 64])]; + tensor qkv_7_cast_fp16 = reshape(shape = var_972, x = linear_13_cast_fp16)[name = tensor("qkv_7_cast_fp16")]; + tensor x_13_begin_0 = const()[name = tensor("x_13_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_13_end_0 = const()[name = tensor("x_13_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_13_end_mask_0 = const()[name = tensor("x_13_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_13_squeeze_mask_0 = const()[name = tensor("x_13_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_13_cast_fp16 = slice_by_index(begin = x_13_begin_0, end = x_13_end_0, end_mask = x_13_end_mask_0, squeeze_mask = x_13_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("x_13_cast_fp16")]; + tensor x_15_begin_0 = const()[name = tensor("x_15_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_15_end_0 = const()[name = tensor("x_15_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_15_end_mask_0 = const()[name = tensor("x_15_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_15_squeeze_mask_0 = const()[name = tensor("x_15_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_15_cast_fp16 = slice_by_index(begin = x_15_begin_0, end = x_15_end_0, end_mask = x_15_end_mask_0, squeeze_mask = x_15_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("x_15_cast_fp16")]; + tensor v_7_begin_0 = const()[name = tensor("v_7_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_7_end_0 = const()[name = tensor("v_7_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_7_end_mask_0 = const()[name = tensor("v_7_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_7_squeeze_mask_0 = const()[name = tensor("v_7_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_7_cast_fp16 = slice_by_index(begin = v_7_begin_0, end = v_7_end_0, end_mask = v_7_end_mask_0, squeeze_mask = v_7_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("v_7_cast_fp16")]; + tensor var_1001 = const()[name = tensor("op_1001"), val = tensor([1, 1, 1])]; + tensor position3_to_fp16_dtype_0 = const()[name = tensor("position3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position3_to_fp16 = cast(dtype = position3_to_fp16_dtype_0, x = position3)[name = tensor("cast_553")]; + tensor var_1002_cast_fp16 = reshape(shape = var_1001, x = position3_to_fp16)[name = tensor("op_1002_cast_fp16")]; + tensor angles_13_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_1002_cast_fp16)[name = tensor("angles_13_cast_fp16")]; + tensor rotr_13_cast_fp16 = cos(x = angles_13_cast_fp16)[name = tensor("rotr_13_cast_fp16")]; + tensor roti_13_cast_fp16 = sin(x = angles_13_cast_fp16)[name = tensor("roti_13_cast_fp16")]; + tensor var_1010 = const()[name = tensor("op_1010"), val = tensor([1, 16, 32, 2])]; + tensor pairs_13_cast_fp16 = reshape(shape = var_1010, x = x_13_cast_fp16)[name = tensor("pairs_13_cast_fp16")]; + tensor xr_13_begin_0 = const()[name = tensor("xr_13_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_13_end_0 = const()[name = tensor("xr_13_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_13_end_mask_0 = const()[name = tensor("xr_13_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_13_squeeze_mask_0 = const()[name = tensor("xr_13_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_13_cast_fp16 = slice_by_index(begin = xr_13_begin_0, end = xr_13_end_0, end_mask = xr_13_end_mask_0, squeeze_mask = xr_13_squeeze_mask_0, x = pairs_13_cast_fp16)[name = tensor("xr_13_cast_fp16")]; + tensor xi_13_begin_0 = const()[name = tensor("xi_13_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_13_end_0 = const()[name = tensor("xi_13_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_13_end_mask_0 = const()[name = tensor("xi_13_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_13_squeeze_mask_0 = const()[name = tensor("xi_13_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_13_cast_fp16 = slice_by_index(begin = xi_13_begin_0, end = xi_13_end_0, end_mask = xi_13_end_mask_0, squeeze_mask = xi_13_squeeze_mask_0, x = pairs_13_cast_fp16)[name = tensor("xi_13_cast_fp16")]; + tensor var_1018_cast_fp16 = mul(x = xr_13_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_1018_cast_fp16")]; + tensor var_1019_cast_fp16 = mul(x = xi_13_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_1019_cast_fp16")]; + tensor xor__13_cast_fp16 = sub(x = var_1018_cast_fp16, y = var_1019_cast_fp16)[name = tensor("xor__13_cast_fp16")]; + tensor var_1022_cast_fp16 = mul(x = xr_13_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_1022_cast_fp16")]; + tensor var_1023_cast_fp16 = mul(x = xi_13_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_1023_cast_fp16")]; + tensor xoi_13_cast_fp16 = add(x = var_1022_cast_fp16, y = var_1023_cast_fp16)[name = tensor("xoi_13_cast_fp16")]; + tensor out_13_axis_0 = const()[name = tensor("out_13_axis_0"), val = tensor(-1)]; + tensor out_13_cast_fp16 = stack(axis = out_13_axis_0, values = (xor__13_cast_fp16, xoi_13_cast_fp16))[name = tensor("out_13_cast_fp16")]; + tensor var_1046 = const()[name = tensor("op_1046"), val = tensor([1, 16, 32, 2])]; + tensor pairs_15_cast_fp16 = reshape(shape = var_1046, x = x_15_cast_fp16)[name = tensor("pairs_15_cast_fp16")]; + tensor xr_15_begin_0 = const()[name = tensor("xr_15_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_15_end_0 = const()[name = tensor("xr_15_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_15_end_mask_0 = const()[name = tensor("xr_15_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_15_squeeze_mask_0 = const()[name = tensor("xr_15_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_15_cast_fp16 = slice_by_index(begin = xr_15_begin_0, end = xr_15_end_0, end_mask = xr_15_end_mask_0, squeeze_mask = xr_15_squeeze_mask_0, x = pairs_15_cast_fp16)[name = tensor("xr_15_cast_fp16")]; + tensor xi_15_begin_0 = const()[name = tensor("xi_15_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_15_end_0 = const()[name = tensor("xi_15_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_15_end_mask_0 = const()[name = tensor("xi_15_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_15_squeeze_mask_0 = const()[name = tensor("xi_15_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_15_cast_fp16 = slice_by_index(begin = xi_15_begin_0, end = xi_15_end_0, end_mask = xi_15_end_mask_0, squeeze_mask = xi_15_squeeze_mask_0, x = pairs_15_cast_fp16)[name = tensor("xi_15_cast_fp16")]; + tensor var_1054_cast_fp16 = mul(x = xr_15_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_1054_cast_fp16")]; + tensor var_1055_cast_fp16 = mul(x = xi_15_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_1055_cast_fp16")]; + tensor xor__15_cast_fp16 = sub(x = var_1054_cast_fp16, y = var_1055_cast_fp16)[name = tensor("xor__15_cast_fp16")]; + tensor var_1058_cast_fp16 = mul(x = xr_15_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_1058_cast_fp16")]; + tensor var_1059_cast_fp16 = mul(x = xi_15_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_1059_cast_fp16")]; + tensor xoi_15_cast_fp16 = add(x = var_1058_cast_fp16, y = var_1059_cast_fp16)[name = tensor("xoi_15_cast_fp16")]; + tensor out_15_axis_0 = const()[name = tensor("out_15_axis_0"), val = tensor(-1)]; + tensor out_15_cast_fp16 = stack(axis = out_15_axis_0, values = (xor__15_cast_fp16, xoi_15_cast_fp16))[name = tensor("out_15_cast_fp16")]; + tensor _inversed_1071_y_0_to_fp16 = const()[name = tensor("_inversed_1071_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1071_cast_fp16 = mul(x = position3_to_fp16, y = _inversed_1071_y_0_to_fp16)[name = tensor("_inversed_1071_cast_fp16")]; + tensor var_1072_cast_fp16 = floor(x = _inversed_1071_cast_fp16)[name = tensor("op_1072_cast_fp16")]; + tensor var_1073_to_fp16 = const()[name = tensor("op_1073_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1074_cast_fp16 = mul(x = var_1072_cast_fp16, y = var_1073_to_fp16)[name = tensor("op_1074_cast_fp16")]; + tensor write_idx_7_cast_fp16 = sub(x = position3_to_fp16, y = var_1074_cast_fp16)[name = tensor("write_idx_7_cast_fp16")]; + tensor var_1077_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_7_cast_fp16)[name = tensor("op_1077_cast_fp16")]; + tensor var_1087 = const()[name = tensor("op_1087"), val = tensor([1, 512, 1, 1])]; + tensor onehot_13_to_fp16_dtype_0 = const()[name = tensor("onehot_13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_1077_cast_fp16_to_fp16 = cast(dtype = onehot_13_to_fp16_dtype_0, x = var_1077_cast_fp16)[name = tensor("cast_552")]; + tensor onehot_15_cast_fp16 = reshape(shape = var_1087, x = var_1077_cast_fp16_to_fp16)[name = tensor("onehot_15_cast_fp16")]; + tensor var_1089_to_fp16 = const()[name = tensor("op_1089_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1091_cast_fp16 = sub(x = var_1089_to_fp16, y = onehot_15_cast_fp16)[name = tensor("op_1091_cast_fp16")]; + tensor k_cache3_to_fp16_dtype_0 = const()[name = tensor("k_cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache3_to_fp16 = cast(dtype = k_cache3_to_fp16_dtype_0, x = k_cache3)[name = tensor("cast_551")]; + tensor var_1092_cast_fp16 = mul(x = k_cache3_to_fp16, y = var_1091_cast_fp16)[name = tensor("op_1092_cast_fp16")]; + tensor var_1097 = const()[name = tensor("op_1097"), val = tensor([1, 1, 16, 64])]; + tensor var_1098_cast_fp16 = reshape(shape = var_1097, x = out_15_cast_fp16)[name = tensor("op_1098_cast_fp16")]; + tensor var_1099_cast_fp16 = mul(x = var_1098_cast_fp16, y = onehot_15_cast_fp16)[name = tensor("op_1099_cast_fp16")]; + tensor new_k_cache_7_cast_fp16 = add(x = var_1092_cast_fp16, y = var_1099_cast_fp16)[name = tensor("new_k_cache_7_cast_fp16")]; + tensor new_k_cache_7_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_7_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache3_to_fp16_dtype_0 = const()[name = tensor("v_cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache3_to_fp16 = cast(dtype = v_cache3_to_fp16_dtype_0, x = v_cache3)[name = tensor("cast_549")]; + tensor var_1105_cast_fp16 = mul(x = v_cache3_to_fp16, y = var_1091_cast_fp16)[name = tensor("op_1105_cast_fp16")]; + tensor var_1110 = const()[name = tensor("op_1110"), val = tensor([1, 1, 16, 64])]; + tensor var_1111_cast_fp16 = reshape(shape = var_1110, x = v_7_cast_fp16)[name = tensor("op_1111_cast_fp16")]; + tensor var_1112_cast_fp16 = mul(x = var_1111_cast_fp16, y = onehot_15_cast_fp16)[name = tensor("op_1112_cast_fp16")]; + tensor new_v_cache_7_cast_fp16 = add(x = var_1105_cast_fp16, y = var_1112_cast_fp16)[name = tensor("new_v_cache_7_cast_fp16")]; + tensor new_v_cache_7_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_7_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_3_perm_0 = const()[name = tensor("transpose_3_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_1125 = const()[name = tensor("op_1125"), val = tensor([0, 2, 1, 3])]; + tensor var_1131 = const()[name = tensor("op_1131"), val = tensor([1, 16, 1, 64])]; + tensor q4_7_cast_fp16 = reshape(shape = var_1131, x = out_13_cast_fp16)[name = tensor("q4_7_cast_fp16")]; + tensor var_1136_transpose_x_0 = const()[name = tensor("op_1136_transpose_x_0"), val = tensor(false)]; + tensor var_1136_transpose_y_0 = const()[name = tensor("op_1136_transpose_y_0"), val = tensor(false)]; + tensor transpose_3_cast_fp16 = transpose(perm = transpose_3_perm_0, x = new_k_cache_7_cast_fp16)[name = tensor("transpose_86")]; + tensor var_1136_cast_fp16 = matmul(transpose_x = var_1136_transpose_x_0, transpose_y = var_1136_transpose_y_0, x = q4_7_cast_fp16, y = transpose_3_cast_fp16)[name = tensor("op_1136_cast_fp16")]; + tensor var_1137_to_fp16 = const()[name = tensor("op_1137_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_13_cast_fp16 = mul(x = var_1136_cast_fp16, y = var_1137_to_fp16)[name = tensor("scores_13_cast_fp16")]; + tensor var_1141 = const()[name = tensor("op_1141"), val = tensor([1, 1])]; + tensor var_1142_cast_fp16 = reshape(shape = var_1141, x = position3_to_fp16)[name = tensor("op_1142_cast_fp16")]; + tensor var_1143_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_1142_cast_fp16)[name = tensor("op_1143_cast_fp16")]; + tensor mask_7_to_fp16_dtype_0 = const()[name = tensor("mask_7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_1150_to_fp16 = const()[name = tensor("op_1150_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1143_cast_fp16_to_fp16 = cast(dtype = mask_7_to_fp16_dtype_0, x = var_1143_cast_fp16)[name = tensor("cast_547")]; + tensor var_1151_cast_fp16 = sub(x = var_1143_cast_fp16_to_fp16, y = var_1150_to_fp16)[name = tensor("op_1151_cast_fp16")]; + tensor var_1156 = const()[name = tensor("op_1156"), val = tensor([1, 1, 1, 512])]; + tensor var_1157_cast_fp16 = reshape(shape = var_1156, x = var_1151_cast_fp16)[name = tensor("op_1157_cast_fp16")]; + tensor var_1158_to_fp16 = const()[name = tensor("op_1158_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_1159_cast_fp16 = mul(x = var_1157_cast_fp16, y = var_1158_to_fp16)[name = tensor("op_1159_cast_fp16")]; + tensor scores_15_cast_fp16 = add(x = scores_13_cast_fp16, y = var_1159_cast_fp16)[name = tensor("scores_15_cast_fp16")]; + tensor var_1162 = const()[name = tensor("op_1162"), val = tensor(-1)]; + tensor weights_7_cast_fp16 = softmax(axis = var_1162, x = scores_15_cast_fp16)[name = tensor("weights_7_cast_fp16")]; + tensor attn_7_transpose_x_0 = const()[name = tensor("attn_7_transpose_x_0"), val = tensor(false)]; + tensor attn_7_transpose_y_0 = const()[name = tensor("attn_7_transpose_y_0"), val = tensor(false)]; + tensor values_7_cast_fp16 = transpose(perm = var_1125, x = new_v_cache_7_cast_fp16)[name = tensor("transpose_85")]; + tensor attn_7_cast_fp16 = matmul(transpose_x = attn_7_transpose_x_0, transpose_y = attn_7_transpose_y_0, x = weights_7_cast_fp16, y = values_7_cast_fp16)[name = tensor("attn_7_cast_fp16")]; + tensor var_1170 = const()[name = tensor("op_1170"), val = tensor([0, 2, 1, 3])]; + tensor var_1175 = const()[name = tensor("op_1175"), val = tensor([1, 1, 1024])]; + tensor var_1171_cast_fp16 = transpose(perm = var_1170, x = attn_7_cast_fp16)[name = tensor("transpose_84")]; + tensor input_43_cast_fp16 = reshape(shape = var_1175, x = var_1171_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor attn3_out_proj_weight_to_fp16 = const()[name = tensor("attn3_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(81902784)))]; + tensor linear_14_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn3_out_proj_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor var_1181_to_fp16 = const()[name = tensor("op_1181_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1182_cast_fp16 = add(x = position3_to_fp16, y = var_1181_to_fp16)[name = tensor("op_1182_cast_fp16")]; + tensor var_1182_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1182_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_45_cast_fp16 = add(x = input_39_cast_fp16, y = linear_14_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor input_47_axes_0 = const()[name = tensor("input_47_axes_0"), val = tensor([-1])]; + tensor norm3_2_weight_to_fp16 = const()[name = tensor("norm3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84000000)))]; + tensor norm3_2_bias_to_fp16 = const()[name = tensor("norm3_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84002112)))]; + tensor var_1186_to_fp16 = const()[name = tensor("op_1186_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_47_cast_fp16 = layer_norm(axes = input_47_axes_0, beta = norm3_2_bias_to_fp16, epsilon = var_1186_to_fp16, gamma = norm3_2_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor linear3_1_weight_to_fp16 = const()[name = tensor("linear3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84004224)))]; + tensor linear_15_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear3_1_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor input_49_mode_0 = const()[name = tensor("input_49_mode_0"), val = tensor("EXACT")]; + tensor input_49_cast_fp16 = gelu(mode = input_49_mode_0, x = linear_15_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor linear3_2_weight_to_fp16 = const()[name = tensor("linear3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(92392896)))]; + tensor linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear3_2_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor input_51_cast_fp16 = add(x = input_45_cast_fp16, y = linear_16_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor input_53_axes_0 = const()[name = tensor("input_53_axes_0"), val = tensor([-1])]; + tensor norm4_1_weight_to_fp16 = const()[name = tensor("norm4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100781568)))]; + tensor norm4_1_bias_to_fp16 = const()[name = tensor("norm4_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100783680)))]; + tensor var_1203_to_fp16 = const()[name = tensor("op_1203_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_53_cast_fp16 = layer_norm(axes = input_53_axes_0, beta = norm4_1_bias_to_fp16, epsilon = var_1203_to_fp16, gamma = norm4_1_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor attn4_in_proj_weight_to_fp16 = const()[name = tensor("attn4_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100785792)))]; + tensor linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn4_in_proj_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor var_1216 = const()[name = tensor("op_1216"), val = tensor([1, 3, 16, 64])]; + tensor qkv_9_cast_fp16 = reshape(shape = var_1216, x = linear_17_cast_fp16)[name = tensor("qkv_9_cast_fp16")]; + tensor x_17_begin_0 = const()[name = tensor("x_17_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_17_end_0 = const()[name = tensor("x_17_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_17_end_mask_0 = const()[name = tensor("x_17_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_17_squeeze_mask_0 = const()[name = tensor("x_17_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_17_cast_fp16 = slice_by_index(begin = x_17_begin_0, end = x_17_end_0, end_mask = x_17_end_mask_0, squeeze_mask = x_17_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("x_17_cast_fp16")]; + tensor x_19_begin_0 = const()[name = tensor("x_19_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_19_end_0 = const()[name = tensor("x_19_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_19_end_mask_0 = const()[name = tensor("x_19_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_19_squeeze_mask_0 = const()[name = tensor("x_19_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_19_cast_fp16 = slice_by_index(begin = x_19_begin_0, end = x_19_end_0, end_mask = x_19_end_mask_0, squeeze_mask = x_19_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("x_19_cast_fp16")]; + tensor v_9_begin_0 = const()[name = tensor("v_9_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_9_end_0 = const()[name = tensor("v_9_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_9_end_mask_0 = const()[name = tensor("v_9_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_9_squeeze_mask_0 = const()[name = tensor("v_9_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_9_cast_fp16 = slice_by_index(begin = v_9_begin_0, end = v_9_end_0, end_mask = v_9_end_mask_0, squeeze_mask = v_9_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("v_9_cast_fp16")]; + tensor var_1245 = const()[name = tensor("op_1245"), val = tensor([1, 1, 1])]; + tensor position4_to_fp16_dtype_0 = const()[name = tensor("position4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position4_to_fp16 = cast(dtype = position4_to_fp16_dtype_0, x = position4)[name = tensor("cast_545")]; + tensor var_1246_cast_fp16 = reshape(shape = var_1245, x = position4_to_fp16)[name = tensor("op_1246_cast_fp16")]; + tensor angles_17_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_1246_cast_fp16)[name = tensor("angles_17_cast_fp16")]; + tensor rotr_17_cast_fp16 = cos(x = angles_17_cast_fp16)[name = tensor("rotr_17_cast_fp16")]; + tensor roti_17_cast_fp16 = sin(x = angles_17_cast_fp16)[name = tensor("roti_17_cast_fp16")]; + tensor var_1254 = const()[name = tensor("op_1254"), val = tensor([1, 16, 32, 2])]; + tensor pairs_17_cast_fp16 = reshape(shape = var_1254, x = x_17_cast_fp16)[name = tensor("pairs_17_cast_fp16")]; + tensor xr_17_begin_0 = const()[name = tensor("xr_17_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_17_end_0 = const()[name = tensor("xr_17_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_17_end_mask_0 = const()[name = tensor("xr_17_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_17_squeeze_mask_0 = const()[name = tensor("xr_17_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_17_cast_fp16 = slice_by_index(begin = xr_17_begin_0, end = xr_17_end_0, end_mask = xr_17_end_mask_0, squeeze_mask = xr_17_squeeze_mask_0, x = pairs_17_cast_fp16)[name = tensor("xr_17_cast_fp16")]; + tensor xi_17_begin_0 = const()[name = tensor("xi_17_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_17_end_0 = const()[name = tensor("xi_17_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_17_end_mask_0 = const()[name = tensor("xi_17_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_17_squeeze_mask_0 = const()[name = tensor("xi_17_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_17_cast_fp16 = slice_by_index(begin = xi_17_begin_0, end = xi_17_end_0, end_mask = xi_17_end_mask_0, squeeze_mask = xi_17_squeeze_mask_0, x = pairs_17_cast_fp16)[name = tensor("xi_17_cast_fp16")]; + tensor var_1262_cast_fp16 = mul(x = xr_17_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_1262_cast_fp16")]; + tensor var_1263_cast_fp16 = mul(x = xi_17_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_1263_cast_fp16")]; + tensor xor__17_cast_fp16 = sub(x = var_1262_cast_fp16, y = var_1263_cast_fp16)[name = tensor("xor__17_cast_fp16")]; + tensor var_1266_cast_fp16 = mul(x = xr_17_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_1266_cast_fp16")]; + tensor var_1267_cast_fp16 = mul(x = xi_17_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_1267_cast_fp16")]; + tensor xoi_17_cast_fp16 = add(x = var_1266_cast_fp16, y = var_1267_cast_fp16)[name = tensor("xoi_17_cast_fp16")]; + tensor out_17_axis_0 = const()[name = tensor("out_17_axis_0"), val = tensor(-1)]; + tensor out_17_cast_fp16 = stack(axis = out_17_axis_0, values = (xor__17_cast_fp16, xoi_17_cast_fp16))[name = tensor("out_17_cast_fp16")]; + tensor var_1290 = const()[name = tensor("op_1290"), val = tensor([1, 16, 32, 2])]; + tensor pairs_19_cast_fp16 = reshape(shape = var_1290, x = x_19_cast_fp16)[name = tensor("pairs_19_cast_fp16")]; + tensor xr_19_begin_0 = const()[name = tensor("xr_19_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_19_end_0 = const()[name = tensor("xr_19_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_19_end_mask_0 = const()[name = tensor("xr_19_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_19_squeeze_mask_0 = const()[name = tensor("xr_19_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_19_cast_fp16 = slice_by_index(begin = xr_19_begin_0, end = xr_19_end_0, end_mask = xr_19_end_mask_0, squeeze_mask = xr_19_squeeze_mask_0, x = pairs_19_cast_fp16)[name = tensor("xr_19_cast_fp16")]; + tensor xi_19_begin_0 = const()[name = tensor("xi_19_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_19_end_0 = const()[name = tensor("xi_19_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_19_end_mask_0 = const()[name = tensor("xi_19_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_19_squeeze_mask_0 = const()[name = tensor("xi_19_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_19_cast_fp16 = slice_by_index(begin = xi_19_begin_0, end = xi_19_end_0, end_mask = xi_19_end_mask_0, squeeze_mask = xi_19_squeeze_mask_0, x = pairs_19_cast_fp16)[name = tensor("xi_19_cast_fp16")]; + tensor var_1298_cast_fp16 = mul(x = xr_19_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_1298_cast_fp16")]; + tensor var_1299_cast_fp16 = mul(x = xi_19_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_1299_cast_fp16")]; + tensor xor__19_cast_fp16 = sub(x = var_1298_cast_fp16, y = var_1299_cast_fp16)[name = tensor("xor__19_cast_fp16")]; + tensor var_1302_cast_fp16 = mul(x = xr_19_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_1302_cast_fp16")]; + tensor var_1303_cast_fp16 = mul(x = xi_19_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_1303_cast_fp16")]; + tensor xoi_19_cast_fp16 = add(x = var_1302_cast_fp16, y = var_1303_cast_fp16)[name = tensor("xoi_19_cast_fp16")]; + tensor out_19_axis_0 = const()[name = tensor("out_19_axis_0"), val = tensor(-1)]; + tensor out_19_cast_fp16 = stack(axis = out_19_axis_0, values = (xor__19_cast_fp16, xoi_19_cast_fp16))[name = tensor("out_19_cast_fp16")]; + tensor _inversed_1315_y_0_to_fp16 = const()[name = tensor("_inversed_1315_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1315_cast_fp16 = mul(x = position4_to_fp16, y = _inversed_1315_y_0_to_fp16)[name = tensor("_inversed_1315_cast_fp16")]; + tensor var_1316_cast_fp16 = floor(x = _inversed_1315_cast_fp16)[name = tensor("op_1316_cast_fp16")]; + tensor var_1317_to_fp16 = const()[name = tensor("op_1317_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1318_cast_fp16 = mul(x = var_1316_cast_fp16, y = var_1317_to_fp16)[name = tensor("op_1318_cast_fp16")]; + tensor write_idx_9_cast_fp16 = sub(x = position4_to_fp16, y = var_1318_cast_fp16)[name = tensor("write_idx_9_cast_fp16")]; + tensor var_1321_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_9_cast_fp16)[name = tensor("op_1321_cast_fp16")]; + tensor var_1331 = const()[name = tensor("op_1331"), val = tensor([1, 512, 1, 1])]; + tensor onehot_17_to_fp16_dtype_0 = const()[name = tensor("onehot_17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_1321_cast_fp16_to_fp16 = cast(dtype = onehot_17_to_fp16_dtype_0, x = var_1321_cast_fp16)[name = tensor("cast_544")]; + tensor onehot_19_cast_fp16 = reshape(shape = var_1331, x = var_1321_cast_fp16_to_fp16)[name = tensor("onehot_19_cast_fp16")]; + tensor var_1333_to_fp16 = const()[name = tensor("op_1333_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1335_cast_fp16 = sub(x = var_1333_to_fp16, y = onehot_19_cast_fp16)[name = tensor("op_1335_cast_fp16")]; + tensor k_cache4_to_fp16_dtype_0 = const()[name = tensor("k_cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache4_to_fp16 = cast(dtype = k_cache4_to_fp16_dtype_0, x = k_cache4)[name = tensor("cast_543")]; + tensor var_1336_cast_fp16 = mul(x = k_cache4_to_fp16, y = var_1335_cast_fp16)[name = tensor("op_1336_cast_fp16")]; + tensor var_1341 = const()[name = tensor("op_1341"), val = tensor([1, 1, 16, 64])]; + tensor var_1342_cast_fp16 = reshape(shape = var_1341, x = out_19_cast_fp16)[name = tensor("op_1342_cast_fp16")]; + tensor var_1343_cast_fp16 = mul(x = var_1342_cast_fp16, y = onehot_19_cast_fp16)[name = tensor("op_1343_cast_fp16")]; + tensor new_k_cache_9_cast_fp16 = add(x = var_1336_cast_fp16, y = var_1343_cast_fp16)[name = tensor("new_k_cache_9_cast_fp16")]; + tensor new_k_cache_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache4_to_fp16_dtype_0 = const()[name = tensor("v_cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache4_to_fp16 = cast(dtype = v_cache4_to_fp16_dtype_0, x = v_cache4)[name = tensor("cast_541")]; + tensor var_1349_cast_fp16 = mul(x = v_cache4_to_fp16, y = var_1335_cast_fp16)[name = tensor("op_1349_cast_fp16")]; + tensor var_1354 = const()[name = tensor("op_1354"), val = tensor([1, 1, 16, 64])]; + tensor var_1355_cast_fp16 = reshape(shape = var_1354, x = v_9_cast_fp16)[name = tensor("op_1355_cast_fp16")]; + tensor var_1356_cast_fp16 = mul(x = var_1355_cast_fp16, y = onehot_19_cast_fp16)[name = tensor("op_1356_cast_fp16")]; + tensor new_v_cache_9_cast_fp16 = add(x = var_1349_cast_fp16, y = var_1356_cast_fp16)[name = tensor("new_v_cache_9_cast_fp16")]; + tensor new_v_cache_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_4_perm_0 = const()[name = tensor("transpose_4_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_1369 = const()[name = tensor("op_1369"), val = tensor([0, 2, 1, 3])]; + tensor var_1375 = const()[name = tensor("op_1375"), val = tensor([1, 16, 1, 64])]; + tensor q4_9_cast_fp16 = reshape(shape = var_1375, x = out_17_cast_fp16)[name = tensor("q4_9_cast_fp16")]; + tensor var_1380_transpose_x_0 = const()[name = tensor("op_1380_transpose_x_0"), val = tensor(false)]; + tensor var_1380_transpose_y_0 = const()[name = tensor("op_1380_transpose_y_0"), val = tensor(false)]; + tensor transpose_4_cast_fp16 = transpose(perm = transpose_4_perm_0, x = new_k_cache_9_cast_fp16)[name = tensor("transpose_83")]; + tensor var_1380_cast_fp16 = matmul(transpose_x = var_1380_transpose_x_0, transpose_y = var_1380_transpose_y_0, x = q4_9_cast_fp16, y = transpose_4_cast_fp16)[name = tensor("op_1380_cast_fp16")]; + tensor var_1381_to_fp16 = const()[name = tensor("op_1381_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_17_cast_fp16 = mul(x = var_1380_cast_fp16, y = var_1381_to_fp16)[name = tensor("scores_17_cast_fp16")]; + tensor var_1385 = const()[name = tensor("op_1385"), val = tensor([1, 1])]; + tensor var_1386_cast_fp16 = reshape(shape = var_1385, x = position4_to_fp16)[name = tensor("op_1386_cast_fp16")]; + tensor var_1387_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_1386_cast_fp16)[name = tensor("op_1387_cast_fp16")]; + tensor mask_9_to_fp16_dtype_0 = const()[name = tensor("mask_9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_1394_to_fp16 = const()[name = tensor("op_1394_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1387_cast_fp16_to_fp16 = cast(dtype = mask_9_to_fp16_dtype_0, x = var_1387_cast_fp16)[name = tensor("cast_539")]; + tensor var_1395_cast_fp16 = sub(x = var_1387_cast_fp16_to_fp16, y = var_1394_to_fp16)[name = tensor("op_1395_cast_fp16")]; + tensor var_1400 = const()[name = tensor("op_1400"), val = tensor([1, 1, 1, 512])]; + tensor var_1401_cast_fp16 = reshape(shape = var_1400, x = var_1395_cast_fp16)[name = tensor("op_1401_cast_fp16")]; + tensor var_1402_to_fp16 = const()[name = tensor("op_1402_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_1403_cast_fp16 = mul(x = var_1401_cast_fp16, y = var_1402_to_fp16)[name = tensor("op_1403_cast_fp16")]; + tensor scores_19_cast_fp16 = add(x = scores_17_cast_fp16, y = var_1403_cast_fp16)[name = tensor("scores_19_cast_fp16")]; + tensor var_1406 = const()[name = tensor("op_1406"), val = tensor(-1)]; + tensor weights_9_cast_fp16 = softmax(axis = var_1406, x = scores_19_cast_fp16)[name = tensor("weights_9_cast_fp16")]; + tensor attn_9_transpose_x_0 = const()[name = tensor("attn_9_transpose_x_0"), val = tensor(false)]; + tensor attn_9_transpose_y_0 = const()[name = tensor("attn_9_transpose_y_0"), val = tensor(false)]; + tensor values_9_cast_fp16 = transpose(perm = var_1369, x = new_v_cache_9_cast_fp16)[name = tensor("transpose_82")]; + tensor attn_9_cast_fp16 = matmul(transpose_x = attn_9_transpose_x_0, transpose_y = attn_9_transpose_y_0, x = weights_9_cast_fp16, y = values_9_cast_fp16)[name = tensor("attn_9_cast_fp16")]; + tensor var_1414 = const()[name = tensor("op_1414"), val = tensor([0, 2, 1, 3])]; + tensor var_1419 = const()[name = tensor("op_1419"), val = tensor([1, 1, 1024])]; + tensor var_1415_cast_fp16 = transpose(perm = var_1414, x = attn_9_cast_fp16)[name = tensor("transpose_81")]; + tensor input_55_cast_fp16 = reshape(shape = var_1419, x = var_1415_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor attn4_out_proj_weight_to_fp16 = const()[name = tensor("attn4_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(107077312)))]; + tensor linear_18_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn4_out_proj_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor var_1425_to_fp16 = const()[name = tensor("op_1425_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1426_cast_fp16 = add(x = position4_to_fp16, y = var_1425_to_fp16)[name = tensor("op_1426_cast_fp16")]; + tensor var_1426_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1426_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_57_cast_fp16 = add(x = input_51_cast_fp16, y = linear_18_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor input_59_axes_0 = const()[name = tensor("input_59_axes_0"), val = tensor([-1])]; + tensor norm4_2_weight_to_fp16 = const()[name = tensor("norm4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109174528)))]; + tensor norm4_2_bias_to_fp16 = const()[name = tensor("norm4_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109176640)))]; + tensor var_1430_to_fp16 = const()[name = tensor("op_1430_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_59_cast_fp16 = layer_norm(axes = input_59_axes_0, beta = norm4_2_bias_to_fp16, epsilon = var_1430_to_fp16, gamma = norm4_2_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor linear4_1_weight_to_fp16 = const()[name = tensor("linear4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109178752)))]; + tensor linear_19_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear4_1_weight_to_fp16, x = input_59_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_61_mode_0 = const()[name = tensor("input_61_mode_0"), val = tensor("EXACT")]; + tensor input_61_cast_fp16 = gelu(mode = input_61_mode_0, x = linear_19_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor linear4_2_weight_to_fp16 = const()[name = tensor("linear4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(117567424)))]; + tensor linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear4_2_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor input_63_cast_fp16 = add(x = input_57_cast_fp16, y = linear_20_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor input_65_axes_0 = const()[name = tensor("input_65_axes_0"), val = tensor([-1])]; + tensor norm5_1_weight_to_fp16 = const()[name = tensor("norm5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125956096)))]; + tensor norm5_1_bias_to_fp16 = const()[name = tensor("norm5_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125958208)))]; + tensor var_1447_to_fp16 = const()[name = tensor("op_1447_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_65_cast_fp16 = layer_norm(axes = input_65_axes_0, beta = norm5_1_bias_to_fp16, epsilon = var_1447_to_fp16, gamma = norm5_1_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("input_65_cast_fp16")]; + tensor attn5_in_proj_weight_to_fp16 = const()[name = tensor("attn5_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125960320)))]; + tensor linear_21_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn5_in_proj_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("linear_21_cast_fp16")]; + tensor var_1460 = const()[name = tensor("op_1460"), val = tensor([1, 3, 16, 64])]; + tensor qkv_11_cast_fp16 = reshape(shape = var_1460, x = linear_21_cast_fp16)[name = tensor("qkv_11_cast_fp16")]; + tensor x_21_begin_0 = const()[name = tensor("x_21_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_21_end_0 = const()[name = tensor("x_21_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_21_end_mask_0 = const()[name = tensor("x_21_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_21_squeeze_mask_0 = const()[name = tensor("x_21_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_21_cast_fp16 = slice_by_index(begin = x_21_begin_0, end = x_21_end_0, end_mask = x_21_end_mask_0, squeeze_mask = x_21_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("x_21_cast_fp16")]; + tensor x_23_begin_0 = const()[name = tensor("x_23_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_23_end_0 = const()[name = tensor("x_23_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_23_end_mask_0 = const()[name = tensor("x_23_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_23_squeeze_mask_0 = const()[name = tensor("x_23_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_23_cast_fp16 = slice_by_index(begin = x_23_begin_0, end = x_23_end_0, end_mask = x_23_end_mask_0, squeeze_mask = x_23_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("x_23_cast_fp16")]; + tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_11_cast_fp16 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("v_11_cast_fp16")]; + tensor var_1489 = const()[name = tensor("op_1489"), val = tensor([1, 1, 1])]; + tensor position5_to_fp16_dtype_0 = const()[name = tensor("position5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position5_to_fp16 = cast(dtype = position5_to_fp16_dtype_0, x = position5)[name = tensor("cast_537")]; + tensor var_1490_cast_fp16 = reshape(shape = var_1489, x = position5_to_fp16)[name = tensor("op_1490_cast_fp16")]; + tensor angles_21_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_1490_cast_fp16)[name = tensor("angles_21_cast_fp16")]; + tensor rotr_21_cast_fp16 = cos(x = angles_21_cast_fp16)[name = tensor("rotr_21_cast_fp16")]; + tensor roti_21_cast_fp16 = sin(x = angles_21_cast_fp16)[name = tensor("roti_21_cast_fp16")]; + tensor var_1498 = const()[name = tensor("op_1498"), val = tensor([1, 16, 32, 2])]; + tensor pairs_21_cast_fp16 = reshape(shape = var_1498, x = x_21_cast_fp16)[name = tensor("pairs_21_cast_fp16")]; + tensor xr_21_begin_0 = const()[name = tensor("xr_21_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_21_end_0 = const()[name = tensor("xr_21_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_21_end_mask_0 = const()[name = tensor("xr_21_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_21_squeeze_mask_0 = const()[name = tensor("xr_21_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_21_cast_fp16 = slice_by_index(begin = xr_21_begin_0, end = xr_21_end_0, end_mask = xr_21_end_mask_0, squeeze_mask = xr_21_squeeze_mask_0, x = pairs_21_cast_fp16)[name = tensor("xr_21_cast_fp16")]; + tensor xi_21_begin_0 = const()[name = tensor("xi_21_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_21_end_0 = const()[name = tensor("xi_21_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_21_end_mask_0 = const()[name = tensor("xi_21_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_21_squeeze_mask_0 = const()[name = tensor("xi_21_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_21_cast_fp16 = slice_by_index(begin = xi_21_begin_0, end = xi_21_end_0, end_mask = xi_21_end_mask_0, squeeze_mask = xi_21_squeeze_mask_0, x = pairs_21_cast_fp16)[name = tensor("xi_21_cast_fp16")]; + tensor var_1506_cast_fp16 = mul(x = xr_21_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_1506_cast_fp16")]; + tensor var_1507_cast_fp16 = mul(x = xi_21_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_1507_cast_fp16")]; + tensor xor__21_cast_fp16 = sub(x = var_1506_cast_fp16, y = var_1507_cast_fp16)[name = tensor("xor__21_cast_fp16")]; + tensor var_1510_cast_fp16 = mul(x = xr_21_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_1510_cast_fp16")]; + tensor var_1511_cast_fp16 = mul(x = xi_21_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_1511_cast_fp16")]; + tensor xoi_21_cast_fp16 = add(x = var_1510_cast_fp16, y = var_1511_cast_fp16)[name = tensor("xoi_21_cast_fp16")]; + tensor out_21_axis_0 = const()[name = tensor("out_21_axis_0"), val = tensor(-1)]; + tensor out_21_cast_fp16 = stack(axis = out_21_axis_0, values = (xor__21_cast_fp16, xoi_21_cast_fp16))[name = tensor("out_21_cast_fp16")]; + tensor var_1534 = const()[name = tensor("op_1534"), val = tensor([1, 16, 32, 2])]; + tensor pairs_23_cast_fp16 = reshape(shape = var_1534, x = x_23_cast_fp16)[name = tensor("pairs_23_cast_fp16")]; + tensor xr_23_begin_0 = const()[name = tensor("xr_23_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_23_end_0 = const()[name = tensor("xr_23_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_23_end_mask_0 = const()[name = tensor("xr_23_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_23_squeeze_mask_0 = const()[name = tensor("xr_23_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_23_cast_fp16 = slice_by_index(begin = xr_23_begin_0, end = xr_23_end_0, end_mask = xr_23_end_mask_0, squeeze_mask = xr_23_squeeze_mask_0, x = pairs_23_cast_fp16)[name = tensor("xr_23_cast_fp16")]; + tensor xi_23_begin_0 = const()[name = tensor("xi_23_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_23_end_0 = const()[name = tensor("xi_23_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_23_end_mask_0 = const()[name = tensor("xi_23_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_23_squeeze_mask_0 = const()[name = tensor("xi_23_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_23_cast_fp16 = slice_by_index(begin = xi_23_begin_0, end = xi_23_end_0, end_mask = xi_23_end_mask_0, squeeze_mask = xi_23_squeeze_mask_0, x = pairs_23_cast_fp16)[name = tensor("xi_23_cast_fp16")]; + tensor var_1542_cast_fp16 = mul(x = xr_23_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_1542_cast_fp16")]; + tensor var_1543_cast_fp16 = mul(x = xi_23_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_1543_cast_fp16")]; + tensor xor__23_cast_fp16 = sub(x = var_1542_cast_fp16, y = var_1543_cast_fp16)[name = tensor("xor__23_cast_fp16")]; + tensor var_1546_cast_fp16 = mul(x = xr_23_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_1546_cast_fp16")]; + tensor var_1547_cast_fp16 = mul(x = xi_23_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_1547_cast_fp16")]; + tensor xoi_23_cast_fp16 = add(x = var_1546_cast_fp16, y = var_1547_cast_fp16)[name = tensor("xoi_23_cast_fp16")]; + tensor out_23_axis_0 = const()[name = tensor("out_23_axis_0"), val = tensor(-1)]; + tensor out_23_cast_fp16 = stack(axis = out_23_axis_0, values = (xor__23_cast_fp16, xoi_23_cast_fp16))[name = tensor("out_23_cast_fp16")]; + tensor _inversed_1559_y_0_to_fp16 = const()[name = tensor("_inversed_1559_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1559_cast_fp16 = mul(x = position5_to_fp16, y = _inversed_1559_y_0_to_fp16)[name = tensor("_inversed_1559_cast_fp16")]; + tensor var_1560_cast_fp16 = floor(x = _inversed_1559_cast_fp16)[name = tensor("op_1560_cast_fp16")]; + tensor var_1561_to_fp16 = const()[name = tensor("op_1561_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1562_cast_fp16 = mul(x = var_1560_cast_fp16, y = var_1561_to_fp16)[name = tensor("op_1562_cast_fp16")]; + tensor write_idx_11_cast_fp16 = sub(x = position5_to_fp16, y = var_1562_cast_fp16)[name = tensor("write_idx_11_cast_fp16")]; + tensor var_1565_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_11_cast_fp16)[name = tensor("op_1565_cast_fp16")]; + tensor var_1575 = const()[name = tensor("op_1575"), val = tensor([1, 512, 1, 1])]; + tensor onehot_21_to_fp16_dtype_0 = const()[name = tensor("onehot_21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_1565_cast_fp16_to_fp16 = cast(dtype = onehot_21_to_fp16_dtype_0, x = var_1565_cast_fp16)[name = tensor("cast_536")]; + tensor onehot_23_cast_fp16 = reshape(shape = var_1575, x = var_1565_cast_fp16_to_fp16)[name = tensor("onehot_23_cast_fp16")]; + tensor var_1577_to_fp16 = const()[name = tensor("op_1577_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1579_cast_fp16 = sub(x = var_1577_to_fp16, y = onehot_23_cast_fp16)[name = tensor("op_1579_cast_fp16")]; + tensor k_cache5_to_fp16_dtype_0 = const()[name = tensor("k_cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache5_to_fp16 = cast(dtype = k_cache5_to_fp16_dtype_0, x = k_cache5)[name = tensor("cast_535")]; + tensor var_1580_cast_fp16 = mul(x = k_cache5_to_fp16, y = var_1579_cast_fp16)[name = tensor("op_1580_cast_fp16")]; + tensor var_1585 = const()[name = tensor("op_1585"), val = tensor([1, 1, 16, 64])]; + tensor var_1586_cast_fp16 = reshape(shape = var_1585, x = out_23_cast_fp16)[name = tensor("op_1586_cast_fp16")]; + tensor var_1587_cast_fp16 = mul(x = var_1586_cast_fp16, y = onehot_23_cast_fp16)[name = tensor("op_1587_cast_fp16")]; + tensor new_k_cache_11_cast_fp16 = add(x = var_1580_cast_fp16, y = var_1587_cast_fp16)[name = tensor("new_k_cache_11_cast_fp16")]; + tensor new_k_cache_11_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_11_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache5_to_fp16_dtype_0 = const()[name = tensor("v_cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache5_to_fp16 = cast(dtype = v_cache5_to_fp16_dtype_0, x = v_cache5)[name = tensor("cast_533")]; + tensor var_1593_cast_fp16 = mul(x = v_cache5_to_fp16, y = var_1579_cast_fp16)[name = tensor("op_1593_cast_fp16")]; + tensor var_1598 = const()[name = tensor("op_1598"), val = tensor([1, 1, 16, 64])]; + tensor var_1599_cast_fp16 = reshape(shape = var_1598, x = v_11_cast_fp16)[name = tensor("op_1599_cast_fp16")]; + tensor var_1600_cast_fp16 = mul(x = var_1599_cast_fp16, y = onehot_23_cast_fp16)[name = tensor("op_1600_cast_fp16")]; + tensor new_v_cache_11_cast_fp16 = add(x = var_1593_cast_fp16, y = var_1600_cast_fp16)[name = tensor("new_v_cache_11_cast_fp16")]; + tensor new_v_cache_11_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_11_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_5_perm_0 = const()[name = tensor("transpose_5_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_1613 = const()[name = tensor("op_1613"), val = tensor([0, 2, 1, 3])]; + tensor var_1619 = const()[name = tensor("op_1619"), val = tensor([1, 16, 1, 64])]; + tensor q4_11_cast_fp16 = reshape(shape = var_1619, x = out_21_cast_fp16)[name = tensor("q4_11_cast_fp16")]; + tensor var_1624_transpose_x_0 = const()[name = tensor("op_1624_transpose_x_0"), val = tensor(false)]; + tensor var_1624_transpose_y_0 = const()[name = tensor("op_1624_transpose_y_0"), val = tensor(false)]; + tensor transpose_5_cast_fp16 = transpose(perm = transpose_5_perm_0, x = new_k_cache_11_cast_fp16)[name = tensor("transpose_80")]; + tensor var_1624_cast_fp16 = matmul(transpose_x = var_1624_transpose_x_0, transpose_y = var_1624_transpose_y_0, x = q4_11_cast_fp16, y = transpose_5_cast_fp16)[name = tensor("op_1624_cast_fp16")]; + tensor var_1625_to_fp16 = const()[name = tensor("op_1625_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_21_cast_fp16 = mul(x = var_1624_cast_fp16, y = var_1625_to_fp16)[name = tensor("scores_21_cast_fp16")]; + tensor var_1629 = const()[name = tensor("op_1629"), val = tensor([1, 1])]; + tensor var_1630_cast_fp16 = reshape(shape = var_1629, x = position5_to_fp16)[name = tensor("op_1630_cast_fp16")]; + tensor var_1631_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_1630_cast_fp16)[name = tensor("op_1631_cast_fp16")]; + tensor mask_11_to_fp16_dtype_0 = const()[name = tensor("mask_11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_1638_to_fp16 = const()[name = tensor("op_1638_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1631_cast_fp16_to_fp16 = cast(dtype = mask_11_to_fp16_dtype_0, x = var_1631_cast_fp16)[name = tensor("cast_531")]; + tensor var_1639_cast_fp16 = sub(x = var_1631_cast_fp16_to_fp16, y = var_1638_to_fp16)[name = tensor("op_1639_cast_fp16")]; + tensor var_1644 = const()[name = tensor("op_1644"), val = tensor([1, 1, 1, 512])]; + tensor var_1645_cast_fp16 = reshape(shape = var_1644, x = var_1639_cast_fp16)[name = tensor("op_1645_cast_fp16")]; + tensor var_1646_to_fp16 = const()[name = tensor("op_1646_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_1647_cast_fp16 = mul(x = var_1645_cast_fp16, y = var_1646_to_fp16)[name = tensor("op_1647_cast_fp16")]; + tensor scores_23_cast_fp16 = add(x = scores_21_cast_fp16, y = var_1647_cast_fp16)[name = tensor("scores_23_cast_fp16")]; + tensor var_1650 = const()[name = tensor("op_1650"), val = tensor(-1)]; + tensor weights_11_cast_fp16 = softmax(axis = var_1650, x = scores_23_cast_fp16)[name = tensor("weights_11_cast_fp16")]; + tensor attn_11_transpose_x_0 = const()[name = tensor("attn_11_transpose_x_0"), val = tensor(false)]; + tensor attn_11_transpose_y_0 = const()[name = tensor("attn_11_transpose_y_0"), val = tensor(false)]; + tensor values_11_cast_fp16 = transpose(perm = var_1613, x = new_v_cache_11_cast_fp16)[name = tensor("transpose_79")]; + tensor attn_11_cast_fp16 = matmul(transpose_x = attn_11_transpose_x_0, transpose_y = attn_11_transpose_y_0, x = weights_11_cast_fp16, y = values_11_cast_fp16)[name = tensor("attn_11_cast_fp16")]; + tensor var_1658 = const()[name = tensor("op_1658"), val = tensor([0, 2, 1, 3])]; + tensor var_1663 = const()[name = tensor("op_1663"), val = tensor([1, 1, 1024])]; + tensor var_1659_cast_fp16 = transpose(perm = var_1658, x = attn_11_cast_fp16)[name = tensor("transpose_78")]; + tensor input_67_cast_fp16 = reshape(shape = var_1663, x = var_1659_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor attn5_out_proj_weight_to_fp16 = const()[name = tensor("attn5_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132251840)))]; + tensor linear_22_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn5_out_proj_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("linear_22_cast_fp16")]; + tensor var_1669_to_fp16 = const()[name = tensor("op_1669_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1670_cast_fp16 = add(x = position5_to_fp16, y = var_1669_to_fp16)[name = tensor("op_1670_cast_fp16")]; + tensor var_1670_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1670_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_69_cast_fp16 = add(x = input_63_cast_fp16, y = linear_22_cast_fp16)[name = tensor("input_69_cast_fp16")]; + tensor input_71_axes_0 = const()[name = tensor("input_71_axes_0"), val = tensor([-1])]; + tensor norm5_2_weight_to_fp16 = const()[name = tensor("norm5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134349056)))]; + tensor norm5_2_bias_to_fp16 = const()[name = tensor("norm5_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134351168)))]; + tensor var_1674_to_fp16 = const()[name = tensor("op_1674_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_71_cast_fp16 = layer_norm(axes = input_71_axes_0, beta = norm5_2_bias_to_fp16, epsilon = var_1674_to_fp16, gamma = norm5_2_weight_to_fp16, x = input_69_cast_fp16)[name = tensor("input_71_cast_fp16")]; + tensor linear5_1_weight_to_fp16 = const()[name = tensor("linear5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134353280)))]; + tensor linear_23_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear5_1_weight_to_fp16, x = input_71_cast_fp16)[name = tensor("linear_23_cast_fp16")]; + tensor input_73_mode_0 = const()[name = tensor("input_73_mode_0"), val = tensor("EXACT")]; + tensor input_73_cast_fp16 = gelu(mode = input_73_mode_0, x = linear_23_cast_fp16)[name = tensor("input_73_cast_fp16")]; + tensor linear5_2_weight_to_fp16 = const()[name = tensor("linear5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142741952)))]; + tensor linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear5_2_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("linear_24_cast_fp16")]; + tensor input_75_cast_fp16 = add(x = input_69_cast_fp16, y = linear_24_cast_fp16)[name = tensor("input_75_cast_fp16")]; + tensor input_77_axes_0 = const()[name = tensor("input_77_axes_0"), val = tensor([-1])]; + tensor norm6_1_weight_to_fp16 = const()[name = tensor("norm6_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151130624)))]; + tensor norm6_1_bias_to_fp16 = const()[name = tensor("norm6_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151132736)))]; + tensor var_1691_to_fp16 = const()[name = tensor("op_1691_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_77_cast_fp16 = layer_norm(axes = input_77_axes_0, beta = norm6_1_bias_to_fp16, epsilon = var_1691_to_fp16, gamma = norm6_1_weight_to_fp16, x = input_75_cast_fp16)[name = tensor("input_77_cast_fp16")]; + tensor attn6_in_proj_weight_to_fp16 = const()[name = tensor("attn6_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151134848)))]; + tensor linear_25_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn6_in_proj_weight_to_fp16, x = input_77_cast_fp16)[name = tensor("linear_25_cast_fp16")]; + tensor var_1704 = const()[name = tensor("op_1704"), val = tensor([1, 3, 16, 64])]; + tensor qkv_13_cast_fp16 = reshape(shape = var_1704, x = linear_25_cast_fp16)[name = tensor("qkv_13_cast_fp16")]; + tensor x_25_begin_0 = const()[name = tensor("x_25_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_25_end_0 = const()[name = tensor("x_25_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_25_end_mask_0 = const()[name = tensor("x_25_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_25_squeeze_mask_0 = const()[name = tensor("x_25_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_25_cast_fp16 = slice_by_index(begin = x_25_begin_0, end = x_25_end_0, end_mask = x_25_end_mask_0, squeeze_mask = x_25_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("x_25_cast_fp16")]; + tensor x_27_begin_0 = const()[name = tensor("x_27_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_27_end_0 = const()[name = tensor("x_27_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_27_end_mask_0 = const()[name = tensor("x_27_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_27_squeeze_mask_0 = const()[name = tensor("x_27_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_27_cast_fp16 = slice_by_index(begin = x_27_begin_0, end = x_27_end_0, end_mask = x_27_end_mask_0, squeeze_mask = x_27_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("x_27_cast_fp16")]; + tensor v_13_begin_0 = const()[name = tensor("v_13_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_13_end_0 = const()[name = tensor("v_13_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_13_end_mask_0 = const()[name = tensor("v_13_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_13_squeeze_mask_0 = const()[name = tensor("v_13_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_13_cast_fp16 = slice_by_index(begin = v_13_begin_0, end = v_13_end_0, end_mask = v_13_end_mask_0, squeeze_mask = v_13_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("v_13_cast_fp16")]; + tensor var_1733 = const()[name = tensor("op_1733"), val = tensor([1, 1, 1])]; + tensor position6_to_fp16_dtype_0 = const()[name = tensor("position6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position6_to_fp16 = cast(dtype = position6_to_fp16_dtype_0, x = position6)[name = tensor("cast_529")]; + tensor var_1734_cast_fp16 = reshape(shape = var_1733, x = position6_to_fp16)[name = tensor("op_1734_cast_fp16")]; + tensor angles_25_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_1734_cast_fp16)[name = tensor("angles_25_cast_fp16")]; + tensor rotr_25_cast_fp16 = cos(x = angles_25_cast_fp16)[name = tensor("rotr_25_cast_fp16")]; + tensor roti_25_cast_fp16 = sin(x = angles_25_cast_fp16)[name = tensor("roti_25_cast_fp16")]; + tensor var_1742 = const()[name = tensor("op_1742"), val = tensor([1, 16, 32, 2])]; + tensor pairs_25_cast_fp16 = reshape(shape = var_1742, x = x_25_cast_fp16)[name = tensor("pairs_25_cast_fp16")]; + tensor xr_25_begin_0 = const()[name = tensor("xr_25_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_25_end_0 = const()[name = tensor("xr_25_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_25_end_mask_0 = const()[name = tensor("xr_25_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_25_squeeze_mask_0 = const()[name = tensor("xr_25_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_25_cast_fp16 = slice_by_index(begin = xr_25_begin_0, end = xr_25_end_0, end_mask = xr_25_end_mask_0, squeeze_mask = xr_25_squeeze_mask_0, x = pairs_25_cast_fp16)[name = tensor("xr_25_cast_fp16")]; + tensor xi_25_begin_0 = const()[name = tensor("xi_25_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_25_end_0 = const()[name = tensor("xi_25_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_25_end_mask_0 = const()[name = tensor("xi_25_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_25_squeeze_mask_0 = const()[name = tensor("xi_25_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_25_cast_fp16 = slice_by_index(begin = xi_25_begin_0, end = xi_25_end_0, end_mask = xi_25_end_mask_0, squeeze_mask = xi_25_squeeze_mask_0, x = pairs_25_cast_fp16)[name = tensor("xi_25_cast_fp16")]; + tensor var_1750_cast_fp16 = mul(x = xr_25_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_1750_cast_fp16")]; + tensor var_1751_cast_fp16 = mul(x = xi_25_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_1751_cast_fp16")]; + tensor xor__25_cast_fp16 = sub(x = var_1750_cast_fp16, y = var_1751_cast_fp16)[name = tensor("xor__25_cast_fp16")]; + tensor var_1754_cast_fp16 = mul(x = xr_25_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_1754_cast_fp16")]; + tensor var_1755_cast_fp16 = mul(x = xi_25_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_1755_cast_fp16")]; + tensor xoi_25_cast_fp16 = add(x = var_1754_cast_fp16, y = var_1755_cast_fp16)[name = tensor("xoi_25_cast_fp16")]; + tensor out_25_axis_0 = const()[name = tensor("out_25_axis_0"), val = tensor(-1)]; + tensor out_25_cast_fp16 = stack(axis = out_25_axis_0, values = (xor__25_cast_fp16, xoi_25_cast_fp16))[name = tensor("out_25_cast_fp16")]; + tensor var_1778 = const()[name = tensor("op_1778"), val = tensor([1, 16, 32, 2])]; + tensor pairs_27_cast_fp16 = reshape(shape = var_1778, x = x_27_cast_fp16)[name = tensor("pairs_27_cast_fp16")]; + tensor xr_27_begin_0 = const()[name = tensor("xr_27_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_27_end_0 = const()[name = tensor("xr_27_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_27_end_mask_0 = const()[name = tensor("xr_27_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_27_squeeze_mask_0 = const()[name = tensor("xr_27_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_27_cast_fp16 = slice_by_index(begin = xr_27_begin_0, end = xr_27_end_0, end_mask = xr_27_end_mask_0, squeeze_mask = xr_27_squeeze_mask_0, x = pairs_27_cast_fp16)[name = tensor("xr_27_cast_fp16")]; + tensor xi_27_begin_0 = const()[name = tensor("xi_27_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_27_end_0 = const()[name = tensor("xi_27_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_27_end_mask_0 = const()[name = tensor("xi_27_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_27_squeeze_mask_0 = const()[name = tensor("xi_27_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_27_cast_fp16 = slice_by_index(begin = xi_27_begin_0, end = xi_27_end_0, end_mask = xi_27_end_mask_0, squeeze_mask = xi_27_squeeze_mask_0, x = pairs_27_cast_fp16)[name = tensor("xi_27_cast_fp16")]; + tensor var_1786_cast_fp16 = mul(x = xr_27_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_1786_cast_fp16")]; + tensor var_1787_cast_fp16 = mul(x = xi_27_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_1787_cast_fp16")]; + tensor xor__27_cast_fp16 = sub(x = var_1786_cast_fp16, y = var_1787_cast_fp16)[name = tensor("xor__27_cast_fp16")]; + tensor var_1790_cast_fp16 = mul(x = xr_27_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_1790_cast_fp16")]; + tensor var_1791_cast_fp16 = mul(x = xi_27_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_1791_cast_fp16")]; + tensor xoi_27_cast_fp16 = add(x = var_1790_cast_fp16, y = var_1791_cast_fp16)[name = tensor("xoi_27_cast_fp16")]; + tensor out_27_axis_0 = const()[name = tensor("out_27_axis_0"), val = tensor(-1)]; + tensor out_27_cast_fp16 = stack(axis = out_27_axis_0, values = (xor__27_cast_fp16, xoi_27_cast_fp16))[name = tensor("out_27_cast_fp16")]; + tensor _inversed_1803_y_0_to_fp16 = const()[name = tensor("_inversed_1803_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1803_cast_fp16 = mul(x = position6_to_fp16, y = _inversed_1803_y_0_to_fp16)[name = tensor("_inversed_1803_cast_fp16")]; + tensor var_1804_cast_fp16 = floor(x = _inversed_1803_cast_fp16)[name = tensor("op_1804_cast_fp16")]; + tensor var_1805_to_fp16 = const()[name = tensor("op_1805_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1806_cast_fp16 = mul(x = var_1804_cast_fp16, y = var_1805_to_fp16)[name = tensor("op_1806_cast_fp16")]; + tensor write_idx_13_cast_fp16 = sub(x = position6_to_fp16, y = var_1806_cast_fp16)[name = tensor("write_idx_13_cast_fp16")]; + tensor var_1809_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_13_cast_fp16)[name = tensor("op_1809_cast_fp16")]; + tensor var_1819 = const()[name = tensor("op_1819"), val = tensor([1, 512, 1, 1])]; + tensor onehot_25_to_fp16_dtype_0 = const()[name = tensor("onehot_25_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_1809_cast_fp16_to_fp16 = cast(dtype = onehot_25_to_fp16_dtype_0, x = var_1809_cast_fp16)[name = tensor("cast_528")]; + tensor onehot_27_cast_fp16 = reshape(shape = var_1819, x = var_1809_cast_fp16_to_fp16)[name = tensor("onehot_27_cast_fp16")]; + tensor var_1821_to_fp16 = const()[name = tensor("op_1821_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1823_cast_fp16 = sub(x = var_1821_to_fp16, y = onehot_27_cast_fp16)[name = tensor("op_1823_cast_fp16")]; + tensor k_cache6_to_fp16_dtype_0 = const()[name = tensor("k_cache6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache6_to_fp16 = cast(dtype = k_cache6_to_fp16_dtype_0, x = k_cache6)[name = tensor("cast_527")]; + tensor var_1824_cast_fp16 = mul(x = k_cache6_to_fp16, y = var_1823_cast_fp16)[name = tensor("op_1824_cast_fp16")]; + tensor var_1829 = const()[name = tensor("op_1829"), val = tensor([1, 1, 16, 64])]; + tensor var_1830_cast_fp16 = reshape(shape = var_1829, x = out_27_cast_fp16)[name = tensor("op_1830_cast_fp16")]; + tensor var_1831_cast_fp16 = mul(x = var_1830_cast_fp16, y = onehot_27_cast_fp16)[name = tensor("op_1831_cast_fp16")]; + tensor new_k_cache_13_cast_fp16 = add(x = var_1824_cast_fp16, y = var_1831_cast_fp16)[name = tensor("new_k_cache_13_cast_fp16")]; + tensor new_k_cache_13_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_13_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache6_to_fp16_dtype_0 = const()[name = tensor("v_cache6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache6_to_fp16 = cast(dtype = v_cache6_to_fp16_dtype_0, x = v_cache6)[name = tensor("cast_525")]; + tensor var_1837_cast_fp16 = mul(x = v_cache6_to_fp16, y = var_1823_cast_fp16)[name = tensor("op_1837_cast_fp16")]; + tensor var_1842 = const()[name = tensor("op_1842"), val = tensor([1, 1, 16, 64])]; + tensor var_1843_cast_fp16 = reshape(shape = var_1842, x = v_13_cast_fp16)[name = tensor("op_1843_cast_fp16")]; + tensor var_1844_cast_fp16 = mul(x = var_1843_cast_fp16, y = onehot_27_cast_fp16)[name = tensor("op_1844_cast_fp16")]; + tensor new_v_cache_13_cast_fp16 = add(x = var_1837_cast_fp16, y = var_1844_cast_fp16)[name = tensor("new_v_cache_13_cast_fp16")]; + tensor new_v_cache_13_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_13_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_6_perm_0 = const()[name = tensor("transpose_6_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_1857 = const()[name = tensor("op_1857"), val = tensor([0, 2, 1, 3])]; + tensor var_1863 = const()[name = tensor("op_1863"), val = tensor([1, 16, 1, 64])]; + tensor q4_13_cast_fp16 = reshape(shape = var_1863, x = out_25_cast_fp16)[name = tensor("q4_13_cast_fp16")]; + tensor var_1868_transpose_x_0 = const()[name = tensor("op_1868_transpose_x_0"), val = tensor(false)]; + tensor var_1868_transpose_y_0 = const()[name = tensor("op_1868_transpose_y_0"), val = tensor(false)]; + tensor transpose_6_cast_fp16 = transpose(perm = transpose_6_perm_0, x = new_k_cache_13_cast_fp16)[name = tensor("transpose_77")]; + tensor var_1868_cast_fp16 = matmul(transpose_x = var_1868_transpose_x_0, transpose_y = var_1868_transpose_y_0, x = q4_13_cast_fp16, y = transpose_6_cast_fp16)[name = tensor("op_1868_cast_fp16")]; + tensor var_1869_to_fp16 = const()[name = tensor("op_1869_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_25_cast_fp16 = mul(x = var_1868_cast_fp16, y = var_1869_to_fp16)[name = tensor("scores_25_cast_fp16")]; + tensor var_1873 = const()[name = tensor("op_1873"), val = tensor([1, 1])]; + tensor var_1874_cast_fp16 = reshape(shape = var_1873, x = position6_to_fp16)[name = tensor("op_1874_cast_fp16")]; + tensor var_1875_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_1874_cast_fp16)[name = tensor("op_1875_cast_fp16")]; + tensor mask_13_to_fp16_dtype_0 = const()[name = tensor("mask_13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_1882_to_fp16 = const()[name = tensor("op_1882_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1875_cast_fp16_to_fp16 = cast(dtype = mask_13_to_fp16_dtype_0, x = var_1875_cast_fp16)[name = tensor("cast_523")]; + tensor var_1883_cast_fp16 = sub(x = var_1875_cast_fp16_to_fp16, y = var_1882_to_fp16)[name = tensor("op_1883_cast_fp16")]; + tensor var_1888 = const()[name = tensor("op_1888"), val = tensor([1, 1, 1, 512])]; + tensor var_1889_cast_fp16 = reshape(shape = var_1888, x = var_1883_cast_fp16)[name = tensor("op_1889_cast_fp16")]; + tensor var_1890_to_fp16 = const()[name = tensor("op_1890_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_1891_cast_fp16 = mul(x = var_1889_cast_fp16, y = var_1890_to_fp16)[name = tensor("op_1891_cast_fp16")]; + tensor scores_27_cast_fp16 = add(x = scores_25_cast_fp16, y = var_1891_cast_fp16)[name = tensor("scores_27_cast_fp16")]; + tensor var_1894 = const()[name = tensor("op_1894"), val = tensor(-1)]; + tensor weights_13_cast_fp16 = softmax(axis = var_1894, x = scores_27_cast_fp16)[name = tensor("weights_13_cast_fp16")]; + tensor attn_13_transpose_x_0 = const()[name = tensor("attn_13_transpose_x_0"), val = tensor(false)]; + tensor attn_13_transpose_y_0 = const()[name = tensor("attn_13_transpose_y_0"), val = tensor(false)]; + tensor values_13_cast_fp16 = transpose(perm = var_1857, x = new_v_cache_13_cast_fp16)[name = tensor("transpose_76")]; + tensor attn_13_cast_fp16 = matmul(transpose_x = attn_13_transpose_x_0, transpose_y = attn_13_transpose_y_0, x = weights_13_cast_fp16, y = values_13_cast_fp16)[name = tensor("attn_13_cast_fp16")]; + tensor var_1902 = const()[name = tensor("op_1902"), val = tensor([0, 2, 1, 3])]; + tensor var_1907 = const()[name = tensor("op_1907"), val = tensor([1, 1, 1024])]; + tensor var_1903_cast_fp16 = transpose(perm = var_1902, x = attn_13_cast_fp16)[name = tensor("transpose_75")]; + tensor input_79_cast_fp16 = reshape(shape = var_1907, x = var_1903_cast_fp16)[name = tensor("input_79_cast_fp16")]; + tensor attn6_out_proj_weight_to_fp16 = const()[name = tensor("attn6_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(157426368)))]; + tensor linear_26_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn6_out_proj_weight_to_fp16, x = input_79_cast_fp16)[name = tensor("linear_26_cast_fp16")]; + tensor var_1913_to_fp16 = const()[name = tensor("op_1913_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1914_cast_fp16 = add(x = position6_to_fp16, y = var_1913_to_fp16)[name = tensor("op_1914_cast_fp16")]; + tensor var_1914_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1914_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_81_cast_fp16 = add(x = input_75_cast_fp16, y = linear_26_cast_fp16)[name = tensor("input_81_cast_fp16")]; + tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([-1])]; + tensor norm6_2_weight_to_fp16 = const()[name = tensor("norm6_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159523584)))]; + tensor norm6_2_bias_to_fp16 = const()[name = tensor("norm6_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159525696)))]; + tensor var_1918_to_fp16 = const()[name = tensor("op_1918_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_83_cast_fp16 = layer_norm(axes = input_83_axes_0, beta = norm6_2_bias_to_fp16, epsilon = var_1918_to_fp16, gamma = norm6_2_weight_to_fp16, x = input_81_cast_fp16)[name = tensor("input_83_cast_fp16")]; + tensor linear6_1_weight_to_fp16 = const()[name = tensor("linear6_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159527808)))]; + tensor linear_27_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear6_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_27_cast_fp16")]; + tensor input_85_mode_0 = const()[name = tensor("input_85_mode_0"), val = tensor("EXACT")]; + tensor input_85_cast_fp16 = gelu(mode = input_85_mode_0, x = linear_27_cast_fp16)[name = tensor("input_85_cast_fp16")]; + tensor linear6_2_weight_to_fp16 = const()[name = tensor("linear6_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(167916480)))]; + tensor linear_28_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear6_2_weight_to_fp16, x = input_85_cast_fp16)[name = tensor("linear_28_cast_fp16")]; + tensor input_87_cast_fp16 = add(x = input_81_cast_fp16, y = linear_28_cast_fp16)[name = tensor("input_87_cast_fp16")]; + tensor input_89_axes_0 = const()[name = tensor("input_89_axes_0"), val = tensor([-1])]; + tensor norm7_1_weight_to_fp16 = const()[name = tensor("norm7_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176305152)))]; + tensor norm7_1_bias_to_fp16 = const()[name = tensor("norm7_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176307264)))]; + tensor var_1935_to_fp16 = const()[name = tensor("op_1935_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_89_cast_fp16 = layer_norm(axes = input_89_axes_0, beta = norm7_1_bias_to_fp16, epsilon = var_1935_to_fp16, gamma = norm7_1_weight_to_fp16, x = input_87_cast_fp16)[name = tensor("input_89_cast_fp16")]; + tensor attn7_in_proj_weight_to_fp16 = const()[name = tensor("attn7_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176309376)))]; + tensor linear_29_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn7_in_proj_weight_to_fp16, x = input_89_cast_fp16)[name = tensor("linear_29_cast_fp16")]; + tensor var_1948 = const()[name = tensor("op_1948"), val = tensor([1, 3, 16, 64])]; + tensor qkv_15_cast_fp16 = reshape(shape = var_1948, x = linear_29_cast_fp16)[name = tensor("qkv_15_cast_fp16")]; + tensor x_29_begin_0 = const()[name = tensor("x_29_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_29_end_0 = const()[name = tensor("x_29_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_29_end_mask_0 = const()[name = tensor("x_29_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_29_squeeze_mask_0 = const()[name = tensor("x_29_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_29_cast_fp16 = slice_by_index(begin = x_29_begin_0, end = x_29_end_0, end_mask = x_29_end_mask_0, squeeze_mask = x_29_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("x_29_cast_fp16")]; + tensor x_31_begin_0 = const()[name = tensor("x_31_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_31_end_0 = const()[name = tensor("x_31_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_31_end_mask_0 = const()[name = tensor("x_31_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_31_squeeze_mask_0 = const()[name = tensor("x_31_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_31_cast_fp16 = slice_by_index(begin = x_31_begin_0, end = x_31_end_0, end_mask = x_31_end_mask_0, squeeze_mask = x_31_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("x_31_cast_fp16")]; + tensor v_15_begin_0 = const()[name = tensor("v_15_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_15_end_0 = const()[name = tensor("v_15_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_15_end_mask_0 = const()[name = tensor("v_15_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_15_squeeze_mask_0 = const()[name = tensor("v_15_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_15_cast_fp16 = slice_by_index(begin = v_15_begin_0, end = v_15_end_0, end_mask = v_15_end_mask_0, squeeze_mask = v_15_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("v_15_cast_fp16")]; + tensor var_1977 = const()[name = tensor("op_1977"), val = tensor([1, 1, 1])]; + tensor position7_to_fp16_dtype_0 = const()[name = tensor("position7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position7_to_fp16 = cast(dtype = position7_to_fp16_dtype_0, x = position7)[name = tensor("cast_521")]; + tensor var_1978_cast_fp16 = reshape(shape = var_1977, x = position7_to_fp16)[name = tensor("op_1978_cast_fp16")]; + tensor angles_29_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_1978_cast_fp16)[name = tensor("angles_29_cast_fp16")]; + tensor rotr_29_cast_fp16 = cos(x = angles_29_cast_fp16)[name = tensor("rotr_29_cast_fp16")]; + tensor roti_29_cast_fp16 = sin(x = angles_29_cast_fp16)[name = tensor("roti_29_cast_fp16")]; + tensor var_1986 = const()[name = tensor("op_1986"), val = tensor([1, 16, 32, 2])]; + tensor pairs_29_cast_fp16 = reshape(shape = var_1986, x = x_29_cast_fp16)[name = tensor("pairs_29_cast_fp16")]; + tensor xr_29_begin_0 = const()[name = tensor("xr_29_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_29_end_0 = const()[name = tensor("xr_29_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_29_end_mask_0 = const()[name = tensor("xr_29_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_29_squeeze_mask_0 = const()[name = tensor("xr_29_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_29_cast_fp16 = slice_by_index(begin = xr_29_begin_0, end = xr_29_end_0, end_mask = xr_29_end_mask_0, squeeze_mask = xr_29_squeeze_mask_0, x = pairs_29_cast_fp16)[name = tensor("xr_29_cast_fp16")]; + tensor xi_29_begin_0 = const()[name = tensor("xi_29_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_29_end_0 = const()[name = tensor("xi_29_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_29_end_mask_0 = const()[name = tensor("xi_29_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_29_squeeze_mask_0 = const()[name = tensor("xi_29_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_29_cast_fp16 = slice_by_index(begin = xi_29_begin_0, end = xi_29_end_0, end_mask = xi_29_end_mask_0, squeeze_mask = xi_29_squeeze_mask_0, x = pairs_29_cast_fp16)[name = tensor("xi_29_cast_fp16")]; + tensor var_1994_cast_fp16 = mul(x = xr_29_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_1994_cast_fp16")]; + tensor var_1995_cast_fp16 = mul(x = xi_29_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_1995_cast_fp16")]; + tensor xor__29_cast_fp16 = sub(x = var_1994_cast_fp16, y = var_1995_cast_fp16)[name = tensor("xor__29_cast_fp16")]; + tensor var_1998_cast_fp16 = mul(x = xr_29_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_1998_cast_fp16")]; + tensor var_1999_cast_fp16 = mul(x = xi_29_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_1999_cast_fp16")]; + tensor xoi_29_cast_fp16 = add(x = var_1998_cast_fp16, y = var_1999_cast_fp16)[name = tensor("xoi_29_cast_fp16")]; + tensor out_29_axis_0 = const()[name = tensor("out_29_axis_0"), val = tensor(-1)]; + tensor out_29_cast_fp16 = stack(axis = out_29_axis_0, values = (xor__29_cast_fp16, xoi_29_cast_fp16))[name = tensor("out_29_cast_fp16")]; + tensor var_2022 = const()[name = tensor("op_2022"), val = tensor([1, 16, 32, 2])]; + tensor pairs_31_cast_fp16 = reshape(shape = var_2022, x = x_31_cast_fp16)[name = tensor("pairs_31_cast_fp16")]; + tensor xr_31_begin_0 = const()[name = tensor("xr_31_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_31_end_0 = const()[name = tensor("xr_31_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_31_end_mask_0 = const()[name = tensor("xr_31_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_31_squeeze_mask_0 = const()[name = tensor("xr_31_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_31_cast_fp16 = slice_by_index(begin = xr_31_begin_0, end = xr_31_end_0, end_mask = xr_31_end_mask_0, squeeze_mask = xr_31_squeeze_mask_0, x = pairs_31_cast_fp16)[name = tensor("xr_31_cast_fp16")]; + tensor xi_31_begin_0 = const()[name = tensor("xi_31_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_31_end_0 = const()[name = tensor("xi_31_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_31_end_mask_0 = const()[name = tensor("xi_31_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_31_squeeze_mask_0 = const()[name = tensor("xi_31_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_31_cast_fp16 = slice_by_index(begin = xi_31_begin_0, end = xi_31_end_0, end_mask = xi_31_end_mask_0, squeeze_mask = xi_31_squeeze_mask_0, x = pairs_31_cast_fp16)[name = tensor("xi_31_cast_fp16")]; + tensor var_2030_cast_fp16 = mul(x = xr_31_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_2030_cast_fp16")]; + tensor var_2031_cast_fp16 = mul(x = xi_31_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_2031_cast_fp16")]; + tensor xor__31_cast_fp16 = sub(x = var_2030_cast_fp16, y = var_2031_cast_fp16)[name = tensor("xor__31_cast_fp16")]; + tensor var_2034_cast_fp16 = mul(x = xr_31_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_2034_cast_fp16")]; + tensor var_2035_cast_fp16 = mul(x = xi_31_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_2035_cast_fp16")]; + tensor xoi_31_cast_fp16 = add(x = var_2034_cast_fp16, y = var_2035_cast_fp16)[name = tensor("xoi_31_cast_fp16")]; + tensor out_31_axis_0 = const()[name = tensor("out_31_axis_0"), val = tensor(-1)]; + tensor out_31_cast_fp16 = stack(axis = out_31_axis_0, values = (xor__31_cast_fp16, xoi_31_cast_fp16))[name = tensor("out_31_cast_fp16")]; + tensor _inversed_2047_y_0_to_fp16 = const()[name = tensor("_inversed_2047_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2047_cast_fp16 = mul(x = position7_to_fp16, y = _inversed_2047_y_0_to_fp16)[name = tensor("_inversed_2047_cast_fp16")]; + tensor var_2048_cast_fp16 = floor(x = _inversed_2047_cast_fp16)[name = tensor("op_2048_cast_fp16")]; + tensor var_2049_to_fp16 = const()[name = tensor("op_2049_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2050_cast_fp16 = mul(x = var_2048_cast_fp16, y = var_2049_to_fp16)[name = tensor("op_2050_cast_fp16")]; + tensor write_idx_15_cast_fp16 = sub(x = position7_to_fp16, y = var_2050_cast_fp16)[name = tensor("write_idx_15_cast_fp16")]; + tensor var_2053_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_15_cast_fp16)[name = tensor("op_2053_cast_fp16")]; + tensor var_2063 = const()[name = tensor("op_2063"), val = tensor([1, 512, 1, 1])]; + tensor onehot_29_to_fp16_dtype_0 = const()[name = tensor("onehot_29_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_2053_cast_fp16_to_fp16 = cast(dtype = onehot_29_to_fp16_dtype_0, x = var_2053_cast_fp16)[name = tensor("cast_520")]; + tensor onehot_31_cast_fp16 = reshape(shape = var_2063, x = var_2053_cast_fp16_to_fp16)[name = tensor("onehot_31_cast_fp16")]; + tensor var_2065_to_fp16 = const()[name = tensor("op_2065_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2067_cast_fp16 = sub(x = var_2065_to_fp16, y = onehot_31_cast_fp16)[name = tensor("op_2067_cast_fp16")]; + tensor k_cache7_to_fp16_dtype_0 = const()[name = tensor("k_cache7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache7_to_fp16 = cast(dtype = k_cache7_to_fp16_dtype_0, x = k_cache7)[name = tensor("cast_519")]; + tensor var_2068_cast_fp16 = mul(x = k_cache7_to_fp16, y = var_2067_cast_fp16)[name = tensor("op_2068_cast_fp16")]; + tensor var_2073 = const()[name = tensor("op_2073"), val = tensor([1, 1, 16, 64])]; + tensor var_2074_cast_fp16 = reshape(shape = var_2073, x = out_31_cast_fp16)[name = tensor("op_2074_cast_fp16")]; + tensor var_2075_cast_fp16 = mul(x = var_2074_cast_fp16, y = onehot_31_cast_fp16)[name = tensor("op_2075_cast_fp16")]; + tensor new_k_cache_15_cast_fp16 = add(x = var_2068_cast_fp16, y = var_2075_cast_fp16)[name = tensor("new_k_cache_15_cast_fp16")]; + tensor new_k_cache_15_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_15_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache7_to_fp16_dtype_0 = const()[name = tensor("v_cache7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache7_to_fp16 = cast(dtype = v_cache7_to_fp16_dtype_0, x = v_cache7)[name = tensor("cast_517")]; + tensor var_2081_cast_fp16 = mul(x = v_cache7_to_fp16, y = var_2067_cast_fp16)[name = tensor("op_2081_cast_fp16")]; + tensor var_2086 = const()[name = tensor("op_2086"), val = tensor([1, 1, 16, 64])]; + tensor var_2087_cast_fp16 = reshape(shape = var_2086, x = v_15_cast_fp16)[name = tensor("op_2087_cast_fp16")]; + tensor var_2088_cast_fp16 = mul(x = var_2087_cast_fp16, y = onehot_31_cast_fp16)[name = tensor("op_2088_cast_fp16")]; + tensor new_v_cache_15_cast_fp16 = add(x = var_2081_cast_fp16, y = var_2088_cast_fp16)[name = tensor("new_v_cache_15_cast_fp16")]; + tensor new_v_cache_15_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_15_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_7_perm_0 = const()[name = tensor("transpose_7_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_2101 = const()[name = tensor("op_2101"), val = tensor([0, 2, 1, 3])]; + tensor var_2107 = const()[name = tensor("op_2107"), val = tensor([1, 16, 1, 64])]; + tensor q4_15_cast_fp16 = reshape(shape = var_2107, x = out_29_cast_fp16)[name = tensor("q4_15_cast_fp16")]; + tensor var_2112_transpose_x_0 = const()[name = tensor("op_2112_transpose_x_0"), val = tensor(false)]; + tensor var_2112_transpose_y_0 = const()[name = tensor("op_2112_transpose_y_0"), val = tensor(false)]; + tensor transpose_7_cast_fp16 = transpose(perm = transpose_7_perm_0, x = new_k_cache_15_cast_fp16)[name = tensor("transpose_74")]; + tensor var_2112_cast_fp16 = matmul(transpose_x = var_2112_transpose_x_0, transpose_y = var_2112_transpose_y_0, x = q4_15_cast_fp16, y = transpose_7_cast_fp16)[name = tensor("op_2112_cast_fp16")]; + tensor var_2113_to_fp16 = const()[name = tensor("op_2113_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_29_cast_fp16 = mul(x = var_2112_cast_fp16, y = var_2113_to_fp16)[name = tensor("scores_29_cast_fp16")]; + tensor var_2117 = const()[name = tensor("op_2117"), val = tensor([1, 1])]; + tensor var_2118_cast_fp16 = reshape(shape = var_2117, x = position7_to_fp16)[name = tensor("op_2118_cast_fp16")]; + tensor var_2119_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_2118_cast_fp16)[name = tensor("op_2119_cast_fp16")]; + tensor mask_15_to_fp16_dtype_0 = const()[name = tensor("mask_15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_2126_to_fp16 = const()[name = tensor("op_2126_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2119_cast_fp16_to_fp16 = cast(dtype = mask_15_to_fp16_dtype_0, x = var_2119_cast_fp16)[name = tensor("cast_515")]; + tensor var_2127_cast_fp16 = sub(x = var_2119_cast_fp16_to_fp16, y = var_2126_to_fp16)[name = tensor("op_2127_cast_fp16")]; + tensor var_2132 = const()[name = tensor("op_2132"), val = tensor([1, 1, 1, 512])]; + tensor var_2133_cast_fp16 = reshape(shape = var_2132, x = var_2127_cast_fp16)[name = tensor("op_2133_cast_fp16")]; + tensor var_2134_to_fp16 = const()[name = tensor("op_2134_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_2135_cast_fp16 = mul(x = var_2133_cast_fp16, y = var_2134_to_fp16)[name = tensor("op_2135_cast_fp16")]; + tensor scores_31_cast_fp16 = add(x = scores_29_cast_fp16, y = var_2135_cast_fp16)[name = tensor("scores_31_cast_fp16")]; + tensor var_2138 = const()[name = tensor("op_2138"), val = tensor(-1)]; + tensor weights_15_cast_fp16 = softmax(axis = var_2138, x = scores_31_cast_fp16)[name = tensor("weights_15_cast_fp16")]; + tensor attn_15_transpose_x_0 = const()[name = tensor("attn_15_transpose_x_0"), val = tensor(false)]; + tensor attn_15_transpose_y_0 = const()[name = tensor("attn_15_transpose_y_0"), val = tensor(false)]; + tensor values_15_cast_fp16 = transpose(perm = var_2101, x = new_v_cache_15_cast_fp16)[name = tensor("transpose_73")]; + tensor attn_15_cast_fp16 = matmul(transpose_x = attn_15_transpose_x_0, transpose_y = attn_15_transpose_y_0, x = weights_15_cast_fp16, y = values_15_cast_fp16)[name = tensor("attn_15_cast_fp16")]; + tensor var_2146 = const()[name = tensor("op_2146"), val = tensor([0, 2, 1, 3])]; + tensor var_2151 = const()[name = tensor("op_2151"), val = tensor([1, 1, 1024])]; + tensor var_2147_cast_fp16 = transpose(perm = var_2146, x = attn_15_cast_fp16)[name = tensor("transpose_72")]; + tensor input_91_cast_fp16 = reshape(shape = var_2151, x = var_2147_cast_fp16)[name = tensor("input_91_cast_fp16")]; + tensor attn7_out_proj_weight_to_fp16 = const()[name = tensor("attn7_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(182600896)))]; + tensor linear_30_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn7_out_proj_weight_to_fp16, x = input_91_cast_fp16)[name = tensor("linear_30_cast_fp16")]; + tensor var_2157_to_fp16 = const()[name = tensor("op_2157_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2158_cast_fp16 = add(x = position7_to_fp16, y = var_2157_to_fp16)[name = tensor("op_2158_cast_fp16")]; + tensor var_2158_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2158_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_93_cast_fp16 = add(x = input_87_cast_fp16, y = linear_30_cast_fp16)[name = tensor("input_93_cast_fp16")]; + tensor input_95_axes_0 = const()[name = tensor("input_95_axes_0"), val = tensor([-1])]; + tensor norm7_2_weight_to_fp16 = const()[name = tensor("norm7_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184698112)))]; + tensor norm7_2_bias_to_fp16 = const()[name = tensor("norm7_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184700224)))]; + tensor var_2162_to_fp16 = const()[name = tensor("op_2162_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_95_cast_fp16 = layer_norm(axes = input_95_axes_0, beta = norm7_2_bias_to_fp16, epsilon = var_2162_to_fp16, gamma = norm7_2_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("input_95_cast_fp16")]; + tensor linear7_1_weight_to_fp16 = const()[name = tensor("linear7_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184702336)))]; + tensor linear_31_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear7_1_weight_to_fp16, x = input_95_cast_fp16)[name = tensor("linear_31_cast_fp16")]; + tensor input_97_mode_0 = const()[name = tensor("input_97_mode_0"), val = tensor("EXACT")]; + tensor input_97_cast_fp16 = gelu(mode = input_97_mode_0, x = linear_31_cast_fp16)[name = tensor("input_97_cast_fp16")]; + tensor linear7_2_weight_to_fp16 = const()[name = tensor("linear7_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(193091008)))]; + tensor linear_32_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear7_2_weight_to_fp16, x = input_97_cast_fp16)[name = tensor("linear_32_cast_fp16")]; + tensor input_99_cast_fp16 = add(x = input_93_cast_fp16, y = linear_32_cast_fp16)[name = tensor("input_99_cast_fp16")]; + tensor input_101_axes_0 = const()[name = tensor("input_101_axes_0"), val = tensor([-1])]; + tensor norm8_1_weight_to_fp16 = const()[name = tensor("norm8_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201479680)))]; + tensor norm8_1_bias_to_fp16 = const()[name = tensor("norm8_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201481792)))]; + tensor var_2179_to_fp16 = const()[name = tensor("op_2179_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_101_cast_fp16 = layer_norm(axes = input_101_axes_0, beta = norm8_1_bias_to_fp16, epsilon = var_2179_to_fp16, gamma = norm8_1_weight_to_fp16, x = input_99_cast_fp16)[name = tensor("input_101_cast_fp16")]; + tensor attn8_in_proj_weight_to_fp16 = const()[name = tensor("attn8_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201483904)))]; + tensor linear_33_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn8_in_proj_weight_to_fp16, x = input_101_cast_fp16)[name = tensor("linear_33_cast_fp16")]; + tensor var_2192 = const()[name = tensor("op_2192"), val = tensor([1, 3, 16, 64])]; + tensor qkv_17_cast_fp16 = reshape(shape = var_2192, x = linear_33_cast_fp16)[name = tensor("qkv_17_cast_fp16")]; + tensor x_33_begin_0 = const()[name = tensor("x_33_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_33_end_0 = const()[name = tensor("x_33_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_33_end_mask_0 = const()[name = tensor("x_33_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_33_squeeze_mask_0 = const()[name = tensor("x_33_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_33_cast_fp16 = slice_by_index(begin = x_33_begin_0, end = x_33_end_0, end_mask = x_33_end_mask_0, squeeze_mask = x_33_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("x_33_cast_fp16")]; + tensor x_35_begin_0 = const()[name = tensor("x_35_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_35_end_0 = const()[name = tensor("x_35_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_35_end_mask_0 = const()[name = tensor("x_35_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_35_squeeze_mask_0 = const()[name = tensor("x_35_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_35_cast_fp16 = slice_by_index(begin = x_35_begin_0, end = x_35_end_0, end_mask = x_35_end_mask_0, squeeze_mask = x_35_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("x_35_cast_fp16")]; + tensor v_17_begin_0 = const()[name = tensor("v_17_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_17_end_0 = const()[name = tensor("v_17_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_17_end_mask_0 = const()[name = tensor("v_17_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_17_squeeze_mask_0 = const()[name = tensor("v_17_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_17_cast_fp16 = slice_by_index(begin = v_17_begin_0, end = v_17_end_0, end_mask = v_17_end_mask_0, squeeze_mask = v_17_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("v_17_cast_fp16")]; + tensor var_2221 = const()[name = tensor("op_2221"), val = tensor([1, 1, 1])]; + tensor position8_to_fp16_dtype_0 = const()[name = tensor("position8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position8_to_fp16 = cast(dtype = position8_to_fp16_dtype_0, x = position8)[name = tensor("cast_513")]; + tensor var_2222_cast_fp16 = reshape(shape = var_2221, x = position8_to_fp16)[name = tensor("op_2222_cast_fp16")]; + tensor angles_33_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_2222_cast_fp16)[name = tensor("angles_33_cast_fp16")]; + tensor rotr_33_cast_fp16 = cos(x = angles_33_cast_fp16)[name = tensor("rotr_33_cast_fp16")]; + tensor roti_33_cast_fp16 = sin(x = angles_33_cast_fp16)[name = tensor("roti_33_cast_fp16")]; + tensor var_2230 = const()[name = tensor("op_2230"), val = tensor([1, 16, 32, 2])]; + tensor pairs_33_cast_fp16 = reshape(shape = var_2230, x = x_33_cast_fp16)[name = tensor("pairs_33_cast_fp16")]; + tensor xr_33_begin_0 = const()[name = tensor("xr_33_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_33_end_0 = const()[name = tensor("xr_33_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_33_end_mask_0 = const()[name = tensor("xr_33_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_33_squeeze_mask_0 = const()[name = tensor("xr_33_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_33_cast_fp16 = slice_by_index(begin = xr_33_begin_0, end = xr_33_end_0, end_mask = xr_33_end_mask_0, squeeze_mask = xr_33_squeeze_mask_0, x = pairs_33_cast_fp16)[name = tensor("xr_33_cast_fp16")]; + tensor xi_33_begin_0 = const()[name = tensor("xi_33_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_33_end_0 = const()[name = tensor("xi_33_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_33_end_mask_0 = const()[name = tensor("xi_33_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_33_squeeze_mask_0 = const()[name = tensor("xi_33_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_33_cast_fp16 = slice_by_index(begin = xi_33_begin_0, end = xi_33_end_0, end_mask = xi_33_end_mask_0, squeeze_mask = xi_33_squeeze_mask_0, x = pairs_33_cast_fp16)[name = tensor("xi_33_cast_fp16")]; + tensor var_2238_cast_fp16 = mul(x = xr_33_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_2238_cast_fp16")]; + tensor var_2239_cast_fp16 = mul(x = xi_33_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_2239_cast_fp16")]; + tensor xor__33_cast_fp16 = sub(x = var_2238_cast_fp16, y = var_2239_cast_fp16)[name = tensor("xor__33_cast_fp16")]; + tensor var_2242_cast_fp16 = mul(x = xr_33_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_2242_cast_fp16")]; + tensor var_2243_cast_fp16 = mul(x = xi_33_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_2243_cast_fp16")]; + tensor xoi_33_cast_fp16 = add(x = var_2242_cast_fp16, y = var_2243_cast_fp16)[name = tensor("xoi_33_cast_fp16")]; + tensor out_33_axis_0 = const()[name = tensor("out_33_axis_0"), val = tensor(-1)]; + tensor out_33_cast_fp16 = stack(axis = out_33_axis_0, values = (xor__33_cast_fp16, xoi_33_cast_fp16))[name = tensor("out_33_cast_fp16")]; + tensor var_2266 = const()[name = tensor("op_2266"), val = tensor([1, 16, 32, 2])]; + tensor pairs_35_cast_fp16 = reshape(shape = var_2266, x = x_35_cast_fp16)[name = tensor("pairs_35_cast_fp16")]; + tensor xr_35_begin_0 = const()[name = tensor("xr_35_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_35_end_0 = const()[name = tensor("xr_35_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_35_end_mask_0 = const()[name = tensor("xr_35_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_35_squeeze_mask_0 = const()[name = tensor("xr_35_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_35_cast_fp16 = slice_by_index(begin = xr_35_begin_0, end = xr_35_end_0, end_mask = xr_35_end_mask_0, squeeze_mask = xr_35_squeeze_mask_0, x = pairs_35_cast_fp16)[name = tensor("xr_35_cast_fp16")]; + tensor xi_35_begin_0 = const()[name = tensor("xi_35_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_35_end_0 = const()[name = tensor("xi_35_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_35_end_mask_0 = const()[name = tensor("xi_35_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_35_squeeze_mask_0 = const()[name = tensor("xi_35_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_35_cast_fp16 = slice_by_index(begin = xi_35_begin_0, end = xi_35_end_0, end_mask = xi_35_end_mask_0, squeeze_mask = xi_35_squeeze_mask_0, x = pairs_35_cast_fp16)[name = tensor("xi_35_cast_fp16")]; + tensor var_2274_cast_fp16 = mul(x = xr_35_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_2274_cast_fp16")]; + tensor var_2275_cast_fp16 = mul(x = xi_35_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_2275_cast_fp16")]; + tensor xor__35_cast_fp16 = sub(x = var_2274_cast_fp16, y = var_2275_cast_fp16)[name = tensor("xor__35_cast_fp16")]; + tensor var_2278_cast_fp16 = mul(x = xr_35_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_2278_cast_fp16")]; + tensor var_2279_cast_fp16 = mul(x = xi_35_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_2279_cast_fp16")]; + tensor xoi_35_cast_fp16 = add(x = var_2278_cast_fp16, y = var_2279_cast_fp16)[name = tensor("xoi_35_cast_fp16")]; + tensor out_35_axis_0 = const()[name = tensor("out_35_axis_0"), val = tensor(-1)]; + tensor out_35_cast_fp16 = stack(axis = out_35_axis_0, values = (xor__35_cast_fp16, xoi_35_cast_fp16))[name = tensor("out_35_cast_fp16")]; + tensor _inversed_2291_y_0_to_fp16 = const()[name = tensor("_inversed_2291_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2291_cast_fp16 = mul(x = position8_to_fp16, y = _inversed_2291_y_0_to_fp16)[name = tensor("_inversed_2291_cast_fp16")]; + tensor var_2292_cast_fp16 = floor(x = _inversed_2291_cast_fp16)[name = tensor("op_2292_cast_fp16")]; + tensor var_2293_to_fp16 = const()[name = tensor("op_2293_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2294_cast_fp16 = mul(x = var_2292_cast_fp16, y = var_2293_to_fp16)[name = tensor("op_2294_cast_fp16")]; + tensor write_idx_17_cast_fp16 = sub(x = position8_to_fp16, y = var_2294_cast_fp16)[name = tensor("write_idx_17_cast_fp16")]; + tensor var_2297_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_17_cast_fp16)[name = tensor("op_2297_cast_fp16")]; + tensor var_2307 = const()[name = tensor("op_2307"), val = tensor([1, 512, 1, 1])]; + tensor onehot_33_to_fp16_dtype_0 = const()[name = tensor("onehot_33_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_2297_cast_fp16_to_fp16 = cast(dtype = onehot_33_to_fp16_dtype_0, x = var_2297_cast_fp16)[name = tensor("cast_512")]; + tensor onehot_35_cast_fp16 = reshape(shape = var_2307, x = var_2297_cast_fp16_to_fp16)[name = tensor("onehot_35_cast_fp16")]; + tensor var_2309_to_fp16 = const()[name = tensor("op_2309_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2311_cast_fp16 = sub(x = var_2309_to_fp16, y = onehot_35_cast_fp16)[name = tensor("op_2311_cast_fp16")]; + tensor k_cache8_to_fp16_dtype_0 = const()[name = tensor("k_cache8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache8_to_fp16 = cast(dtype = k_cache8_to_fp16_dtype_0, x = k_cache8)[name = tensor("cast_511")]; + tensor var_2312_cast_fp16 = mul(x = k_cache8_to_fp16, y = var_2311_cast_fp16)[name = tensor("op_2312_cast_fp16")]; + tensor var_2317 = const()[name = tensor("op_2317"), val = tensor([1, 1, 16, 64])]; + tensor var_2318_cast_fp16 = reshape(shape = var_2317, x = out_35_cast_fp16)[name = tensor("op_2318_cast_fp16")]; + tensor var_2319_cast_fp16 = mul(x = var_2318_cast_fp16, y = onehot_35_cast_fp16)[name = tensor("op_2319_cast_fp16")]; + tensor new_k_cache_17_cast_fp16 = add(x = var_2312_cast_fp16, y = var_2319_cast_fp16)[name = tensor("new_k_cache_17_cast_fp16")]; + tensor new_k_cache_17_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_17_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache8_to_fp16_dtype_0 = const()[name = tensor("v_cache8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache8_to_fp16 = cast(dtype = v_cache8_to_fp16_dtype_0, x = v_cache8)[name = tensor("cast_509")]; + tensor var_2325_cast_fp16 = mul(x = v_cache8_to_fp16, y = var_2311_cast_fp16)[name = tensor("op_2325_cast_fp16")]; + tensor var_2330 = const()[name = tensor("op_2330"), val = tensor([1, 1, 16, 64])]; + tensor var_2331_cast_fp16 = reshape(shape = var_2330, x = v_17_cast_fp16)[name = tensor("op_2331_cast_fp16")]; + tensor var_2332_cast_fp16 = mul(x = var_2331_cast_fp16, y = onehot_35_cast_fp16)[name = tensor("op_2332_cast_fp16")]; + tensor new_v_cache_17_cast_fp16 = add(x = var_2325_cast_fp16, y = var_2332_cast_fp16)[name = tensor("new_v_cache_17_cast_fp16")]; + tensor new_v_cache_17_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_17_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_8_perm_0 = const()[name = tensor("transpose_8_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_2345 = const()[name = tensor("op_2345"), val = tensor([0, 2, 1, 3])]; + tensor var_2351 = const()[name = tensor("op_2351"), val = tensor([1, 16, 1, 64])]; + tensor q4_17_cast_fp16 = reshape(shape = var_2351, x = out_33_cast_fp16)[name = tensor("q4_17_cast_fp16")]; + tensor var_2356_transpose_x_0 = const()[name = tensor("op_2356_transpose_x_0"), val = tensor(false)]; + tensor var_2356_transpose_y_0 = const()[name = tensor("op_2356_transpose_y_0"), val = tensor(false)]; + tensor transpose_8_cast_fp16 = transpose(perm = transpose_8_perm_0, x = new_k_cache_17_cast_fp16)[name = tensor("transpose_71")]; + tensor var_2356_cast_fp16 = matmul(transpose_x = var_2356_transpose_x_0, transpose_y = var_2356_transpose_y_0, x = q4_17_cast_fp16, y = transpose_8_cast_fp16)[name = tensor("op_2356_cast_fp16")]; + tensor var_2357_to_fp16 = const()[name = tensor("op_2357_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_33_cast_fp16 = mul(x = var_2356_cast_fp16, y = var_2357_to_fp16)[name = tensor("scores_33_cast_fp16")]; + tensor var_2361 = const()[name = tensor("op_2361"), val = tensor([1, 1])]; + tensor var_2362_cast_fp16 = reshape(shape = var_2361, x = position8_to_fp16)[name = tensor("op_2362_cast_fp16")]; + tensor var_2363_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_2362_cast_fp16)[name = tensor("op_2363_cast_fp16")]; + tensor mask_17_to_fp16_dtype_0 = const()[name = tensor("mask_17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_2370_to_fp16 = const()[name = tensor("op_2370_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2363_cast_fp16_to_fp16 = cast(dtype = mask_17_to_fp16_dtype_0, x = var_2363_cast_fp16)[name = tensor("cast_507")]; + tensor var_2371_cast_fp16 = sub(x = var_2363_cast_fp16_to_fp16, y = var_2370_to_fp16)[name = tensor("op_2371_cast_fp16")]; + tensor var_2376 = const()[name = tensor("op_2376"), val = tensor([1, 1, 1, 512])]; + tensor var_2377_cast_fp16 = reshape(shape = var_2376, x = var_2371_cast_fp16)[name = tensor("op_2377_cast_fp16")]; + tensor var_2378_to_fp16 = const()[name = tensor("op_2378_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_2379_cast_fp16 = mul(x = var_2377_cast_fp16, y = var_2378_to_fp16)[name = tensor("op_2379_cast_fp16")]; + tensor scores_35_cast_fp16 = add(x = scores_33_cast_fp16, y = var_2379_cast_fp16)[name = tensor("scores_35_cast_fp16")]; + tensor var_2382 = const()[name = tensor("op_2382"), val = tensor(-1)]; + tensor weights_17_cast_fp16 = softmax(axis = var_2382, x = scores_35_cast_fp16)[name = tensor("weights_17_cast_fp16")]; + tensor attn_17_transpose_x_0 = const()[name = tensor("attn_17_transpose_x_0"), val = tensor(false)]; + tensor attn_17_transpose_y_0 = const()[name = tensor("attn_17_transpose_y_0"), val = tensor(false)]; + tensor values_17_cast_fp16 = transpose(perm = var_2345, x = new_v_cache_17_cast_fp16)[name = tensor("transpose_70")]; + tensor attn_17_cast_fp16 = matmul(transpose_x = attn_17_transpose_x_0, transpose_y = attn_17_transpose_y_0, x = weights_17_cast_fp16, y = values_17_cast_fp16)[name = tensor("attn_17_cast_fp16")]; + tensor var_2390 = const()[name = tensor("op_2390"), val = tensor([0, 2, 1, 3])]; + tensor var_2395 = const()[name = tensor("op_2395"), val = tensor([1, 1, 1024])]; + tensor var_2391_cast_fp16 = transpose(perm = var_2390, x = attn_17_cast_fp16)[name = tensor("transpose_69")]; + tensor input_103_cast_fp16 = reshape(shape = var_2395, x = var_2391_cast_fp16)[name = tensor("input_103_cast_fp16")]; + tensor attn8_out_proj_weight_to_fp16 = const()[name = tensor("attn8_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(207775424)))]; + tensor linear_34_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn8_out_proj_weight_to_fp16, x = input_103_cast_fp16)[name = tensor("linear_34_cast_fp16")]; + tensor var_2401_to_fp16 = const()[name = tensor("op_2401_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2402_cast_fp16 = add(x = position8_to_fp16, y = var_2401_to_fp16)[name = tensor("op_2402_cast_fp16")]; + tensor var_2402_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2402_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_105_cast_fp16 = add(x = input_99_cast_fp16, y = linear_34_cast_fp16)[name = tensor("input_105_cast_fp16")]; + tensor input_107_axes_0 = const()[name = tensor("input_107_axes_0"), val = tensor([-1])]; + tensor norm8_2_weight_to_fp16 = const()[name = tensor("norm8_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209872640)))]; + tensor norm8_2_bias_to_fp16 = const()[name = tensor("norm8_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209874752)))]; + tensor var_2406_to_fp16 = const()[name = tensor("op_2406_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_107_cast_fp16 = layer_norm(axes = input_107_axes_0, beta = norm8_2_bias_to_fp16, epsilon = var_2406_to_fp16, gamma = norm8_2_weight_to_fp16, x = input_105_cast_fp16)[name = tensor("input_107_cast_fp16")]; + tensor linear8_1_weight_to_fp16 = const()[name = tensor("linear8_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209876864)))]; + tensor linear_35_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear8_1_weight_to_fp16, x = input_107_cast_fp16)[name = tensor("linear_35_cast_fp16")]; + tensor input_109_mode_0 = const()[name = tensor("input_109_mode_0"), val = tensor("EXACT")]; + tensor input_109_cast_fp16 = gelu(mode = input_109_mode_0, x = linear_35_cast_fp16)[name = tensor("input_109_cast_fp16")]; + tensor linear8_2_weight_to_fp16 = const()[name = tensor("linear8_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(218265536)))]; + tensor linear_36_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear8_2_weight_to_fp16, x = input_109_cast_fp16)[name = tensor("linear_36_cast_fp16")]; + tensor input_111_cast_fp16 = add(x = input_105_cast_fp16, y = linear_36_cast_fp16)[name = tensor("input_111_cast_fp16")]; + tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([-1])]; + tensor norm9_1_weight_to_fp16 = const()[name = tensor("norm9_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226654208)))]; + tensor norm9_1_bias_to_fp16 = const()[name = tensor("norm9_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226656320)))]; + tensor var_2423_to_fp16 = const()[name = tensor("op_2423_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_113_cast_fp16 = layer_norm(axes = input_113_axes_0, beta = norm9_1_bias_to_fp16, epsilon = var_2423_to_fp16, gamma = norm9_1_weight_to_fp16, x = input_111_cast_fp16)[name = tensor("input_113_cast_fp16")]; + tensor attn9_in_proj_weight_to_fp16 = const()[name = tensor("attn9_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226658432)))]; + tensor linear_37_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn9_in_proj_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("linear_37_cast_fp16")]; + tensor var_2436 = const()[name = tensor("op_2436"), val = tensor([1, 3, 16, 64])]; + tensor qkv_19_cast_fp16 = reshape(shape = var_2436, x = linear_37_cast_fp16)[name = tensor("qkv_19_cast_fp16")]; + tensor x_37_begin_0 = const()[name = tensor("x_37_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_37_end_0 = const()[name = tensor("x_37_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_37_end_mask_0 = const()[name = tensor("x_37_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_37_squeeze_mask_0 = const()[name = tensor("x_37_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_37_cast_fp16 = slice_by_index(begin = x_37_begin_0, end = x_37_end_0, end_mask = x_37_end_mask_0, squeeze_mask = x_37_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("x_37_cast_fp16")]; + tensor x_39_begin_0 = const()[name = tensor("x_39_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_39_end_0 = const()[name = tensor("x_39_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_39_end_mask_0 = const()[name = tensor("x_39_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_39_squeeze_mask_0 = const()[name = tensor("x_39_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_39_cast_fp16 = slice_by_index(begin = x_39_begin_0, end = x_39_end_0, end_mask = x_39_end_mask_0, squeeze_mask = x_39_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("x_39_cast_fp16")]; + tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_19_cast_fp16 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("v_19_cast_fp16")]; + tensor var_2465 = const()[name = tensor("op_2465"), val = tensor([1, 1, 1])]; + tensor position9_to_fp16_dtype_0 = const()[name = tensor("position9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position9_to_fp16 = cast(dtype = position9_to_fp16_dtype_0, x = position9)[name = tensor("cast_505")]; + tensor var_2466_cast_fp16 = reshape(shape = var_2465, x = position9_to_fp16)[name = tensor("op_2466_cast_fp16")]; + tensor angles_37_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_2466_cast_fp16)[name = tensor("angles_37_cast_fp16")]; + tensor rotr_37_cast_fp16 = cos(x = angles_37_cast_fp16)[name = tensor("rotr_37_cast_fp16")]; + tensor roti_37_cast_fp16 = sin(x = angles_37_cast_fp16)[name = tensor("roti_37_cast_fp16")]; + tensor var_2474 = const()[name = tensor("op_2474"), val = tensor([1, 16, 32, 2])]; + tensor pairs_37_cast_fp16 = reshape(shape = var_2474, x = x_37_cast_fp16)[name = tensor("pairs_37_cast_fp16")]; + tensor xr_37_begin_0 = const()[name = tensor("xr_37_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_37_end_0 = const()[name = tensor("xr_37_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_37_end_mask_0 = const()[name = tensor("xr_37_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_37_squeeze_mask_0 = const()[name = tensor("xr_37_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_37_cast_fp16 = slice_by_index(begin = xr_37_begin_0, end = xr_37_end_0, end_mask = xr_37_end_mask_0, squeeze_mask = xr_37_squeeze_mask_0, x = pairs_37_cast_fp16)[name = tensor("xr_37_cast_fp16")]; + tensor xi_37_begin_0 = const()[name = tensor("xi_37_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_37_end_0 = const()[name = tensor("xi_37_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_37_end_mask_0 = const()[name = tensor("xi_37_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_37_squeeze_mask_0 = const()[name = tensor("xi_37_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_37_cast_fp16 = slice_by_index(begin = xi_37_begin_0, end = xi_37_end_0, end_mask = xi_37_end_mask_0, squeeze_mask = xi_37_squeeze_mask_0, x = pairs_37_cast_fp16)[name = tensor("xi_37_cast_fp16")]; + tensor var_2482_cast_fp16 = mul(x = xr_37_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_2482_cast_fp16")]; + tensor var_2483_cast_fp16 = mul(x = xi_37_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_2483_cast_fp16")]; + tensor xor__37_cast_fp16 = sub(x = var_2482_cast_fp16, y = var_2483_cast_fp16)[name = tensor("xor__37_cast_fp16")]; + tensor var_2486_cast_fp16 = mul(x = xr_37_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_2486_cast_fp16")]; + tensor var_2487_cast_fp16 = mul(x = xi_37_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_2487_cast_fp16")]; + tensor xoi_37_cast_fp16 = add(x = var_2486_cast_fp16, y = var_2487_cast_fp16)[name = tensor("xoi_37_cast_fp16")]; + tensor out_37_axis_0 = const()[name = tensor("out_37_axis_0"), val = tensor(-1)]; + tensor out_37_cast_fp16 = stack(axis = out_37_axis_0, values = (xor__37_cast_fp16, xoi_37_cast_fp16))[name = tensor("out_37_cast_fp16")]; + tensor var_2510 = const()[name = tensor("op_2510"), val = tensor([1, 16, 32, 2])]; + tensor pairs_39_cast_fp16 = reshape(shape = var_2510, x = x_39_cast_fp16)[name = tensor("pairs_39_cast_fp16")]; + tensor xr_39_begin_0 = const()[name = tensor("xr_39_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_39_end_0 = const()[name = tensor("xr_39_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_39_end_mask_0 = const()[name = tensor("xr_39_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_39_squeeze_mask_0 = const()[name = tensor("xr_39_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_39_cast_fp16 = slice_by_index(begin = xr_39_begin_0, end = xr_39_end_0, end_mask = xr_39_end_mask_0, squeeze_mask = xr_39_squeeze_mask_0, x = pairs_39_cast_fp16)[name = tensor("xr_39_cast_fp16")]; + tensor xi_39_begin_0 = const()[name = tensor("xi_39_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_39_end_0 = const()[name = tensor("xi_39_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_39_end_mask_0 = const()[name = tensor("xi_39_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_39_squeeze_mask_0 = const()[name = tensor("xi_39_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_39_cast_fp16 = slice_by_index(begin = xi_39_begin_0, end = xi_39_end_0, end_mask = xi_39_end_mask_0, squeeze_mask = xi_39_squeeze_mask_0, x = pairs_39_cast_fp16)[name = tensor("xi_39_cast_fp16")]; + tensor var_2518_cast_fp16 = mul(x = xr_39_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_2518_cast_fp16")]; + tensor var_2519_cast_fp16 = mul(x = xi_39_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_2519_cast_fp16")]; + tensor xor__39_cast_fp16 = sub(x = var_2518_cast_fp16, y = var_2519_cast_fp16)[name = tensor("xor__39_cast_fp16")]; + tensor var_2522_cast_fp16 = mul(x = xr_39_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_2522_cast_fp16")]; + tensor var_2523_cast_fp16 = mul(x = xi_39_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_2523_cast_fp16")]; + tensor xoi_39_cast_fp16 = add(x = var_2522_cast_fp16, y = var_2523_cast_fp16)[name = tensor("xoi_39_cast_fp16")]; + tensor out_39_axis_0 = const()[name = tensor("out_39_axis_0"), val = tensor(-1)]; + tensor out_39_cast_fp16 = stack(axis = out_39_axis_0, values = (xor__39_cast_fp16, xoi_39_cast_fp16))[name = tensor("out_39_cast_fp16")]; + tensor _inversed_2535_y_0_to_fp16 = const()[name = tensor("_inversed_2535_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2535_cast_fp16 = mul(x = position9_to_fp16, y = _inversed_2535_y_0_to_fp16)[name = tensor("_inversed_2535_cast_fp16")]; + tensor var_2536_cast_fp16 = floor(x = _inversed_2535_cast_fp16)[name = tensor("op_2536_cast_fp16")]; + tensor var_2537_to_fp16 = const()[name = tensor("op_2537_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2538_cast_fp16 = mul(x = var_2536_cast_fp16, y = var_2537_to_fp16)[name = tensor("op_2538_cast_fp16")]; + tensor write_idx_19_cast_fp16 = sub(x = position9_to_fp16, y = var_2538_cast_fp16)[name = tensor("write_idx_19_cast_fp16")]; + tensor var_2541_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_19_cast_fp16)[name = tensor("op_2541_cast_fp16")]; + tensor var_2551 = const()[name = tensor("op_2551"), val = tensor([1, 512, 1, 1])]; + tensor onehot_37_to_fp16_dtype_0 = const()[name = tensor("onehot_37_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_2541_cast_fp16_to_fp16 = cast(dtype = onehot_37_to_fp16_dtype_0, x = var_2541_cast_fp16)[name = tensor("cast_504")]; + tensor onehot_39_cast_fp16 = reshape(shape = var_2551, x = var_2541_cast_fp16_to_fp16)[name = tensor("onehot_39_cast_fp16")]; + tensor var_2553_to_fp16 = const()[name = tensor("op_2553_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2555_cast_fp16 = sub(x = var_2553_to_fp16, y = onehot_39_cast_fp16)[name = tensor("op_2555_cast_fp16")]; + tensor k_cache9_to_fp16_dtype_0 = const()[name = tensor("k_cache9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache9_to_fp16 = cast(dtype = k_cache9_to_fp16_dtype_0, x = k_cache9)[name = tensor("cast_503")]; + tensor var_2556_cast_fp16 = mul(x = k_cache9_to_fp16, y = var_2555_cast_fp16)[name = tensor("op_2556_cast_fp16")]; + tensor var_2561 = const()[name = tensor("op_2561"), val = tensor([1, 1, 16, 64])]; + tensor var_2562_cast_fp16 = reshape(shape = var_2561, x = out_39_cast_fp16)[name = tensor("op_2562_cast_fp16")]; + tensor var_2563_cast_fp16 = mul(x = var_2562_cast_fp16, y = onehot_39_cast_fp16)[name = tensor("op_2563_cast_fp16")]; + tensor new_k_cache_19_cast_fp16 = add(x = var_2556_cast_fp16, y = var_2563_cast_fp16)[name = tensor("new_k_cache_19_cast_fp16")]; + tensor new_k_cache_19_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_19_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache9_to_fp16_dtype_0 = const()[name = tensor("v_cache9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache9_to_fp16 = cast(dtype = v_cache9_to_fp16_dtype_0, x = v_cache9)[name = tensor("cast_501")]; + tensor var_2569_cast_fp16 = mul(x = v_cache9_to_fp16, y = var_2555_cast_fp16)[name = tensor("op_2569_cast_fp16")]; + tensor var_2574 = const()[name = tensor("op_2574"), val = tensor([1, 1, 16, 64])]; + tensor var_2575_cast_fp16 = reshape(shape = var_2574, x = v_19_cast_fp16)[name = tensor("op_2575_cast_fp16")]; + tensor var_2576_cast_fp16 = mul(x = var_2575_cast_fp16, y = onehot_39_cast_fp16)[name = tensor("op_2576_cast_fp16")]; + tensor new_v_cache_19_cast_fp16 = add(x = var_2569_cast_fp16, y = var_2576_cast_fp16)[name = tensor("new_v_cache_19_cast_fp16")]; + tensor new_v_cache_19_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_19_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_9_perm_0 = const()[name = tensor("transpose_9_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_2589 = const()[name = tensor("op_2589"), val = tensor([0, 2, 1, 3])]; + tensor var_2595 = const()[name = tensor("op_2595"), val = tensor([1, 16, 1, 64])]; + tensor q4_19_cast_fp16 = reshape(shape = var_2595, x = out_37_cast_fp16)[name = tensor("q4_19_cast_fp16")]; + tensor var_2600_transpose_x_0 = const()[name = tensor("op_2600_transpose_x_0"), val = tensor(false)]; + tensor var_2600_transpose_y_0 = const()[name = tensor("op_2600_transpose_y_0"), val = tensor(false)]; + tensor transpose_9_cast_fp16 = transpose(perm = transpose_9_perm_0, x = new_k_cache_19_cast_fp16)[name = tensor("transpose_68")]; + tensor var_2600_cast_fp16 = matmul(transpose_x = var_2600_transpose_x_0, transpose_y = var_2600_transpose_y_0, x = q4_19_cast_fp16, y = transpose_9_cast_fp16)[name = tensor("op_2600_cast_fp16")]; + tensor var_2601_to_fp16 = const()[name = tensor("op_2601_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_37_cast_fp16 = mul(x = var_2600_cast_fp16, y = var_2601_to_fp16)[name = tensor("scores_37_cast_fp16")]; + tensor var_2605 = const()[name = tensor("op_2605"), val = tensor([1, 1])]; + tensor var_2606_cast_fp16 = reshape(shape = var_2605, x = position9_to_fp16)[name = tensor("op_2606_cast_fp16")]; + tensor var_2607_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_2606_cast_fp16)[name = tensor("op_2607_cast_fp16")]; + tensor mask_19_to_fp16_dtype_0 = const()[name = tensor("mask_19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_2614_to_fp16 = const()[name = tensor("op_2614_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2607_cast_fp16_to_fp16 = cast(dtype = mask_19_to_fp16_dtype_0, x = var_2607_cast_fp16)[name = tensor("cast_499")]; + tensor var_2615_cast_fp16 = sub(x = var_2607_cast_fp16_to_fp16, y = var_2614_to_fp16)[name = tensor("op_2615_cast_fp16")]; + tensor var_2620 = const()[name = tensor("op_2620"), val = tensor([1, 1, 1, 512])]; + tensor var_2621_cast_fp16 = reshape(shape = var_2620, x = var_2615_cast_fp16)[name = tensor("op_2621_cast_fp16")]; + tensor var_2622_to_fp16 = const()[name = tensor("op_2622_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_2623_cast_fp16 = mul(x = var_2621_cast_fp16, y = var_2622_to_fp16)[name = tensor("op_2623_cast_fp16")]; + tensor scores_39_cast_fp16 = add(x = scores_37_cast_fp16, y = var_2623_cast_fp16)[name = tensor("scores_39_cast_fp16")]; + tensor var_2626 = const()[name = tensor("op_2626"), val = tensor(-1)]; + tensor weights_19_cast_fp16 = softmax(axis = var_2626, x = scores_39_cast_fp16)[name = tensor("weights_19_cast_fp16")]; + tensor attn_19_transpose_x_0 = const()[name = tensor("attn_19_transpose_x_0"), val = tensor(false)]; + tensor attn_19_transpose_y_0 = const()[name = tensor("attn_19_transpose_y_0"), val = tensor(false)]; + tensor values_19_cast_fp16 = transpose(perm = var_2589, x = new_v_cache_19_cast_fp16)[name = tensor("transpose_67")]; + tensor attn_19_cast_fp16 = matmul(transpose_x = attn_19_transpose_x_0, transpose_y = attn_19_transpose_y_0, x = weights_19_cast_fp16, y = values_19_cast_fp16)[name = tensor("attn_19_cast_fp16")]; + tensor var_2634 = const()[name = tensor("op_2634"), val = tensor([0, 2, 1, 3])]; + tensor var_2639 = const()[name = tensor("op_2639"), val = tensor([1, 1, 1024])]; + tensor var_2635_cast_fp16 = transpose(perm = var_2634, x = attn_19_cast_fp16)[name = tensor("transpose_66")]; + tensor input_115_cast_fp16 = reshape(shape = var_2639, x = var_2635_cast_fp16)[name = tensor("input_115_cast_fp16")]; + tensor attn9_out_proj_weight_to_fp16 = const()[name = tensor("attn9_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(232949952)))]; + tensor linear_38_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn9_out_proj_weight_to_fp16, x = input_115_cast_fp16)[name = tensor("linear_38_cast_fp16")]; + tensor var_2645_to_fp16 = const()[name = tensor("op_2645_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2646_cast_fp16 = add(x = position9_to_fp16, y = var_2645_to_fp16)[name = tensor("op_2646_cast_fp16")]; + tensor var_2646_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2646_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_117_cast_fp16 = add(x = input_111_cast_fp16, y = linear_38_cast_fp16)[name = tensor("input_117_cast_fp16")]; + tensor input_119_axes_0 = const()[name = tensor("input_119_axes_0"), val = tensor([-1])]; + tensor norm9_2_weight_to_fp16 = const()[name = tensor("norm9_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(235047168)))]; + tensor norm9_2_bias_to_fp16 = const()[name = tensor("norm9_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(235049280)))]; + tensor var_2650_to_fp16 = const()[name = tensor("op_2650_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_119_cast_fp16 = layer_norm(axes = input_119_axes_0, beta = norm9_2_bias_to_fp16, epsilon = var_2650_to_fp16, gamma = norm9_2_weight_to_fp16, x = input_117_cast_fp16)[name = tensor("input_119_cast_fp16")]; + tensor linear9_1_weight_to_fp16 = const()[name = tensor("linear9_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(235051392)))]; + tensor linear_39_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear9_1_weight_to_fp16, x = input_119_cast_fp16)[name = tensor("linear_39_cast_fp16")]; + tensor input_121_mode_0 = const()[name = tensor("input_121_mode_0"), val = tensor("EXACT")]; + tensor input_121_cast_fp16 = gelu(mode = input_121_mode_0, x = linear_39_cast_fp16)[name = tensor("input_121_cast_fp16")]; + tensor linear9_2_weight_to_fp16 = const()[name = tensor("linear9_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(243440064)))]; + tensor linear_40_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear9_2_weight_to_fp16, x = input_121_cast_fp16)[name = tensor("linear_40_cast_fp16")]; + tensor input_123_cast_fp16 = add(x = input_117_cast_fp16, y = linear_40_cast_fp16)[name = tensor("input_123_cast_fp16")]; + tensor input_125_axes_0 = const()[name = tensor("input_125_axes_0"), val = tensor([-1])]; + tensor norm10_1_weight_to_fp16 = const()[name = tensor("norm10_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(251828736)))]; + tensor norm10_1_bias_to_fp16 = const()[name = tensor("norm10_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(251830848)))]; + tensor var_2667_to_fp16 = const()[name = tensor("op_2667_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_125_cast_fp16 = layer_norm(axes = input_125_axes_0, beta = norm10_1_bias_to_fp16, epsilon = var_2667_to_fp16, gamma = norm10_1_weight_to_fp16, x = input_123_cast_fp16)[name = tensor("input_125_cast_fp16")]; + tensor attn10_in_proj_weight_to_fp16 = const()[name = tensor("attn10_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(251832960)))]; + tensor linear_41_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn10_in_proj_weight_to_fp16, x = input_125_cast_fp16)[name = tensor("linear_41_cast_fp16")]; + tensor var_2680 = const()[name = tensor("op_2680"), val = tensor([1, 3, 16, 64])]; + tensor qkv_21_cast_fp16 = reshape(shape = var_2680, x = linear_41_cast_fp16)[name = tensor("qkv_21_cast_fp16")]; + tensor x_41_begin_0 = const()[name = tensor("x_41_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_41_end_0 = const()[name = tensor("x_41_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_41_end_mask_0 = const()[name = tensor("x_41_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_41_squeeze_mask_0 = const()[name = tensor("x_41_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_41_cast_fp16 = slice_by_index(begin = x_41_begin_0, end = x_41_end_0, end_mask = x_41_end_mask_0, squeeze_mask = x_41_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("x_41_cast_fp16")]; + tensor x_43_begin_0 = const()[name = tensor("x_43_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_43_end_0 = const()[name = tensor("x_43_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_43_end_mask_0 = const()[name = tensor("x_43_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_43_squeeze_mask_0 = const()[name = tensor("x_43_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_43_cast_fp16 = slice_by_index(begin = x_43_begin_0, end = x_43_end_0, end_mask = x_43_end_mask_0, squeeze_mask = x_43_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("x_43_cast_fp16")]; + tensor v_21_begin_0 = const()[name = tensor("v_21_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_21_end_0 = const()[name = tensor("v_21_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_21_end_mask_0 = const()[name = tensor("v_21_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_21_squeeze_mask_0 = const()[name = tensor("v_21_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_21_cast_fp16 = slice_by_index(begin = v_21_begin_0, end = v_21_end_0, end_mask = v_21_end_mask_0, squeeze_mask = v_21_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("v_21_cast_fp16")]; + tensor var_2709 = const()[name = tensor("op_2709"), val = tensor([1, 1, 1])]; + tensor position10_to_fp16_dtype_0 = const()[name = tensor("position10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position10_to_fp16 = cast(dtype = position10_to_fp16_dtype_0, x = position10)[name = tensor("cast_497")]; + tensor var_2710_cast_fp16 = reshape(shape = var_2709, x = position10_to_fp16)[name = tensor("op_2710_cast_fp16")]; + tensor angles_41_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_2710_cast_fp16)[name = tensor("angles_41_cast_fp16")]; + tensor rotr_41_cast_fp16 = cos(x = angles_41_cast_fp16)[name = tensor("rotr_41_cast_fp16")]; + tensor roti_41_cast_fp16 = sin(x = angles_41_cast_fp16)[name = tensor("roti_41_cast_fp16")]; + tensor var_2718 = const()[name = tensor("op_2718"), val = tensor([1, 16, 32, 2])]; + tensor pairs_41_cast_fp16 = reshape(shape = var_2718, x = x_41_cast_fp16)[name = tensor("pairs_41_cast_fp16")]; + tensor xr_41_begin_0 = const()[name = tensor("xr_41_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_41_end_0 = const()[name = tensor("xr_41_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_41_end_mask_0 = const()[name = tensor("xr_41_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_41_squeeze_mask_0 = const()[name = tensor("xr_41_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_41_cast_fp16 = slice_by_index(begin = xr_41_begin_0, end = xr_41_end_0, end_mask = xr_41_end_mask_0, squeeze_mask = xr_41_squeeze_mask_0, x = pairs_41_cast_fp16)[name = tensor("xr_41_cast_fp16")]; + tensor xi_41_begin_0 = const()[name = tensor("xi_41_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_41_end_0 = const()[name = tensor("xi_41_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_41_end_mask_0 = const()[name = tensor("xi_41_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_41_squeeze_mask_0 = const()[name = tensor("xi_41_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_41_cast_fp16 = slice_by_index(begin = xi_41_begin_0, end = xi_41_end_0, end_mask = xi_41_end_mask_0, squeeze_mask = xi_41_squeeze_mask_0, x = pairs_41_cast_fp16)[name = tensor("xi_41_cast_fp16")]; + tensor var_2726_cast_fp16 = mul(x = xr_41_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_2726_cast_fp16")]; + tensor var_2727_cast_fp16 = mul(x = xi_41_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_2727_cast_fp16")]; + tensor xor__41_cast_fp16 = sub(x = var_2726_cast_fp16, y = var_2727_cast_fp16)[name = tensor("xor__41_cast_fp16")]; + tensor var_2730_cast_fp16 = mul(x = xr_41_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_2730_cast_fp16")]; + tensor var_2731_cast_fp16 = mul(x = xi_41_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_2731_cast_fp16")]; + tensor xoi_41_cast_fp16 = add(x = var_2730_cast_fp16, y = var_2731_cast_fp16)[name = tensor("xoi_41_cast_fp16")]; + tensor out_41_axis_0 = const()[name = tensor("out_41_axis_0"), val = tensor(-1)]; + tensor out_41_cast_fp16 = stack(axis = out_41_axis_0, values = (xor__41_cast_fp16, xoi_41_cast_fp16))[name = tensor("out_41_cast_fp16")]; + tensor var_2754 = const()[name = tensor("op_2754"), val = tensor([1, 16, 32, 2])]; + tensor pairs_43_cast_fp16 = reshape(shape = var_2754, x = x_43_cast_fp16)[name = tensor("pairs_43_cast_fp16")]; + tensor xr_43_begin_0 = const()[name = tensor("xr_43_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_43_end_0 = const()[name = tensor("xr_43_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_43_end_mask_0 = const()[name = tensor("xr_43_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_43_squeeze_mask_0 = const()[name = tensor("xr_43_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_43_cast_fp16 = slice_by_index(begin = xr_43_begin_0, end = xr_43_end_0, end_mask = xr_43_end_mask_0, squeeze_mask = xr_43_squeeze_mask_0, x = pairs_43_cast_fp16)[name = tensor("xr_43_cast_fp16")]; + tensor xi_43_begin_0 = const()[name = tensor("xi_43_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_43_end_0 = const()[name = tensor("xi_43_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_43_end_mask_0 = const()[name = tensor("xi_43_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_43_squeeze_mask_0 = const()[name = tensor("xi_43_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_43_cast_fp16 = slice_by_index(begin = xi_43_begin_0, end = xi_43_end_0, end_mask = xi_43_end_mask_0, squeeze_mask = xi_43_squeeze_mask_0, x = pairs_43_cast_fp16)[name = tensor("xi_43_cast_fp16")]; + tensor var_2762_cast_fp16 = mul(x = xr_43_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_2762_cast_fp16")]; + tensor var_2763_cast_fp16 = mul(x = xi_43_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_2763_cast_fp16")]; + tensor xor__43_cast_fp16 = sub(x = var_2762_cast_fp16, y = var_2763_cast_fp16)[name = tensor("xor__43_cast_fp16")]; + tensor var_2766_cast_fp16 = mul(x = xr_43_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_2766_cast_fp16")]; + tensor var_2767_cast_fp16 = mul(x = xi_43_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_2767_cast_fp16")]; + tensor xoi_43_cast_fp16 = add(x = var_2766_cast_fp16, y = var_2767_cast_fp16)[name = tensor("xoi_43_cast_fp16")]; + tensor out_43_axis_0 = const()[name = tensor("out_43_axis_0"), val = tensor(-1)]; + tensor out_43_cast_fp16 = stack(axis = out_43_axis_0, values = (xor__43_cast_fp16, xoi_43_cast_fp16))[name = tensor("out_43_cast_fp16")]; + tensor _inversed_2779_y_0_to_fp16 = const()[name = tensor("_inversed_2779_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2779_cast_fp16 = mul(x = position10_to_fp16, y = _inversed_2779_y_0_to_fp16)[name = tensor("_inversed_2779_cast_fp16")]; + tensor var_2780_cast_fp16 = floor(x = _inversed_2779_cast_fp16)[name = tensor("op_2780_cast_fp16")]; + tensor var_2781_to_fp16 = const()[name = tensor("op_2781_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2782_cast_fp16 = mul(x = var_2780_cast_fp16, y = var_2781_to_fp16)[name = tensor("op_2782_cast_fp16")]; + tensor write_idx_21_cast_fp16 = sub(x = position10_to_fp16, y = var_2782_cast_fp16)[name = tensor("write_idx_21_cast_fp16")]; + tensor var_2785_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_21_cast_fp16)[name = tensor("op_2785_cast_fp16")]; + tensor var_2795 = const()[name = tensor("op_2795"), val = tensor([1, 512, 1, 1])]; + tensor onehot_41_to_fp16_dtype_0 = const()[name = tensor("onehot_41_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_2785_cast_fp16_to_fp16 = cast(dtype = onehot_41_to_fp16_dtype_0, x = var_2785_cast_fp16)[name = tensor("cast_496")]; + tensor onehot_43_cast_fp16 = reshape(shape = var_2795, x = var_2785_cast_fp16_to_fp16)[name = tensor("onehot_43_cast_fp16")]; + tensor var_2797_to_fp16 = const()[name = tensor("op_2797_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2799_cast_fp16 = sub(x = var_2797_to_fp16, y = onehot_43_cast_fp16)[name = tensor("op_2799_cast_fp16")]; + tensor k_cache10_to_fp16_dtype_0 = const()[name = tensor("k_cache10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache10_to_fp16 = cast(dtype = k_cache10_to_fp16_dtype_0, x = k_cache10)[name = tensor("cast_495")]; + tensor var_2800_cast_fp16 = mul(x = k_cache10_to_fp16, y = var_2799_cast_fp16)[name = tensor("op_2800_cast_fp16")]; + tensor var_2805 = const()[name = tensor("op_2805"), val = tensor([1, 1, 16, 64])]; + tensor var_2806_cast_fp16 = reshape(shape = var_2805, x = out_43_cast_fp16)[name = tensor("op_2806_cast_fp16")]; + tensor var_2807_cast_fp16 = mul(x = var_2806_cast_fp16, y = onehot_43_cast_fp16)[name = tensor("op_2807_cast_fp16")]; + tensor new_k_cache_21_cast_fp16 = add(x = var_2800_cast_fp16, y = var_2807_cast_fp16)[name = tensor("new_k_cache_21_cast_fp16")]; + tensor new_k_cache_21_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_21_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache10_to_fp16_dtype_0 = const()[name = tensor("v_cache10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache10_to_fp16 = cast(dtype = v_cache10_to_fp16_dtype_0, x = v_cache10)[name = tensor("cast_493")]; + tensor var_2813_cast_fp16 = mul(x = v_cache10_to_fp16, y = var_2799_cast_fp16)[name = tensor("op_2813_cast_fp16")]; + tensor var_2818 = const()[name = tensor("op_2818"), val = tensor([1, 1, 16, 64])]; + tensor var_2819_cast_fp16 = reshape(shape = var_2818, x = v_21_cast_fp16)[name = tensor("op_2819_cast_fp16")]; + tensor var_2820_cast_fp16 = mul(x = var_2819_cast_fp16, y = onehot_43_cast_fp16)[name = tensor("op_2820_cast_fp16")]; + tensor new_v_cache_21_cast_fp16 = add(x = var_2813_cast_fp16, y = var_2820_cast_fp16)[name = tensor("new_v_cache_21_cast_fp16")]; + tensor new_v_cache_21_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_21_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_10_perm_0 = const()[name = tensor("transpose_10_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_2833 = const()[name = tensor("op_2833"), val = tensor([0, 2, 1, 3])]; + tensor var_2839 = const()[name = tensor("op_2839"), val = tensor([1, 16, 1, 64])]; + tensor q4_21_cast_fp16 = reshape(shape = var_2839, x = out_41_cast_fp16)[name = tensor("q4_21_cast_fp16")]; + tensor var_2844_transpose_x_0 = const()[name = tensor("op_2844_transpose_x_0"), val = tensor(false)]; + tensor var_2844_transpose_y_0 = const()[name = tensor("op_2844_transpose_y_0"), val = tensor(false)]; + tensor transpose_10_cast_fp16 = transpose(perm = transpose_10_perm_0, x = new_k_cache_21_cast_fp16)[name = tensor("transpose_65")]; + tensor var_2844_cast_fp16 = matmul(transpose_x = var_2844_transpose_x_0, transpose_y = var_2844_transpose_y_0, x = q4_21_cast_fp16, y = transpose_10_cast_fp16)[name = tensor("op_2844_cast_fp16")]; + tensor var_2845_to_fp16 = const()[name = tensor("op_2845_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_41_cast_fp16 = mul(x = var_2844_cast_fp16, y = var_2845_to_fp16)[name = tensor("scores_41_cast_fp16")]; + tensor var_2849 = const()[name = tensor("op_2849"), val = tensor([1, 1])]; + tensor var_2850_cast_fp16 = reshape(shape = var_2849, x = position10_to_fp16)[name = tensor("op_2850_cast_fp16")]; + tensor var_2851_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_2850_cast_fp16)[name = tensor("op_2851_cast_fp16")]; + tensor mask_21_to_fp16_dtype_0 = const()[name = tensor("mask_21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_2858_to_fp16 = const()[name = tensor("op_2858_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2851_cast_fp16_to_fp16 = cast(dtype = mask_21_to_fp16_dtype_0, x = var_2851_cast_fp16)[name = tensor("cast_491")]; + tensor var_2859_cast_fp16 = sub(x = var_2851_cast_fp16_to_fp16, y = var_2858_to_fp16)[name = tensor("op_2859_cast_fp16")]; + tensor var_2864 = const()[name = tensor("op_2864"), val = tensor([1, 1, 1, 512])]; + tensor var_2865_cast_fp16 = reshape(shape = var_2864, x = var_2859_cast_fp16)[name = tensor("op_2865_cast_fp16")]; + tensor var_2866_to_fp16 = const()[name = tensor("op_2866_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_2867_cast_fp16 = mul(x = var_2865_cast_fp16, y = var_2866_to_fp16)[name = tensor("op_2867_cast_fp16")]; + tensor scores_43_cast_fp16 = add(x = scores_41_cast_fp16, y = var_2867_cast_fp16)[name = tensor("scores_43_cast_fp16")]; + tensor var_2870 = const()[name = tensor("op_2870"), val = tensor(-1)]; + tensor weights_21_cast_fp16 = softmax(axis = var_2870, x = scores_43_cast_fp16)[name = tensor("weights_21_cast_fp16")]; + tensor attn_21_transpose_x_0 = const()[name = tensor("attn_21_transpose_x_0"), val = tensor(false)]; + tensor attn_21_transpose_y_0 = const()[name = tensor("attn_21_transpose_y_0"), val = tensor(false)]; + tensor values_21_cast_fp16 = transpose(perm = var_2833, x = new_v_cache_21_cast_fp16)[name = tensor("transpose_64")]; + tensor attn_21_cast_fp16 = matmul(transpose_x = attn_21_transpose_x_0, transpose_y = attn_21_transpose_y_0, x = weights_21_cast_fp16, y = values_21_cast_fp16)[name = tensor("attn_21_cast_fp16")]; + tensor var_2878 = const()[name = tensor("op_2878"), val = tensor([0, 2, 1, 3])]; + tensor var_2883 = const()[name = tensor("op_2883"), val = tensor([1, 1, 1024])]; + tensor var_2879_cast_fp16 = transpose(perm = var_2878, x = attn_21_cast_fp16)[name = tensor("transpose_63")]; + tensor input_127_cast_fp16 = reshape(shape = var_2883, x = var_2879_cast_fp16)[name = tensor("input_127_cast_fp16")]; + tensor attn10_out_proj_weight_to_fp16 = const()[name = tensor("attn10_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(258124480)))]; + tensor linear_42_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn10_out_proj_weight_to_fp16, x = input_127_cast_fp16)[name = tensor("linear_42_cast_fp16")]; + tensor var_2889_to_fp16 = const()[name = tensor("op_2889_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2890_cast_fp16 = add(x = position10_to_fp16, y = var_2889_to_fp16)[name = tensor("op_2890_cast_fp16")]; + tensor var_2890_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2890_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_129_cast_fp16 = add(x = input_123_cast_fp16, y = linear_42_cast_fp16)[name = tensor("input_129_cast_fp16")]; + tensor input_131_axes_0 = const()[name = tensor("input_131_axes_0"), val = tensor([-1])]; + tensor norm10_2_weight_to_fp16 = const()[name = tensor("norm10_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260221696)))]; + tensor norm10_2_bias_to_fp16 = const()[name = tensor("norm10_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260223808)))]; + tensor var_2894_to_fp16 = const()[name = tensor("op_2894_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_131_cast_fp16 = layer_norm(axes = input_131_axes_0, beta = norm10_2_bias_to_fp16, epsilon = var_2894_to_fp16, gamma = norm10_2_weight_to_fp16, x = input_129_cast_fp16)[name = tensor("input_131_cast_fp16")]; + tensor linear10_1_weight_to_fp16 = const()[name = tensor("linear10_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260225920)))]; + tensor linear_43_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear10_1_weight_to_fp16, x = input_131_cast_fp16)[name = tensor("linear_43_cast_fp16")]; + tensor input_133_mode_0 = const()[name = tensor("input_133_mode_0"), val = tensor("EXACT")]; + tensor input_133_cast_fp16 = gelu(mode = input_133_mode_0, x = linear_43_cast_fp16)[name = tensor("input_133_cast_fp16")]; + tensor linear10_2_weight_to_fp16 = const()[name = tensor("linear10_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(268614592)))]; + tensor linear_44_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear10_2_weight_to_fp16, x = input_133_cast_fp16)[name = tensor("linear_44_cast_fp16")]; + tensor input_135_cast_fp16 = add(x = input_129_cast_fp16, y = linear_44_cast_fp16)[name = tensor("input_135_cast_fp16")]; + tensor input_137_axes_0 = const()[name = tensor("input_137_axes_0"), val = tensor([-1])]; + tensor norm11_1_weight_to_fp16 = const()[name = tensor("norm11_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277003264)))]; + tensor norm11_1_bias_to_fp16 = const()[name = tensor("norm11_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277005376)))]; + tensor var_2911_to_fp16 = const()[name = tensor("op_2911_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_137_cast_fp16 = layer_norm(axes = input_137_axes_0, beta = norm11_1_bias_to_fp16, epsilon = var_2911_to_fp16, gamma = norm11_1_weight_to_fp16, x = input_135_cast_fp16)[name = tensor("input_137_cast_fp16")]; + tensor attn11_in_proj_weight_to_fp16 = const()[name = tensor("attn11_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277007488)))]; + tensor linear_45_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn11_in_proj_weight_to_fp16, x = input_137_cast_fp16)[name = tensor("linear_45_cast_fp16")]; + tensor var_2924 = const()[name = tensor("op_2924"), val = tensor([1, 3, 16, 64])]; + tensor qkv_23_cast_fp16 = reshape(shape = var_2924, x = linear_45_cast_fp16)[name = tensor("qkv_23_cast_fp16")]; + tensor x_45_begin_0 = const()[name = tensor("x_45_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_45_end_0 = const()[name = tensor("x_45_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_45_end_mask_0 = const()[name = tensor("x_45_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_45_squeeze_mask_0 = const()[name = tensor("x_45_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_45_cast_fp16 = slice_by_index(begin = x_45_begin_0, end = x_45_end_0, end_mask = x_45_end_mask_0, squeeze_mask = x_45_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("x_45_cast_fp16")]; + tensor x_47_begin_0 = const()[name = tensor("x_47_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_47_end_0 = const()[name = tensor("x_47_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_47_end_mask_0 = const()[name = tensor("x_47_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_47_squeeze_mask_0 = const()[name = tensor("x_47_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_47_cast_fp16 = slice_by_index(begin = x_47_begin_0, end = x_47_end_0, end_mask = x_47_end_mask_0, squeeze_mask = x_47_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("x_47_cast_fp16")]; + tensor v_23_begin_0 = const()[name = tensor("v_23_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_23_end_0 = const()[name = tensor("v_23_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_23_end_mask_0 = const()[name = tensor("v_23_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_23_squeeze_mask_0 = const()[name = tensor("v_23_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_23_cast_fp16 = slice_by_index(begin = v_23_begin_0, end = v_23_end_0, end_mask = v_23_end_mask_0, squeeze_mask = v_23_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("v_23_cast_fp16")]; + tensor var_2953 = const()[name = tensor("op_2953"), val = tensor([1, 1, 1])]; + tensor position11_to_fp16_dtype_0 = const()[name = tensor("position11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position11_to_fp16 = cast(dtype = position11_to_fp16_dtype_0, x = position11)[name = tensor("cast_489")]; + tensor var_2954_cast_fp16 = reshape(shape = var_2953, x = position11_to_fp16)[name = tensor("op_2954_cast_fp16")]; + tensor angles_45_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_2954_cast_fp16)[name = tensor("angles_45_cast_fp16")]; + tensor rotr_45_cast_fp16 = cos(x = angles_45_cast_fp16)[name = tensor("rotr_45_cast_fp16")]; + tensor roti_45_cast_fp16 = sin(x = angles_45_cast_fp16)[name = tensor("roti_45_cast_fp16")]; + tensor var_2962 = const()[name = tensor("op_2962"), val = tensor([1, 16, 32, 2])]; + tensor pairs_45_cast_fp16 = reshape(shape = var_2962, x = x_45_cast_fp16)[name = tensor("pairs_45_cast_fp16")]; + tensor xr_45_begin_0 = const()[name = tensor("xr_45_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_45_end_0 = const()[name = tensor("xr_45_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_45_end_mask_0 = const()[name = tensor("xr_45_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_45_squeeze_mask_0 = const()[name = tensor("xr_45_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_45_cast_fp16 = slice_by_index(begin = xr_45_begin_0, end = xr_45_end_0, end_mask = xr_45_end_mask_0, squeeze_mask = xr_45_squeeze_mask_0, x = pairs_45_cast_fp16)[name = tensor("xr_45_cast_fp16")]; + tensor xi_45_begin_0 = const()[name = tensor("xi_45_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_45_end_0 = const()[name = tensor("xi_45_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_45_end_mask_0 = const()[name = tensor("xi_45_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_45_squeeze_mask_0 = const()[name = tensor("xi_45_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_45_cast_fp16 = slice_by_index(begin = xi_45_begin_0, end = xi_45_end_0, end_mask = xi_45_end_mask_0, squeeze_mask = xi_45_squeeze_mask_0, x = pairs_45_cast_fp16)[name = tensor("xi_45_cast_fp16")]; + tensor var_2970_cast_fp16 = mul(x = xr_45_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_2970_cast_fp16")]; + tensor var_2971_cast_fp16 = mul(x = xi_45_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_2971_cast_fp16")]; + tensor xor__45_cast_fp16 = sub(x = var_2970_cast_fp16, y = var_2971_cast_fp16)[name = tensor("xor__45_cast_fp16")]; + tensor var_2974_cast_fp16 = mul(x = xr_45_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_2974_cast_fp16")]; + tensor var_2975_cast_fp16 = mul(x = xi_45_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_2975_cast_fp16")]; + tensor xoi_45_cast_fp16 = add(x = var_2974_cast_fp16, y = var_2975_cast_fp16)[name = tensor("xoi_45_cast_fp16")]; + tensor out_45_axis_0 = const()[name = tensor("out_45_axis_0"), val = tensor(-1)]; + tensor out_45_cast_fp16 = stack(axis = out_45_axis_0, values = (xor__45_cast_fp16, xoi_45_cast_fp16))[name = tensor("out_45_cast_fp16")]; + tensor var_2998 = const()[name = tensor("op_2998"), val = tensor([1, 16, 32, 2])]; + tensor pairs_47_cast_fp16 = reshape(shape = var_2998, x = x_47_cast_fp16)[name = tensor("pairs_47_cast_fp16")]; + tensor xr_47_begin_0 = const()[name = tensor("xr_47_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_47_end_0 = const()[name = tensor("xr_47_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_47_end_mask_0 = const()[name = tensor("xr_47_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_47_squeeze_mask_0 = const()[name = tensor("xr_47_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_47_cast_fp16 = slice_by_index(begin = xr_47_begin_0, end = xr_47_end_0, end_mask = xr_47_end_mask_0, squeeze_mask = xr_47_squeeze_mask_0, x = pairs_47_cast_fp16)[name = tensor("xr_47_cast_fp16")]; + tensor xi_47_begin_0 = const()[name = tensor("xi_47_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_47_end_0 = const()[name = tensor("xi_47_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_47_end_mask_0 = const()[name = tensor("xi_47_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_47_squeeze_mask_0 = const()[name = tensor("xi_47_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_47_cast_fp16 = slice_by_index(begin = xi_47_begin_0, end = xi_47_end_0, end_mask = xi_47_end_mask_0, squeeze_mask = xi_47_squeeze_mask_0, x = pairs_47_cast_fp16)[name = tensor("xi_47_cast_fp16")]; + tensor var_3006_cast_fp16 = mul(x = xr_47_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_3006_cast_fp16")]; + tensor var_3007_cast_fp16 = mul(x = xi_47_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_3007_cast_fp16")]; + tensor xor__47_cast_fp16 = sub(x = var_3006_cast_fp16, y = var_3007_cast_fp16)[name = tensor("xor__47_cast_fp16")]; + tensor var_3010_cast_fp16 = mul(x = xr_47_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_3010_cast_fp16")]; + tensor var_3011_cast_fp16 = mul(x = xi_47_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_3011_cast_fp16")]; + tensor xoi_47_cast_fp16 = add(x = var_3010_cast_fp16, y = var_3011_cast_fp16)[name = tensor("xoi_47_cast_fp16")]; + tensor out_47_axis_0 = const()[name = tensor("out_47_axis_0"), val = tensor(-1)]; + tensor out_47_cast_fp16 = stack(axis = out_47_axis_0, values = (xor__47_cast_fp16, xoi_47_cast_fp16))[name = tensor("out_47_cast_fp16")]; + tensor _inversed_3023_y_0_to_fp16 = const()[name = tensor("_inversed_3023_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3023_cast_fp16 = mul(x = position11_to_fp16, y = _inversed_3023_y_0_to_fp16)[name = tensor("_inversed_3023_cast_fp16")]; + tensor var_3024_cast_fp16 = floor(x = _inversed_3023_cast_fp16)[name = tensor("op_3024_cast_fp16")]; + tensor var_3025_to_fp16 = const()[name = tensor("op_3025_to_fp16"), val = tensor(0x1p+9)]; + tensor var_3026_cast_fp16 = mul(x = var_3024_cast_fp16, y = var_3025_to_fp16)[name = tensor("op_3026_cast_fp16")]; + tensor write_idx_23_cast_fp16 = sub(x = position11_to_fp16, y = var_3026_cast_fp16)[name = tensor("write_idx_23_cast_fp16")]; + tensor var_3029_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_23_cast_fp16)[name = tensor("op_3029_cast_fp16")]; + tensor var_3039 = const()[name = tensor("op_3039"), val = tensor([1, 512, 1, 1])]; + tensor onehot_45_to_fp16_dtype_0 = const()[name = tensor("onehot_45_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_3029_cast_fp16_to_fp16 = cast(dtype = onehot_45_to_fp16_dtype_0, x = var_3029_cast_fp16)[name = tensor("cast_488")]; + tensor onehot_47_cast_fp16 = reshape(shape = var_3039, x = var_3029_cast_fp16_to_fp16)[name = tensor("onehot_47_cast_fp16")]; + tensor var_3041_to_fp16 = const()[name = tensor("op_3041_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3043_cast_fp16 = sub(x = var_3041_to_fp16, y = onehot_47_cast_fp16)[name = tensor("op_3043_cast_fp16")]; + tensor k_cache11_to_fp16_dtype_0 = const()[name = tensor("k_cache11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache11_to_fp16 = cast(dtype = k_cache11_to_fp16_dtype_0, x = k_cache11)[name = tensor("cast_487")]; + tensor var_3044_cast_fp16 = mul(x = k_cache11_to_fp16, y = var_3043_cast_fp16)[name = tensor("op_3044_cast_fp16")]; + tensor var_3049 = const()[name = tensor("op_3049"), val = tensor([1, 1, 16, 64])]; + tensor var_3050_cast_fp16 = reshape(shape = var_3049, x = out_47_cast_fp16)[name = tensor("op_3050_cast_fp16")]; + tensor var_3051_cast_fp16 = mul(x = var_3050_cast_fp16, y = onehot_47_cast_fp16)[name = tensor("op_3051_cast_fp16")]; + tensor new_k_cache_23_cast_fp16 = add(x = var_3044_cast_fp16, y = var_3051_cast_fp16)[name = tensor("new_k_cache_23_cast_fp16")]; + tensor new_k_cache_23_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_23_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache11_to_fp16_dtype_0 = const()[name = tensor("v_cache11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache11_to_fp16 = cast(dtype = v_cache11_to_fp16_dtype_0, x = v_cache11)[name = tensor("cast_485")]; + tensor var_3057_cast_fp16 = mul(x = v_cache11_to_fp16, y = var_3043_cast_fp16)[name = tensor("op_3057_cast_fp16")]; + tensor var_3062 = const()[name = tensor("op_3062"), val = tensor([1, 1, 16, 64])]; + tensor var_3063_cast_fp16 = reshape(shape = var_3062, x = v_23_cast_fp16)[name = tensor("op_3063_cast_fp16")]; + tensor var_3064_cast_fp16 = mul(x = var_3063_cast_fp16, y = onehot_47_cast_fp16)[name = tensor("op_3064_cast_fp16")]; + tensor new_v_cache_23_cast_fp16 = add(x = var_3057_cast_fp16, y = var_3064_cast_fp16)[name = tensor("new_v_cache_23_cast_fp16")]; + tensor new_v_cache_23_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_23_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_11_perm_0 = const()[name = tensor("transpose_11_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_3077 = const()[name = tensor("op_3077"), val = tensor([0, 2, 1, 3])]; + tensor var_3083 = const()[name = tensor("op_3083"), val = tensor([1, 16, 1, 64])]; + tensor q4_23_cast_fp16 = reshape(shape = var_3083, x = out_45_cast_fp16)[name = tensor("q4_23_cast_fp16")]; + tensor var_3088_transpose_x_0 = const()[name = tensor("op_3088_transpose_x_0"), val = tensor(false)]; + tensor var_3088_transpose_y_0 = const()[name = tensor("op_3088_transpose_y_0"), val = tensor(false)]; + tensor transpose_11_cast_fp16 = transpose(perm = transpose_11_perm_0, x = new_k_cache_23_cast_fp16)[name = tensor("transpose_62")]; + tensor var_3088_cast_fp16 = matmul(transpose_x = var_3088_transpose_x_0, transpose_y = var_3088_transpose_y_0, x = q4_23_cast_fp16, y = transpose_11_cast_fp16)[name = tensor("op_3088_cast_fp16")]; + tensor var_3089_to_fp16 = const()[name = tensor("op_3089_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_45_cast_fp16 = mul(x = var_3088_cast_fp16, y = var_3089_to_fp16)[name = tensor("scores_45_cast_fp16")]; + tensor var_3093 = const()[name = tensor("op_3093"), val = tensor([1, 1])]; + tensor var_3094_cast_fp16 = reshape(shape = var_3093, x = position11_to_fp16)[name = tensor("op_3094_cast_fp16")]; + tensor var_3095_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_3094_cast_fp16)[name = tensor("op_3095_cast_fp16")]; + tensor mask_23_to_fp16_dtype_0 = const()[name = tensor("mask_23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_3102_to_fp16 = const()[name = tensor("op_3102_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3095_cast_fp16_to_fp16 = cast(dtype = mask_23_to_fp16_dtype_0, x = var_3095_cast_fp16)[name = tensor("cast_483")]; + tensor var_3103_cast_fp16 = sub(x = var_3095_cast_fp16_to_fp16, y = var_3102_to_fp16)[name = tensor("op_3103_cast_fp16")]; + tensor var_3108 = const()[name = tensor("op_3108"), val = tensor([1, 1, 1, 512])]; + tensor var_3109_cast_fp16 = reshape(shape = var_3108, x = var_3103_cast_fp16)[name = tensor("op_3109_cast_fp16")]; + tensor var_3110_to_fp16 = const()[name = tensor("op_3110_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_3111_cast_fp16 = mul(x = var_3109_cast_fp16, y = var_3110_to_fp16)[name = tensor("op_3111_cast_fp16")]; + tensor scores_47_cast_fp16 = add(x = scores_45_cast_fp16, y = var_3111_cast_fp16)[name = tensor("scores_47_cast_fp16")]; + tensor var_3114 = const()[name = tensor("op_3114"), val = tensor(-1)]; + tensor weights_23_cast_fp16 = softmax(axis = var_3114, x = scores_47_cast_fp16)[name = tensor("weights_23_cast_fp16")]; + tensor attn_23_transpose_x_0 = const()[name = tensor("attn_23_transpose_x_0"), val = tensor(false)]; + tensor attn_23_transpose_y_0 = const()[name = tensor("attn_23_transpose_y_0"), val = tensor(false)]; + tensor values_23_cast_fp16 = transpose(perm = var_3077, x = new_v_cache_23_cast_fp16)[name = tensor("transpose_61")]; + tensor attn_23_cast_fp16 = matmul(transpose_x = attn_23_transpose_x_0, transpose_y = attn_23_transpose_y_0, x = weights_23_cast_fp16, y = values_23_cast_fp16)[name = tensor("attn_23_cast_fp16")]; + tensor var_3122 = const()[name = tensor("op_3122"), val = tensor([0, 2, 1, 3])]; + tensor var_3127 = const()[name = tensor("op_3127"), val = tensor([1, 1, 1024])]; + tensor var_3123_cast_fp16 = transpose(perm = var_3122, x = attn_23_cast_fp16)[name = tensor("transpose_60")]; + tensor input_139_cast_fp16 = reshape(shape = var_3127, x = var_3123_cast_fp16)[name = tensor("input_139_cast_fp16")]; + tensor attn11_out_proj_weight_to_fp16 = const()[name = tensor("attn11_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(283299008)))]; + tensor linear_46_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn11_out_proj_weight_to_fp16, x = input_139_cast_fp16)[name = tensor("linear_46_cast_fp16")]; + tensor var_3133_to_fp16 = const()[name = tensor("op_3133_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3134_cast_fp16 = add(x = position11_to_fp16, y = var_3133_to_fp16)[name = tensor("op_3134_cast_fp16")]; + tensor var_3134_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3134_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_141_cast_fp16 = add(x = input_135_cast_fp16, y = linear_46_cast_fp16)[name = tensor("input_141_cast_fp16")]; + tensor input_143_axes_0 = const()[name = tensor("input_143_axes_0"), val = tensor([-1])]; + tensor norm11_2_weight_to_fp16 = const()[name = tensor("norm11_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285396224)))]; + tensor norm11_2_bias_to_fp16 = const()[name = tensor("norm11_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285398336)))]; + tensor var_3138_to_fp16 = const()[name = tensor("op_3138_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_143_cast_fp16 = layer_norm(axes = input_143_axes_0, beta = norm11_2_bias_to_fp16, epsilon = var_3138_to_fp16, gamma = norm11_2_weight_to_fp16, x = input_141_cast_fp16)[name = tensor("input_143_cast_fp16")]; + tensor linear11_1_weight_to_fp16 = const()[name = tensor("linear11_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285400448)))]; + tensor linear_47_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear11_1_weight_to_fp16, x = input_143_cast_fp16)[name = tensor("linear_47_cast_fp16")]; + tensor input_145_mode_0 = const()[name = tensor("input_145_mode_0"), val = tensor("EXACT")]; + tensor input_145_cast_fp16 = gelu(mode = input_145_mode_0, x = linear_47_cast_fp16)[name = tensor("input_145_cast_fp16")]; + tensor linear11_2_weight_to_fp16 = const()[name = tensor("linear11_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293789120)))]; + tensor linear_48_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear11_2_weight_to_fp16, x = input_145_cast_fp16)[name = tensor("linear_48_cast_fp16")]; + tensor input_147_cast_fp16 = add(x = input_141_cast_fp16, y = linear_48_cast_fp16)[name = tensor("input_147_cast_fp16")]; + tensor input_149_axes_0 = const()[name = tensor("input_149_axes_0"), val = tensor([-1])]; + tensor norm12_1_weight_to_fp16 = const()[name = tensor("norm12_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302177792)))]; + tensor norm12_1_bias_to_fp16 = const()[name = tensor("norm12_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302179904)))]; + tensor var_3155_to_fp16 = const()[name = tensor("op_3155_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_149_cast_fp16 = layer_norm(axes = input_149_axes_0, beta = norm12_1_bias_to_fp16, epsilon = var_3155_to_fp16, gamma = norm12_1_weight_to_fp16, x = input_147_cast_fp16)[name = tensor("input_149_cast_fp16")]; + tensor attn12_in_proj_weight_to_fp16 = const()[name = tensor("attn12_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302182016)))]; + tensor linear_49_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn12_in_proj_weight_to_fp16, x = input_149_cast_fp16)[name = tensor("linear_49_cast_fp16")]; + tensor var_3168 = const()[name = tensor("op_3168"), val = tensor([1, 3, 16, 64])]; + tensor qkv_25_cast_fp16 = reshape(shape = var_3168, x = linear_49_cast_fp16)[name = tensor("qkv_25_cast_fp16")]; + tensor x_49_begin_0 = const()[name = tensor("x_49_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_49_end_0 = const()[name = tensor("x_49_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_49_end_mask_0 = const()[name = tensor("x_49_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_49_squeeze_mask_0 = const()[name = tensor("x_49_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_49_cast_fp16 = slice_by_index(begin = x_49_begin_0, end = x_49_end_0, end_mask = x_49_end_mask_0, squeeze_mask = x_49_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("x_49_cast_fp16")]; + tensor x_51_begin_0 = const()[name = tensor("x_51_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_51_end_0 = const()[name = tensor("x_51_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_51_end_mask_0 = const()[name = tensor("x_51_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_51_squeeze_mask_0 = const()[name = tensor("x_51_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_51_cast_fp16 = slice_by_index(begin = x_51_begin_0, end = x_51_end_0, end_mask = x_51_end_mask_0, squeeze_mask = x_51_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("x_51_cast_fp16")]; + tensor v_25_begin_0 = const()[name = tensor("v_25_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_25_end_0 = const()[name = tensor("v_25_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_25_end_mask_0 = const()[name = tensor("v_25_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_25_squeeze_mask_0 = const()[name = tensor("v_25_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_25_cast_fp16 = slice_by_index(begin = v_25_begin_0, end = v_25_end_0, end_mask = v_25_end_mask_0, squeeze_mask = v_25_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("v_25_cast_fp16")]; + tensor var_3197 = const()[name = tensor("op_3197"), val = tensor([1, 1, 1])]; + tensor position12_to_fp16_dtype_0 = const()[name = tensor("position12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position12_to_fp16 = cast(dtype = position12_to_fp16_dtype_0, x = position12)[name = tensor("cast_481")]; + tensor var_3198_cast_fp16 = reshape(shape = var_3197, x = position12_to_fp16)[name = tensor("op_3198_cast_fp16")]; + tensor angles_49_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_3198_cast_fp16)[name = tensor("angles_49_cast_fp16")]; + tensor rotr_49_cast_fp16 = cos(x = angles_49_cast_fp16)[name = tensor("rotr_49_cast_fp16")]; + tensor roti_49_cast_fp16 = sin(x = angles_49_cast_fp16)[name = tensor("roti_49_cast_fp16")]; + tensor var_3206 = const()[name = tensor("op_3206"), val = tensor([1, 16, 32, 2])]; + tensor pairs_49_cast_fp16 = reshape(shape = var_3206, x = x_49_cast_fp16)[name = tensor("pairs_49_cast_fp16")]; + tensor xr_49_begin_0 = const()[name = tensor("xr_49_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_49_end_0 = const()[name = tensor("xr_49_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_49_end_mask_0 = const()[name = tensor("xr_49_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_49_squeeze_mask_0 = const()[name = tensor("xr_49_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_49_cast_fp16 = slice_by_index(begin = xr_49_begin_0, end = xr_49_end_0, end_mask = xr_49_end_mask_0, squeeze_mask = xr_49_squeeze_mask_0, x = pairs_49_cast_fp16)[name = tensor("xr_49_cast_fp16")]; + tensor xi_49_begin_0 = const()[name = tensor("xi_49_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_49_end_0 = const()[name = tensor("xi_49_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_49_end_mask_0 = const()[name = tensor("xi_49_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_49_squeeze_mask_0 = const()[name = tensor("xi_49_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_49_cast_fp16 = slice_by_index(begin = xi_49_begin_0, end = xi_49_end_0, end_mask = xi_49_end_mask_0, squeeze_mask = xi_49_squeeze_mask_0, x = pairs_49_cast_fp16)[name = tensor("xi_49_cast_fp16")]; + tensor var_3214_cast_fp16 = mul(x = xr_49_cast_fp16, y = rotr_49_cast_fp16)[name = tensor("op_3214_cast_fp16")]; + tensor var_3215_cast_fp16 = mul(x = xi_49_cast_fp16, y = roti_49_cast_fp16)[name = tensor("op_3215_cast_fp16")]; + tensor xor__49_cast_fp16 = sub(x = var_3214_cast_fp16, y = var_3215_cast_fp16)[name = tensor("xor__49_cast_fp16")]; + tensor var_3218_cast_fp16 = mul(x = xr_49_cast_fp16, y = roti_49_cast_fp16)[name = tensor("op_3218_cast_fp16")]; + tensor var_3219_cast_fp16 = mul(x = xi_49_cast_fp16, y = rotr_49_cast_fp16)[name = tensor("op_3219_cast_fp16")]; + tensor xoi_49_cast_fp16 = add(x = var_3218_cast_fp16, y = var_3219_cast_fp16)[name = tensor("xoi_49_cast_fp16")]; + tensor out_49_axis_0 = const()[name = tensor("out_49_axis_0"), val = tensor(-1)]; + tensor out_49_cast_fp16 = stack(axis = out_49_axis_0, values = (xor__49_cast_fp16, xoi_49_cast_fp16))[name = tensor("out_49_cast_fp16")]; + tensor var_3242 = const()[name = tensor("op_3242"), val = tensor([1, 16, 32, 2])]; + tensor pairs_51_cast_fp16 = reshape(shape = var_3242, x = x_51_cast_fp16)[name = tensor("pairs_51_cast_fp16")]; + tensor xr_51_begin_0 = const()[name = tensor("xr_51_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_51_end_0 = const()[name = tensor("xr_51_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_51_end_mask_0 = const()[name = tensor("xr_51_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_51_squeeze_mask_0 = const()[name = tensor("xr_51_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_51_cast_fp16 = slice_by_index(begin = xr_51_begin_0, end = xr_51_end_0, end_mask = xr_51_end_mask_0, squeeze_mask = xr_51_squeeze_mask_0, x = pairs_51_cast_fp16)[name = tensor("xr_51_cast_fp16")]; + tensor xi_51_begin_0 = const()[name = tensor("xi_51_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_51_end_0 = const()[name = tensor("xi_51_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_51_end_mask_0 = const()[name = tensor("xi_51_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_51_squeeze_mask_0 = const()[name = tensor("xi_51_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_51_cast_fp16 = slice_by_index(begin = xi_51_begin_0, end = xi_51_end_0, end_mask = xi_51_end_mask_0, squeeze_mask = xi_51_squeeze_mask_0, x = pairs_51_cast_fp16)[name = tensor("xi_51_cast_fp16")]; + tensor var_3250_cast_fp16 = mul(x = xr_51_cast_fp16, y = rotr_49_cast_fp16)[name = tensor("op_3250_cast_fp16")]; + tensor var_3251_cast_fp16 = mul(x = xi_51_cast_fp16, y = roti_49_cast_fp16)[name = tensor("op_3251_cast_fp16")]; + tensor xor__51_cast_fp16 = sub(x = var_3250_cast_fp16, y = var_3251_cast_fp16)[name = tensor("xor__51_cast_fp16")]; + tensor var_3254_cast_fp16 = mul(x = xr_51_cast_fp16, y = roti_49_cast_fp16)[name = tensor("op_3254_cast_fp16")]; + tensor var_3255_cast_fp16 = mul(x = xi_51_cast_fp16, y = rotr_49_cast_fp16)[name = tensor("op_3255_cast_fp16")]; + tensor xoi_51_cast_fp16 = add(x = var_3254_cast_fp16, y = var_3255_cast_fp16)[name = tensor("xoi_51_cast_fp16")]; + tensor out_51_axis_0 = const()[name = tensor("out_51_axis_0"), val = tensor(-1)]; + tensor out_51_cast_fp16 = stack(axis = out_51_axis_0, values = (xor__51_cast_fp16, xoi_51_cast_fp16))[name = tensor("out_51_cast_fp16")]; + tensor _inversed_3267_y_0_to_fp16 = const()[name = tensor("_inversed_3267_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3267_cast_fp16 = mul(x = position12_to_fp16, y = _inversed_3267_y_0_to_fp16)[name = tensor("_inversed_3267_cast_fp16")]; + tensor var_3268_cast_fp16 = floor(x = _inversed_3267_cast_fp16)[name = tensor("op_3268_cast_fp16")]; + tensor var_3269_to_fp16 = const()[name = tensor("op_3269_to_fp16"), val = tensor(0x1p+9)]; + tensor var_3270_cast_fp16 = mul(x = var_3268_cast_fp16, y = var_3269_to_fp16)[name = tensor("op_3270_cast_fp16")]; + tensor write_idx_25_cast_fp16 = sub(x = position12_to_fp16, y = var_3270_cast_fp16)[name = tensor("write_idx_25_cast_fp16")]; + tensor var_3273_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_25_cast_fp16)[name = tensor("op_3273_cast_fp16")]; + tensor var_3283 = const()[name = tensor("op_3283"), val = tensor([1, 512, 1, 1])]; + tensor onehot_49_to_fp16_dtype_0 = const()[name = tensor("onehot_49_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_3273_cast_fp16_to_fp16 = cast(dtype = onehot_49_to_fp16_dtype_0, x = var_3273_cast_fp16)[name = tensor("cast_480")]; + tensor onehot_51_cast_fp16 = reshape(shape = var_3283, x = var_3273_cast_fp16_to_fp16)[name = tensor("onehot_51_cast_fp16")]; + tensor var_3285_to_fp16 = const()[name = tensor("op_3285_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3287_cast_fp16 = sub(x = var_3285_to_fp16, y = onehot_51_cast_fp16)[name = tensor("op_3287_cast_fp16")]; + tensor k_cache12_to_fp16_dtype_0 = const()[name = tensor("k_cache12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache12_to_fp16 = cast(dtype = k_cache12_to_fp16_dtype_0, x = k_cache12)[name = tensor("cast_479")]; + tensor var_3288_cast_fp16 = mul(x = k_cache12_to_fp16, y = var_3287_cast_fp16)[name = tensor("op_3288_cast_fp16")]; + tensor var_3293 = const()[name = tensor("op_3293"), val = tensor([1, 1, 16, 64])]; + tensor var_3294_cast_fp16 = reshape(shape = var_3293, x = out_51_cast_fp16)[name = tensor("op_3294_cast_fp16")]; + tensor var_3295_cast_fp16 = mul(x = var_3294_cast_fp16, y = onehot_51_cast_fp16)[name = tensor("op_3295_cast_fp16")]; + tensor new_k_cache_25_cast_fp16 = add(x = var_3288_cast_fp16, y = var_3295_cast_fp16)[name = tensor("new_k_cache_25_cast_fp16")]; + tensor new_k_cache_25_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_25_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache12_to_fp16_dtype_0 = const()[name = tensor("v_cache12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache12_to_fp16 = cast(dtype = v_cache12_to_fp16_dtype_0, x = v_cache12)[name = tensor("cast_477")]; + tensor var_3301_cast_fp16 = mul(x = v_cache12_to_fp16, y = var_3287_cast_fp16)[name = tensor("op_3301_cast_fp16")]; + tensor var_3306 = const()[name = tensor("op_3306"), val = tensor([1, 1, 16, 64])]; + tensor var_3307_cast_fp16 = reshape(shape = var_3306, x = v_25_cast_fp16)[name = tensor("op_3307_cast_fp16")]; + tensor var_3308_cast_fp16 = mul(x = var_3307_cast_fp16, y = onehot_51_cast_fp16)[name = tensor("op_3308_cast_fp16")]; + tensor new_v_cache_25_cast_fp16 = add(x = var_3301_cast_fp16, y = var_3308_cast_fp16)[name = tensor("new_v_cache_25_cast_fp16")]; + tensor new_v_cache_25_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_25_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_12_perm_0 = const()[name = tensor("transpose_12_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_3321 = const()[name = tensor("op_3321"), val = tensor([0, 2, 1, 3])]; + tensor var_3327 = const()[name = tensor("op_3327"), val = tensor([1, 16, 1, 64])]; + tensor q4_25_cast_fp16 = reshape(shape = var_3327, x = out_49_cast_fp16)[name = tensor("q4_25_cast_fp16")]; + tensor var_3332_transpose_x_0 = const()[name = tensor("op_3332_transpose_x_0"), val = tensor(false)]; + tensor var_3332_transpose_y_0 = const()[name = tensor("op_3332_transpose_y_0"), val = tensor(false)]; + tensor transpose_12_cast_fp16 = transpose(perm = transpose_12_perm_0, x = new_k_cache_25_cast_fp16)[name = tensor("transpose_59")]; + tensor var_3332_cast_fp16 = matmul(transpose_x = var_3332_transpose_x_0, transpose_y = var_3332_transpose_y_0, x = q4_25_cast_fp16, y = transpose_12_cast_fp16)[name = tensor("op_3332_cast_fp16")]; + tensor var_3333_to_fp16 = const()[name = tensor("op_3333_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_49_cast_fp16 = mul(x = var_3332_cast_fp16, y = var_3333_to_fp16)[name = tensor("scores_49_cast_fp16")]; + tensor var_3337 = const()[name = tensor("op_3337"), val = tensor([1, 1])]; + tensor var_3338_cast_fp16 = reshape(shape = var_3337, x = position12_to_fp16)[name = tensor("op_3338_cast_fp16")]; + tensor var_3339_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_3338_cast_fp16)[name = tensor("op_3339_cast_fp16")]; + tensor mask_25_to_fp16_dtype_0 = const()[name = tensor("mask_25_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_3346_to_fp16 = const()[name = tensor("op_3346_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3339_cast_fp16_to_fp16 = cast(dtype = mask_25_to_fp16_dtype_0, x = var_3339_cast_fp16)[name = tensor("cast_475")]; + tensor var_3347_cast_fp16 = sub(x = var_3339_cast_fp16_to_fp16, y = var_3346_to_fp16)[name = tensor("op_3347_cast_fp16")]; + tensor var_3352 = const()[name = tensor("op_3352"), val = tensor([1, 1, 1, 512])]; + tensor var_3353_cast_fp16 = reshape(shape = var_3352, x = var_3347_cast_fp16)[name = tensor("op_3353_cast_fp16")]; + tensor var_3354_to_fp16 = const()[name = tensor("op_3354_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_3355_cast_fp16 = mul(x = var_3353_cast_fp16, y = var_3354_to_fp16)[name = tensor("op_3355_cast_fp16")]; + tensor scores_51_cast_fp16 = add(x = scores_49_cast_fp16, y = var_3355_cast_fp16)[name = tensor("scores_51_cast_fp16")]; + tensor var_3358 = const()[name = tensor("op_3358"), val = tensor(-1)]; + tensor weights_25_cast_fp16 = softmax(axis = var_3358, x = scores_51_cast_fp16)[name = tensor("weights_25_cast_fp16")]; + tensor attn_25_transpose_x_0 = const()[name = tensor("attn_25_transpose_x_0"), val = tensor(false)]; + tensor attn_25_transpose_y_0 = const()[name = tensor("attn_25_transpose_y_0"), val = tensor(false)]; + tensor values_25_cast_fp16 = transpose(perm = var_3321, x = new_v_cache_25_cast_fp16)[name = tensor("transpose_58")]; + tensor attn_25_cast_fp16 = matmul(transpose_x = attn_25_transpose_x_0, transpose_y = attn_25_transpose_y_0, x = weights_25_cast_fp16, y = values_25_cast_fp16)[name = tensor("attn_25_cast_fp16")]; + tensor var_3366 = const()[name = tensor("op_3366"), val = tensor([0, 2, 1, 3])]; + tensor var_3371 = const()[name = tensor("op_3371"), val = tensor([1, 1, 1024])]; + tensor var_3367_cast_fp16 = transpose(perm = var_3366, x = attn_25_cast_fp16)[name = tensor("transpose_57")]; + tensor input_151_cast_fp16 = reshape(shape = var_3371, x = var_3367_cast_fp16)[name = tensor("input_151_cast_fp16")]; + tensor attn12_out_proj_weight_to_fp16 = const()[name = tensor("attn12_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(308473536)))]; + tensor linear_50_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn12_out_proj_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_50_cast_fp16")]; + tensor var_3377_to_fp16 = const()[name = tensor("op_3377_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3378_cast_fp16 = add(x = position12_to_fp16, y = var_3377_to_fp16)[name = tensor("op_3378_cast_fp16")]; + tensor var_3378_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3378_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_153_cast_fp16 = add(x = input_147_cast_fp16, y = linear_50_cast_fp16)[name = tensor("input_153_cast_fp16")]; + tensor input_155_axes_0 = const()[name = tensor("input_155_axes_0"), val = tensor([-1])]; + tensor norm12_2_weight_to_fp16 = const()[name = tensor("norm12_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(310570752)))]; + tensor norm12_2_bias_to_fp16 = const()[name = tensor("norm12_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(310572864)))]; + tensor var_3382_to_fp16 = const()[name = tensor("op_3382_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_155_cast_fp16 = layer_norm(axes = input_155_axes_0, beta = norm12_2_bias_to_fp16, epsilon = var_3382_to_fp16, gamma = norm12_2_weight_to_fp16, x = input_153_cast_fp16)[name = tensor("input_155_cast_fp16")]; + tensor linear12_1_weight_to_fp16 = const()[name = tensor("linear12_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(310574976)))]; + tensor linear_51_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear12_1_weight_to_fp16, x = input_155_cast_fp16)[name = tensor("linear_51_cast_fp16")]; + tensor input_157_mode_0 = const()[name = tensor("input_157_mode_0"), val = tensor("EXACT")]; + tensor input_157_cast_fp16 = gelu(mode = input_157_mode_0, x = linear_51_cast_fp16)[name = tensor("input_157_cast_fp16")]; + tensor linear12_2_weight_to_fp16 = const()[name = tensor("linear12_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(318963648)))]; + tensor linear_52_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear12_2_weight_to_fp16, x = input_157_cast_fp16)[name = tensor("linear_52_cast_fp16")]; + tensor input_159_cast_fp16 = add(x = input_153_cast_fp16, y = linear_52_cast_fp16)[name = tensor("input_159_cast_fp16")]; + tensor input_161_axes_0 = const()[name = tensor("input_161_axes_0"), val = tensor([-1])]; + tensor norm13_1_weight_to_fp16 = const()[name = tensor("norm13_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(327352320)))]; + tensor norm13_1_bias_to_fp16 = const()[name = tensor("norm13_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(327354432)))]; + tensor var_3399_to_fp16 = const()[name = tensor("op_3399_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_161_cast_fp16 = layer_norm(axes = input_161_axes_0, beta = norm13_1_bias_to_fp16, epsilon = var_3399_to_fp16, gamma = norm13_1_weight_to_fp16, x = input_159_cast_fp16)[name = tensor("input_161_cast_fp16")]; + tensor attn13_in_proj_weight_to_fp16 = const()[name = tensor("attn13_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(327356544)))]; + tensor linear_53_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn13_in_proj_weight_to_fp16, x = input_161_cast_fp16)[name = tensor("linear_53_cast_fp16")]; + tensor var_3412 = const()[name = tensor("op_3412"), val = tensor([1, 3, 16, 64])]; + tensor qkv_27_cast_fp16 = reshape(shape = var_3412, x = linear_53_cast_fp16)[name = tensor("qkv_27_cast_fp16")]; + tensor x_53_begin_0 = const()[name = tensor("x_53_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_53_end_0 = const()[name = tensor("x_53_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_53_end_mask_0 = const()[name = tensor("x_53_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_53_squeeze_mask_0 = const()[name = tensor("x_53_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_53_cast_fp16 = slice_by_index(begin = x_53_begin_0, end = x_53_end_0, end_mask = x_53_end_mask_0, squeeze_mask = x_53_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("x_53_cast_fp16")]; + tensor x_55_begin_0 = const()[name = tensor("x_55_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_55_end_0 = const()[name = tensor("x_55_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_55_end_mask_0 = const()[name = tensor("x_55_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_55_squeeze_mask_0 = const()[name = tensor("x_55_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_55_cast_fp16 = slice_by_index(begin = x_55_begin_0, end = x_55_end_0, end_mask = x_55_end_mask_0, squeeze_mask = x_55_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("x_55_cast_fp16")]; + tensor v_27_begin_0 = const()[name = tensor("v_27_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_27_end_0 = const()[name = tensor("v_27_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_27_end_mask_0 = const()[name = tensor("v_27_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_27_squeeze_mask_0 = const()[name = tensor("v_27_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_27_cast_fp16 = slice_by_index(begin = v_27_begin_0, end = v_27_end_0, end_mask = v_27_end_mask_0, squeeze_mask = v_27_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("v_27_cast_fp16")]; + tensor var_3441 = const()[name = tensor("op_3441"), val = tensor([1, 1, 1])]; + tensor position13_to_fp16_dtype_0 = const()[name = tensor("position13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position13_to_fp16 = cast(dtype = position13_to_fp16_dtype_0, x = position13)[name = tensor("cast_473")]; + tensor var_3442_cast_fp16 = reshape(shape = var_3441, x = position13_to_fp16)[name = tensor("op_3442_cast_fp16")]; + tensor angles_53_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_3442_cast_fp16)[name = tensor("angles_53_cast_fp16")]; + tensor rotr_53_cast_fp16 = cos(x = angles_53_cast_fp16)[name = tensor("rotr_53_cast_fp16")]; + tensor roti_53_cast_fp16 = sin(x = angles_53_cast_fp16)[name = tensor("roti_53_cast_fp16")]; + tensor var_3450 = const()[name = tensor("op_3450"), val = tensor([1, 16, 32, 2])]; + tensor pairs_53_cast_fp16 = reshape(shape = var_3450, x = x_53_cast_fp16)[name = tensor("pairs_53_cast_fp16")]; + tensor xr_53_begin_0 = const()[name = tensor("xr_53_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_53_end_0 = const()[name = tensor("xr_53_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_53_end_mask_0 = const()[name = tensor("xr_53_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_53_squeeze_mask_0 = const()[name = tensor("xr_53_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_53_cast_fp16 = slice_by_index(begin = xr_53_begin_0, end = xr_53_end_0, end_mask = xr_53_end_mask_0, squeeze_mask = xr_53_squeeze_mask_0, x = pairs_53_cast_fp16)[name = tensor("xr_53_cast_fp16")]; + tensor xi_53_begin_0 = const()[name = tensor("xi_53_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_53_end_0 = const()[name = tensor("xi_53_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_53_end_mask_0 = const()[name = tensor("xi_53_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_53_squeeze_mask_0 = const()[name = tensor("xi_53_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_53_cast_fp16 = slice_by_index(begin = xi_53_begin_0, end = xi_53_end_0, end_mask = xi_53_end_mask_0, squeeze_mask = xi_53_squeeze_mask_0, x = pairs_53_cast_fp16)[name = tensor("xi_53_cast_fp16")]; + tensor var_3458_cast_fp16 = mul(x = xr_53_cast_fp16, y = rotr_53_cast_fp16)[name = tensor("op_3458_cast_fp16")]; + tensor var_3459_cast_fp16 = mul(x = xi_53_cast_fp16, y = roti_53_cast_fp16)[name = tensor("op_3459_cast_fp16")]; + tensor xor__53_cast_fp16 = sub(x = var_3458_cast_fp16, y = var_3459_cast_fp16)[name = tensor("xor__53_cast_fp16")]; + tensor var_3462_cast_fp16 = mul(x = xr_53_cast_fp16, y = roti_53_cast_fp16)[name = tensor("op_3462_cast_fp16")]; + tensor var_3463_cast_fp16 = mul(x = xi_53_cast_fp16, y = rotr_53_cast_fp16)[name = tensor("op_3463_cast_fp16")]; + tensor xoi_53_cast_fp16 = add(x = var_3462_cast_fp16, y = var_3463_cast_fp16)[name = tensor("xoi_53_cast_fp16")]; + tensor out_53_axis_0 = const()[name = tensor("out_53_axis_0"), val = tensor(-1)]; + tensor out_53_cast_fp16 = stack(axis = out_53_axis_0, values = (xor__53_cast_fp16, xoi_53_cast_fp16))[name = tensor("out_53_cast_fp16")]; + tensor var_3486 = const()[name = tensor("op_3486"), val = tensor([1, 16, 32, 2])]; + tensor pairs_55_cast_fp16 = reshape(shape = var_3486, x = x_55_cast_fp16)[name = tensor("pairs_55_cast_fp16")]; + tensor xr_55_begin_0 = const()[name = tensor("xr_55_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_55_end_0 = const()[name = tensor("xr_55_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_55_end_mask_0 = const()[name = tensor("xr_55_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_55_squeeze_mask_0 = const()[name = tensor("xr_55_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_55_cast_fp16 = slice_by_index(begin = xr_55_begin_0, end = xr_55_end_0, end_mask = xr_55_end_mask_0, squeeze_mask = xr_55_squeeze_mask_0, x = pairs_55_cast_fp16)[name = tensor("xr_55_cast_fp16")]; + tensor xi_55_begin_0 = const()[name = tensor("xi_55_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_55_end_0 = const()[name = tensor("xi_55_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_55_end_mask_0 = const()[name = tensor("xi_55_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_55_squeeze_mask_0 = const()[name = tensor("xi_55_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_55_cast_fp16 = slice_by_index(begin = xi_55_begin_0, end = xi_55_end_0, end_mask = xi_55_end_mask_0, squeeze_mask = xi_55_squeeze_mask_0, x = pairs_55_cast_fp16)[name = tensor("xi_55_cast_fp16")]; + tensor var_3494_cast_fp16 = mul(x = xr_55_cast_fp16, y = rotr_53_cast_fp16)[name = tensor("op_3494_cast_fp16")]; + tensor var_3495_cast_fp16 = mul(x = xi_55_cast_fp16, y = roti_53_cast_fp16)[name = tensor("op_3495_cast_fp16")]; + tensor xor__55_cast_fp16 = sub(x = var_3494_cast_fp16, y = var_3495_cast_fp16)[name = tensor("xor__55_cast_fp16")]; + tensor var_3498_cast_fp16 = mul(x = xr_55_cast_fp16, y = roti_53_cast_fp16)[name = tensor("op_3498_cast_fp16")]; + tensor var_3499_cast_fp16 = mul(x = xi_55_cast_fp16, y = rotr_53_cast_fp16)[name = tensor("op_3499_cast_fp16")]; + tensor xoi_55_cast_fp16 = add(x = var_3498_cast_fp16, y = var_3499_cast_fp16)[name = tensor("xoi_55_cast_fp16")]; + tensor out_55_axis_0 = const()[name = tensor("out_55_axis_0"), val = tensor(-1)]; + tensor out_55_cast_fp16 = stack(axis = out_55_axis_0, values = (xor__55_cast_fp16, xoi_55_cast_fp16))[name = tensor("out_55_cast_fp16")]; + tensor _inversed_3511_y_0_to_fp16 = const()[name = tensor("_inversed_3511_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3511_cast_fp16 = mul(x = position13_to_fp16, y = _inversed_3511_y_0_to_fp16)[name = tensor("_inversed_3511_cast_fp16")]; + tensor var_3512_cast_fp16 = floor(x = _inversed_3511_cast_fp16)[name = tensor("op_3512_cast_fp16")]; + tensor var_3513_to_fp16 = const()[name = tensor("op_3513_to_fp16"), val = tensor(0x1p+9)]; + tensor var_3514_cast_fp16 = mul(x = var_3512_cast_fp16, y = var_3513_to_fp16)[name = tensor("op_3514_cast_fp16")]; + tensor write_idx_27_cast_fp16 = sub(x = position13_to_fp16, y = var_3514_cast_fp16)[name = tensor("write_idx_27_cast_fp16")]; + tensor var_3517_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_27_cast_fp16)[name = tensor("op_3517_cast_fp16")]; + tensor var_3527 = const()[name = tensor("op_3527"), val = tensor([1, 512, 1, 1])]; + tensor onehot_53_to_fp16_dtype_0 = const()[name = tensor("onehot_53_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_3517_cast_fp16_to_fp16 = cast(dtype = onehot_53_to_fp16_dtype_0, x = var_3517_cast_fp16)[name = tensor("cast_472")]; + tensor onehot_55_cast_fp16 = reshape(shape = var_3527, x = var_3517_cast_fp16_to_fp16)[name = tensor("onehot_55_cast_fp16")]; + tensor var_3529_to_fp16 = const()[name = tensor("op_3529_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3531_cast_fp16 = sub(x = var_3529_to_fp16, y = onehot_55_cast_fp16)[name = tensor("op_3531_cast_fp16")]; + tensor k_cache13_to_fp16_dtype_0 = const()[name = tensor("k_cache13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache13_to_fp16 = cast(dtype = k_cache13_to_fp16_dtype_0, x = k_cache13)[name = tensor("cast_471")]; + tensor var_3532_cast_fp16 = mul(x = k_cache13_to_fp16, y = var_3531_cast_fp16)[name = tensor("op_3532_cast_fp16")]; + tensor var_3537 = const()[name = tensor("op_3537"), val = tensor([1, 1, 16, 64])]; + tensor var_3538_cast_fp16 = reshape(shape = var_3537, x = out_55_cast_fp16)[name = tensor("op_3538_cast_fp16")]; + tensor var_3539_cast_fp16 = mul(x = var_3538_cast_fp16, y = onehot_55_cast_fp16)[name = tensor("op_3539_cast_fp16")]; + tensor new_k_cache_27_cast_fp16 = add(x = var_3532_cast_fp16, y = var_3539_cast_fp16)[name = tensor("new_k_cache_27_cast_fp16")]; + tensor new_k_cache_27_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_27_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache13_to_fp16_dtype_0 = const()[name = tensor("v_cache13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache13_to_fp16 = cast(dtype = v_cache13_to_fp16_dtype_0, x = v_cache13)[name = tensor("cast_469")]; + tensor var_3545_cast_fp16 = mul(x = v_cache13_to_fp16, y = var_3531_cast_fp16)[name = tensor("op_3545_cast_fp16")]; + tensor var_3550 = const()[name = tensor("op_3550"), val = tensor([1, 1, 16, 64])]; + tensor var_3551_cast_fp16 = reshape(shape = var_3550, x = v_27_cast_fp16)[name = tensor("op_3551_cast_fp16")]; + tensor var_3552_cast_fp16 = mul(x = var_3551_cast_fp16, y = onehot_55_cast_fp16)[name = tensor("op_3552_cast_fp16")]; + tensor new_v_cache_27_cast_fp16 = add(x = var_3545_cast_fp16, y = var_3552_cast_fp16)[name = tensor("new_v_cache_27_cast_fp16")]; + tensor new_v_cache_27_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_27_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_13_perm_0 = const()[name = tensor("transpose_13_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_3565 = const()[name = tensor("op_3565"), val = tensor([0, 2, 1, 3])]; + tensor var_3571 = const()[name = tensor("op_3571"), val = tensor([1, 16, 1, 64])]; + tensor q4_27_cast_fp16 = reshape(shape = var_3571, x = out_53_cast_fp16)[name = tensor("q4_27_cast_fp16")]; + tensor var_3576_transpose_x_0 = const()[name = tensor("op_3576_transpose_x_0"), val = tensor(false)]; + tensor var_3576_transpose_y_0 = const()[name = tensor("op_3576_transpose_y_0"), val = tensor(false)]; + tensor transpose_13_cast_fp16 = transpose(perm = transpose_13_perm_0, x = new_k_cache_27_cast_fp16)[name = tensor("transpose_56")]; + tensor var_3576_cast_fp16 = matmul(transpose_x = var_3576_transpose_x_0, transpose_y = var_3576_transpose_y_0, x = q4_27_cast_fp16, y = transpose_13_cast_fp16)[name = tensor("op_3576_cast_fp16")]; + tensor var_3577_to_fp16 = const()[name = tensor("op_3577_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_53_cast_fp16 = mul(x = var_3576_cast_fp16, y = var_3577_to_fp16)[name = tensor("scores_53_cast_fp16")]; + tensor var_3581 = const()[name = tensor("op_3581"), val = tensor([1, 1])]; + tensor var_3582_cast_fp16 = reshape(shape = var_3581, x = position13_to_fp16)[name = tensor("op_3582_cast_fp16")]; + tensor var_3583_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_3582_cast_fp16)[name = tensor("op_3583_cast_fp16")]; + tensor mask_27_to_fp16_dtype_0 = const()[name = tensor("mask_27_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_3590_to_fp16 = const()[name = tensor("op_3590_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3583_cast_fp16_to_fp16 = cast(dtype = mask_27_to_fp16_dtype_0, x = var_3583_cast_fp16)[name = tensor("cast_467")]; + tensor var_3591_cast_fp16 = sub(x = var_3583_cast_fp16_to_fp16, y = var_3590_to_fp16)[name = tensor("op_3591_cast_fp16")]; + tensor var_3596 = const()[name = tensor("op_3596"), val = tensor([1, 1, 1, 512])]; + tensor var_3597_cast_fp16 = reshape(shape = var_3596, x = var_3591_cast_fp16)[name = tensor("op_3597_cast_fp16")]; + tensor var_3598_to_fp16 = const()[name = tensor("op_3598_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_3599_cast_fp16 = mul(x = var_3597_cast_fp16, y = var_3598_to_fp16)[name = tensor("op_3599_cast_fp16")]; + tensor scores_55_cast_fp16 = add(x = scores_53_cast_fp16, y = var_3599_cast_fp16)[name = tensor("scores_55_cast_fp16")]; + tensor var_3602 = const()[name = tensor("op_3602"), val = tensor(-1)]; + tensor weights_27_cast_fp16 = softmax(axis = var_3602, x = scores_55_cast_fp16)[name = tensor("weights_27_cast_fp16")]; + tensor attn_27_transpose_x_0 = const()[name = tensor("attn_27_transpose_x_0"), val = tensor(false)]; + tensor attn_27_transpose_y_0 = const()[name = tensor("attn_27_transpose_y_0"), val = tensor(false)]; + tensor values_27_cast_fp16 = transpose(perm = var_3565, x = new_v_cache_27_cast_fp16)[name = tensor("transpose_55")]; + tensor attn_27_cast_fp16 = matmul(transpose_x = attn_27_transpose_x_0, transpose_y = attn_27_transpose_y_0, x = weights_27_cast_fp16, y = values_27_cast_fp16)[name = tensor("attn_27_cast_fp16")]; + tensor var_3610 = const()[name = tensor("op_3610"), val = tensor([0, 2, 1, 3])]; + tensor var_3615 = const()[name = tensor("op_3615"), val = tensor([1, 1, 1024])]; + tensor var_3611_cast_fp16 = transpose(perm = var_3610, x = attn_27_cast_fp16)[name = tensor("transpose_54")]; + tensor input_163_cast_fp16 = reshape(shape = var_3615, x = var_3611_cast_fp16)[name = tensor("input_163_cast_fp16")]; + tensor attn13_out_proj_weight_to_fp16 = const()[name = tensor("attn13_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(333648064)))]; + tensor linear_54_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn13_out_proj_weight_to_fp16, x = input_163_cast_fp16)[name = tensor("linear_54_cast_fp16")]; + tensor var_3621_to_fp16 = const()[name = tensor("op_3621_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3622_cast_fp16 = add(x = position13_to_fp16, y = var_3621_to_fp16)[name = tensor("op_3622_cast_fp16")]; + tensor var_3622_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3622_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_165_cast_fp16 = add(x = input_159_cast_fp16, y = linear_54_cast_fp16)[name = tensor("input_165_cast_fp16")]; + tensor input_167_axes_0 = const()[name = tensor("input_167_axes_0"), val = tensor([-1])]; + tensor norm13_2_weight_to_fp16 = const()[name = tensor("norm13_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335745280)))]; + tensor norm13_2_bias_to_fp16 = const()[name = tensor("norm13_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335747392)))]; + tensor var_3626_to_fp16 = const()[name = tensor("op_3626_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_167_cast_fp16 = layer_norm(axes = input_167_axes_0, beta = norm13_2_bias_to_fp16, epsilon = var_3626_to_fp16, gamma = norm13_2_weight_to_fp16, x = input_165_cast_fp16)[name = tensor("input_167_cast_fp16")]; + tensor linear13_1_weight_to_fp16 = const()[name = tensor("linear13_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335749504)))]; + tensor linear_55_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear13_1_weight_to_fp16, x = input_167_cast_fp16)[name = tensor("linear_55_cast_fp16")]; + tensor input_169_mode_0 = const()[name = tensor("input_169_mode_0"), val = tensor("EXACT")]; + tensor input_169_cast_fp16 = gelu(mode = input_169_mode_0, x = linear_55_cast_fp16)[name = tensor("input_169_cast_fp16")]; + tensor linear13_2_weight_to_fp16 = const()[name = tensor("linear13_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(344138176)))]; + tensor linear_56_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear13_2_weight_to_fp16, x = input_169_cast_fp16)[name = tensor("linear_56_cast_fp16")]; + tensor input_171_cast_fp16 = add(x = input_165_cast_fp16, y = linear_56_cast_fp16)[name = tensor("input_171_cast_fp16")]; + tensor input_173_axes_0 = const()[name = tensor("input_173_axes_0"), val = tensor([-1])]; + tensor norm14_1_weight_to_fp16 = const()[name = tensor("norm14_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(352526848)))]; + tensor norm14_1_bias_to_fp16 = const()[name = tensor("norm14_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(352528960)))]; + tensor var_3643_to_fp16 = const()[name = tensor("op_3643_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_173_cast_fp16 = layer_norm(axes = input_173_axes_0, beta = norm14_1_bias_to_fp16, epsilon = var_3643_to_fp16, gamma = norm14_1_weight_to_fp16, x = input_171_cast_fp16)[name = tensor("input_173_cast_fp16")]; + tensor attn14_in_proj_weight_to_fp16 = const()[name = tensor("attn14_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(352531072)))]; + tensor linear_57_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn14_in_proj_weight_to_fp16, x = input_173_cast_fp16)[name = tensor("linear_57_cast_fp16")]; + tensor var_3656 = const()[name = tensor("op_3656"), val = tensor([1, 3, 16, 64])]; + tensor qkv_29_cast_fp16 = reshape(shape = var_3656, x = linear_57_cast_fp16)[name = tensor("qkv_29_cast_fp16")]; + tensor x_57_begin_0 = const()[name = tensor("x_57_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_57_end_0 = const()[name = tensor("x_57_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_57_end_mask_0 = const()[name = tensor("x_57_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_57_squeeze_mask_0 = const()[name = tensor("x_57_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_57_cast_fp16 = slice_by_index(begin = x_57_begin_0, end = x_57_end_0, end_mask = x_57_end_mask_0, squeeze_mask = x_57_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("x_57_cast_fp16")]; + tensor x_59_begin_0 = const()[name = tensor("x_59_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_59_end_0 = const()[name = tensor("x_59_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_59_end_mask_0 = const()[name = tensor("x_59_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_59_squeeze_mask_0 = const()[name = tensor("x_59_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_59_cast_fp16 = slice_by_index(begin = x_59_begin_0, end = x_59_end_0, end_mask = x_59_end_mask_0, squeeze_mask = x_59_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("x_59_cast_fp16")]; + tensor v_29_begin_0 = const()[name = tensor("v_29_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_29_end_0 = const()[name = tensor("v_29_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_29_end_mask_0 = const()[name = tensor("v_29_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_29_squeeze_mask_0 = const()[name = tensor("v_29_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_29_cast_fp16 = slice_by_index(begin = v_29_begin_0, end = v_29_end_0, end_mask = v_29_end_mask_0, squeeze_mask = v_29_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("v_29_cast_fp16")]; + tensor var_3685 = const()[name = tensor("op_3685"), val = tensor([1, 1, 1])]; + tensor position14_to_fp16_dtype_0 = const()[name = tensor("position14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position14_to_fp16 = cast(dtype = position14_to_fp16_dtype_0, x = position14)[name = tensor("cast_465")]; + tensor var_3686_cast_fp16 = reshape(shape = var_3685, x = position14_to_fp16)[name = tensor("op_3686_cast_fp16")]; + tensor angles_57_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_3686_cast_fp16)[name = tensor("angles_57_cast_fp16")]; + tensor rotr_57_cast_fp16 = cos(x = angles_57_cast_fp16)[name = tensor("rotr_57_cast_fp16")]; + tensor roti_57_cast_fp16 = sin(x = angles_57_cast_fp16)[name = tensor("roti_57_cast_fp16")]; + tensor var_3694 = const()[name = tensor("op_3694"), val = tensor([1, 16, 32, 2])]; + tensor pairs_57_cast_fp16 = reshape(shape = var_3694, x = x_57_cast_fp16)[name = tensor("pairs_57_cast_fp16")]; + tensor xr_57_begin_0 = const()[name = tensor("xr_57_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_57_end_0 = const()[name = tensor("xr_57_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_57_end_mask_0 = const()[name = tensor("xr_57_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_57_squeeze_mask_0 = const()[name = tensor("xr_57_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_57_cast_fp16 = slice_by_index(begin = xr_57_begin_0, end = xr_57_end_0, end_mask = xr_57_end_mask_0, squeeze_mask = xr_57_squeeze_mask_0, x = pairs_57_cast_fp16)[name = tensor("xr_57_cast_fp16")]; + tensor xi_57_begin_0 = const()[name = tensor("xi_57_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_57_end_0 = const()[name = tensor("xi_57_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_57_end_mask_0 = const()[name = tensor("xi_57_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_57_squeeze_mask_0 = const()[name = tensor("xi_57_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_57_cast_fp16 = slice_by_index(begin = xi_57_begin_0, end = xi_57_end_0, end_mask = xi_57_end_mask_0, squeeze_mask = xi_57_squeeze_mask_0, x = pairs_57_cast_fp16)[name = tensor("xi_57_cast_fp16")]; + tensor var_3702_cast_fp16 = mul(x = xr_57_cast_fp16, y = rotr_57_cast_fp16)[name = tensor("op_3702_cast_fp16")]; + tensor var_3703_cast_fp16 = mul(x = xi_57_cast_fp16, y = roti_57_cast_fp16)[name = tensor("op_3703_cast_fp16")]; + tensor xor__57_cast_fp16 = sub(x = var_3702_cast_fp16, y = var_3703_cast_fp16)[name = tensor("xor__57_cast_fp16")]; + tensor var_3706_cast_fp16 = mul(x = xr_57_cast_fp16, y = roti_57_cast_fp16)[name = tensor("op_3706_cast_fp16")]; + tensor var_3707_cast_fp16 = mul(x = xi_57_cast_fp16, y = rotr_57_cast_fp16)[name = tensor("op_3707_cast_fp16")]; + tensor xoi_57_cast_fp16 = add(x = var_3706_cast_fp16, y = var_3707_cast_fp16)[name = tensor("xoi_57_cast_fp16")]; + tensor out_57_axis_0 = const()[name = tensor("out_57_axis_0"), val = tensor(-1)]; + tensor out_57_cast_fp16 = stack(axis = out_57_axis_0, values = (xor__57_cast_fp16, xoi_57_cast_fp16))[name = tensor("out_57_cast_fp16")]; + tensor var_3730 = const()[name = tensor("op_3730"), val = tensor([1, 16, 32, 2])]; + tensor pairs_59_cast_fp16 = reshape(shape = var_3730, x = x_59_cast_fp16)[name = tensor("pairs_59_cast_fp16")]; + tensor xr_59_begin_0 = const()[name = tensor("xr_59_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_59_end_0 = const()[name = tensor("xr_59_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_59_end_mask_0 = const()[name = tensor("xr_59_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_59_squeeze_mask_0 = const()[name = tensor("xr_59_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_59_cast_fp16 = slice_by_index(begin = xr_59_begin_0, end = xr_59_end_0, end_mask = xr_59_end_mask_0, squeeze_mask = xr_59_squeeze_mask_0, x = pairs_59_cast_fp16)[name = tensor("xr_59_cast_fp16")]; + tensor xi_59_begin_0 = const()[name = tensor("xi_59_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_59_end_0 = const()[name = tensor("xi_59_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_59_end_mask_0 = const()[name = tensor("xi_59_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_59_squeeze_mask_0 = const()[name = tensor("xi_59_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_59_cast_fp16 = slice_by_index(begin = xi_59_begin_0, end = xi_59_end_0, end_mask = xi_59_end_mask_0, squeeze_mask = xi_59_squeeze_mask_0, x = pairs_59_cast_fp16)[name = tensor("xi_59_cast_fp16")]; + tensor var_3738_cast_fp16 = mul(x = xr_59_cast_fp16, y = rotr_57_cast_fp16)[name = tensor("op_3738_cast_fp16")]; + tensor var_3739_cast_fp16 = mul(x = xi_59_cast_fp16, y = roti_57_cast_fp16)[name = tensor("op_3739_cast_fp16")]; + tensor xor__59_cast_fp16 = sub(x = var_3738_cast_fp16, y = var_3739_cast_fp16)[name = tensor("xor__59_cast_fp16")]; + tensor var_3742_cast_fp16 = mul(x = xr_59_cast_fp16, y = roti_57_cast_fp16)[name = tensor("op_3742_cast_fp16")]; + tensor var_3743_cast_fp16 = mul(x = xi_59_cast_fp16, y = rotr_57_cast_fp16)[name = tensor("op_3743_cast_fp16")]; + tensor xoi_59_cast_fp16 = add(x = var_3742_cast_fp16, y = var_3743_cast_fp16)[name = tensor("xoi_59_cast_fp16")]; + tensor out_59_axis_0 = const()[name = tensor("out_59_axis_0"), val = tensor(-1)]; + tensor out_59_cast_fp16 = stack(axis = out_59_axis_0, values = (xor__59_cast_fp16, xoi_59_cast_fp16))[name = tensor("out_59_cast_fp16")]; + tensor _inversed_3755_y_0_to_fp16 = const()[name = tensor("_inversed_3755_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3755_cast_fp16 = mul(x = position14_to_fp16, y = _inversed_3755_y_0_to_fp16)[name = tensor("_inversed_3755_cast_fp16")]; + tensor var_3756_cast_fp16 = floor(x = _inversed_3755_cast_fp16)[name = tensor("op_3756_cast_fp16")]; + tensor var_3757_to_fp16 = const()[name = tensor("op_3757_to_fp16"), val = tensor(0x1p+9)]; + tensor var_3758_cast_fp16 = mul(x = var_3756_cast_fp16, y = var_3757_to_fp16)[name = tensor("op_3758_cast_fp16")]; + tensor write_idx_29_cast_fp16 = sub(x = position14_to_fp16, y = var_3758_cast_fp16)[name = tensor("write_idx_29_cast_fp16")]; + tensor var_3761_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_29_cast_fp16)[name = tensor("op_3761_cast_fp16")]; + tensor var_3771 = const()[name = tensor("op_3771"), val = tensor([1, 512, 1, 1])]; + tensor onehot_57_to_fp16_dtype_0 = const()[name = tensor("onehot_57_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_3761_cast_fp16_to_fp16 = cast(dtype = onehot_57_to_fp16_dtype_0, x = var_3761_cast_fp16)[name = tensor("cast_464")]; + tensor onehot_59_cast_fp16 = reshape(shape = var_3771, x = var_3761_cast_fp16_to_fp16)[name = tensor("onehot_59_cast_fp16")]; + tensor var_3773_to_fp16 = const()[name = tensor("op_3773_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3775_cast_fp16 = sub(x = var_3773_to_fp16, y = onehot_59_cast_fp16)[name = tensor("op_3775_cast_fp16")]; + tensor k_cache14_to_fp16_dtype_0 = const()[name = tensor("k_cache14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache14_to_fp16 = cast(dtype = k_cache14_to_fp16_dtype_0, x = k_cache14)[name = tensor("cast_463")]; + tensor var_3776_cast_fp16 = mul(x = k_cache14_to_fp16, y = var_3775_cast_fp16)[name = tensor("op_3776_cast_fp16")]; + tensor var_3781 = const()[name = tensor("op_3781"), val = tensor([1, 1, 16, 64])]; + tensor var_3782_cast_fp16 = reshape(shape = var_3781, x = out_59_cast_fp16)[name = tensor("op_3782_cast_fp16")]; + tensor var_3783_cast_fp16 = mul(x = var_3782_cast_fp16, y = onehot_59_cast_fp16)[name = tensor("op_3783_cast_fp16")]; + tensor new_k_cache_29_cast_fp16 = add(x = var_3776_cast_fp16, y = var_3783_cast_fp16)[name = tensor("new_k_cache_29_cast_fp16")]; + tensor new_k_cache_29_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_29_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache14_to_fp16_dtype_0 = const()[name = tensor("v_cache14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache14_to_fp16 = cast(dtype = v_cache14_to_fp16_dtype_0, x = v_cache14)[name = tensor("cast_461")]; + tensor var_3789_cast_fp16 = mul(x = v_cache14_to_fp16, y = var_3775_cast_fp16)[name = tensor("op_3789_cast_fp16")]; + tensor var_3794 = const()[name = tensor("op_3794"), val = tensor([1, 1, 16, 64])]; + tensor var_3795_cast_fp16 = reshape(shape = var_3794, x = v_29_cast_fp16)[name = tensor("op_3795_cast_fp16")]; + tensor var_3796_cast_fp16 = mul(x = var_3795_cast_fp16, y = onehot_59_cast_fp16)[name = tensor("op_3796_cast_fp16")]; + tensor new_v_cache_29_cast_fp16 = add(x = var_3789_cast_fp16, y = var_3796_cast_fp16)[name = tensor("new_v_cache_29_cast_fp16")]; + tensor new_v_cache_29_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_29_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_14_perm_0 = const()[name = tensor("transpose_14_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_3809 = const()[name = tensor("op_3809"), val = tensor([0, 2, 1, 3])]; + tensor var_3815 = const()[name = tensor("op_3815"), val = tensor([1, 16, 1, 64])]; + tensor q4_29_cast_fp16 = reshape(shape = var_3815, x = out_57_cast_fp16)[name = tensor("q4_29_cast_fp16")]; + tensor var_3820_transpose_x_0 = const()[name = tensor("op_3820_transpose_x_0"), val = tensor(false)]; + tensor var_3820_transpose_y_0 = const()[name = tensor("op_3820_transpose_y_0"), val = tensor(false)]; + tensor transpose_14_cast_fp16 = transpose(perm = transpose_14_perm_0, x = new_k_cache_29_cast_fp16)[name = tensor("transpose_53")]; + tensor var_3820_cast_fp16 = matmul(transpose_x = var_3820_transpose_x_0, transpose_y = var_3820_transpose_y_0, x = q4_29_cast_fp16, y = transpose_14_cast_fp16)[name = tensor("op_3820_cast_fp16")]; + tensor var_3821_to_fp16 = const()[name = tensor("op_3821_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_57_cast_fp16 = mul(x = var_3820_cast_fp16, y = var_3821_to_fp16)[name = tensor("scores_57_cast_fp16")]; + tensor var_3825 = const()[name = tensor("op_3825"), val = tensor([1, 1])]; + tensor var_3826_cast_fp16 = reshape(shape = var_3825, x = position14_to_fp16)[name = tensor("op_3826_cast_fp16")]; + tensor var_3827_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_3826_cast_fp16)[name = tensor("op_3827_cast_fp16")]; + tensor mask_29_to_fp16_dtype_0 = const()[name = tensor("mask_29_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_3834_to_fp16 = const()[name = tensor("op_3834_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3827_cast_fp16_to_fp16 = cast(dtype = mask_29_to_fp16_dtype_0, x = var_3827_cast_fp16)[name = tensor("cast_459")]; + tensor var_3835_cast_fp16 = sub(x = var_3827_cast_fp16_to_fp16, y = var_3834_to_fp16)[name = tensor("op_3835_cast_fp16")]; + tensor var_3840 = const()[name = tensor("op_3840"), val = tensor([1, 1, 1, 512])]; + tensor var_3841_cast_fp16 = reshape(shape = var_3840, x = var_3835_cast_fp16)[name = tensor("op_3841_cast_fp16")]; + tensor var_3842_to_fp16 = const()[name = tensor("op_3842_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_3843_cast_fp16 = mul(x = var_3841_cast_fp16, y = var_3842_to_fp16)[name = tensor("op_3843_cast_fp16")]; + tensor scores_59_cast_fp16 = add(x = scores_57_cast_fp16, y = var_3843_cast_fp16)[name = tensor("scores_59_cast_fp16")]; + tensor var_3846 = const()[name = tensor("op_3846"), val = tensor(-1)]; + tensor weights_29_cast_fp16 = softmax(axis = var_3846, x = scores_59_cast_fp16)[name = tensor("weights_29_cast_fp16")]; + tensor attn_29_transpose_x_0 = const()[name = tensor("attn_29_transpose_x_0"), val = tensor(false)]; + tensor attn_29_transpose_y_0 = const()[name = tensor("attn_29_transpose_y_0"), val = tensor(false)]; + tensor values_29_cast_fp16 = transpose(perm = var_3809, x = new_v_cache_29_cast_fp16)[name = tensor("transpose_52")]; + tensor attn_29_cast_fp16 = matmul(transpose_x = attn_29_transpose_x_0, transpose_y = attn_29_transpose_y_0, x = weights_29_cast_fp16, y = values_29_cast_fp16)[name = tensor("attn_29_cast_fp16")]; + tensor var_3854 = const()[name = tensor("op_3854"), val = tensor([0, 2, 1, 3])]; + tensor var_3859 = const()[name = tensor("op_3859"), val = tensor([1, 1, 1024])]; + tensor var_3855_cast_fp16 = transpose(perm = var_3854, x = attn_29_cast_fp16)[name = tensor("transpose_51")]; + tensor input_175_cast_fp16 = reshape(shape = var_3859, x = var_3855_cast_fp16)[name = tensor("input_175_cast_fp16")]; + tensor attn14_out_proj_weight_to_fp16 = const()[name = tensor("attn14_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(358822592)))]; + tensor linear_58_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn14_out_proj_weight_to_fp16, x = input_175_cast_fp16)[name = tensor("linear_58_cast_fp16")]; + tensor var_3865_to_fp16 = const()[name = tensor("op_3865_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3866_cast_fp16 = add(x = position14_to_fp16, y = var_3865_to_fp16)[name = tensor("op_3866_cast_fp16")]; + tensor var_3866_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3866_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_177_cast_fp16 = add(x = input_171_cast_fp16, y = linear_58_cast_fp16)[name = tensor("input_177_cast_fp16")]; + tensor input_179_axes_0 = const()[name = tensor("input_179_axes_0"), val = tensor([-1])]; + tensor norm14_2_weight_to_fp16 = const()[name = tensor("norm14_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(360919808)))]; + tensor norm14_2_bias_to_fp16 = const()[name = tensor("norm14_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(360921920)))]; + tensor var_3870_to_fp16 = const()[name = tensor("op_3870_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_179_cast_fp16 = layer_norm(axes = input_179_axes_0, beta = norm14_2_bias_to_fp16, epsilon = var_3870_to_fp16, gamma = norm14_2_weight_to_fp16, x = input_177_cast_fp16)[name = tensor("input_179_cast_fp16")]; + tensor linear14_1_weight_to_fp16 = const()[name = tensor("linear14_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(360924032)))]; + tensor linear_59_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear14_1_weight_to_fp16, x = input_179_cast_fp16)[name = tensor("linear_59_cast_fp16")]; + tensor input_181_mode_0 = const()[name = tensor("input_181_mode_0"), val = tensor("EXACT")]; + tensor input_181_cast_fp16 = gelu(mode = input_181_mode_0, x = linear_59_cast_fp16)[name = tensor("input_181_cast_fp16")]; + tensor linear14_2_weight_to_fp16 = const()[name = tensor("linear14_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(369312704)))]; + tensor linear_60_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear14_2_weight_to_fp16, x = input_181_cast_fp16)[name = tensor("linear_60_cast_fp16")]; + tensor input_183_cast_fp16 = add(x = input_177_cast_fp16, y = linear_60_cast_fp16)[name = tensor("input_183_cast_fp16")]; + tensor input_185_axes_0 = const()[name = tensor("input_185_axes_0"), val = tensor([-1])]; + tensor norm15_1_weight_to_fp16 = const()[name = tensor("norm15_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(377701376)))]; + tensor norm15_1_bias_to_fp16 = const()[name = tensor("norm15_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(377703488)))]; + tensor var_3887_to_fp16 = const()[name = tensor("op_3887_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_185_cast_fp16 = layer_norm(axes = input_185_axes_0, beta = norm15_1_bias_to_fp16, epsilon = var_3887_to_fp16, gamma = norm15_1_weight_to_fp16, x = input_183_cast_fp16)[name = tensor("input_185_cast_fp16")]; + tensor attn15_in_proj_weight_to_fp16 = const()[name = tensor("attn15_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(377705600)))]; + tensor linear_61_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn15_in_proj_weight_to_fp16, x = input_185_cast_fp16)[name = tensor("linear_61_cast_fp16")]; + tensor var_3900 = const()[name = tensor("op_3900"), val = tensor([1, 3, 16, 64])]; + tensor qkv_31_cast_fp16 = reshape(shape = var_3900, x = linear_61_cast_fp16)[name = tensor("qkv_31_cast_fp16")]; + tensor x_61_begin_0 = const()[name = tensor("x_61_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_61_end_0 = const()[name = tensor("x_61_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_61_end_mask_0 = const()[name = tensor("x_61_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_61_squeeze_mask_0 = const()[name = tensor("x_61_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_61_cast_fp16 = slice_by_index(begin = x_61_begin_0, end = x_61_end_0, end_mask = x_61_end_mask_0, squeeze_mask = x_61_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("x_61_cast_fp16")]; + tensor x_63_begin_0 = const()[name = tensor("x_63_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_63_end_0 = const()[name = tensor("x_63_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_63_end_mask_0 = const()[name = tensor("x_63_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_63_squeeze_mask_0 = const()[name = tensor("x_63_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_63_cast_fp16 = slice_by_index(begin = x_63_begin_0, end = x_63_end_0, end_mask = x_63_end_mask_0, squeeze_mask = x_63_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("x_63_cast_fp16")]; + tensor v_31_begin_0 = const()[name = tensor("v_31_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_31_end_0 = const()[name = tensor("v_31_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_31_end_mask_0 = const()[name = tensor("v_31_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_31_squeeze_mask_0 = const()[name = tensor("v_31_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_31_cast_fp16 = slice_by_index(begin = v_31_begin_0, end = v_31_end_0, end_mask = v_31_end_mask_0, squeeze_mask = v_31_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("v_31_cast_fp16")]; + tensor var_3929 = const()[name = tensor("op_3929"), val = tensor([1, 1, 1])]; + tensor position15_to_fp16_dtype_0 = const()[name = tensor("position15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position15_to_fp16 = cast(dtype = position15_to_fp16_dtype_0, x = position15)[name = tensor("cast_457")]; + tensor var_3930_cast_fp16 = reshape(shape = var_3929, x = position15_to_fp16)[name = tensor("op_3930_cast_fp16")]; + tensor angles_61_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_3930_cast_fp16)[name = tensor("angles_61_cast_fp16")]; + tensor rotr_61_cast_fp16 = cos(x = angles_61_cast_fp16)[name = tensor("rotr_61_cast_fp16")]; + tensor roti_61_cast_fp16 = sin(x = angles_61_cast_fp16)[name = tensor("roti_61_cast_fp16")]; + tensor var_3938 = const()[name = tensor("op_3938"), val = tensor([1, 16, 32, 2])]; + tensor pairs_61_cast_fp16 = reshape(shape = var_3938, x = x_61_cast_fp16)[name = tensor("pairs_61_cast_fp16")]; + tensor xr_61_begin_0 = const()[name = tensor("xr_61_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_61_end_0 = const()[name = tensor("xr_61_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_61_end_mask_0 = const()[name = tensor("xr_61_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_61_squeeze_mask_0 = const()[name = tensor("xr_61_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_61_cast_fp16 = slice_by_index(begin = xr_61_begin_0, end = xr_61_end_0, end_mask = xr_61_end_mask_0, squeeze_mask = xr_61_squeeze_mask_0, x = pairs_61_cast_fp16)[name = tensor("xr_61_cast_fp16")]; + tensor xi_61_begin_0 = const()[name = tensor("xi_61_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_61_end_0 = const()[name = tensor("xi_61_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_61_end_mask_0 = const()[name = tensor("xi_61_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_61_squeeze_mask_0 = const()[name = tensor("xi_61_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_61_cast_fp16 = slice_by_index(begin = xi_61_begin_0, end = xi_61_end_0, end_mask = xi_61_end_mask_0, squeeze_mask = xi_61_squeeze_mask_0, x = pairs_61_cast_fp16)[name = tensor("xi_61_cast_fp16")]; + tensor var_3946_cast_fp16 = mul(x = xr_61_cast_fp16, y = rotr_61_cast_fp16)[name = tensor("op_3946_cast_fp16")]; + tensor var_3947_cast_fp16 = mul(x = xi_61_cast_fp16, y = roti_61_cast_fp16)[name = tensor("op_3947_cast_fp16")]; + tensor xor__61_cast_fp16 = sub(x = var_3946_cast_fp16, y = var_3947_cast_fp16)[name = tensor("xor__61_cast_fp16")]; + tensor var_3950_cast_fp16 = mul(x = xr_61_cast_fp16, y = roti_61_cast_fp16)[name = tensor("op_3950_cast_fp16")]; + tensor var_3951_cast_fp16 = mul(x = xi_61_cast_fp16, y = rotr_61_cast_fp16)[name = tensor("op_3951_cast_fp16")]; + tensor xoi_61_cast_fp16 = add(x = var_3950_cast_fp16, y = var_3951_cast_fp16)[name = tensor("xoi_61_cast_fp16")]; + tensor out_61_axis_0 = const()[name = tensor("out_61_axis_0"), val = tensor(-1)]; + tensor out_61_cast_fp16 = stack(axis = out_61_axis_0, values = (xor__61_cast_fp16, xoi_61_cast_fp16))[name = tensor("out_61_cast_fp16")]; + tensor var_3974 = const()[name = tensor("op_3974"), val = tensor([1, 16, 32, 2])]; + tensor pairs_63_cast_fp16 = reshape(shape = var_3974, x = x_63_cast_fp16)[name = tensor("pairs_63_cast_fp16")]; + tensor xr_63_begin_0 = const()[name = tensor("xr_63_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_63_end_0 = const()[name = tensor("xr_63_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_63_end_mask_0 = const()[name = tensor("xr_63_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_63_squeeze_mask_0 = const()[name = tensor("xr_63_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_63_cast_fp16 = slice_by_index(begin = xr_63_begin_0, end = xr_63_end_0, end_mask = xr_63_end_mask_0, squeeze_mask = xr_63_squeeze_mask_0, x = pairs_63_cast_fp16)[name = tensor("xr_63_cast_fp16")]; + tensor xi_63_begin_0 = const()[name = tensor("xi_63_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_63_end_0 = const()[name = tensor("xi_63_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_63_end_mask_0 = const()[name = tensor("xi_63_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_63_squeeze_mask_0 = const()[name = tensor("xi_63_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_63_cast_fp16 = slice_by_index(begin = xi_63_begin_0, end = xi_63_end_0, end_mask = xi_63_end_mask_0, squeeze_mask = xi_63_squeeze_mask_0, x = pairs_63_cast_fp16)[name = tensor("xi_63_cast_fp16")]; + tensor var_3982_cast_fp16 = mul(x = xr_63_cast_fp16, y = rotr_61_cast_fp16)[name = tensor("op_3982_cast_fp16")]; + tensor var_3983_cast_fp16 = mul(x = xi_63_cast_fp16, y = roti_61_cast_fp16)[name = tensor("op_3983_cast_fp16")]; + tensor xor__63_cast_fp16 = sub(x = var_3982_cast_fp16, y = var_3983_cast_fp16)[name = tensor("xor__63_cast_fp16")]; + tensor var_3986_cast_fp16 = mul(x = xr_63_cast_fp16, y = roti_61_cast_fp16)[name = tensor("op_3986_cast_fp16")]; + tensor var_3987_cast_fp16 = mul(x = xi_63_cast_fp16, y = rotr_61_cast_fp16)[name = tensor("op_3987_cast_fp16")]; + tensor xoi_63_cast_fp16 = add(x = var_3986_cast_fp16, y = var_3987_cast_fp16)[name = tensor("xoi_63_cast_fp16")]; + tensor out_63_axis_0 = const()[name = tensor("out_63_axis_0"), val = tensor(-1)]; + tensor out_63_cast_fp16 = stack(axis = out_63_axis_0, values = (xor__63_cast_fp16, xoi_63_cast_fp16))[name = tensor("out_63_cast_fp16")]; + tensor _inversed_3999_y_0_to_fp16 = const()[name = tensor("_inversed_3999_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3999_cast_fp16 = mul(x = position15_to_fp16, y = _inversed_3999_y_0_to_fp16)[name = tensor("_inversed_3999_cast_fp16")]; + tensor var_4000_cast_fp16 = floor(x = _inversed_3999_cast_fp16)[name = tensor("op_4000_cast_fp16")]; + tensor var_4001_to_fp16 = const()[name = tensor("op_4001_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4002_cast_fp16 = mul(x = var_4000_cast_fp16, y = var_4001_to_fp16)[name = tensor("op_4002_cast_fp16")]; + tensor write_idx_31_cast_fp16 = sub(x = position15_to_fp16, y = var_4002_cast_fp16)[name = tensor("write_idx_31_cast_fp16")]; + tensor var_4005_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_31_cast_fp16)[name = tensor("op_4005_cast_fp16")]; + tensor var_4015 = const()[name = tensor("op_4015"), val = tensor([1, 512, 1, 1])]; + tensor onehot_61_to_fp16_dtype_0 = const()[name = tensor("onehot_61_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_4005_cast_fp16_to_fp16 = cast(dtype = onehot_61_to_fp16_dtype_0, x = var_4005_cast_fp16)[name = tensor("cast_456")]; + tensor onehot_63_cast_fp16 = reshape(shape = var_4015, x = var_4005_cast_fp16_to_fp16)[name = tensor("onehot_63_cast_fp16")]; + tensor var_4017_to_fp16 = const()[name = tensor("op_4017_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4019_cast_fp16 = sub(x = var_4017_to_fp16, y = onehot_63_cast_fp16)[name = tensor("op_4019_cast_fp16")]; + tensor k_cache15_to_fp16_dtype_0 = const()[name = tensor("k_cache15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache15_to_fp16 = cast(dtype = k_cache15_to_fp16_dtype_0, x = k_cache15)[name = tensor("cast_455")]; + tensor var_4020_cast_fp16 = mul(x = k_cache15_to_fp16, y = var_4019_cast_fp16)[name = tensor("op_4020_cast_fp16")]; + tensor var_4025 = const()[name = tensor("op_4025"), val = tensor([1, 1, 16, 64])]; + tensor var_4026_cast_fp16 = reshape(shape = var_4025, x = out_63_cast_fp16)[name = tensor("op_4026_cast_fp16")]; + tensor var_4027_cast_fp16 = mul(x = var_4026_cast_fp16, y = onehot_63_cast_fp16)[name = tensor("op_4027_cast_fp16")]; + tensor new_k_cache_31_cast_fp16 = add(x = var_4020_cast_fp16, y = var_4027_cast_fp16)[name = tensor("new_k_cache_31_cast_fp16")]; + tensor new_k_cache_31_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_31_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache15_to_fp16_dtype_0 = const()[name = tensor("v_cache15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache15_to_fp16 = cast(dtype = v_cache15_to_fp16_dtype_0, x = v_cache15)[name = tensor("cast_453")]; + tensor var_4033_cast_fp16 = mul(x = v_cache15_to_fp16, y = var_4019_cast_fp16)[name = tensor("op_4033_cast_fp16")]; + tensor var_4038 = const()[name = tensor("op_4038"), val = tensor([1, 1, 16, 64])]; + tensor var_4039_cast_fp16 = reshape(shape = var_4038, x = v_31_cast_fp16)[name = tensor("op_4039_cast_fp16")]; + tensor var_4040_cast_fp16 = mul(x = var_4039_cast_fp16, y = onehot_63_cast_fp16)[name = tensor("op_4040_cast_fp16")]; + tensor new_v_cache_31_cast_fp16 = add(x = var_4033_cast_fp16, y = var_4040_cast_fp16)[name = tensor("new_v_cache_31_cast_fp16")]; + tensor new_v_cache_31_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_31_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_15_perm_0 = const()[name = tensor("transpose_15_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_4053 = const()[name = tensor("op_4053"), val = tensor([0, 2, 1, 3])]; + tensor var_4059 = const()[name = tensor("op_4059"), val = tensor([1, 16, 1, 64])]; + tensor q4_31_cast_fp16 = reshape(shape = var_4059, x = out_61_cast_fp16)[name = tensor("q4_31_cast_fp16")]; + tensor var_4064_transpose_x_0 = const()[name = tensor("op_4064_transpose_x_0"), val = tensor(false)]; + tensor var_4064_transpose_y_0 = const()[name = tensor("op_4064_transpose_y_0"), val = tensor(false)]; + tensor transpose_15_cast_fp16 = transpose(perm = transpose_15_perm_0, x = new_k_cache_31_cast_fp16)[name = tensor("transpose_50")]; + tensor var_4064_cast_fp16 = matmul(transpose_x = var_4064_transpose_x_0, transpose_y = var_4064_transpose_y_0, x = q4_31_cast_fp16, y = transpose_15_cast_fp16)[name = tensor("op_4064_cast_fp16")]; + tensor var_4065_to_fp16 = const()[name = tensor("op_4065_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_61_cast_fp16 = mul(x = var_4064_cast_fp16, y = var_4065_to_fp16)[name = tensor("scores_61_cast_fp16")]; + tensor var_4069 = const()[name = tensor("op_4069"), val = tensor([1, 1])]; + tensor var_4070_cast_fp16 = reshape(shape = var_4069, x = position15_to_fp16)[name = tensor("op_4070_cast_fp16")]; + tensor var_4071_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_4070_cast_fp16)[name = tensor("op_4071_cast_fp16")]; + tensor mask_31_to_fp16_dtype_0 = const()[name = tensor("mask_31_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_4078_to_fp16 = const()[name = tensor("op_4078_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4071_cast_fp16_to_fp16 = cast(dtype = mask_31_to_fp16_dtype_0, x = var_4071_cast_fp16)[name = tensor("cast_451")]; + tensor var_4079_cast_fp16 = sub(x = var_4071_cast_fp16_to_fp16, y = var_4078_to_fp16)[name = tensor("op_4079_cast_fp16")]; + tensor var_4084 = const()[name = tensor("op_4084"), val = tensor([1, 1, 1, 512])]; + tensor var_4085_cast_fp16 = reshape(shape = var_4084, x = var_4079_cast_fp16)[name = tensor("op_4085_cast_fp16")]; + tensor var_4086_to_fp16 = const()[name = tensor("op_4086_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_4087_cast_fp16 = mul(x = var_4085_cast_fp16, y = var_4086_to_fp16)[name = tensor("op_4087_cast_fp16")]; + tensor scores_63_cast_fp16 = add(x = scores_61_cast_fp16, y = var_4087_cast_fp16)[name = tensor("scores_63_cast_fp16")]; + tensor var_4090 = const()[name = tensor("op_4090"), val = tensor(-1)]; + tensor weights_31_cast_fp16 = softmax(axis = var_4090, x = scores_63_cast_fp16)[name = tensor("weights_31_cast_fp16")]; + tensor attn_31_transpose_x_0 = const()[name = tensor("attn_31_transpose_x_0"), val = tensor(false)]; + tensor attn_31_transpose_y_0 = const()[name = tensor("attn_31_transpose_y_0"), val = tensor(false)]; + tensor values_31_cast_fp16 = transpose(perm = var_4053, x = new_v_cache_31_cast_fp16)[name = tensor("transpose_49")]; + tensor attn_31_cast_fp16 = matmul(transpose_x = attn_31_transpose_x_0, transpose_y = attn_31_transpose_y_0, x = weights_31_cast_fp16, y = values_31_cast_fp16)[name = tensor("attn_31_cast_fp16")]; + tensor var_4098 = const()[name = tensor("op_4098"), val = tensor([0, 2, 1, 3])]; + tensor var_4103 = const()[name = tensor("op_4103"), val = tensor([1, 1, 1024])]; + tensor var_4099_cast_fp16 = transpose(perm = var_4098, x = attn_31_cast_fp16)[name = tensor("transpose_48")]; + tensor input_187_cast_fp16 = reshape(shape = var_4103, x = var_4099_cast_fp16)[name = tensor("input_187_cast_fp16")]; + tensor attn15_out_proj_weight_to_fp16 = const()[name = tensor("attn15_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(383997120)))]; + tensor linear_62_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn15_out_proj_weight_to_fp16, x = input_187_cast_fp16)[name = tensor("linear_62_cast_fp16")]; + tensor var_4109_to_fp16 = const()[name = tensor("op_4109_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4110_cast_fp16 = add(x = position15_to_fp16, y = var_4109_to_fp16)[name = tensor("op_4110_cast_fp16")]; + tensor var_4110_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4110_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_189_cast_fp16 = add(x = input_183_cast_fp16, y = linear_62_cast_fp16)[name = tensor("input_189_cast_fp16")]; + tensor input_191_axes_0 = const()[name = tensor("input_191_axes_0"), val = tensor([-1])]; + tensor norm15_2_weight_to_fp16 = const()[name = tensor("norm15_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(386094336)))]; + tensor norm15_2_bias_to_fp16 = const()[name = tensor("norm15_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(386096448)))]; + tensor var_4114_to_fp16 = const()[name = tensor("op_4114_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_191_cast_fp16 = layer_norm(axes = input_191_axes_0, beta = norm15_2_bias_to_fp16, epsilon = var_4114_to_fp16, gamma = norm15_2_weight_to_fp16, x = input_189_cast_fp16)[name = tensor("input_191_cast_fp16")]; + tensor linear15_1_weight_to_fp16 = const()[name = tensor("linear15_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(386098560)))]; + tensor linear_63_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear15_1_weight_to_fp16, x = input_191_cast_fp16)[name = tensor("linear_63_cast_fp16")]; + tensor input_193_mode_0 = const()[name = tensor("input_193_mode_0"), val = tensor("EXACT")]; + tensor input_193_cast_fp16 = gelu(mode = input_193_mode_0, x = linear_63_cast_fp16)[name = tensor("input_193_cast_fp16")]; + tensor linear15_2_weight_to_fp16 = const()[name = tensor("linear15_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(394487232)))]; + tensor linear_64_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear15_2_weight_to_fp16, x = input_193_cast_fp16)[name = tensor("linear_64_cast_fp16")]; + tensor input_195_cast_fp16 = add(x = input_189_cast_fp16, y = linear_64_cast_fp16)[name = tensor("input_195_cast_fp16")]; + tensor input_197_axes_0 = const()[name = tensor("input_197_axes_0"), val = tensor([-1])]; + tensor norm16_1_weight_to_fp16 = const()[name = tensor("norm16_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(402875904)))]; + tensor norm16_1_bias_to_fp16 = const()[name = tensor("norm16_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(402878016)))]; + tensor var_4131_to_fp16 = const()[name = tensor("op_4131_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_197_cast_fp16 = layer_norm(axes = input_197_axes_0, beta = norm16_1_bias_to_fp16, epsilon = var_4131_to_fp16, gamma = norm16_1_weight_to_fp16, x = input_195_cast_fp16)[name = tensor("input_197_cast_fp16")]; + tensor attn16_in_proj_weight_to_fp16 = const()[name = tensor("attn16_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(402880128)))]; + tensor linear_65_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn16_in_proj_weight_to_fp16, x = input_197_cast_fp16)[name = tensor("linear_65_cast_fp16")]; + tensor var_4144 = const()[name = tensor("op_4144"), val = tensor([1, 3, 16, 64])]; + tensor qkv_33_cast_fp16 = reshape(shape = var_4144, x = linear_65_cast_fp16)[name = tensor("qkv_33_cast_fp16")]; + tensor x_65_begin_0 = const()[name = tensor("x_65_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_65_end_0 = const()[name = tensor("x_65_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_65_end_mask_0 = const()[name = tensor("x_65_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_65_squeeze_mask_0 = const()[name = tensor("x_65_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_65_cast_fp16 = slice_by_index(begin = x_65_begin_0, end = x_65_end_0, end_mask = x_65_end_mask_0, squeeze_mask = x_65_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("x_65_cast_fp16")]; + tensor x_67_begin_0 = const()[name = tensor("x_67_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_67_end_0 = const()[name = tensor("x_67_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_67_end_mask_0 = const()[name = tensor("x_67_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_67_squeeze_mask_0 = const()[name = tensor("x_67_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_67_cast_fp16 = slice_by_index(begin = x_67_begin_0, end = x_67_end_0, end_mask = x_67_end_mask_0, squeeze_mask = x_67_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("x_67_cast_fp16")]; + tensor v_33_begin_0 = const()[name = tensor("v_33_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_33_end_0 = const()[name = tensor("v_33_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_33_end_mask_0 = const()[name = tensor("v_33_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_33_squeeze_mask_0 = const()[name = tensor("v_33_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_33_cast_fp16 = slice_by_index(begin = v_33_begin_0, end = v_33_end_0, end_mask = v_33_end_mask_0, squeeze_mask = v_33_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("v_33_cast_fp16")]; + tensor var_4173 = const()[name = tensor("op_4173"), val = tensor([1, 1, 1])]; + tensor position16_to_fp16_dtype_0 = const()[name = tensor("position16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position16_to_fp16 = cast(dtype = position16_to_fp16_dtype_0, x = position16)[name = tensor("cast_449")]; + tensor var_4174_cast_fp16 = reshape(shape = var_4173, x = position16_to_fp16)[name = tensor("op_4174_cast_fp16")]; + tensor angles_65_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_4174_cast_fp16)[name = tensor("angles_65_cast_fp16")]; + tensor rotr_65_cast_fp16 = cos(x = angles_65_cast_fp16)[name = tensor("rotr_65_cast_fp16")]; + tensor roti_65_cast_fp16 = sin(x = angles_65_cast_fp16)[name = tensor("roti_65_cast_fp16")]; + tensor var_4182 = const()[name = tensor("op_4182"), val = tensor([1, 16, 32, 2])]; + tensor pairs_65_cast_fp16 = reshape(shape = var_4182, x = x_65_cast_fp16)[name = tensor("pairs_65_cast_fp16")]; + tensor xr_65_begin_0 = const()[name = tensor("xr_65_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_65_end_0 = const()[name = tensor("xr_65_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_65_end_mask_0 = const()[name = tensor("xr_65_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_65_squeeze_mask_0 = const()[name = tensor("xr_65_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_65_cast_fp16 = slice_by_index(begin = xr_65_begin_0, end = xr_65_end_0, end_mask = xr_65_end_mask_0, squeeze_mask = xr_65_squeeze_mask_0, x = pairs_65_cast_fp16)[name = tensor("xr_65_cast_fp16")]; + tensor xi_65_begin_0 = const()[name = tensor("xi_65_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_65_end_0 = const()[name = tensor("xi_65_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_65_end_mask_0 = const()[name = tensor("xi_65_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_65_squeeze_mask_0 = const()[name = tensor("xi_65_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_65_cast_fp16 = slice_by_index(begin = xi_65_begin_0, end = xi_65_end_0, end_mask = xi_65_end_mask_0, squeeze_mask = xi_65_squeeze_mask_0, x = pairs_65_cast_fp16)[name = tensor("xi_65_cast_fp16")]; + tensor var_4190_cast_fp16 = mul(x = xr_65_cast_fp16, y = rotr_65_cast_fp16)[name = tensor("op_4190_cast_fp16")]; + tensor var_4191_cast_fp16 = mul(x = xi_65_cast_fp16, y = roti_65_cast_fp16)[name = tensor("op_4191_cast_fp16")]; + tensor xor__65_cast_fp16 = sub(x = var_4190_cast_fp16, y = var_4191_cast_fp16)[name = tensor("xor__65_cast_fp16")]; + tensor var_4194_cast_fp16 = mul(x = xr_65_cast_fp16, y = roti_65_cast_fp16)[name = tensor("op_4194_cast_fp16")]; + tensor var_4195_cast_fp16 = mul(x = xi_65_cast_fp16, y = rotr_65_cast_fp16)[name = tensor("op_4195_cast_fp16")]; + tensor xoi_65_cast_fp16 = add(x = var_4194_cast_fp16, y = var_4195_cast_fp16)[name = tensor("xoi_65_cast_fp16")]; + tensor out_65_axis_0 = const()[name = tensor("out_65_axis_0"), val = tensor(-1)]; + tensor out_65_cast_fp16 = stack(axis = out_65_axis_0, values = (xor__65_cast_fp16, xoi_65_cast_fp16))[name = tensor("out_65_cast_fp16")]; + tensor var_4218 = const()[name = tensor("op_4218"), val = tensor([1, 16, 32, 2])]; + tensor pairs_67_cast_fp16 = reshape(shape = var_4218, x = x_67_cast_fp16)[name = tensor("pairs_67_cast_fp16")]; + tensor xr_67_begin_0 = const()[name = tensor("xr_67_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_67_end_0 = const()[name = tensor("xr_67_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_67_end_mask_0 = const()[name = tensor("xr_67_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_67_squeeze_mask_0 = const()[name = tensor("xr_67_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_67_cast_fp16 = slice_by_index(begin = xr_67_begin_0, end = xr_67_end_0, end_mask = xr_67_end_mask_0, squeeze_mask = xr_67_squeeze_mask_0, x = pairs_67_cast_fp16)[name = tensor("xr_67_cast_fp16")]; + tensor xi_67_begin_0 = const()[name = tensor("xi_67_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_67_end_0 = const()[name = tensor("xi_67_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_67_end_mask_0 = const()[name = tensor("xi_67_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_67_squeeze_mask_0 = const()[name = tensor("xi_67_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_67_cast_fp16 = slice_by_index(begin = xi_67_begin_0, end = xi_67_end_0, end_mask = xi_67_end_mask_0, squeeze_mask = xi_67_squeeze_mask_0, x = pairs_67_cast_fp16)[name = tensor("xi_67_cast_fp16")]; + tensor var_4226_cast_fp16 = mul(x = xr_67_cast_fp16, y = rotr_65_cast_fp16)[name = tensor("op_4226_cast_fp16")]; + tensor var_4227_cast_fp16 = mul(x = xi_67_cast_fp16, y = roti_65_cast_fp16)[name = tensor("op_4227_cast_fp16")]; + tensor xor__67_cast_fp16 = sub(x = var_4226_cast_fp16, y = var_4227_cast_fp16)[name = tensor("xor__67_cast_fp16")]; + tensor var_4230_cast_fp16 = mul(x = xr_67_cast_fp16, y = roti_65_cast_fp16)[name = tensor("op_4230_cast_fp16")]; + tensor var_4231_cast_fp16 = mul(x = xi_67_cast_fp16, y = rotr_65_cast_fp16)[name = tensor("op_4231_cast_fp16")]; + tensor xoi_67_cast_fp16 = add(x = var_4230_cast_fp16, y = var_4231_cast_fp16)[name = tensor("xoi_67_cast_fp16")]; + tensor out_67_axis_0 = const()[name = tensor("out_67_axis_0"), val = tensor(-1)]; + tensor out_67_cast_fp16 = stack(axis = out_67_axis_0, values = (xor__67_cast_fp16, xoi_67_cast_fp16))[name = tensor("out_67_cast_fp16")]; + tensor _inversed_4243_y_0_to_fp16 = const()[name = tensor("_inversed_4243_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_4243_cast_fp16 = mul(x = position16_to_fp16, y = _inversed_4243_y_0_to_fp16)[name = tensor("_inversed_4243_cast_fp16")]; + tensor var_4244_cast_fp16 = floor(x = _inversed_4243_cast_fp16)[name = tensor("op_4244_cast_fp16")]; + tensor var_4245_to_fp16 = const()[name = tensor("op_4245_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4246_cast_fp16 = mul(x = var_4244_cast_fp16, y = var_4245_to_fp16)[name = tensor("op_4246_cast_fp16")]; + tensor write_idx_33_cast_fp16 = sub(x = position16_to_fp16, y = var_4246_cast_fp16)[name = tensor("write_idx_33_cast_fp16")]; + tensor var_4249_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_33_cast_fp16)[name = tensor("op_4249_cast_fp16")]; + tensor var_4259 = const()[name = tensor("op_4259"), val = tensor([1, 512, 1, 1])]; + tensor onehot_65_to_fp16_dtype_0 = const()[name = tensor("onehot_65_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_4249_cast_fp16_to_fp16 = cast(dtype = onehot_65_to_fp16_dtype_0, x = var_4249_cast_fp16)[name = tensor("cast_448")]; + tensor onehot_67_cast_fp16 = reshape(shape = var_4259, x = var_4249_cast_fp16_to_fp16)[name = tensor("onehot_67_cast_fp16")]; + tensor var_4261_to_fp16 = const()[name = tensor("op_4261_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4263_cast_fp16 = sub(x = var_4261_to_fp16, y = onehot_67_cast_fp16)[name = tensor("op_4263_cast_fp16")]; + tensor k_cache16_to_fp16_dtype_0 = const()[name = tensor("k_cache16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache16_to_fp16 = cast(dtype = k_cache16_to_fp16_dtype_0, x = k_cache16)[name = tensor("cast_447")]; + tensor var_4264_cast_fp16 = mul(x = k_cache16_to_fp16, y = var_4263_cast_fp16)[name = tensor("op_4264_cast_fp16")]; + tensor var_4269 = const()[name = tensor("op_4269"), val = tensor([1, 1, 16, 64])]; + tensor var_4270_cast_fp16 = reshape(shape = var_4269, x = out_67_cast_fp16)[name = tensor("op_4270_cast_fp16")]; + tensor var_4271_cast_fp16 = mul(x = var_4270_cast_fp16, y = onehot_67_cast_fp16)[name = tensor("op_4271_cast_fp16")]; + tensor new_k_cache_33_cast_fp16 = add(x = var_4264_cast_fp16, y = var_4271_cast_fp16)[name = tensor("new_k_cache_33_cast_fp16")]; + tensor new_k_cache_33_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_33_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache16_to_fp16_dtype_0 = const()[name = tensor("v_cache16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache16_to_fp16 = cast(dtype = v_cache16_to_fp16_dtype_0, x = v_cache16)[name = tensor("cast_445")]; + tensor var_4277_cast_fp16 = mul(x = v_cache16_to_fp16, y = var_4263_cast_fp16)[name = tensor("op_4277_cast_fp16")]; + tensor var_4282 = const()[name = tensor("op_4282"), val = tensor([1, 1, 16, 64])]; + tensor var_4283_cast_fp16 = reshape(shape = var_4282, x = v_33_cast_fp16)[name = tensor("op_4283_cast_fp16")]; + tensor var_4284_cast_fp16 = mul(x = var_4283_cast_fp16, y = onehot_67_cast_fp16)[name = tensor("op_4284_cast_fp16")]; + tensor new_v_cache_33_cast_fp16 = add(x = var_4277_cast_fp16, y = var_4284_cast_fp16)[name = tensor("new_v_cache_33_cast_fp16")]; + tensor new_v_cache_33_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_33_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_16_perm_0 = const()[name = tensor("transpose_16_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_4297 = const()[name = tensor("op_4297"), val = tensor([0, 2, 1, 3])]; + tensor var_4303 = const()[name = tensor("op_4303"), val = tensor([1, 16, 1, 64])]; + tensor q4_33_cast_fp16 = reshape(shape = var_4303, x = out_65_cast_fp16)[name = tensor("q4_33_cast_fp16")]; + tensor var_4308_transpose_x_0 = const()[name = tensor("op_4308_transpose_x_0"), val = tensor(false)]; + tensor var_4308_transpose_y_0 = const()[name = tensor("op_4308_transpose_y_0"), val = tensor(false)]; + tensor transpose_16_cast_fp16 = transpose(perm = transpose_16_perm_0, x = new_k_cache_33_cast_fp16)[name = tensor("transpose_47")]; + tensor var_4308_cast_fp16 = matmul(transpose_x = var_4308_transpose_x_0, transpose_y = var_4308_transpose_y_0, x = q4_33_cast_fp16, y = transpose_16_cast_fp16)[name = tensor("op_4308_cast_fp16")]; + tensor var_4309_to_fp16 = const()[name = tensor("op_4309_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_65_cast_fp16 = mul(x = var_4308_cast_fp16, y = var_4309_to_fp16)[name = tensor("scores_65_cast_fp16")]; + tensor var_4313 = const()[name = tensor("op_4313"), val = tensor([1, 1])]; + tensor var_4314_cast_fp16 = reshape(shape = var_4313, x = position16_to_fp16)[name = tensor("op_4314_cast_fp16")]; + tensor var_4315_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_4314_cast_fp16)[name = tensor("op_4315_cast_fp16")]; + tensor mask_33_to_fp16_dtype_0 = const()[name = tensor("mask_33_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_4322_to_fp16 = const()[name = tensor("op_4322_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4315_cast_fp16_to_fp16 = cast(dtype = mask_33_to_fp16_dtype_0, x = var_4315_cast_fp16)[name = tensor("cast_443")]; + tensor var_4323_cast_fp16 = sub(x = var_4315_cast_fp16_to_fp16, y = var_4322_to_fp16)[name = tensor("op_4323_cast_fp16")]; + tensor var_4328 = const()[name = tensor("op_4328"), val = tensor([1, 1, 1, 512])]; + tensor var_4329_cast_fp16 = reshape(shape = var_4328, x = var_4323_cast_fp16)[name = tensor("op_4329_cast_fp16")]; + tensor var_4330_to_fp16 = const()[name = tensor("op_4330_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_4331_cast_fp16 = mul(x = var_4329_cast_fp16, y = var_4330_to_fp16)[name = tensor("op_4331_cast_fp16")]; + tensor scores_67_cast_fp16 = add(x = scores_65_cast_fp16, y = var_4331_cast_fp16)[name = tensor("scores_67_cast_fp16")]; + tensor var_4334 = const()[name = tensor("op_4334"), val = tensor(-1)]; + tensor weights_33_cast_fp16 = softmax(axis = var_4334, x = scores_67_cast_fp16)[name = tensor("weights_33_cast_fp16")]; + tensor attn_33_transpose_x_0 = const()[name = tensor("attn_33_transpose_x_0"), val = tensor(false)]; + tensor attn_33_transpose_y_0 = const()[name = tensor("attn_33_transpose_y_0"), val = tensor(false)]; + tensor values_33_cast_fp16 = transpose(perm = var_4297, x = new_v_cache_33_cast_fp16)[name = tensor("transpose_46")]; + tensor attn_33_cast_fp16 = matmul(transpose_x = attn_33_transpose_x_0, transpose_y = attn_33_transpose_y_0, x = weights_33_cast_fp16, y = values_33_cast_fp16)[name = tensor("attn_33_cast_fp16")]; + tensor var_4342 = const()[name = tensor("op_4342"), val = tensor([0, 2, 1, 3])]; + tensor var_4347 = const()[name = tensor("op_4347"), val = tensor([1, 1, 1024])]; + tensor var_4343_cast_fp16 = transpose(perm = var_4342, x = attn_33_cast_fp16)[name = tensor("transpose_45")]; + tensor input_199_cast_fp16 = reshape(shape = var_4347, x = var_4343_cast_fp16)[name = tensor("input_199_cast_fp16")]; + tensor attn16_out_proj_weight_to_fp16 = const()[name = tensor("attn16_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(409171648)))]; + tensor linear_66_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn16_out_proj_weight_to_fp16, x = input_199_cast_fp16)[name = tensor("linear_66_cast_fp16")]; + tensor var_4353_to_fp16 = const()[name = tensor("op_4353_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4354_cast_fp16 = add(x = position16_to_fp16, y = var_4353_to_fp16)[name = tensor("op_4354_cast_fp16")]; + tensor var_4354_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4354_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_201_cast_fp16 = add(x = input_195_cast_fp16, y = linear_66_cast_fp16)[name = tensor("input_201_cast_fp16")]; + tensor input_203_axes_0 = const()[name = tensor("input_203_axes_0"), val = tensor([-1])]; + tensor norm16_2_weight_to_fp16 = const()[name = tensor("norm16_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411268864)))]; + tensor norm16_2_bias_to_fp16 = const()[name = tensor("norm16_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411270976)))]; + tensor var_4358_to_fp16 = const()[name = tensor("op_4358_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_203_cast_fp16 = layer_norm(axes = input_203_axes_0, beta = norm16_2_bias_to_fp16, epsilon = var_4358_to_fp16, gamma = norm16_2_weight_to_fp16, x = input_201_cast_fp16)[name = tensor("input_203_cast_fp16")]; + tensor linear16_1_weight_to_fp16 = const()[name = tensor("linear16_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411273088)))]; + tensor linear_67_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear16_1_weight_to_fp16, x = input_203_cast_fp16)[name = tensor("linear_67_cast_fp16")]; + tensor input_205_mode_0 = const()[name = tensor("input_205_mode_0"), val = tensor("EXACT")]; + tensor input_205_cast_fp16 = gelu(mode = input_205_mode_0, x = linear_67_cast_fp16)[name = tensor("input_205_cast_fp16")]; + tensor linear16_2_weight_to_fp16 = const()[name = tensor("linear16_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(419661760)))]; + tensor linear_68_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear16_2_weight_to_fp16, x = input_205_cast_fp16)[name = tensor("linear_68_cast_fp16")]; + tensor input_207_cast_fp16 = add(x = input_201_cast_fp16, y = linear_68_cast_fp16)[name = tensor("input_207_cast_fp16")]; + tensor input_209_axes_0 = const()[name = tensor("input_209_axes_0"), val = tensor([-1])]; + tensor norm17_1_weight_to_fp16 = const()[name = tensor("norm17_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(428050432)))]; + tensor norm17_1_bias_to_fp16 = const()[name = tensor("norm17_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(428052544)))]; + tensor var_4375_to_fp16 = const()[name = tensor("op_4375_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_209_cast_fp16 = layer_norm(axes = input_209_axes_0, beta = norm17_1_bias_to_fp16, epsilon = var_4375_to_fp16, gamma = norm17_1_weight_to_fp16, x = input_207_cast_fp16)[name = tensor("input_209_cast_fp16")]; + tensor attn17_in_proj_weight_to_fp16 = const()[name = tensor("attn17_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(428054656)))]; + tensor linear_69_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn17_in_proj_weight_to_fp16, x = input_209_cast_fp16)[name = tensor("linear_69_cast_fp16")]; + tensor var_4388 = const()[name = tensor("op_4388"), val = tensor([1, 3, 16, 64])]; + tensor qkv_35_cast_fp16 = reshape(shape = var_4388, x = linear_69_cast_fp16)[name = tensor("qkv_35_cast_fp16")]; + tensor x_69_begin_0 = const()[name = tensor("x_69_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_69_end_0 = const()[name = tensor("x_69_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_69_end_mask_0 = const()[name = tensor("x_69_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_69_squeeze_mask_0 = const()[name = tensor("x_69_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_69_cast_fp16 = slice_by_index(begin = x_69_begin_0, end = x_69_end_0, end_mask = x_69_end_mask_0, squeeze_mask = x_69_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("x_69_cast_fp16")]; + tensor x_71_begin_0 = const()[name = tensor("x_71_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_71_end_0 = const()[name = tensor("x_71_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_71_end_mask_0 = const()[name = tensor("x_71_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_71_squeeze_mask_0 = const()[name = tensor("x_71_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_71_cast_fp16 = slice_by_index(begin = x_71_begin_0, end = x_71_end_0, end_mask = x_71_end_mask_0, squeeze_mask = x_71_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("x_71_cast_fp16")]; + tensor v_35_begin_0 = const()[name = tensor("v_35_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_35_end_0 = const()[name = tensor("v_35_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_35_end_mask_0 = const()[name = tensor("v_35_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_35_squeeze_mask_0 = const()[name = tensor("v_35_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_35_cast_fp16 = slice_by_index(begin = v_35_begin_0, end = v_35_end_0, end_mask = v_35_end_mask_0, squeeze_mask = v_35_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("v_35_cast_fp16")]; + tensor var_4417 = const()[name = tensor("op_4417"), val = tensor([1, 1, 1])]; + tensor position17_to_fp16_dtype_0 = const()[name = tensor("position17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position17_to_fp16 = cast(dtype = position17_to_fp16_dtype_0, x = position17)[name = tensor("cast_441")]; + tensor var_4418_cast_fp16 = reshape(shape = var_4417, x = position17_to_fp16)[name = tensor("op_4418_cast_fp16")]; + tensor angles_69_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_4418_cast_fp16)[name = tensor("angles_69_cast_fp16")]; + tensor rotr_69_cast_fp16 = cos(x = angles_69_cast_fp16)[name = tensor("rotr_69_cast_fp16")]; + tensor roti_69_cast_fp16 = sin(x = angles_69_cast_fp16)[name = tensor("roti_69_cast_fp16")]; + tensor var_4426 = const()[name = tensor("op_4426"), val = tensor([1, 16, 32, 2])]; + tensor pairs_69_cast_fp16 = reshape(shape = var_4426, x = x_69_cast_fp16)[name = tensor("pairs_69_cast_fp16")]; + tensor xr_69_begin_0 = const()[name = tensor("xr_69_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_69_end_0 = const()[name = tensor("xr_69_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_69_end_mask_0 = const()[name = tensor("xr_69_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_69_squeeze_mask_0 = const()[name = tensor("xr_69_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_69_cast_fp16 = slice_by_index(begin = xr_69_begin_0, end = xr_69_end_0, end_mask = xr_69_end_mask_0, squeeze_mask = xr_69_squeeze_mask_0, x = pairs_69_cast_fp16)[name = tensor("xr_69_cast_fp16")]; + tensor xi_69_begin_0 = const()[name = tensor("xi_69_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_69_end_0 = const()[name = tensor("xi_69_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_69_end_mask_0 = const()[name = tensor("xi_69_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_69_squeeze_mask_0 = const()[name = tensor("xi_69_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_69_cast_fp16 = slice_by_index(begin = xi_69_begin_0, end = xi_69_end_0, end_mask = xi_69_end_mask_0, squeeze_mask = xi_69_squeeze_mask_0, x = pairs_69_cast_fp16)[name = tensor("xi_69_cast_fp16")]; + tensor var_4434_cast_fp16 = mul(x = xr_69_cast_fp16, y = rotr_69_cast_fp16)[name = tensor("op_4434_cast_fp16")]; + tensor var_4435_cast_fp16 = mul(x = xi_69_cast_fp16, y = roti_69_cast_fp16)[name = tensor("op_4435_cast_fp16")]; + tensor xor__69_cast_fp16 = sub(x = var_4434_cast_fp16, y = var_4435_cast_fp16)[name = tensor("xor__69_cast_fp16")]; + tensor var_4438_cast_fp16 = mul(x = xr_69_cast_fp16, y = roti_69_cast_fp16)[name = tensor("op_4438_cast_fp16")]; + tensor var_4439_cast_fp16 = mul(x = xi_69_cast_fp16, y = rotr_69_cast_fp16)[name = tensor("op_4439_cast_fp16")]; + tensor xoi_69_cast_fp16 = add(x = var_4438_cast_fp16, y = var_4439_cast_fp16)[name = tensor("xoi_69_cast_fp16")]; + tensor out_69_axis_0 = const()[name = tensor("out_69_axis_0"), val = tensor(-1)]; + tensor out_69_cast_fp16 = stack(axis = out_69_axis_0, values = (xor__69_cast_fp16, xoi_69_cast_fp16))[name = tensor("out_69_cast_fp16")]; + tensor var_4462 = const()[name = tensor("op_4462"), val = tensor([1, 16, 32, 2])]; + tensor pairs_71_cast_fp16 = reshape(shape = var_4462, x = x_71_cast_fp16)[name = tensor("pairs_71_cast_fp16")]; + tensor xr_71_begin_0 = const()[name = tensor("xr_71_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_71_end_0 = const()[name = tensor("xr_71_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_71_end_mask_0 = const()[name = tensor("xr_71_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_71_squeeze_mask_0 = const()[name = tensor("xr_71_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_71_cast_fp16 = slice_by_index(begin = xr_71_begin_0, end = xr_71_end_0, end_mask = xr_71_end_mask_0, squeeze_mask = xr_71_squeeze_mask_0, x = pairs_71_cast_fp16)[name = tensor("xr_71_cast_fp16")]; + tensor xi_71_begin_0 = const()[name = tensor("xi_71_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_71_end_0 = const()[name = tensor("xi_71_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_71_end_mask_0 = const()[name = tensor("xi_71_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_71_squeeze_mask_0 = const()[name = tensor("xi_71_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_71_cast_fp16 = slice_by_index(begin = xi_71_begin_0, end = xi_71_end_0, end_mask = xi_71_end_mask_0, squeeze_mask = xi_71_squeeze_mask_0, x = pairs_71_cast_fp16)[name = tensor("xi_71_cast_fp16")]; + tensor var_4470_cast_fp16 = mul(x = xr_71_cast_fp16, y = rotr_69_cast_fp16)[name = tensor("op_4470_cast_fp16")]; + tensor var_4471_cast_fp16 = mul(x = xi_71_cast_fp16, y = roti_69_cast_fp16)[name = tensor("op_4471_cast_fp16")]; + tensor xor__71_cast_fp16 = sub(x = var_4470_cast_fp16, y = var_4471_cast_fp16)[name = tensor("xor__71_cast_fp16")]; + tensor var_4474_cast_fp16 = mul(x = xr_71_cast_fp16, y = roti_69_cast_fp16)[name = tensor("op_4474_cast_fp16")]; + tensor var_4475_cast_fp16 = mul(x = xi_71_cast_fp16, y = rotr_69_cast_fp16)[name = tensor("op_4475_cast_fp16")]; + tensor xoi_71_cast_fp16 = add(x = var_4474_cast_fp16, y = var_4475_cast_fp16)[name = tensor("xoi_71_cast_fp16")]; + tensor out_71_axis_0 = const()[name = tensor("out_71_axis_0"), val = tensor(-1)]; + tensor out_71_cast_fp16 = stack(axis = out_71_axis_0, values = (xor__71_cast_fp16, xoi_71_cast_fp16))[name = tensor("out_71_cast_fp16")]; + tensor _inversed_4487_y_0_to_fp16 = const()[name = tensor("_inversed_4487_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_4487_cast_fp16 = mul(x = position17_to_fp16, y = _inversed_4487_y_0_to_fp16)[name = tensor("_inversed_4487_cast_fp16")]; + tensor var_4488_cast_fp16 = floor(x = _inversed_4487_cast_fp16)[name = tensor("op_4488_cast_fp16")]; + tensor var_4489_to_fp16 = const()[name = tensor("op_4489_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4490_cast_fp16 = mul(x = var_4488_cast_fp16, y = var_4489_to_fp16)[name = tensor("op_4490_cast_fp16")]; + tensor write_idx_35_cast_fp16 = sub(x = position17_to_fp16, y = var_4490_cast_fp16)[name = tensor("write_idx_35_cast_fp16")]; + tensor var_4493_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_35_cast_fp16)[name = tensor("op_4493_cast_fp16")]; + tensor var_4503 = const()[name = tensor("op_4503"), val = tensor([1, 512, 1, 1])]; + tensor onehot_69_to_fp16_dtype_0 = const()[name = tensor("onehot_69_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_4493_cast_fp16_to_fp16 = cast(dtype = onehot_69_to_fp16_dtype_0, x = var_4493_cast_fp16)[name = tensor("cast_440")]; + tensor onehot_71_cast_fp16 = reshape(shape = var_4503, x = var_4493_cast_fp16_to_fp16)[name = tensor("onehot_71_cast_fp16")]; + tensor var_4505_to_fp16 = const()[name = tensor("op_4505_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4507_cast_fp16 = sub(x = var_4505_to_fp16, y = onehot_71_cast_fp16)[name = tensor("op_4507_cast_fp16")]; + tensor k_cache17_to_fp16_dtype_0 = const()[name = tensor("k_cache17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache17_to_fp16 = cast(dtype = k_cache17_to_fp16_dtype_0, x = k_cache17)[name = tensor("cast_439")]; + tensor var_4508_cast_fp16 = mul(x = k_cache17_to_fp16, y = var_4507_cast_fp16)[name = tensor("op_4508_cast_fp16")]; + tensor var_4513 = const()[name = tensor("op_4513"), val = tensor([1, 1, 16, 64])]; + tensor var_4514_cast_fp16 = reshape(shape = var_4513, x = out_71_cast_fp16)[name = tensor("op_4514_cast_fp16")]; + tensor var_4515_cast_fp16 = mul(x = var_4514_cast_fp16, y = onehot_71_cast_fp16)[name = tensor("op_4515_cast_fp16")]; + tensor new_k_cache_35_cast_fp16 = add(x = var_4508_cast_fp16, y = var_4515_cast_fp16)[name = tensor("new_k_cache_35_cast_fp16")]; + tensor new_k_cache_35_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_35_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache17_to_fp16_dtype_0 = const()[name = tensor("v_cache17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache17_to_fp16 = cast(dtype = v_cache17_to_fp16_dtype_0, x = v_cache17)[name = tensor("cast_437")]; + tensor var_4521_cast_fp16 = mul(x = v_cache17_to_fp16, y = var_4507_cast_fp16)[name = tensor("op_4521_cast_fp16")]; + tensor var_4526 = const()[name = tensor("op_4526"), val = tensor([1, 1, 16, 64])]; + tensor var_4527_cast_fp16 = reshape(shape = var_4526, x = v_35_cast_fp16)[name = tensor("op_4527_cast_fp16")]; + tensor var_4528_cast_fp16 = mul(x = var_4527_cast_fp16, y = onehot_71_cast_fp16)[name = tensor("op_4528_cast_fp16")]; + tensor new_v_cache_35_cast_fp16 = add(x = var_4521_cast_fp16, y = var_4528_cast_fp16)[name = tensor("new_v_cache_35_cast_fp16")]; + tensor new_v_cache_35_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_35_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_17_perm_0 = const()[name = tensor("transpose_17_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_4541 = const()[name = tensor("op_4541"), val = tensor([0, 2, 1, 3])]; + tensor var_4547 = const()[name = tensor("op_4547"), val = tensor([1, 16, 1, 64])]; + tensor q4_35_cast_fp16 = reshape(shape = var_4547, x = out_69_cast_fp16)[name = tensor("q4_35_cast_fp16")]; + tensor var_4552_transpose_x_0 = const()[name = tensor("op_4552_transpose_x_0"), val = tensor(false)]; + tensor var_4552_transpose_y_0 = const()[name = tensor("op_4552_transpose_y_0"), val = tensor(false)]; + tensor transpose_17_cast_fp16 = transpose(perm = transpose_17_perm_0, x = new_k_cache_35_cast_fp16)[name = tensor("transpose_44")]; + tensor var_4552_cast_fp16 = matmul(transpose_x = var_4552_transpose_x_0, transpose_y = var_4552_transpose_y_0, x = q4_35_cast_fp16, y = transpose_17_cast_fp16)[name = tensor("op_4552_cast_fp16")]; + tensor var_4553_to_fp16 = const()[name = tensor("op_4553_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_69_cast_fp16 = mul(x = var_4552_cast_fp16, y = var_4553_to_fp16)[name = tensor("scores_69_cast_fp16")]; + tensor var_4557 = const()[name = tensor("op_4557"), val = tensor([1, 1])]; + tensor var_4558_cast_fp16 = reshape(shape = var_4557, x = position17_to_fp16)[name = tensor("op_4558_cast_fp16")]; + tensor var_4559_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_4558_cast_fp16)[name = tensor("op_4559_cast_fp16")]; + tensor mask_35_to_fp16_dtype_0 = const()[name = tensor("mask_35_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_4566_to_fp16 = const()[name = tensor("op_4566_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4559_cast_fp16_to_fp16 = cast(dtype = mask_35_to_fp16_dtype_0, x = var_4559_cast_fp16)[name = tensor("cast_435")]; + tensor var_4567_cast_fp16 = sub(x = var_4559_cast_fp16_to_fp16, y = var_4566_to_fp16)[name = tensor("op_4567_cast_fp16")]; + tensor var_4572 = const()[name = tensor("op_4572"), val = tensor([1, 1, 1, 512])]; + tensor var_4573_cast_fp16 = reshape(shape = var_4572, x = var_4567_cast_fp16)[name = tensor("op_4573_cast_fp16")]; + tensor var_4574_to_fp16 = const()[name = tensor("op_4574_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_4575_cast_fp16 = mul(x = var_4573_cast_fp16, y = var_4574_to_fp16)[name = tensor("op_4575_cast_fp16")]; + tensor scores_71_cast_fp16 = add(x = scores_69_cast_fp16, y = var_4575_cast_fp16)[name = tensor("scores_71_cast_fp16")]; + tensor var_4578 = const()[name = tensor("op_4578"), val = tensor(-1)]; + tensor weights_35_cast_fp16 = softmax(axis = var_4578, x = scores_71_cast_fp16)[name = tensor("weights_35_cast_fp16")]; + tensor attn_35_transpose_x_0 = const()[name = tensor("attn_35_transpose_x_0"), val = tensor(false)]; + tensor attn_35_transpose_y_0 = const()[name = tensor("attn_35_transpose_y_0"), val = tensor(false)]; + tensor values_35_cast_fp16 = transpose(perm = var_4541, x = new_v_cache_35_cast_fp16)[name = tensor("transpose_43")]; + tensor attn_35_cast_fp16 = matmul(transpose_x = attn_35_transpose_x_0, transpose_y = attn_35_transpose_y_0, x = weights_35_cast_fp16, y = values_35_cast_fp16)[name = tensor("attn_35_cast_fp16")]; + tensor var_4586 = const()[name = tensor("op_4586"), val = tensor([0, 2, 1, 3])]; + tensor var_4591 = const()[name = tensor("op_4591"), val = tensor([1, 1, 1024])]; + tensor var_4587_cast_fp16 = transpose(perm = var_4586, x = attn_35_cast_fp16)[name = tensor("transpose_42")]; + tensor input_211_cast_fp16 = reshape(shape = var_4591, x = var_4587_cast_fp16)[name = tensor("input_211_cast_fp16")]; + tensor attn17_out_proj_weight_to_fp16 = const()[name = tensor("attn17_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(434346176)))]; + tensor linear_70_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn17_out_proj_weight_to_fp16, x = input_211_cast_fp16)[name = tensor("linear_70_cast_fp16")]; + tensor var_4597_to_fp16 = const()[name = tensor("op_4597_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4598_cast_fp16 = add(x = position17_to_fp16, y = var_4597_to_fp16)[name = tensor("op_4598_cast_fp16")]; + tensor var_4598_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4598_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_213_cast_fp16 = add(x = input_207_cast_fp16, y = linear_70_cast_fp16)[name = tensor("input_213_cast_fp16")]; + tensor input_215_axes_0 = const()[name = tensor("input_215_axes_0"), val = tensor([-1])]; + tensor norm17_2_weight_to_fp16 = const()[name = tensor("norm17_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(436443392)))]; + tensor norm17_2_bias_to_fp16 = const()[name = tensor("norm17_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(436445504)))]; + tensor var_4602_to_fp16 = const()[name = tensor("op_4602_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_215_cast_fp16 = layer_norm(axes = input_215_axes_0, beta = norm17_2_bias_to_fp16, epsilon = var_4602_to_fp16, gamma = norm17_2_weight_to_fp16, x = input_213_cast_fp16)[name = tensor("input_215_cast_fp16")]; + tensor linear17_1_weight_to_fp16 = const()[name = tensor("linear17_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(436447616)))]; + tensor linear_71_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear17_1_weight_to_fp16, x = input_215_cast_fp16)[name = tensor("linear_71_cast_fp16")]; + tensor input_217_mode_0 = const()[name = tensor("input_217_mode_0"), val = tensor("EXACT")]; + tensor input_217_cast_fp16 = gelu(mode = input_217_mode_0, x = linear_71_cast_fp16)[name = tensor("input_217_cast_fp16")]; + tensor linear17_2_weight_to_fp16 = const()[name = tensor("linear17_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(444836288)))]; + tensor linear_72_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear17_2_weight_to_fp16, x = input_217_cast_fp16)[name = tensor("linear_72_cast_fp16")]; + tensor input_219_cast_fp16 = add(x = input_213_cast_fp16, y = linear_72_cast_fp16)[name = tensor("input_219_cast_fp16")]; + tensor input_221_axes_0 = const()[name = tensor("input_221_axes_0"), val = tensor([-1])]; + tensor norm18_1_weight_to_fp16 = const()[name = tensor("norm18_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(453224960)))]; + tensor norm18_1_bias_to_fp16 = const()[name = tensor("norm18_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(453227072)))]; + tensor var_4619_to_fp16 = const()[name = tensor("op_4619_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_221_cast_fp16 = layer_norm(axes = input_221_axes_0, beta = norm18_1_bias_to_fp16, epsilon = var_4619_to_fp16, gamma = norm18_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("input_221_cast_fp16")]; + tensor attn18_in_proj_weight_to_fp16 = const()[name = tensor("attn18_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(453229184)))]; + tensor linear_73_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn18_in_proj_weight_to_fp16, x = input_221_cast_fp16)[name = tensor("linear_73_cast_fp16")]; + tensor var_4632 = const()[name = tensor("op_4632"), val = tensor([1, 3, 16, 64])]; + tensor qkv_37_cast_fp16 = reshape(shape = var_4632, x = linear_73_cast_fp16)[name = tensor("qkv_37_cast_fp16")]; + tensor x_73_begin_0 = const()[name = tensor("x_73_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_73_end_0 = const()[name = tensor("x_73_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_73_end_mask_0 = const()[name = tensor("x_73_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_73_squeeze_mask_0 = const()[name = tensor("x_73_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_73_cast_fp16 = slice_by_index(begin = x_73_begin_0, end = x_73_end_0, end_mask = x_73_end_mask_0, squeeze_mask = x_73_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("x_73_cast_fp16")]; + tensor x_75_begin_0 = const()[name = tensor("x_75_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_75_end_0 = const()[name = tensor("x_75_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_75_end_mask_0 = const()[name = tensor("x_75_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_75_squeeze_mask_0 = const()[name = tensor("x_75_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_75_cast_fp16 = slice_by_index(begin = x_75_begin_0, end = x_75_end_0, end_mask = x_75_end_mask_0, squeeze_mask = x_75_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("x_75_cast_fp16")]; + tensor v_37_begin_0 = const()[name = tensor("v_37_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_37_end_0 = const()[name = tensor("v_37_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_37_end_mask_0 = const()[name = tensor("v_37_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_37_squeeze_mask_0 = const()[name = tensor("v_37_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_37_cast_fp16 = slice_by_index(begin = v_37_begin_0, end = v_37_end_0, end_mask = v_37_end_mask_0, squeeze_mask = v_37_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("v_37_cast_fp16")]; + tensor var_4661 = const()[name = tensor("op_4661"), val = tensor([1, 1, 1])]; + tensor position18_to_fp16_dtype_0 = const()[name = tensor("position18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position18_to_fp16 = cast(dtype = position18_to_fp16_dtype_0, x = position18)[name = tensor("cast_433")]; + tensor var_4662_cast_fp16 = reshape(shape = var_4661, x = position18_to_fp16)[name = tensor("op_4662_cast_fp16")]; + tensor angles_73_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_4662_cast_fp16)[name = tensor("angles_73_cast_fp16")]; + tensor rotr_73_cast_fp16 = cos(x = angles_73_cast_fp16)[name = tensor("rotr_73_cast_fp16")]; + tensor roti_73_cast_fp16 = sin(x = angles_73_cast_fp16)[name = tensor("roti_73_cast_fp16")]; + tensor var_4670 = const()[name = tensor("op_4670"), val = tensor([1, 16, 32, 2])]; + tensor pairs_73_cast_fp16 = reshape(shape = var_4670, x = x_73_cast_fp16)[name = tensor("pairs_73_cast_fp16")]; + tensor xr_73_begin_0 = const()[name = tensor("xr_73_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_73_end_0 = const()[name = tensor("xr_73_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_73_end_mask_0 = const()[name = tensor("xr_73_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_73_squeeze_mask_0 = const()[name = tensor("xr_73_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_73_cast_fp16 = slice_by_index(begin = xr_73_begin_0, end = xr_73_end_0, end_mask = xr_73_end_mask_0, squeeze_mask = xr_73_squeeze_mask_0, x = pairs_73_cast_fp16)[name = tensor("xr_73_cast_fp16")]; + tensor xi_73_begin_0 = const()[name = tensor("xi_73_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_73_end_0 = const()[name = tensor("xi_73_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_73_end_mask_0 = const()[name = tensor("xi_73_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_73_squeeze_mask_0 = const()[name = tensor("xi_73_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_73_cast_fp16 = slice_by_index(begin = xi_73_begin_0, end = xi_73_end_0, end_mask = xi_73_end_mask_0, squeeze_mask = xi_73_squeeze_mask_0, x = pairs_73_cast_fp16)[name = tensor("xi_73_cast_fp16")]; + tensor var_4678_cast_fp16 = mul(x = xr_73_cast_fp16, y = rotr_73_cast_fp16)[name = tensor("op_4678_cast_fp16")]; + tensor var_4679_cast_fp16 = mul(x = xi_73_cast_fp16, y = roti_73_cast_fp16)[name = tensor("op_4679_cast_fp16")]; + tensor xor__73_cast_fp16 = sub(x = var_4678_cast_fp16, y = var_4679_cast_fp16)[name = tensor("xor__73_cast_fp16")]; + tensor var_4682_cast_fp16 = mul(x = xr_73_cast_fp16, y = roti_73_cast_fp16)[name = tensor("op_4682_cast_fp16")]; + tensor var_4683_cast_fp16 = mul(x = xi_73_cast_fp16, y = rotr_73_cast_fp16)[name = tensor("op_4683_cast_fp16")]; + tensor xoi_73_cast_fp16 = add(x = var_4682_cast_fp16, y = var_4683_cast_fp16)[name = tensor("xoi_73_cast_fp16")]; + tensor out_73_axis_0 = const()[name = tensor("out_73_axis_0"), val = tensor(-1)]; + tensor out_73_cast_fp16 = stack(axis = out_73_axis_0, values = (xor__73_cast_fp16, xoi_73_cast_fp16))[name = tensor("out_73_cast_fp16")]; + tensor var_4706 = const()[name = tensor("op_4706"), val = tensor([1, 16, 32, 2])]; + tensor pairs_75_cast_fp16 = reshape(shape = var_4706, x = x_75_cast_fp16)[name = tensor("pairs_75_cast_fp16")]; + tensor xr_75_begin_0 = const()[name = tensor("xr_75_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_75_end_0 = const()[name = tensor("xr_75_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_75_end_mask_0 = const()[name = tensor("xr_75_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_75_squeeze_mask_0 = const()[name = tensor("xr_75_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_75_cast_fp16 = slice_by_index(begin = xr_75_begin_0, end = xr_75_end_0, end_mask = xr_75_end_mask_0, squeeze_mask = xr_75_squeeze_mask_0, x = pairs_75_cast_fp16)[name = tensor("xr_75_cast_fp16")]; + tensor xi_75_begin_0 = const()[name = tensor("xi_75_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_75_end_0 = const()[name = tensor("xi_75_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_75_end_mask_0 = const()[name = tensor("xi_75_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_75_squeeze_mask_0 = const()[name = tensor("xi_75_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_75_cast_fp16 = slice_by_index(begin = xi_75_begin_0, end = xi_75_end_0, end_mask = xi_75_end_mask_0, squeeze_mask = xi_75_squeeze_mask_0, x = pairs_75_cast_fp16)[name = tensor("xi_75_cast_fp16")]; + tensor var_4714_cast_fp16 = mul(x = xr_75_cast_fp16, y = rotr_73_cast_fp16)[name = tensor("op_4714_cast_fp16")]; + tensor var_4715_cast_fp16 = mul(x = xi_75_cast_fp16, y = roti_73_cast_fp16)[name = tensor("op_4715_cast_fp16")]; + tensor xor__75_cast_fp16 = sub(x = var_4714_cast_fp16, y = var_4715_cast_fp16)[name = tensor("xor__75_cast_fp16")]; + tensor var_4718_cast_fp16 = mul(x = xr_75_cast_fp16, y = roti_73_cast_fp16)[name = tensor("op_4718_cast_fp16")]; + tensor var_4719_cast_fp16 = mul(x = xi_75_cast_fp16, y = rotr_73_cast_fp16)[name = tensor("op_4719_cast_fp16")]; + tensor xoi_75_cast_fp16 = add(x = var_4718_cast_fp16, y = var_4719_cast_fp16)[name = tensor("xoi_75_cast_fp16")]; + tensor out_75_axis_0 = const()[name = tensor("out_75_axis_0"), val = tensor(-1)]; + tensor out_75_cast_fp16 = stack(axis = out_75_axis_0, values = (xor__75_cast_fp16, xoi_75_cast_fp16))[name = tensor("out_75_cast_fp16")]; + tensor _inversed_4731_y_0_to_fp16 = const()[name = tensor("_inversed_4731_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_4731_cast_fp16 = mul(x = position18_to_fp16, y = _inversed_4731_y_0_to_fp16)[name = tensor("_inversed_4731_cast_fp16")]; + tensor var_4732_cast_fp16 = floor(x = _inversed_4731_cast_fp16)[name = tensor("op_4732_cast_fp16")]; + tensor var_4733_to_fp16 = const()[name = tensor("op_4733_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4734_cast_fp16 = mul(x = var_4732_cast_fp16, y = var_4733_to_fp16)[name = tensor("op_4734_cast_fp16")]; + tensor write_idx_37_cast_fp16 = sub(x = position18_to_fp16, y = var_4734_cast_fp16)[name = tensor("write_idx_37_cast_fp16")]; + tensor var_4737_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_37_cast_fp16)[name = tensor("op_4737_cast_fp16")]; + tensor var_4747 = const()[name = tensor("op_4747"), val = tensor([1, 512, 1, 1])]; + tensor onehot_73_to_fp16_dtype_0 = const()[name = tensor("onehot_73_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_4737_cast_fp16_to_fp16 = cast(dtype = onehot_73_to_fp16_dtype_0, x = var_4737_cast_fp16)[name = tensor("cast_432")]; + tensor onehot_75_cast_fp16 = reshape(shape = var_4747, x = var_4737_cast_fp16_to_fp16)[name = tensor("onehot_75_cast_fp16")]; + tensor var_4749_to_fp16 = const()[name = tensor("op_4749_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4751_cast_fp16 = sub(x = var_4749_to_fp16, y = onehot_75_cast_fp16)[name = tensor("op_4751_cast_fp16")]; + tensor k_cache18_to_fp16_dtype_0 = const()[name = tensor("k_cache18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache18_to_fp16 = cast(dtype = k_cache18_to_fp16_dtype_0, x = k_cache18)[name = tensor("cast_431")]; + tensor var_4752_cast_fp16 = mul(x = k_cache18_to_fp16, y = var_4751_cast_fp16)[name = tensor("op_4752_cast_fp16")]; + tensor var_4757 = const()[name = tensor("op_4757"), val = tensor([1, 1, 16, 64])]; + tensor var_4758_cast_fp16 = reshape(shape = var_4757, x = out_75_cast_fp16)[name = tensor("op_4758_cast_fp16")]; + tensor var_4759_cast_fp16 = mul(x = var_4758_cast_fp16, y = onehot_75_cast_fp16)[name = tensor("op_4759_cast_fp16")]; + tensor new_k_cache_37_cast_fp16 = add(x = var_4752_cast_fp16, y = var_4759_cast_fp16)[name = tensor("new_k_cache_37_cast_fp16")]; + tensor new_k_cache_37_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_37_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache18_to_fp16_dtype_0 = const()[name = tensor("v_cache18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache18_to_fp16 = cast(dtype = v_cache18_to_fp16_dtype_0, x = v_cache18)[name = tensor("cast_429")]; + tensor var_4765_cast_fp16 = mul(x = v_cache18_to_fp16, y = var_4751_cast_fp16)[name = tensor("op_4765_cast_fp16")]; + tensor var_4770 = const()[name = tensor("op_4770"), val = tensor([1, 1, 16, 64])]; + tensor var_4771_cast_fp16 = reshape(shape = var_4770, x = v_37_cast_fp16)[name = tensor("op_4771_cast_fp16")]; + tensor var_4772_cast_fp16 = mul(x = var_4771_cast_fp16, y = onehot_75_cast_fp16)[name = tensor("op_4772_cast_fp16")]; + tensor new_v_cache_37_cast_fp16 = add(x = var_4765_cast_fp16, y = var_4772_cast_fp16)[name = tensor("new_v_cache_37_cast_fp16")]; + tensor new_v_cache_37_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_37_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_18_perm_0 = const()[name = tensor("transpose_18_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_4785 = const()[name = tensor("op_4785"), val = tensor([0, 2, 1, 3])]; + tensor var_4791 = const()[name = tensor("op_4791"), val = tensor([1, 16, 1, 64])]; + tensor q4_37_cast_fp16 = reshape(shape = var_4791, x = out_73_cast_fp16)[name = tensor("q4_37_cast_fp16")]; + tensor var_4796_transpose_x_0 = const()[name = tensor("op_4796_transpose_x_0"), val = tensor(false)]; + tensor var_4796_transpose_y_0 = const()[name = tensor("op_4796_transpose_y_0"), val = tensor(false)]; + tensor transpose_18_cast_fp16 = transpose(perm = transpose_18_perm_0, x = new_k_cache_37_cast_fp16)[name = tensor("transpose_41")]; + tensor var_4796_cast_fp16 = matmul(transpose_x = var_4796_transpose_x_0, transpose_y = var_4796_transpose_y_0, x = q4_37_cast_fp16, y = transpose_18_cast_fp16)[name = tensor("op_4796_cast_fp16")]; + tensor var_4797_to_fp16 = const()[name = tensor("op_4797_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_73_cast_fp16 = mul(x = var_4796_cast_fp16, y = var_4797_to_fp16)[name = tensor("scores_73_cast_fp16")]; + tensor var_4801 = const()[name = tensor("op_4801"), val = tensor([1, 1])]; + tensor var_4802_cast_fp16 = reshape(shape = var_4801, x = position18_to_fp16)[name = tensor("op_4802_cast_fp16")]; + tensor var_4803_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_4802_cast_fp16)[name = tensor("op_4803_cast_fp16")]; + tensor mask_37_to_fp16_dtype_0 = const()[name = tensor("mask_37_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_4810_to_fp16 = const()[name = tensor("op_4810_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4803_cast_fp16_to_fp16 = cast(dtype = mask_37_to_fp16_dtype_0, x = var_4803_cast_fp16)[name = tensor("cast_427")]; + tensor var_4811_cast_fp16 = sub(x = var_4803_cast_fp16_to_fp16, y = var_4810_to_fp16)[name = tensor("op_4811_cast_fp16")]; + tensor var_4816 = const()[name = tensor("op_4816"), val = tensor([1, 1, 1, 512])]; + tensor var_4817_cast_fp16 = reshape(shape = var_4816, x = var_4811_cast_fp16)[name = tensor("op_4817_cast_fp16")]; + tensor var_4818_to_fp16 = const()[name = tensor("op_4818_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_4819_cast_fp16 = mul(x = var_4817_cast_fp16, y = var_4818_to_fp16)[name = tensor("op_4819_cast_fp16")]; + tensor scores_75_cast_fp16 = add(x = scores_73_cast_fp16, y = var_4819_cast_fp16)[name = tensor("scores_75_cast_fp16")]; + tensor var_4822 = const()[name = tensor("op_4822"), val = tensor(-1)]; + tensor weights_37_cast_fp16 = softmax(axis = var_4822, x = scores_75_cast_fp16)[name = tensor("weights_37_cast_fp16")]; + tensor attn_37_transpose_x_0 = const()[name = tensor("attn_37_transpose_x_0"), val = tensor(false)]; + tensor attn_37_transpose_y_0 = const()[name = tensor("attn_37_transpose_y_0"), val = tensor(false)]; + tensor values_37_cast_fp16 = transpose(perm = var_4785, x = new_v_cache_37_cast_fp16)[name = tensor("transpose_40")]; + tensor attn_37_cast_fp16 = matmul(transpose_x = attn_37_transpose_x_0, transpose_y = attn_37_transpose_y_0, x = weights_37_cast_fp16, y = values_37_cast_fp16)[name = tensor("attn_37_cast_fp16")]; + tensor var_4830 = const()[name = tensor("op_4830"), val = tensor([0, 2, 1, 3])]; + tensor var_4835 = const()[name = tensor("op_4835"), val = tensor([1, 1, 1024])]; + tensor var_4831_cast_fp16 = transpose(perm = var_4830, x = attn_37_cast_fp16)[name = tensor("transpose_39")]; + tensor input_223_cast_fp16 = reshape(shape = var_4835, x = var_4831_cast_fp16)[name = tensor("input_223_cast_fp16")]; + tensor attn18_out_proj_weight_to_fp16 = const()[name = tensor("attn18_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(459520704)))]; + tensor linear_74_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn18_out_proj_weight_to_fp16, x = input_223_cast_fp16)[name = tensor("linear_74_cast_fp16")]; + tensor var_4841_to_fp16 = const()[name = tensor("op_4841_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4842_cast_fp16 = add(x = position18_to_fp16, y = var_4841_to_fp16)[name = tensor("op_4842_cast_fp16")]; + tensor var_4842_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4842_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_225_cast_fp16 = add(x = input_219_cast_fp16, y = linear_74_cast_fp16)[name = tensor("input_225_cast_fp16")]; + tensor input_227_axes_0 = const()[name = tensor("input_227_axes_0"), val = tensor([-1])]; + tensor norm18_2_weight_to_fp16 = const()[name = tensor("norm18_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(461617920)))]; + tensor norm18_2_bias_to_fp16 = const()[name = tensor("norm18_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(461620032)))]; + tensor var_4846_to_fp16 = const()[name = tensor("op_4846_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_227_cast_fp16 = layer_norm(axes = input_227_axes_0, beta = norm18_2_bias_to_fp16, epsilon = var_4846_to_fp16, gamma = norm18_2_weight_to_fp16, x = input_225_cast_fp16)[name = tensor("input_227_cast_fp16")]; + tensor linear18_1_weight_to_fp16 = const()[name = tensor("linear18_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(461622144)))]; + tensor linear_75_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear18_1_weight_to_fp16, x = input_227_cast_fp16)[name = tensor("linear_75_cast_fp16")]; + tensor input_229_mode_0 = const()[name = tensor("input_229_mode_0"), val = tensor("EXACT")]; + tensor input_229_cast_fp16 = gelu(mode = input_229_mode_0, x = linear_75_cast_fp16)[name = tensor("input_229_cast_fp16")]; + tensor linear18_2_weight_to_fp16 = const()[name = tensor("linear18_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(470010816)))]; + tensor linear_76_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear18_2_weight_to_fp16, x = input_229_cast_fp16)[name = tensor("linear_76_cast_fp16")]; + tensor input_231_cast_fp16 = add(x = input_225_cast_fp16, y = linear_76_cast_fp16)[name = tensor("input_231_cast_fp16")]; + tensor input_233_axes_0 = const()[name = tensor("input_233_axes_0"), val = tensor([-1])]; + tensor norm19_1_weight_to_fp16 = const()[name = tensor("norm19_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(478399488)))]; + tensor norm19_1_bias_to_fp16 = const()[name = tensor("norm19_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(478401600)))]; + tensor var_4863_to_fp16 = const()[name = tensor("op_4863_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_233_cast_fp16 = layer_norm(axes = input_233_axes_0, beta = norm19_1_bias_to_fp16, epsilon = var_4863_to_fp16, gamma = norm19_1_weight_to_fp16, x = input_231_cast_fp16)[name = tensor("input_233_cast_fp16")]; + tensor attn19_in_proj_weight_to_fp16 = const()[name = tensor("attn19_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(478403712)))]; + tensor linear_77_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn19_in_proj_weight_to_fp16, x = input_233_cast_fp16)[name = tensor("linear_77_cast_fp16")]; + tensor var_4876 = const()[name = tensor("op_4876"), val = tensor([1, 3, 16, 64])]; + tensor qkv_39_cast_fp16 = reshape(shape = var_4876, x = linear_77_cast_fp16)[name = tensor("qkv_39_cast_fp16")]; + tensor x_77_begin_0 = const()[name = tensor("x_77_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_77_end_0 = const()[name = tensor("x_77_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_77_end_mask_0 = const()[name = tensor("x_77_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_77_squeeze_mask_0 = const()[name = tensor("x_77_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_77_cast_fp16 = slice_by_index(begin = x_77_begin_0, end = x_77_end_0, end_mask = x_77_end_mask_0, squeeze_mask = x_77_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("x_77_cast_fp16")]; + tensor x_79_begin_0 = const()[name = tensor("x_79_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_79_end_0 = const()[name = tensor("x_79_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_79_end_mask_0 = const()[name = tensor("x_79_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_79_squeeze_mask_0 = const()[name = tensor("x_79_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_79_cast_fp16 = slice_by_index(begin = x_79_begin_0, end = x_79_end_0, end_mask = x_79_end_mask_0, squeeze_mask = x_79_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("x_79_cast_fp16")]; + tensor v_39_begin_0 = const()[name = tensor("v_39_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_39_end_0 = const()[name = tensor("v_39_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_39_end_mask_0 = const()[name = tensor("v_39_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_39_squeeze_mask_0 = const()[name = tensor("v_39_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_39_cast_fp16 = slice_by_index(begin = v_39_begin_0, end = v_39_end_0, end_mask = v_39_end_mask_0, squeeze_mask = v_39_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("v_39_cast_fp16")]; + tensor var_4905 = const()[name = tensor("op_4905"), val = tensor([1, 1, 1])]; + tensor position19_to_fp16_dtype_0 = const()[name = tensor("position19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position19_to_fp16 = cast(dtype = position19_to_fp16_dtype_0, x = position19)[name = tensor("cast_425")]; + tensor var_4906_cast_fp16 = reshape(shape = var_4905, x = position19_to_fp16)[name = tensor("op_4906_cast_fp16")]; + tensor angles_77_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_4906_cast_fp16)[name = tensor("angles_77_cast_fp16")]; + tensor rotr_77_cast_fp16 = cos(x = angles_77_cast_fp16)[name = tensor("rotr_77_cast_fp16")]; + tensor roti_77_cast_fp16 = sin(x = angles_77_cast_fp16)[name = tensor("roti_77_cast_fp16")]; + tensor var_4914 = const()[name = tensor("op_4914"), val = tensor([1, 16, 32, 2])]; + tensor pairs_77_cast_fp16 = reshape(shape = var_4914, x = x_77_cast_fp16)[name = tensor("pairs_77_cast_fp16")]; + tensor xr_77_begin_0 = const()[name = tensor("xr_77_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_77_end_0 = const()[name = tensor("xr_77_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_77_end_mask_0 = const()[name = tensor("xr_77_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_77_squeeze_mask_0 = const()[name = tensor("xr_77_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_77_cast_fp16 = slice_by_index(begin = xr_77_begin_0, end = xr_77_end_0, end_mask = xr_77_end_mask_0, squeeze_mask = xr_77_squeeze_mask_0, x = pairs_77_cast_fp16)[name = tensor("xr_77_cast_fp16")]; + tensor xi_77_begin_0 = const()[name = tensor("xi_77_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_77_end_0 = const()[name = tensor("xi_77_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_77_end_mask_0 = const()[name = tensor("xi_77_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_77_squeeze_mask_0 = const()[name = tensor("xi_77_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_77_cast_fp16 = slice_by_index(begin = xi_77_begin_0, end = xi_77_end_0, end_mask = xi_77_end_mask_0, squeeze_mask = xi_77_squeeze_mask_0, x = pairs_77_cast_fp16)[name = tensor("xi_77_cast_fp16")]; + tensor var_4922_cast_fp16 = mul(x = xr_77_cast_fp16, y = rotr_77_cast_fp16)[name = tensor("op_4922_cast_fp16")]; + tensor var_4923_cast_fp16 = mul(x = xi_77_cast_fp16, y = roti_77_cast_fp16)[name = tensor("op_4923_cast_fp16")]; + tensor xor__77_cast_fp16 = sub(x = var_4922_cast_fp16, y = var_4923_cast_fp16)[name = tensor("xor__77_cast_fp16")]; + tensor var_4926_cast_fp16 = mul(x = xr_77_cast_fp16, y = roti_77_cast_fp16)[name = tensor("op_4926_cast_fp16")]; + tensor var_4927_cast_fp16 = mul(x = xi_77_cast_fp16, y = rotr_77_cast_fp16)[name = tensor("op_4927_cast_fp16")]; + tensor xoi_77_cast_fp16 = add(x = var_4926_cast_fp16, y = var_4927_cast_fp16)[name = tensor("xoi_77_cast_fp16")]; + tensor out_77_axis_0 = const()[name = tensor("out_77_axis_0"), val = tensor(-1)]; + tensor out_77_cast_fp16 = stack(axis = out_77_axis_0, values = (xor__77_cast_fp16, xoi_77_cast_fp16))[name = tensor("out_77_cast_fp16")]; + tensor var_4950 = const()[name = tensor("op_4950"), val = tensor([1, 16, 32, 2])]; + tensor pairs_79_cast_fp16 = reshape(shape = var_4950, x = x_79_cast_fp16)[name = tensor("pairs_79_cast_fp16")]; + tensor xr_79_begin_0 = const()[name = tensor("xr_79_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_79_end_0 = const()[name = tensor("xr_79_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_79_end_mask_0 = const()[name = tensor("xr_79_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_79_squeeze_mask_0 = const()[name = tensor("xr_79_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_79_cast_fp16 = slice_by_index(begin = xr_79_begin_0, end = xr_79_end_0, end_mask = xr_79_end_mask_0, squeeze_mask = xr_79_squeeze_mask_0, x = pairs_79_cast_fp16)[name = tensor("xr_79_cast_fp16")]; + tensor xi_79_begin_0 = const()[name = tensor("xi_79_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_79_end_0 = const()[name = tensor("xi_79_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_79_end_mask_0 = const()[name = tensor("xi_79_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_79_squeeze_mask_0 = const()[name = tensor("xi_79_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_79_cast_fp16 = slice_by_index(begin = xi_79_begin_0, end = xi_79_end_0, end_mask = xi_79_end_mask_0, squeeze_mask = xi_79_squeeze_mask_0, x = pairs_79_cast_fp16)[name = tensor("xi_79_cast_fp16")]; + tensor var_4958_cast_fp16 = mul(x = xr_79_cast_fp16, y = rotr_77_cast_fp16)[name = tensor("op_4958_cast_fp16")]; + tensor var_4959_cast_fp16 = mul(x = xi_79_cast_fp16, y = roti_77_cast_fp16)[name = tensor("op_4959_cast_fp16")]; + tensor xor__79_cast_fp16 = sub(x = var_4958_cast_fp16, y = var_4959_cast_fp16)[name = tensor("xor__79_cast_fp16")]; + tensor var_4962_cast_fp16 = mul(x = xr_79_cast_fp16, y = roti_77_cast_fp16)[name = tensor("op_4962_cast_fp16")]; + tensor var_4963_cast_fp16 = mul(x = xi_79_cast_fp16, y = rotr_77_cast_fp16)[name = tensor("op_4963_cast_fp16")]; + tensor xoi_79_cast_fp16 = add(x = var_4962_cast_fp16, y = var_4963_cast_fp16)[name = tensor("xoi_79_cast_fp16")]; + tensor out_79_axis_0 = const()[name = tensor("out_79_axis_0"), val = tensor(-1)]; + tensor out_79_cast_fp16 = stack(axis = out_79_axis_0, values = (xor__79_cast_fp16, xoi_79_cast_fp16))[name = tensor("out_79_cast_fp16")]; + tensor _inversed_4975_y_0_to_fp16 = const()[name = tensor("_inversed_4975_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_4975_cast_fp16 = mul(x = position19_to_fp16, y = _inversed_4975_y_0_to_fp16)[name = tensor("_inversed_4975_cast_fp16")]; + tensor var_4976_cast_fp16 = floor(x = _inversed_4975_cast_fp16)[name = tensor("op_4976_cast_fp16")]; + tensor var_4977_to_fp16 = const()[name = tensor("op_4977_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4978_cast_fp16 = mul(x = var_4976_cast_fp16, y = var_4977_to_fp16)[name = tensor("op_4978_cast_fp16")]; + tensor write_idx_39_cast_fp16 = sub(x = position19_to_fp16, y = var_4978_cast_fp16)[name = tensor("write_idx_39_cast_fp16")]; + tensor var_4981_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_39_cast_fp16)[name = tensor("op_4981_cast_fp16")]; + tensor var_4991 = const()[name = tensor("op_4991"), val = tensor([1, 512, 1, 1])]; + tensor onehot_77_to_fp16_dtype_0 = const()[name = tensor("onehot_77_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_4981_cast_fp16_to_fp16 = cast(dtype = onehot_77_to_fp16_dtype_0, x = var_4981_cast_fp16)[name = tensor("cast_424")]; + tensor onehot_79_cast_fp16 = reshape(shape = var_4991, x = var_4981_cast_fp16_to_fp16)[name = tensor("onehot_79_cast_fp16")]; + tensor var_4993_to_fp16 = const()[name = tensor("op_4993_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4995_cast_fp16 = sub(x = var_4993_to_fp16, y = onehot_79_cast_fp16)[name = tensor("op_4995_cast_fp16")]; + tensor k_cache19_to_fp16_dtype_0 = const()[name = tensor("k_cache19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache19_to_fp16 = cast(dtype = k_cache19_to_fp16_dtype_0, x = k_cache19)[name = tensor("cast_423")]; + tensor var_4996_cast_fp16 = mul(x = k_cache19_to_fp16, y = var_4995_cast_fp16)[name = tensor("op_4996_cast_fp16")]; + tensor var_5001 = const()[name = tensor("op_5001"), val = tensor([1, 1, 16, 64])]; + tensor var_5002_cast_fp16 = reshape(shape = var_5001, x = out_79_cast_fp16)[name = tensor("op_5002_cast_fp16")]; + tensor var_5003_cast_fp16 = mul(x = var_5002_cast_fp16, y = onehot_79_cast_fp16)[name = tensor("op_5003_cast_fp16")]; + tensor new_k_cache_39_cast_fp16 = add(x = var_4996_cast_fp16, y = var_5003_cast_fp16)[name = tensor("new_k_cache_39_cast_fp16")]; + tensor new_k_cache_39_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_39_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache19_to_fp16_dtype_0 = const()[name = tensor("v_cache19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache19_to_fp16 = cast(dtype = v_cache19_to_fp16_dtype_0, x = v_cache19)[name = tensor("cast_421")]; + tensor var_5009_cast_fp16 = mul(x = v_cache19_to_fp16, y = var_4995_cast_fp16)[name = tensor("op_5009_cast_fp16")]; + tensor var_5014 = const()[name = tensor("op_5014"), val = tensor([1, 1, 16, 64])]; + tensor var_5015_cast_fp16 = reshape(shape = var_5014, x = v_39_cast_fp16)[name = tensor("op_5015_cast_fp16")]; + tensor var_5016_cast_fp16 = mul(x = var_5015_cast_fp16, y = onehot_79_cast_fp16)[name = tensor("op_5016_cast_fp16")]; + tensor new_v_cache_39_cast_fp16 = add(x = var_5009_cast_fp16, y = var_5016_cast_fp16)[name = tensor("new_v_cache_39_cast_fp16")]; + tensor new_v_cache_39_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_39_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_19_perm_0 = const()[name = tensor("transpose_19_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_5029 = const()[name = tensor("op_5029"), val = tensor([0, 2, 1, 3])]; + tensor var_5035 = const()[name = tensor("op_5035"), val = tensor([1, 16, 1, 64])]; + tensor q4_39_cast_fp16 = reshape(shape = var_5035, x = out_77_cast_fp16)[name = tensor("q4_39_cast_fp16")]; + tensor var_5040_transpose_x_0 = const()[name = tensor("op_5040_transpose_x_0"), val = tensor(false)]; + tensor var_5040_transpose_y_0 = const()[name = tensor("op_5040_transpose_y_0"), val = tensor(false)]; + tensor transpose_19_cast_fp16 = transpose(perm = transpose_19_perm_0, x = new_k_cache_39_cast_fp16)[name = tensor("transpose_38")]; + tensor var_5040_cast_fp16 = matmul(transpose_x = var_5040_transpose_x_0, transpose_y = var_5040_transpose_y_0, x = q4_39_cast_fp16, y = transpose_19_cast_fp16)[name = tensor("op_5040_cast_fp16")]; + tensor var_5041_to_fp16 = const()[name = tensor("op_5041_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_77_cast_fp16 = mul(x = var_5040_cast_fp16, y = var_5041_to_fp16)[name = tensor("scores_77_cast_fp16")]; + tensor var_5045 = const()[name = tensor("op_5045"), val = tensor([1, 1])]; + tensor var_5046_cast_fp16 = reshape(shape = var_5045, x = position19_to_fp16)[name = tensor("op_5046_cast_fp16")]; + tensor var_5047_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_5046_cast_fp16)[name = tensor("op_5047_cast_fp16")]; + tensor mask_39_to_fp16_dtype_0 = const()[name = tensor("mask_39_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_5054_to_fp16 = const()[name = tensor("op_5054_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5047_cast_fp16_to_fp16 = cast(dtype = mask_39_to_fp16_dtype_0, x = var_5047_cast_fp16)[name = tensor("cast_419")]; + tensor var_5055_cast_fp16 = sub(x = var_5047_cast_fp16_to_fp16, y = var_5054_to_fp16)[name = tensor("op_5055_cast_fp16")]; + tensor var_5060 = const()[name = tensor("op_5060"), val = tensor([1, 1, 1, 512])]; + tensor var_5061_cast_fp16 = reshape(shape = var_5060, x = var_5055_cast_fp16)[name = tensor("op_5061_cast_fp16")]; + tensor var_5062_to_fp16 = const()[name = tensor("op_5062_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_5063_cast_fp16 = mul(x = var_5061_cast_fp16, y = var_5062_to_fp16)[name = tensor("op_5063_cast_fp16")]; + tensor scores_79_cast_fp16 = add(x = scores_77_cast_fp16, y = var_5063_cast_fp16)[name = tensor("scores_79_cast_fp16")]; + tensor var_5066 = const()[name = tensor("op_5066"), val = tensor(-1)]; + tensor weights_39_cast_fp16 = softmax(axis = var_5066, x = scores_79_cast_fp16)[name = tensor("weights_39_cast_fp16")]; + tensor attn_39_transpose_x_0 = const()[name = tensor("attn_39_transpose_x_0"), val = tensor(false)]; + tensor attn_39_transpose_y_0 = const()[name = tensor("attn_39_transpose_y_0"), val = tensor(false)]; + tensor values_39_cast_fp16 = transpose(perm = var_5029, x = new_v_cache_39_cast_fp16)[name = tensor("transpose_37")]; + tensor attn_39_cast_fp16 = matmul(transpose_x = attn_39_transpose_x_0, transpose_y = attn_39_transpose_y_0, x = weights_39_cast_fp16, y = values_39_cast_fp16)[name = tensor("attn_39_cast_fp16")]; + tensor var_5074 = const()[name = tensor("op_5074"), val = tensor([0, 2, 1, 3])]; + tensor var_5079 = const()[name = tensor("op_5079"), val = tensor([1, 1, 1024])]; + tensor var_5075_cast_fp16 = transpose(perm = var_5074, x = attn_39_cast_fp16)[name = tensor("transpose_36")]; + tensor input_235_cast_fp16 = reshape(shape = var_5079, x = var_5075_cast_fp16)[name = tensor("input_235_cast_fp16")]; + tensor attn19_out_proj_weight_to_fp16 = const()[name = tensor("attn19_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(484695232)))]; + tensor linear_78_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn19_out_proj_weight_to_fp16, x = input_235_cast_fp16)[name = tensor("linear_78_cast_fp16")]; + tensor var_5085_to_fp16 = const()[name = tensor("op_5085_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5086_cast_fp16 = add(x = position19_to_fp16, y = var_5085_to_fp16)[name = tensor("op_5086_cast_fp16")]; + tensor var_5086_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5086_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_237_cast_fp16 = add(x = input_231_cast_fp16, y = linear_78_cast_fp16)[name = tensor("input_237_cast_fp16")]; + tensor input_239_axes_0 = const()[name = tensor("input_239_axes_0"), val = tensor([-1])]; + tensor norm19_2_weight_to_fp16 = const()[name = tensor("norm19_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486792448)))]; + tensor norm19_2_bias_to_fp16 = const()[name = tensor("norm19_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486794560)))]; + tensor var_5090_to_fp16 = const()[name = tensor("op_5090_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_239_cast_fp16 = layer_norm(axes = input_239_axes_0, beta = norm19_2_bias_to_fp16, epsilon = var_5090_to_fp16, gamma = norm19_2_weight_to_fp16, x = input_237_cast_fp16)[name = tensor("input_239_cast_fp16")]; + tensor linear19_1_weight_to_fp16 = const()[name = tensor("linear19_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486796672)))]; + tensor linear_79_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear19_1_weight_to_fp16, x = input_239_cast_fp16)[name = tensor("linear_79_cast_fp16")]; + tensor input_241_mode_0 = const()[name = tensor("input_241_mode_0"), val = tensor("EXACT")]; + tensor input_241_cast_fp16 = gelu(mode = input_241_mode_0, x = linear_79_cast_fp16)[name = tensor("input_241_cast_fp16")]; + tensor linear19_2_weight_to_fp16 = const()[name = tensor("linear19_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(495185344)))]; + tensor linear_80_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear19_2_weight_to_fp16, x = input_241_cast_fp16)[name = tensor("linear_80_cast_fp16")]; + tensor input_243_cast_fp16 = add(x = input_237_cast_fp16, y = linear_80_cast_fp16)[name = tensor("input_243_cast_fp16")]; + tensor input_245_axes_0 = const()[name = tensor("input_245_axes_0"), val = tensor([-1])]; + tensor norm20_1_weight_to_fp16 = const()[name = tensor("norm20_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(503574016)))]; + tensor norm20_1_bias_to_fp16 = const()[name = tensor("norm20_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(503576128)))]; + tensor var_5107_to_fp16 = const()[name = tensor("op_5107_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_245_cast_fp16 = layer_norm(axes = input_245_axes_0, beta = norm20_1_bias_to_fp16, epsilon = var_5107_to_fp16, gamma = norm20_1_weight_to_fp16, x = input_243_cast_fp16)[name = tensor("input_245_cast_fp16")]; + tensor attn20_in_proj_weight_to_fp16 = const()[name = tensor("attn20_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(503578240)))]; + tensor linear_81_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn20_in_proj_weight_to_fp16, x = input_245_cast_fp16)[name = tensor("linear_81_cast_fp16")]; + tensor var_5120 = const()[name = tensor("op_5120"), val = tensor([1, 3, 16, 64])]; + tensor qkv_41_cast_fp16 = reshape(shape = var_5120, x = linear_81_cast_fp16)[name = tensor("qkv_41_cast_fp16")]; + tensor x_81_begin_0 = const()[name = tensor("x_81_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_81_end_0 = const()[name = tensor("x_81_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_81_end_mask_0 = const()[name = tensor("x_81_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_81_squeeze_mask_0 = const()[name = tensor("x_81_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_81_cast_fp16 = slice_by_index(begin = x_81_begin_0, end = x_81_end_0, end_mask = x_81_end_mask_0, squeeze_mask = x_81_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("x_81_cast_fp16")]; + tensor x_83_begin_0 = const()[name = tensor("x_83_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_83_end_0 = const()[name = tensor("x_83_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_83_end_mask_0 = const()[name = tensor("x_83_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_83_squeeze_mask_0 = const()[name = tensor("x_83_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_83_cast_fp16 = slice_by_index(begin = x_83_begin_0, end = x_83_end_0, end_mask = x_83_end_mask_0, squeeze_mask = x_83_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("x_83_cast_fp16")]; + tensor v_41_begin_0 = const()[name = tensor("v_41_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_41_end_0 = const()[name = tensor("v_41_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_41_end_mask_0 = const()[name = tensor("v_41_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_41_squeeze_mask_0 = const()[name = tensor("v_41_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_41_cast_fp16 = slice_by_index(begin = v_41_begin_0, end = v_41_end_0, end_mask = v_41_end_mask_0, squeeze_mask = v_41_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("v_41_cast_fp16")]; + tensor var_5149 = const()[name = tensor("op_5149"), val = tensor([1, 1, 1])]; + tensor position20_to_fp16_dtype_0 = const()[name = tensor("position20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position20_to_fp16 = cast(dtype = position20_to_fp16_dtype_0, x = position20)[name = tensor("cast_417")]; + tensor var_5150_cast_fp16 = reshape(shape = var_5149, x = position20_to_fp16)[name = tensor("op_5150_cast_fp16")]; + tensor angles_81_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_5150_cast_fp16)[name = tensor("angles_81_cast_fp16")]; + tensor rotr_81_cast_fp16 = cos(x = angles_81_cast_fp16)[name = tensor("rotr_81_cast_fp16")]; + tensor roti_81_cast_fp16 = sin(x = angles_81_cast_fp16)[name = tensor("roti_81_cast_fp16")]; + tensor var_5158 = const()[name = tensor("op_5158"), val = tensor([1, 16, 32, 2])]; + tensor pairs_81_cast_fp16 = reshape(shape = var_5158, x = x_81_cast_fp16)[name = tensor("pairs_81_cast_fp16")]; + tensor xr_81_begin_0 = const()[name = tensor("xr_81_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_81_end_0 = const()[name = tensor("xr_81_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_81_end_mask_0 = const()[name = tensor("xr_81_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_81_squeeze_mask_0 = const()[name = tensor("xr_81_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_81_cast_fp16 = slice_by_index(begin = xr_81_begin_0, end = xr_81_end_0, end_mask = xr_81_end_mask_0, squeeze_mask = xr_81_squeeze_mask_0, x = pairs_81_cast_fp16)[name = tensor("xr_81_cast_fp16")]; + tensor xi_81_begin_0 = const()[name = tensor("xi_81_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_81_end_0 = const()[name = tensor("xi_81_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_81_end_mask_0 = const()[name = tensor("xi_81_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_81_squeeze_mask_0 = const()[name = tensor("xi_81_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_81_cast_fp16 = slice_by_index(begin = xi_81_begin_0, end = xi_81_end_0, end_mask = xi_81_end_mask_0, squeeze_mask = xi_81_squeeze_mask_0, x = pairs_81_cast_fp16)[name = tensor("xi_81_cast_fp16")]; + tensor var_5166_cast_fp16 = mul(x = xr_81_cast_fp16, y = rotr_81_cast_fp16)[name = tensor("op_5166_cast_fp16")]; + tensor var_5167_cast_fp16 = mul(x = xi_81_cast_fp16, y = roti_81_cast_fp16)[name = tensor("op_5167_cast_fp16")]; + tensor xor__81_cast_fp16 = sub(x = var_5166_cast_fp16, y = var_5167_cast_fp16)[name = tensor("xor__81_cast_fp16")]; + tensor var_5170_cast_fp16 = mul(x = xr_81_cast_fp16, y = roti_81_cast_fp16)[name = tensor("op_5170_cast_fp16")]; + tensor var_5171_cast_fp16 = mul(x = xi_81_cast_fp16, y = rotr_81_cast_fp16)[name = tensor("op_5171_cast_fp16")]; + tensor xoi_81_cast_fp16 = add(x = var_5170_cast_fp16, y = var_5171_cast_fp16)[name = tensor("xoi_81_cast_fp16")]; + tensor out_81_axis_0 = const()[name = tensor("out_81_axis_0"), val = tensor(-1)]; + tensor out_81_cast_fp16 = stack(axis = out_81_axis_0, values = (xor__81_cast_fp16, xoi_81_cast_fp16))[name = tensor("out_81_cast_fp16")]; + tensor var_5194 = const()[name = tensor("op_5194"), val = tensor([1, 16, 32, 2])]; + tensor pairs_83_cast_fp16 = reshape(shape = var_5194, x = x_83_cast_fp16)[name = tensor("pairs_83_cast_fp16")]; + tensor xr_83_begin_0 = const()[name = tensor("xr_83_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_83_end_0 = const()[name = tensor("xr_83_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_83_end_mask_0 = const()[name = tensor("xr_83_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_83_squeeze_mask_0 = const()[name = tensor("xr_83_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_83_cast_fp16 = slice_by_index(begin = xr_83_begin_0, end = xr_83_end_0, end_mask = xr_83_end_mask_0, squeeze_mask = xr_83_squeeze_mask_0, x = pairs_83_cast_fp16)[name = tensor("xr_83_cast_fp16")]; + tensor xi_83_begin_0 = const()[name = tensor("xi_83_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_83_end_0 = const()[name = tensor("xi_83_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_83_end_mask_0 = const()[name = tensor("xi_83_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_83_squeeze_mask_0 = const()[name = tensor("xi_83_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_83_cast_fp16 = slice_by_index(begin = xi_83_begin_0, end = xi_83_end_0, end_mask = xi_83_end_mask_0, squeeze_mask = xi_83_squeeze_mask_0, x = pairs_83_cast_fp16)[name = tensor("xi_83_cast_fp16")]; + tensor var_5202_cast_fp16 = mul(x = xr_83_cast_fp16, y = rotr_81_cast_fp16)[name = tensor("op_5202_cast_fp16")]; + tensor var_5203_cast_fp16 = mul(x = xi_83_cast_fp16, y = roti_81_cast_fp16)[name = tensor("op_5203_cast_fp16")]; + tensor xor__83_cast_fp16 = sub(x = var_5202_cast_fp16, y = var_5203_cast_fp16)[name = tensor("xor__83_cast_fp16")]; + tensor var_5206_cast_fp16 = mul(x = xr_83_cast_fp16, y = roti_81_cast_fp16)[name = tensor("op_5206_cast_fp16")]; + tensor var_5207_cast_fp16 = mul(x = xi_83_cast_fp16, y = rotr_81_cast_fp16)[name = tensor("op_5207_cast_fp16")]; + tensor xoi_83_cast_fp16 = add(x = var_5206_cast_fp16, y = var_5207_cast_fp16)[name = tensor("xoi_83_cast_fp16")]; + tensor out_83_axis_0 = const()[name = tensor("out_83_axis_0"), val = tensor(-1)]; + tensor out_83_cast_fp16 = stack(axis = out_83_axis_0, values = (xor__83_cast_fp16, xoi_83_cast_fp16))[name = tensor("out_83_cast_fp16")]; + tensor _inversed_5219_y_0_to_fp16 = const()[name = tensor("_inversed_5219_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5219_cast_fp16 = mul(x = position20_to_fp16, y = _inversed_5219_y_0_to_fp16)[name = tensor("_inversed_5219_cast_fp16")]; + tensor var_5220_cast_fp16 = floor(x = _inversed_5219_cast_fp16)[name = tensor("op_5220_cast_fp16")]; + tensor var_5221_to_fp16 = const()[name = tensor("op_5221_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5222_cast_fp16 = mul(x = var_5220_cast_fp16, y = var_5221_to_fp16)[name = tensor("op_5222_cast_fp16")]; + tensor write_idx_41_cast_fp16 = sub(x = position20_to_fp16, y = var_5222_cast_fp16)[name = tensor("write_idx_41_cast_fp16")]; + tensor var_5225_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_41_cast_fp16)[name = tensor("op_5225_cast_fp16")]; + tensor var_5235 = const()[name = tensor("op_5235"), val = tensor([1, 512, 1, 1])]; + tensor onehot_81_to_fp16_dtype_0 = const()[name = tensor("onehot_81_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_5225_cast_fp16_to_fp16 = cast(dtype = onehot_81_to_fp16_dtype_0, x = var_5225_cast_fp16)[name = tensor("cast_416")]; + tensor onehot_83_cast_fp16 = reshape(shape = var_5235, x = var_5225_cast_fp16_to_fp16)[name = tensor("onehot_83_cast_fp16")]; + tensor var_5237_to_fp16 = const()[name = tensor("op_5237_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5239_cast_fp16 = sub(x = var_5237_to_fp16, y = onehot_83_cast_fp16)[name = tensor("op_5239_cast_fp16")]; + tensor k_cache20_to_fp16_dtype_0 = const()[name = tensor("k_cache20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache20_to_fp16 = cast(dtype = k_cache20_to_fp16_dtype_0, x = k_cache20)[name = tensor("cast_415")]; + tensor var_5240_cast_fp16 = mul(x = k_cache20_to_fp16, y = var_5239_cast_fp16)[name = tensor("op_5240_cast_fp16")]; + tensor var_5245 = const()[name = tensor("op_5245"), val = tensor([1, 1, 16, 64])]; + tensor var_5246_cast_fp16 = reshape(shape = var_5245, x = out_83_cast_fp16)[name = tensor("op_5246_cast_fp16")]; + tensor var_5247_cast_fp16 = mul(x = var_5246_cast_fp16, y = onehot_83_cast_fp16)[name = tensor("op_5247_cast_fp16")]; + tensor new_k_cache_41_cast_fp16 = add(x = var_5240_cast_fp16, y = var_5247_cast_fp16)[name = tensor("new_k_cache_41_cast_fp16")]; + tensor new_k_cache_41_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_41_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache20_to_fp16_dtype_0 = const()[name = tensor("v_cache20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache20_to_fp16 = cast(dtype = v_cache20_to_fp16_dtype_0, x = v_cache20)[name = tensor("cast_413")]; + tensor var_5253_cast_fp16 = mul(x = v_cache20_to_fp16, y = var_5239_cast_fp16)[name = tensor("op_5253_cast_fp16")]; + tensor var_5258 = const()[name = tensor("op_5258"), val = tensor([1, 1, 16, 64])]; + tensor var_5259_cast_fp16 = reshape(shape = var_5258, x = v_41_cast_fp16)[name = tensor("op_5259_cast_fp16")]; + tensor var_5260_cast_fp16 = mul(x = var_5259_cast_fp16, y = onehot_83_cast_fp16)[name = tensor("op_5260_cast_fp16")]; + tensor new_v_cache_41_cast_fp16 = add(x = var_5253_cast_fp16, y = var_5260_cast_fp16)[name = tensor("new_v_cache_41_cast_fp16")]; + tensor new_v_cache_41_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_41_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_20_perm_0 = const()[name = tensor("transpose_20_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_5273 = const()[name = tensor("op_5273"), val = tensor([0, 2, 1, 3])]; + tensor var_5279 = const()[name = tensor("op_5279"), val = tensor([1, 16, 1, 64])]; + tensor q4_41_cast_fp16 = reshape(shape = var_5279, x = out_81_cast_fp16)[name = tensor("q4_41_cast_fp16")]; + tensor var_5284_transpose_x_0 = const()[name = tensor("op_5284_transpose_x_0"), val = tensor(false)]; + tensor var_5284_transpose_y_0 = const()[name = tensor("op_5284_transpose_y_0"), val = tensor(false)]; + tensor transpose_20_cast_fp16 = transpose(perm = transpose_20_perm_0, x = new_k_cache_41_cast_fp16)[name = tensor("transpose_35")]; + tensor var_5284_cast_fp16 = matmul(transpose_x = var_5284_transpose_x_0, transpose_y = var_5284_transpose_y_0, x = q4_41_cast_fp16, y = transpose_20_cast_fp16)[name = tensor("op_5284_cast_fp16")]; + tensor var_5285_to_fp16 = const()[name = tensor("op_5285_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_81_cast_fp16 = mul(x = var_5284_cast_fp16, y = var_5285_to_fp16)[name = tensor("scores_81_cast_fp16")]; + tensor var_5289 = const()[name = tensor("op_5289"), val = tensor([1, 1])]; + tensor var_5290_cast_fp16 = reshape(shape = var_5289, x = position20_to_fp16)[name = tensor("op_5290_cast_fp16")]; + tensor var_5291_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_5290_cast_fp16)[name = tensor("op_5291_cast_fp16")]; + tensor mask_41_to_fp16_dtype_0 = const()[name = tensor("mask_41_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_5298_to_fp16 = const()[name = tensor("op_5298_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5291_cast_fp16_to_fp16 = cast(dtype = mask_41_to_fp16_dtype_0, x = var_5291_cast_fp16)[name = tensor("cast_411")]; + tensor var_5299_cast_fp16 = sub(x = var_5291_cast_fp16_to_fp16, y = var_5298_to_fp16)[name = tensor("op_5299_cast_fp16")]; + tensor var_5304 = const()[name = tensor("op_5304"), val = tensor([1, 1, 1, 512])]; + tensor var_5305_cast_fp16 = reshape(shape = var_5304, x = var_5299_cast_fp16)[name = tensor("op_5305_cast_fp16")]; + tensor var_5306_to_fp16 = const()[name = tensor("op_5306_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_5307_cast_fp16 = mul(x = var_5305_cast_fp16, y = var_5306_to_fp16)[name = tensor("op_5307_cast_fp16")]; + tensor scores_83_cast_fp16 = add(x = scores_81_cast_fp16, y = var_5307_cast_fp16)[name = tensor("scores_83_cast_fp16")]; + tensor var_5310 = const()[name = tensor("op_5310"), val = tensor(-1)]; + tensor weights_41_cast_fp16 = softmax(axis = var_5310, x = scores_83_cast_fp16)[name = tensor("weights_41_cast_fp16")]; + tensor attn_41_transpose_x_0 = const()[name = tensor("attn_41_transpose_x_0"), val = tensor(false)]; + tensor attn_41_transpose_y_0 = const()[name = tensor("attn_41_transpose_y_0"), val = tensor(false)]; + tensor values_41_cast_fp16 = transpose(perm = var_5273, x = new_v_cache_41_cast_fp16)[name = tensor("transpose_34")]; + tensor attn_41_cast_fp16 = matmul(transpose_x = attn_41_transpose_x_0, transpose_y = attn_41_transpose_y_0, x = weights_41_cast_fp16, y = values_41_cast_fp16)[name = tensor("attn_41_cast_fp16")]; + tensor var_5318 = const()[name = tensor("op_5318"), val = tensor([0, 2, 1, 3])]; + tensor var_5323 = const()[name = tensor("op_5323"), val = tensor([1, 1, 1024])]; + tensor var_5319_cast_fp16 = transpose(perm = var_5318, x = attn_41_cast_fp16)[name = tensor("transpose_33")]; + tensor input_247_cast_fp16 = reshape(shape = var_5323, x = var_5319_cast_fp16)[name = tensor("input_247_cast_fp16")]; + tensor attn20_out_proj_weight_to_fp16 = const()[name = tensor("attn20_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(509869760)))]; + tensor linear_82_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn20_out_proj_weight_to_fp16, x = input_247_cast_fp16)[name = tensor("linear_82_cast_fp16")]; + tensor var_5329_to_fp16 = const()[name = tensor("op_5329_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5330_cast_fp16 = add(x = position20_to_fp16, y = var_5329_to_fp16)[name = tensor("op_5330_cast_fp16")]; + tensor var_5330_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5330_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_249_cast_fp16 = add(x = input_243_cast_fp16, y = linear_82_cast_fp16)[name = tensor("input_249_cast_fp16")]; + tensor input_251_axes_0 = const()[name = tensor("input_251_axes_0"), val = tensor([-1])]; + tensor norm20_2_weight_to_fp16 = const()[name = tensor("norm20_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(511966976)))]; + tensor norm20_2_bias_to_fp16 = const()[name = tensor("norm20_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(511969088)))]; + tensor var_5334_to_fp16 = const()[name = tensor("op_5334_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_251_cast_fp16 = layer_norm(axes = input_251_axes_0, beta = norm20_2_bias_to_fp16, epsilon = var_5334_to_fp16, gamma = norm20_2_weight_to_fp16, x = input_249_cast_fp16)[name = tensor("input_251_cast_fp16")]; + tensor linear20_1_weight_to_fp16 = const()[name = tensor("linear20_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(511971200)))]; + tensor linear_83_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear20_1_weight_to_fp16, x = input_251_cast_fp16)[name = tensor("linear_83_cast_fp16")]; + tensor input_253_mode_0 = const()[name = tensor("input_253_mode_0"), val = tensor("EXACT")]; + tensor input_253_cast_fp16 = gelu(mode = input_253_mode_0, x = linear_83_cast_fp16)[name = tensor("input_253_cast_fp16")]; + tensor linear20_2_weight_to_fp16 = const()[name = tensor("linear20_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(520359872)))]; + tensor linear_84_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear20_2_weight_to_fp16, x = input_253_cast_fp16)[name = tensor("linear_84_cast_fp16")]; + tensor input_255_cast_fp16 = add(x = input_249_cast_fp16, y = linear_84_cast_fp16)[name = tensor("input_255_cast_fp16")]; + tensor input_257_axes_0 = const()[name = tensor("input_257_axes_0"), val = tensor([-1])]; + tensor norm21_1_weight_to_fp16 = const()[name = tensor("norm21_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(528748544)))]; + tensor norm21_1_bias_to_fp16 = const()[name = tensor("norm21_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(528750656)))]; + tensor var_5351_to_fp16 = const()[name = tensor("op_5351_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_257_cast_fp16 = layer_norm(axes = input_257_axes_0, beta = norm21_1_bias_to_fp16, epsilon = var_5351_to_fp16, gamma = norm21_1_weight_to_fp16, x = input_255_cast_fp16)[name = tensor("input_257_cast_fp16")]; + tensor attn21_in_proj_weight_to_fp16 = const()[name = tensor("attn21_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(528752768)))]; + tensor linear_85_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn21_in_proj_weight_to_fp16, x = input_257_cast_fp16)[name = tensor("linear_85_cast_fp16")]; + tensor var_5364 = const()[name = tensor("op_5364"), val = tensor([1, 3, 16, 64])]; + tensor qkv_43_cast_fp16 = reshape(shape = var_5364, x = linear_85_cast_fp16)[name = tensor("qkv_43_cast_fp16")]; + tensor x_85_begin_0 = const()[name = tensor("x_85_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_85_end_0 = const()[name = tensor("x_85_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_85_end_mask_0 = const()[name = tensor("x_85_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_85_squeeze_mask_0 = const()[name = tensor("x_85_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_85_cast_fp16 = slice_by_index(begin = x_85_begin_0, end = x_85_end_0, end_mask = x_85_end_mask_0, squeeze_mask = x_85_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("x_85_cast_fp16")]; + tensor x_87_begin_0 = const()[name = tensor("x_87_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_87_end_0 = const()[name = tensor("x_87_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_87_end_mask_0 = const()[name = tensor("x_87_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_87_squeeze_mask_0 = const()[name = tensor("x_87_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_87_cast_fp16 = slice_by_index(begin = x_87_begin_0, end = x_87_end_0, end_mask = x_87_end_mask_0, squeeze_mask = x_87_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("x_87_cast_fp16")]; + tensor v_43_begin_0 = const()[name = tensor("v_43_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_43_end_0 = const()[name = tensor("v_43_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_43_end_mask_0 = const()[name = tensor("v_43_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_43_squeeze_mask_0 = const()[name = tensor("v_43_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_43_cast_fp16 = slice_by_index(begin = v_43_begin_0, end = v_43_end_0, end_mask = v_43_end_mask_0, squeeze_mask = v_43_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("v_43_cast_fp16")]; + tensor var_5393 = const()[name = tensor("op_5393"), val = tensor([1, 1, 1])]; + tensor position21_to_fp16_dtype_0 = const()[name = tensor("position21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position21_to_fp16 = cast(dtype = position21_to_fp16_dtype_0, x = position21)[name = tensor("cast_409")]; + tensor var_5394_cast_fp16 = reshape(shape = var_5393, x = position21_to_fp16)[name = tensor("op_5394_cast_fp16")]; + tensor angles_85_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_5394_cast_fp16)[name = tensor("angles_85_cast_fp16")]; + tensor rotr_85_cast_fp16 = cos(x = angles_85_cast_fp16)[name = tensor("rotr_85_cast_fp16")]; + tensor roti_85_cast_fp16 = sin(x = angles_85_cast_fp16)[name = tensor("roti_85_cast_fp16")]; + tensor var_5402 = const()[name = tensor("op_5402"), val = tensor([1, 16, 32, 2])]; + tensor pairs_85_cast_fp16 = reshape(shape = var_5402, x = x_85_cast_fp16)[name = tensor("pairs_85_cast_fp16")]; + tensor xr_85_begin_0 = const()[name = tensor("xr_85_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_85_end_0 = const()[name = tensor("xr_85_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_85_end_mask_0 = const()[name = tensor("xr_85_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_85_squeeze_mask_0 = const()[name = tensor("xr_85_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_85_cast_fp16 = slice_by_index(begin = xr_85_begin_0, end = xr_85_end_0, end_mask = xr_85_end_mask_0, squeeze_mask = xr_85_squeeze_mask_0, x = pairs_85_cast_fp16)[name = tensor("xr_85_cast_fp16")]; + tensor xi_85_begin_0 = const()[name = tensor("xi_85_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_85_end_0 = const()[name = tensor("xi_85_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_85_end_mask_0 = const()[name = tensor("xi_85_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_85_squeeze_mask_0 = const()[name = tensor("xi_85_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_85_cast_fp16 = slice_by_index(begin = xi_85_begin_0, end = xi_85_end_0, end_mask = xi_85_end_mask_0, squeeze_mask = xi_85_squeeze_mask_0, x = pairs_85_cast_fp16)[name = tensor("xi_85_cast_fp16")]; + tensor var_5410_cast_fp16 = mul(x = xr_85_cast_fp16, y = rotr_85_cast_fp16)[name = tensor("op_5410_cast_fp16")]; + tensor var_5411_cast_fp16 = mul(x = xi_85_cast_fp16, y = roti_85_cast_fp16)[name = tensor("op_5411_cast_fp16")]; + tensor xor__85_cast_fp16 = sub(x = var_5410_cast_fp16, y = var_5411_cast_fp16)[name = tensor("xor__85_cast_fp16")]; + tensor var_5414_cast_fp16 = mul(x = xr_85_cast_fp16, y = roti_85_cast_fp16)[name = tensor("op_5414_cast_fp16")]; + tensor var_5415_cast_fp16 = mul(x = xi_85_cast_fp16, y = rotr_85_cast_fp16)[name = tensor("op_5415_cast_fp16")]; + tensor xoi_85_cast_fp16 = add(x = var_5414_cast_fp16, y = var_5415_cast_fp16)[name = tensor("xoi_85_cast_fp16")]; + tensor out_85_axis_0 = const()[name = tensor("out_85_axis_0"), val = tensor(-1)]; + tensor out_85_cast_fp16 = stack(axis = out_85_axis_0, values = (xor__85_cast_fp16, xoi_85_cast_fp16))[name = tensor("out_85_cast_fp16")]; + tensor var_5438 = const()[name = tensor("op_5438"), val = tensor([1, 16, 32, 2])]; + tensor pairs_87_cast_fp16 = reshape(shape = var_5438, x = x_87_cast_fp16)[name = tensor("pairs_87_cast_fp16")]; + tensor xr_87_begin_0 = const()[name = tensor("xr_87_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_87_end_0 = const()[name = tensor("xr_87_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_87_end_mask_0 = const()[name = tensor("xr_87_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_87_squeeze_mask_0 = const()[name = tensor("xr_87_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_87_cast_fp16 = slice_by_index(begin = xr_87_begin_0, end = xr_87_end_0, end_mask = xr_87_end_mask_0, squeeze_mask = xr_87_squeeze_mask_0, x = pairs_87_cast_fp16)[name = tensor("xr_87_cast_fp16")]; + tensor xi_87_begin_0 = const()[name = tensor("xi_87_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_87_end_0 = const()[name = tensor("xi_87_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_87_end_mask_0 = const()[name = tensor("xi_87_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_87_squeeze_mask_0 = const()[name = tensor("xi_87_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_87_cast_fp16 = slice_by_index(begin = xi_87_begin_0, end = xi_87_end_0, end_mask = xi_87_end_mask_0, squeeze_mask = xi_87_squeeze_mask_0, x = pairs_87_cast_fp16)[name = tensor("xi_87_cast_fp16")]; + tensor var_5446_cast_fp16 = mul(x = xr_87_cast_fp16, y = rotr_85_cast_fp16)[name = tensor("op_5446_cast_fp16")]; + tensor var_5447_cast_fp16 = mul(x = xi_87_cast_fp16, y = roti_85_cast_fp16)[name = tensor("op_5447_cast_fp16")]; + tensor xor__87_cast_fp16 = sub(x = var_5446_cast_fp16, y = var_5447_cast_fp16)[name = tensor("xor__87_cast_fp16")]; + tensor var_5450_cast_fp16 = mul(x = xr_87_cast_fp16, y = roti_85_cast_fp16)[name = tensor("op_5450_cast_fp16")]; + tensor var_5451_cast_fp16 = mul(x = xi_87_cast_fp16, y = rotr_85_cast_fp16)[name = tensor("op_5451_cast_fp16")]; + tensor xoi_87_cast_fp16 = add(x = var_5450_cast_fp16, y = var_5451_cast_fp16)[name = tensor("xoi_87_cast_fp16")]; + tensor out_87_axis_0 = const()[name = tensor("out_87_axis_0"), val = tensor(-1)]; + tensor out_87_cast_fp16 = stack(axis = out_87_axis_0, values = (xor__87_cast_fp16, xoi_87_cast_fp16))[name = tensor("out_87_cast_fp16")]; + tensor _inversed_5463_y_0_to_fp16 = const()[name = tensor("_inversed_5463_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5463_cast_fp16 = mul(x = position21_to_fp16, y = _inversed_5463_y_0_to_fp16)[name = tensor("_inversed_5463_cast_fp16")]; + tensor var_5464_cast_fp16 = floor(x = _inversed_5463_cast_fp16)[name = tensor("op_5464_cast_fp16")]; + tensor var_5465_to_fp16 = const()[name = tensor("op_5465_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5466_cast_fp16 = mul(x = var_5464_cast_fp16, y = var_5465_to_fp16)[name = tensor("op_5466_cast_fp16")]; + tensor write_idx_43_cast_fp16 = sub(x = position21_to_fp16, y = var_5466_cast_fp16)[name = tensor("write_idx_43_cast_fp16")]; + tensor var_5469_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_43_cast_fp16)[name = tensor("op_5469_cast_fp16")]; + tensor var_5479 = const()[name = tensor("op_5479"), val = tensor([1, 512, 1, 1])]; + tensor onehot_85_to_fp16_dtype_0 = const()[name = tensor("onehot_85_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_5469_cast_fp16_to_fp16 = cast(dtype = onehot_85_to_fp16_dtype_0, x = var_5469_cast_fp16)[name = tensor("cast_408")]; + tensor onehot_87_cast_fp16 = reshape(shape = var_5479, x = var_5469_cast_fp16_to_fp16)[name = tensor("onehot_87_cast_fp16")]; + tensor var_5481_to_fp16 = const()[name = tensor("op_5481_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5483_cast_fp16 = sub(x = var_5481_to_fp16, y = onehot_87_cast_fp16)[name = tensor("op_5483_cast_fp16")]; + tensor k_cache21_to_fp16_dtype_0 = const()[name = tensor("k_cache21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache21_to_fp16 = cast(dtype = k_cache21_to_fp16_dtype_0, x = k_cache21)[name = tensor("cast_407")]; + tensor var_5484_cast_fp16 = mul(x = k_cache21_to_fp16, y = var_5483_cast_fp16)[name = tensor("op_5484_cast_fp16")]; + tensor var_5489 = const()[name = tensor("op_5489"), val = tensor([1, 1, 16, 64])]; + tensor var_5490_cast_fp16 = reshape(shape = var_5489, x = out_87_cast_fp16)[name = tensor("op_5490_cast_fp16")]; + tensor var_5491_cast_fp16 = mul(x = var_5490_cast_fp16, y = onehot_87_cast_fp16)[name = tensor("op_5491_cast_fp16")]; + tensor new_k_cache_43_cast_fp16 = add(x = var_5484_cast_fp16, y = var_5491_cast_fp16)[name = tensor("new_k_cache_43_cast_fp16")]; + tensor new_k_cache_43_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_43_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache21_to_fp16_dtype_0 = const()[name = tensor("v_cache21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache21_to_fp16 = cast(dtype = v_cache21_to_fp16_dtype_0, x = v_cache21)[name = tensor("cast_405")]; + tensor var_5497_cast_fp16 = mul(x = v_cache21_to_fp16, y = var_5483_cast_fp16)[name = tensor("op_5497_cast_fp16")]; + tensor var_5502 = const()[name = tensor("op_5502"), val = tensor([1, 1, 16, 64])]; + tensor var_5503_cast_fp16 = reshape(shape = var_5502, x = v_43_cast_fp16)[name = tensor("op_5503_cast_fp16")]; + tensor var_5504_cast_fp16 = mul(x = var_5503_cast_fp16, y = onehot_87_cast_fp16)[name = tensor("op_5504_cast_fp16")]; + tensor new_v_cache_43_cast_fp16 = add(x = var_5497_cast_fp16, y = var_5504_cast_fp16)[name = tensor("new_v_cache_43_cast_fp16")]; + tensor new_v_cache_43_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_43_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_21_perm_0 = const()[name = tensor("transpose_21_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_5517 = const()[name = tensor("op_5517"), val = tensor([0, 2, 1, 3])]; + tensor var_5523 = const()[name = tensor("op_5523"), val = tensor([1, 16, 1, 64])]; + tensor q4_43_cast_fp16 = reshape(shape = var_5523, x = out_85_cast_fp16)[name = tensor("q4_43_cast_fp16")]; + tensor var_5528_transpose_x_0 = const()[name = tensor("op_5528_transpose_x_0"), val = tensor(false)]; + tensor var_5528_transpose_y_0 = const()[name = tensor("op_5528_transpose_y_0"), val = tensor(false)]; + tensor transpose_21_cast_fp16 = transpose(perm = transpose_21_perm_0, x = new_k_cache_43_cast_fp16)[name = tensor("transpose_32")]; + tensor var_5528_cast_fp16 = matmul(transpose_x = var_5528_transpose_x_0, transpose_y = var_5528_transpose_y_0, x = q4_43_cast_fp16, y = transpose_21_cast_fp16)[name = tensor("op_5528_cast_fp16")]; + tensor var_5529_to_fp16 = const()[name = tensor("op_5529_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_85_cast_fp16 = mul(x = var_5528_cast_fp16, y = var_5529_to_fp16)[name = tensor("scores_85_cast_fp16")]; + tensor var_5533 = const()[name = tensor("op_5533"), val = tensor([1, 1])]; + tensor var_5534_cast_fp16 = reshape(shape = var_5533, x = position21_to_fp16)[name = tensor("op_5534_cast_fp16")]; + tensor var_5535_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_5534_cast_fp16)[name = tensor("op_5535_cast_fp16")]; + tensor mask_43_to_fp16_dtype_0 = const()[name = tensor("mask_43_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_5542_to_fp16 = const()[name = tensor("op_5542_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5535_cast_fp16_to_fp16 = cast(dtype = mask_43_to_fp16_dtype_0, x = var_5535_cast_fp16)[name = tensor("cast_403")]; + tensor var_5543_cast_fp16 = sub(x = var_5535_cast_fp16_to_fp16, y = var_5542_to_fp16)[name = tensor("op_5543_cast_fp16")]; + tensor var_5548 = const()[name = tensor("op_5548"), val = tensor([1, 1, 1, 512])]; + tensor var_5549_cast_fp16 = reshape(shape = var_5548, x = var_5543_cast_fp16)[name = tensor("op_5549_cast_fp16")]; + tensor var_5550_to_fp16 = const()[name = tensor("op_5550_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_5551_cast_fp16 = mul(x = var_5549_cast_fp16, y = var_5550_to_fp16)[name = tensor("op_5551_cast_fp16")]; + tensor scores_87_cast_fp16 = add(x = scores_85_cast_fp16, y = var_5551_cast_fp16)[name = tensor("scores_87_cast_fp16")]; + tensor var_5554 = const()[name = tensor("op_5554"), val = tensor(-1)]; + tensor weights_43_cast_fp16 = softmax(axis = var_5554, x = scores_87_cast_fp16)[name = tensor("weights_43_cast_fp16")]; + tensor attn_43_transpose_x_0 = const()[name = tensor("attn_43_transpose_x_0"), val = tensor(false)]; + tensor attn_43_transpose_y_0 = const()[name = tensor("attn_43_transpose_y_0"), val = tensor(false)]; + tensor values_43_cast_fp16 = transpose(perm = var_5517, x = new_v_cache_43_cast_fp16)[name = tensor("transpose_31")]; + tensor attn_43_cast_fp16 = matmul(transpose_x = attn_43_transpose_x_0, transpose_y = attn_43_transpose_y_0, x = weights_43_cast_fp16, y = values_43_cast_fp16)[name = tensor("attn_43_cast_fp16")]; + tensor var_5562 = const()[name = tensor("op_5562"), val = tensor([0, 2, 1, 3])]; + tensor var_5567 = const()[name = tensor("op_5567"), val = tensor([1, 1, 1024])]; + tensor var_5563_cast_fp16 = transpose(perm = var_5562, x = attn_43_cast_fp16)[name = tensor("transpose_30")]; + tensor input_259_cast_fp16 = reshape(shape = var_5567, x = var_5563_cast_fp16)[name = tensor("input_259_cast_fp16")]; + tensor attn21_out_proj_weight_to_fp16 = const()[name = tensor("attn21_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(535044288)))]; + tensor linear_86_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn21_out_proj_weight_to_fp16, x = input_259_cast_fp16)[name = tensor("linear_86_cast_fp16")]; + tensor var_5573_to_fp16 = const()[name = tensor("op_5573_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5574_cast_fp16 = add(x = position21_to_fp16, y = var_5573_to_fp16)[name = tensor("op_5574_cast_fp16")]; + tensor var_5574_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5574_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_261_cast_fp16 = add(x = input_255_cast_fp16, y = linear_86_cast_fp16)[name = tensor("input_261_cast_fp16")]; + tensor input_263_axes_0 = const()[name = tensor("input_263_axes_0"), val = tensor([-1])]; + tensor norm21_2_weight_to_fp16 = const()[name = tensor("norm21_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(537141504)))]; + tensor norm21_2_bias_to_fp16 = const()[name = tensor("norm21_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(537143616)))]; + tensor var_5578_to_fp16 = const()[name = tensor("op_5578_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_263_cast_fp16 = layer_norm(axes = input_263_axes_0, beta = norm21_2_bias_to_fp16, epsilon = var_5578_to_fp16, gamma = norm21_2_weight_to_fp16, x = input_261_cast_fp16)[name = tensor("input_263_cast_fp16")]; + tensor linear21_1_weight_to_fp16 = const()[name = tensor("linear21_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(537145728)))]; + tensor linear_87_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear21_1_weight_to_fp16, x = input_263_cast_fp16)[name = tensor("linear_87_cast_fp16")]; + tensor input_265_mode_0 = const()[name = tensor("input_265_mode_0"), val = tensor("EXACT")]; + tensor input_265_cast_fp16 = gelu(mode = input_265_mode_0, x = linear_87_cast_fp16)[name = tensor("input_265_cast_fp16")]; + tensor linear21_2_weight_to_fp16 = const()[name = tensor("linear21_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(545534400)))]; + tensor linear_88_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear21_2_weight_to_fp16, x = input_265_cast_fp16)[name = tensor("linear_88_cast_fp16")]; + tensor input_267_cast_fp16 = add(x = input_261_cast_fp16, y = linear_88_cast_fp16)[name = tensor("input_267_cast_fp16")]; + tensor input_269_axes_0 = const()[name = tensor("input_269_axes_0"), val = tensor([-1])]; + tensor norm22_1_weight_to_fp16 = const()[name = tensor("norm22_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(553923072)))]; + tensor norm22_1_bias_to_fp16 = const()[name = tensor("norm22_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(553925184)))]; + tensor var_5595_to_fp16 = const()[name = tensor("op_5595_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_269_cast_fp16 = layer_norm(axes = input_269_axes_0, beta = norm22_1_bias_to_fp16, epsilon = var_5595_to_fp16, gamma = norm22_1_weight_to_fp16, x = input_267_cast_fp16)[name = tensor("input_269_cast_fp16")]; + tensor attn22_in_proj_weight_to_fp16 = const()[name = tensor("attn22_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(553927296)))]; + tensor linear_89_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn22_in_proj_weight_to_fp16, x = input_269_cast_fp16)[name = tensor("linear_89_cast_fp16")]; + tensor var_5608 = const()[name = tensor("op_5608"), val = tensor([1, 3, 16, 64])]; + tensor qkv_45_cast_fp16 = reshape(shape = var_5608, x = linear_89_cast_fp16)[name = tensor("qkv_45_cast_fp16")]; + tensor x_89_begin_0 = const()[name = tensor("x_89_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_89_end_0 = const()[name = tensor("x_89_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_89_end_mask_0 = const()[name = tensor("x_89_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_89_squeeze_mask_0 = const()[name = tensor("x_89_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_89_cast_fp16 = slice_by_index(begin = x_89_begin_0, end = x_89_end_0, end_mask = x_89_end_mask_0, squeeze_mask = x_89_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("x_89_cast_fp16")]; + tensor x_91_begin_0 = const()[name = tensor("x_91_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_91_end_0 = const()[name = tensor("x_91_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_91_end_mask_0 = const()[name = tensor("x_91_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_91_squeeze_mask_0 = const()[name = tensor("x_91_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_91_cast_fp16 = slice_by_index(begin = x_91_begin_0, end = x_91_end_0, end_mask = x_91_end_mask_0, squeeze_mask = x_91_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("x_91_cast_fp16")]; + tensor v_45_begin_0 = const()[name = tensor("v_45_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_45_end_0 = const()[name = tensor("v_45_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_45_end_mask_0 = const()[name = tensor("v_45_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_45_squeeze_mask_0 = const()[name = tensor("v_45_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_45_cast_fp16 = slice_by_index(begin = v_45_begin_0, end = v_45_end_0, end_mask = v_45_end_mask_0, squeeze_mask = v_45_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("v_45_cast_fp16")]; + tensor var_5637 = const()[name = tensor("op_5637"), val = tensor([1, 1, 1])]; + tensor position22_to_fp16_dtype_0 = const()[name = tensor("position22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position22_to_fp16 = cast(dtype = position22_to_fp16_dtype_0, x = position22)[name = tensor("cast_401")]; + tensor var_5638_cast_fp16 = reshape(shape = var_5637, x = position22_to_fp16)[name = tensor("op_5638_cast_fp16")]; + tensor angles_89_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_5638_cast_fp16)[name = tensor("angles_89_cast_fp16")]; + tensor rotr_89_cast_fp16 = cos(x = angles_89_cast_fp16)[name = tensor("rotr_89_cast_fp16")]; + tensor roti_89_cast_fp16 = sin(x = angles_89_cast_fp16)[name = tensor("roti_89_cast_fp16")]; + tensor var_5646 = const()[name = tensor("op_5646"), val = tensor([1, 16, 32, 2])]; + tensor pairs_89_cast_fp16 = reshape(shape = var_5646, x = x_89_cast_fp16)[name = tensor("pairs_89_cast_fp16")]; + tensor xr_89_begin_0 = const()[name = tensor("xr_89_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_89_end_0 = const()[name = tensor("xr_89_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_89_end_mask_0 = const()[name = tensor("xr_89_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_89_squeeze_mask_0 = const()[name = tensor("xr_89_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_89_cast_fp16 = slice_by_index(begin = xr_89_begin_0, end = xr_89_end_0, end_mask = xr_89_end_mask_0, squeeze_mask = xr_89_squeeze_mask_0, x = pairs_89_cast_fp16)[name = tensor("xr_89_cast_fp16")]; + tensor xi_89_begin_0 = const()[name = tensor("xi_89_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_89_end_0 = const()[name = tensor("xi_89_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_89_end_mask_0 = const()[name = tensor("xi_89_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_89_squeeze_mask_0 = const()[name = tensor("xi_89_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_89_cast_fp16 = slice_by_index(begin = xi_89_begin_0, end = xi_89_end_0, end_mask = xi_89_end_mask_0, squeeze_mask = xi_89_squeeze_mask_0, x = pairs_89_cast_fp16)[name = tensor("xi_89_cast_fp16")]; + tensor var_5654_cast_fp16 = mul(x = xr_89_cast_fp16, y = rotr_89_cast_fp16)[name = tensor("op_5654_cast_fp16")]; + tensor var_5655_cast_fp16 = mul(x = xi_89_cast_fp16, y = roti_89_cast_fp16)[name = tensor("op_5655_cast_fp16")]; + tensor xor__89_cast_fp16 = sub(x = var_5654_cast_fp16, y = var_5655_cast_fp16)[name = tensor("xor__89_cast_fp16")]; + tensor var_5658_cast_fp16 = mul(x = xr_89_cast_fp16, y = roti_89_cast_fp16)[name = tensor("op_5658_cast_fp16")]; + tensor var_5659_cast_fp16 = mul(x = xi_89_cast_fp16, y = rotr_89_cast_fp16)[name = tensor("op_5659_cast_fp16")]; + tensor xoi_89_cast_fp16 = add(x = var_5658_cast_fp16, y = var_5659_cast_fp16)[name = tensor("xoi_89_cast_fp16")]; + tensor out_89_axis_0 = const()[name = tensor("out_89_axis_0"), val = tensor(-1)]; + tensor out_89_cast_fp16 = stack(axis = out_89_axis_0, values = (xor__89_cast_fp16, xoi_89_cast_fp16))[name = tensor("out_89_cast_fp16")]; + tensor var_5682 = const()[name = tensor("op_5682"), val = tensor([1, 16, 32, 2])]; + tensor pairs_91_cast_fp16 = reshape(shape = var_5682, x = x_91_cast_fp16)[name = tensor("pairs_91_cast_fp16")]; + tensor xr_91_begin_0 = const()[name = tensor("xr_91_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_91_end_0 = const()[name = tensor("xr_91_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_91_end_mask_0 = const()[name = tensor("xr_91_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_91_squeeze_mask_0 = const()[name = tensor("xr_91_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_91_cast_fp16 = slice_by_index(begin = xr_91_begin_0, end = xr_91_end_0, end_mask = xr_91_end_mask_0, squeeze_mask = xr_91_squeeze_mask_0, x = pairs_91_cast_fp16)[name = tensor("xr_91_cast_fp16")]; + tensor xi_91_begin_0 = const()[name = tensor("xi_91_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_91_end_0 = const()[name = tensor("xi_91_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_91_end_mask_0 = const()[name = tensor("xi_91_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_91_squeeze_mask_0 = const()[name = tensor("xi_91_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_91_cast_fp16 = slice_by_index(begin = xi_91_begin_0, end = xi_91_end_0, end_mask = xi_91_end_mask_0, squeeze_mask = xi_91_squeeze_mask_0, x = pairs_91_cast_fp16)[name = tensor("xi_91_cast_fp16")]; + tensor var_5690_cast_fp16 = mul(x = xr_91_cast_fp16, y = rotr_89_cast_fp16)[name = tensor("op_5690_cast_fp16")]; + tensor var_5691_cast_fp16 = mul(x = xi_91_cast_fp16, y = roti_89_cast_fp16)[name = tensor("op_5691_cast_fp16")]; + tensor xor__91_cast_fp16 = sub(x = var_5690_cast_fp16, y = var_5691_cast_fp16)[name = tensor("xor__91_cast_fp16")]; + tensor var_5694_cast_fp16 = mul(x = xr_91_cast_fp16, y = roti_89_cast_fp16)[name = tensor("op_5694_cast_fp16")]; + tensor var_5695_cast_fp16 = mul(x = xi_91_cast_fp16, y = rotr_89_cast_fp16)[name = tensor("op_5695_cast_fp16")]; + tensor xoi_91_cast_fp16 = add(x = var_5694_cast_fp16, y = var_5695_cast_fp16)[name = tensor("xoi_91_cast_fp16")]; + tensor out_91_axis_0 = const()[name = tensor("out_91_axis_0"), val = tensor(-1)]; + tensor out_91_cast_fp16 = stack(axis = out_91_axis_0, values = (xor__91_cast_fp16, xoi_91_cast_fp16))[name = tensor("out_91_cast_fp16")]; + tensor _inversed_5707_y_0_to_fp16 = const()[name = tensor("_inversed_5707_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5707_cast_fp16 = mul(x = position22_to_fp16, y = _inversed_5707_y_0_to_fp16)[name = tensor("_inversed_5707_cast_fp16")]; + tensor var_5708_cast_fp16 = floor(x = _inversed_5707_cast_fp16)[name = tensor("op_5708_cast_fp16")]; + tensor var_5709_to_fp16 = const()[name = tensor("op_5709_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5710_cast_fp16 = mul(x = var_5708_cast_fp16, y = var_5709_to_fp16)[name = tensor("op_5710_cast_fp16")]; + tensor write_idx_45_cast_fp16 = sub(x = position22_to_fp16, y = var_5710_cast_fp16)[name = tensor("write_idx_45_cast_fp16")]; + tensor var_5713_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_45_cast_fp16)[name = tensor("op_5713_cast_fp16")]; + tensor var_5723 = const()[name = tensor("op_5723"), val = tensor([1, 512, 1, 1])]; + tensor onehot_89_to_fp16_dtype_0 = const()[name = tensor("onehot_89_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_5713_cast_fp16_to_fp16 = cast(dtype = onehot_89_to_fp16_dtype_0, x = var_5713_cast_fp16)[name = tensor("cast_400")]; + tensor onehot_91_cast_fp16 = reshape(shape = var_5723, x = var_5713_cast_fp16_to_fp16)[name = tensor("onehot_91_cast_fp16")]; + tensor var_5725_to_fp16 = const()[name = tensor("op_5725_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5727_cast_fp16 = sub(x = var_5725_to_fp16, y = onehot_91_cast_fp16)[name = tensor("op_5727_cast_fp16")]; + tensor k_cache22_to_fp16_dtype_0 = const()[name = tensor("k_cache22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache22_to_fp16 = cast(dtype = k_cache22_to_fp16_dtype_0, x = k_cache22)[name = tensor("cast_399")]; + tensor var_5728_cast_fp16 = mul(x = k_cache22_to_fp16, y = var_5727_cast_fp16)[name = tensor("op_5728_cast_fp16")]; + tensor var_5733 = const()[name = tensor("op_5733"), val = tensor([1, 1, 16, 64])]; + tensor var_5734_cast_fp16 = reshape(shape = var_5733, x = out_91_cast_fp16)[name = tensor("op_5734_cast_fp16")]; + tensor var_5735_cast_fp16 = mul(x = var_5734_cast_fp16, y = onehot_91_cast_fp16)[name = tensor("op_5735_cast_fp16")]; + tensor new_k_cache_45_cast_fp16 = add(x = var_5728_cast_fp16, y = var_5735_cast_fp16)[name = tensor("new_k_cache_45_cast_fp16")]; + tensor new_k_cache_45_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_45_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache22_to_fp16_dtype_0 = const()[name = tensor("v_cache22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache22_to_fp16 = cast(dtype = v_cache22_to_fp16_dtype_0, x = v_cache22)[name = tensor("cast_397")]; + tensor var_5741_cast_fp16 = mul(x = v_cache22_to_fp16, y = var_5727_cast_fp16)[name = tensor("op_5741_cast_fp16")]; + tensor var_5746 = const()[name = tensor("op_5746"), val = tensor([1, 1, 16, 64])]; + tensor var_5747_cast_fp16 = reshape(shape = var_5746, x = v_45_cast_fp16)[name = tensor("op_5747_cast_fp16")]; + tensor var_5748_cast_fp16 = mul(x = var_5747_cast_fp16, y = onehot_91_cast_fp16)[name = tensor("op_5748_cast_fp16")]; + tensor new_v_cache_45_cast_fp16 = add(x = var_5741_cast_fp16, y = var_5748_cast_fp16)[name = tensor("new_v_cache_45_cast_fp16")]; + tensor new_v_cache_45_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_45_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_22_perm_0 = const()[name = tensor("transpose_22_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_5761 = const()[name = tensor("op_5761"), val = tensor([0, 2, 1, 3])]; + tensor var_5767 = const()[name = tensor("op_5767"), val = tensor([1, 16, 1, 64])]; + tensor q4_45_cast_fp16 = reshape(shape = var_5767, x = out_89_cast_fp16)[name = tensor("q4_45_cast_fp16")]; + tensor var_5772_transpose_x_0 = const()[name = tensor("op_5772_transpose_x_0"), val = tensor(false)]; + tensor var_5772_transpose_y_0 = const()[name = tensor("op_5772_transpose_y_0"), val = tensor(false)]; + tensor transpose_22_cast_fp16 = transpose(perm = transpose_22_perm_0, x = new_k_cache_45_cast_fp16)[name = tensor("transpose_29")]; + tensor var_5772_cast_fp16 = matmul(transpose_x = var_5772_transpose_x_0, transpose_y = var_5772_transpose_y_0, x = q4_45_cast_fp16, y = transpose_22_cast_fp16)[name = tensor("op_5772_cast_fp16")]; + tensor var_5773_to_fp16 = const()[name = tensor("op_5773_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_89_cast_fp16 = mul(x = var_5772_cast_fp16, y = var_5773_to_fp16)[name = tensor("scores_89_cast_fp16")]; + tensor var_5777 = const()[name = tensor("op_5777"), val = tensor([1, 1])]; + tensor var_5778_cast_fp16 = reshape(shape = var_5777, x = position22_to_fp16)[name = tensor("op_5778_cast_fp16")]; + tensor var_5779_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_5778_cast_fp16)[name = tensor("op_5779_cast_fp16")]; + tensor mask_45_to_fp16_dtype_0 = const()[name = tensor("mask_45_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_5786_to_fp16 = const()[name = tensor("op_5786_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5779_cast_fp16_to_fp16 = cast(dtype = mask_45_to_fp16_dtype_0, x = var_5779_cast_fp16)[name = tensor("cast_395")]; + tensor var_5787_cast_fp16 = sub(x = var_5779_cast_fp16_to_fp16, y = var_5786_to_fp16)[name = tensor("op_5787_cast_fp16")]; + tensor var_5792 = const()[name = tensor("op_5792"), val = tensor([1, 1, 1, 512])]; + tensor var_5793_cast_fp16 = reshape(shape = var_5792, x = var_5787_cast_fp16)[name = tensor("op_5793_cast_fp16")]; + tensor var_5794_to_fp16 = const()[name = tensor("op_5794_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_5795_cast_fp16 = mul(x = var_5793_cast_fp16, y = var_5794_to_fp16)[name = tensor("op_5795_cast_fp16")]; + tensor scores_91_cast_fp16 = add(x = scores_89_cast_fp16, y = var_5795_cast_fp16)[name = tensor("scores_91_cast_fp16")]; + tensor var_5798 = const()[name = tensor("op_5798"), val = tensor(-1)]; + tensor weights_45_cast_fp16 = softmax(axis = var_5798, x = scores_91_cast_fp16)[name = tensor("weights_45_cast_fp16")]; + tensor attn_45_transpose_x_0 = const()[name = tensor("attn_45_transpose_x_0"), val = tensor(false)]; + tensor attn_45_transpose_y_0 = const()[name = tensor("attn_45_transpose_y_0"), val = tensor(false)]; + tensor values_45_cast_fp16 = transpose(perm = var_5761, x = new_v_cache_45_cast_fp16)[name = tensor("transpose_28")]; + tensor attn_45_cast_fp16 = matmul(transpose_x = attn_45_transpose_x_0, transpose_y = attn_45_transpose_y_0, x = weights_45_cast_fp16, y = values_45_cast_fp16)[name = tensor("attn_45_cast_fp16")]; + tensor var_5806 = const()[name = tensor("op_5806"), val = tensor([0, 2, 1, 3])]; + tensor var_5811 = const()[name = tensor("op_5811"), val = tensor([1, 1, 1024])]; + tensor var_5807_cast_fp16 = transpose(perm = var_5806, x = attn_45_cast_fp16)[name = tensor("transpose_27")]; + tensor input_271_cast_fp16 = reshape(shape = var_5811, x = var_5807_cast_fp16)[name = tensor("input_271_cast_fp16")]; + tensor attn22_out_proj_weight_to_fp16 = const()[name = tensor("attn22_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(560218816)))]; + tensor linear_90_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn22_out_proj_weight_to_fp16, x = input_271_cast_fp16)[name = tensor("linear_90_cast_fp16")]; + tensor var_5817_to_fp16 = const()[name = tensor("op_5817_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5818_cast_fp16 = add(x = position22_to_fp16, y = var_5817_to_fp16)[name = tensor("op_5818_cast_fp16")]; + tensor var_5818_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5818_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_273_cast_fp16 = add(x = input_267_cast_fp16, y = linear_90_cast_fp16)[name = tensor("input_273_cast_fp16")]; + tensor input_275_axes_0 = const()[name = tensor("input_275_axes_0"), val = tensor([-1])]; + tensor norm22_2_weight_to_fp16 = const()[name = tensor("norm22_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(562316032)))]; + tensor norm22_2_bias_to_fp16 = const()[name = tensor("norm22_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(562318144)))]; + tensor var_5822_to_fp16 = const()[name = tensor("op_5822_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_275_cast_fp16 = layer_norm(axes = input_275_axes_0, beta = norm22_2_bias_to_fp16, epsilon = var_5822_to_fp16, gamma = norm22_2_weight_to_fp16, x = input_273_cast_fp16)[name = tensor("input_275_cast_fp16")]; + tensor linear22_1_weight_to_fp16 = const()[name = tensor("linear22_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(562320256)))]; + tensor linear_91_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear22_1_weight_to_fp16, x = input_275_cast_fp16)[name = tensor("linear_91_cast_fp16")]; + tensor input_277_mode_0 = const()[name = tensor("input_277_mode_0"), val = tensor("EXACT")]; + tensor input_277_cast_fp16 = gelu(mode = input_277_mode_0, x = linear_91_cast_fp16)[name = tensor("input_277_cast_fp16")]; + tensor linear22_2_weight_to_fp16 = const()[name = tensor("linear22_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(570708928)))]; + tensor linear_92_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear22_2_weight_to_fp16, x = input_277_cast_fp16)[name = tensor("linear_92_cast_fp16")]; + tensor input_279_cast_fp16 = add(x = input_273_cast_fp16, y = linear_92_cast_fp16)[name = tensor("input_279_cast_fp16")]; + tensor input_281_axes_0 = const()[name = tensor("input_281_axes_0"), val = tensor([-1])]; + tensor norm23_1_weight_to_fp16 = const()[name = tensor("norm23_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(579097600)))]; + tensor norm23_1_bias_to_fp16 = const()[name = tensor("norm23_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(579099712)))]; + tensor var_5839_to_fp16 = const()[name = tensor("op_5839_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_281_cast_fp16 = layer_norm(axes = input_281_axes_0, beta = norm23_1_bias_to_fp16, epsilon = var_5839_to_fp16, gamma = norm23_1_weight_to_fp16, x = input_279_cast_fp16)[name = tensor("input_281_cast_fp16")]; + tensor attn23_in_proj_weight_to_fp16 = const()[name = tensor("attn23_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(579101824)))]; + tensor linear_93_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn23_in_proj_weight_to_fp16, x = input_281_cast_fp16)[name = tensor("linear_93_cast_fp16")]; + tensor var_5852 = const()[name = tensor("op_5852"), val = tensor([1, 3, 16, 64])]; + tensor qkv_cast_fp16 = reshape(shape = var_5852, x = linear_93_cast_fp16)[name = tensor("qkv_cast_fp16")]; + tensor x_93_begin_0 = const()[name = tensor("x_93_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_93_end_0 = const()[name = tensor("x_93_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_93_end_mask_0 = const()[name = tensor("x_93_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_93_squeeze_mask_0 = const()[name = tensor("x_93_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_93_cast_fp16 = slice_by_index(begin = x_93_begin_0, end = x_93_end_0, end_mask = x_93_end_mask_0, squeeze_mask = x_93_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("x_93_cast_fp16")]; + tensor x_begin_0 = const()[name = tensor("x_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_end_0 = const()[name = tensor("x_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_end_mask_0 = const()[name = tensor("x_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_squeeze_mask_0 = const()[name = tensor("x_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_cast_fp16 = slice_by_index(begin = x_begin_0, end = x_end_0, end_mask = x_end_mask_0, squeeze_mask = x_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("x_cast_fp16")]; + tensor v_begin_0 = const()[name = tensor("v_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_end_0 = const()[name = tensor("v_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_end_mask_0 = const()[name = tensor("v_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_squeeze_mask_0 = const()[name = tensor("v_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_cast_fp16 = slice_by_index(begin = v_begin_0, end = v_end_0, end_mask = v_end_mask_0, squeeze_mask = v_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("v_cast_fp16")]; + tensor var_5881 = const()[name = tensor("op_5881"), val = tensor([1, 1, 1])]; + tensor position23_to_fp16_dtype_0 = const()[name = tensor("position23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position23_to_fp16 = cast(dtype = position23_to_fp16_dtype_0, x = position23)[name = tensor("cast_393")]; + tensor var_5882_cast_fp16 = reshape(shape = var_5881, x = position23_to_fp16)[name = tensor("op_5882_cast_fp16")]; + tensor angles_93_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_5882_cast_fp16)[name = tensor("angles_93_cast_fp16")]; + tensor rotr_93_cast_fp16 = cos(x = angles_93_cast_fp16)[name = tensor("rotr_93_cast_fp16")]; + tensor roti_93_cast_fp16 = sin(x = angles_93_cast_fp16)[name = tensor("roti_93_cast_fp16")]; + tensor var_5890 = const()[name = tensor("op_5890"), val = tensor([1, 16, 32, 2])]; + tensor pairs_93_cast_fp16 = reshape(shape = var_5890, x = x_93_cast_fp16)[name = tensor("pairs_93_cast_fp16")]; + tensor xr_93_begin_0 = const()[name = tensor("xr_93_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_93_end_0 = const()[name = tensor("xr_93_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_93_end_mask_0 = const()[name = tensor("xr_93_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_93_squeeze_mask_0 = const()[name = tensor("xr_93_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_93_cast_fp16 = slice_by_index(begin = xr_93_begin_0, end = xr_93_end_0, end_mask = xr_93_end_mask_0, squeeze_mask = xr_93_squeeze_mask_0, x = pairs_93_cast_fp16)[name = tensor("xr_93_cast_fp16")]; + tensor xi_93_begin_0 = const()[name = tensor("xi_93_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_93_end_0 = const()[name = tensor("xi_93_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_93_end_mask_0 = const()[name = tensor("xi_93_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_93_squeeze_mask_0 = const()[name = tensor("xi_93_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_93_cast_fp16 = slice_by_index(begin = xi_93_begin_0, end = xi_93_end_0, end_mask = xi_93_end_mask_0, squeeze_mask = xi_93_squeeze_mask_0, x = pairs_93_cast_fp16)[name = tensor("xi_93_cast_fp16")]; + tensor var_5898_cast_fp16 = mul(x = xr_93_cast_fp16, y = rotr_93_cast_fp16)[name = tensor("op_5898_cast_fp16")]; + tensor var_5899_cast_fp16 = mul(x = xi_93_cast_fp16, y = roti_93_cast_fp16)[name = tensor("op_5899_cast_fp16")]; + tensor xor__93_cast_fp16 = sub(x = var_5898_cast_fp16, y = var_5899_cast_fp16)[name = tensor("xor__93_cast_fp16")]; + tensor var_5902_cast_fp16 = mul(x = xr_93_cast_fp16, y = roti_93_cast_fp16)[name = tensor("op_5902_cast_fp16")]; + tensor var_5903_cast_fp16 = mul(x = xi_93_cast_fp16, y = rotr_93_cast_fp16)[name = tensor("op_5903_cast_fp16")]; + tensor xoi_93_cast_fp16 = add(x = var_5902_cast_fp16, y = var_5903_cast_fp16)[name = tensor("xoi_93_cast_fp16")]; + tensor out_93_axis_0 = const()[name = tensor("out_93_axis_0"), val = tensor(-1)]; + tensor out_93_cast_fp16 = stack(axis = out_93_axis_0, values = (xor__93_cast_fp16, xoi_93_cast_fp16))[name = tensor("out_93_cast_fp16")]; + tensor var_5926 = const()[name = tensor("op_5926"), val = tensor([1, 16, 32, 2])]; + tensor pairs_cast_fp16 = reshape(shape = var_5926, x = x_cast_fp16)[name = tensor("pairs_cast_fp16")]; + tensor xr_begin_0 = const()[name = tensor("xr_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_end_0 = const()[name = tensor("xr_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_end_mask_0 = const()[name = tensor("xr_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_squeeze_mask_0 = const()[name = tensor("xr_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_cast_fp16 = slice_by_index(begin = xr_begin_0, end = xr_end_0, end_mask = xr_end_mask_0, squeeze_mask = xr_squeeze_mask_0, x = pairs_cast_fp16)[name = tensor("xr_cast_fp16")]; + tensor xi_begin_0 = const()[name = tensor("xi_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_end_0 = const()[name = tensor("xi_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_end_mask_0 = const()[name = tensor("xi_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_squeeze_mask_0 = const()[name = tensor("xi_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_cast_fp16 = slice_by_index(begin = xi_begin_0, end = xi_end_0, end_mask = xi_end_mask_0, squeeze_mask = xi_squeeze_mask_0, x = pairs_cast_fp16)[name = tensor("xi_cast_fp16")]; + tensor var_5934_cast_fp16 = mul(x = xr_cast_fp16, y = rotr_93_cast_fp16)[name = tensor("op_5934_cast_fp16")]; + tensor var_5935_cast_fp16 = mul(x = xi_cast_fp16, y = roti_93_cast_fp16)[name = tensor("op_5935_cast_fp16")]; + tensor xor__cast_fp16 = sub(x = var_5934_cast_fp16, y = var_5935_cast_fp16)[name = tensor("xor__cast_fp16")]; + tensor var_5938_cast_fp16 = mul(x = xr_cast_fp16, y = roti_93_cast_fp16)[name = tensor("op_5938_cast_fp16")]; + tensor var_5939_cast_fp16 = mul(x = xi_cast_fp16, y = rotr_93_cast_fp16)[name = tensor("op_5939_cast_fp16")]; + tensor xoi_cast_fp16 = add(x = var_5938_cast_fp16, y = var_5939_cast_fp16)[name = tensor("xoi_cast_fp16")]; + tensor out_axis_0 = const()[name = tensor("out_axis_0"), val = tensor(-1)]; + tensor out_cast_fp16 = stack(axis = out_axis_0, values = (xor__cast_fp16, xoi_cast_fp16))[name = tensor("out_cast_fp16")]; + tensor _inversed_5951_y_0_to_fp16 = const()[name = tensor("_inversed_5951_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5951_cast_fp16 = mul(x = position23_to_fp16, y = _inversed_5951_y_0_to_fp16)[name = tensor("_inversed_5951_cast_fp16")]; + tensor var_5952_cast_fp16 = floor(x = _inversed_5951_cast_fp16)[name = tensor("op_5952_cast_fp16")]; + tensor var_5953_to_fp16 = const()[name = tensor("op_5953_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5954_cast_fp16 = mul(x = var_5952_cast_fp16, y = var_5953_to_fp16)[name = tensor("op_5954_cast_fp16")]; + tensor write_idx_cast_fp16 = sub(x = position23_to_fp16, y = var_5954_cast_fp16)[name = tensor("write_idx_cast_fp16")]; + tensor var_5957_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_cast_fp16)[name = tensor("op_5957_cast_fp16")]; + tensor var_5967 = const()[name = tensor("op_5967"), val = tensor([1, 512, 1, 1])]; + tensor onehot_93_to_fp16_dtype_0 = const()[name = tensor("onehot_93_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_5957_cast_fp16_to_fp16 = cast(dtype = onehot_93_to_fp16_dtype_0, x = var_5957_cast_fp16)[name = tensor("cast_392")]; + tensor onehot_cast_fp16 = reshape(shape = var_5967, x = var_5957_cast_fp16_to_fp16)[name = tensor("onehot_cast_fp16")]; + tensor var_5969_to_fp16 = const()[name = tensor("op_5969_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5971_cast_fp16 = sub(x = var_5969_to_fp16, y = onehot_cast_fp16)[name = tensor("op_5971_cast_fp16")]; + tensor k_cache23_to_fp16_dtype_0 = const()[name = tensor("k_cache23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache23_to_fp16 = cast(dtype = k_cache23_to_fp16_dtype_0, x = k_cache23)[name = tensor("cast_391")]; + tensor var_5972_cast_fp16 = mul(x = k_cache23_to_fp16, y = var_5971_cast_fp16)[name = tensor("op_5972_cast_fp16")]; + tensor var_5977 = const()[name = tensor("op_5977"), val = tensor([1, 1, 16, 64])]; + tensor var_5978_cast_fp16 = reshape(shape = var_5977, x = out_cast_fp16)[name = tensor("op_5978_cast_fp16")]; + tensor var_5979_cast_fp16 = mul(x = var_5978_cast_fp16, y = onehot_cast_fp16)[name = tensor("op_5979_cast_fp16")]; + tensor new_k_cache_cast_fp16 = add(x = var_5972_cast_fp16, y = var_5979_cast_fp16)[name = tensor("new_k_cache_cast_fp16")]; + tensor new_k_cache_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache23_to_fp16_dtype_0 = const()[name = tensor("v_cache23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache23_to_fp16 = cast(dtype = v_cache23_to_fp16_dtype_0, x = v_cache23)[name = tensor("cast_389")]; + tensor var_5985_cast_fp16 = mul(x = v_cache23_to_fp16, y = var_5971_cast_fp16)[name = tensor("op_5985_cast_fp16")]; + tensor var_5990 = const()[name = tensor("op_5990"), val = tensor([1, 1, 16, 64])]; + tensor var_5991_cast_fp16 = reshape(shape = var_5990, x = v_cast_fp16)[name = tensor("op_5991_cast_fp16")]; + tensor var_5992_cast_fp16 = mul(x = var_5991_cast_fp16, y = onehot_cast_fp16)[name = tensor("op_5992_cast_fp16")]; + tensor new_v_cache_cast_fp16 = add(x = var_5985_cast_fp16, y = var_5992_cast_fp16)[name = tensor("new_v_cache_cast_fp16")]; + tensor new_v_cache_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_23_perm_0 = const()[name = tensor("transpose_23_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_6005 = const()[name = tensor("op_6005"), val = tensor([0, 2, 1, 3])]; + tensor var_6011 = const()[name = tensor("op_6011"), val = tensor([1, 16, 1, 64])]; + tensor q4_cast_fp16 = reshape(shape = var_6011, x = out_93_cast_fp16)[name = tensor("q4_cast_fp16")]; + tensor var_6016_transpose_x_0 = const()[name = tensor("op_6016_transpose_x_0"), val = tensor(false)]; + tensor var_6016_transpose_y_0 = const()[name = tensor("op_6016_transpose_y_0"), val = tensor(false)]; + tensor transpose_23_cast_fp16 = transpose(perm = transpose_23_perm_0, x = new_k_cache_cast_fp16)[name = tensor("transpose_26")]; + tensor var_6016_cast_fp16 = matmul(transpose_x = var_6016_transpose_x_0, transpose_y = var_6016_transpose_y_0, x = q4_cast_fp16, y = transpose_23_cast_fp16)[name = tensor("op_6016_cast_fp16")]; + tensor var_6017_to_fp16 = const()[name = tensor("op_6017_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_93_cast_fp16 = mul(x = var_6016_cast_fp16, y = var_6017_to_fp16)[name = tensor("scores_93_cast_fp16")]; + tensor var_6021 = const()[name = tensor("op_6021"), val = tensor([1, 1])]; + tensor var_6022_cast_fp16 = reshape(shape = var_6021, x = position23_to_fp16)[name = tensor("op_6022_cast_fp16")]; + tensor var_6023_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_6022_cast_fp16)[name = tensor("op_6023_cast_fp16")]; + tensor mask_to_fp16_dtype_0 = const()[name = tensor("mask_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_6030_to_fp16 = const()[name = tensor("op_6030_to_fp16"), val = tensor(0x1p+0)]; + tensor var_6023_cast_fp16_to_fp16 = cast(dtype = mask_to_fp16_dtype_0, x = var_6023_cast_fp16)[name = tensor("cast_387")]; + tensor var_6031_cast_fp16 = sub(x = var_6023_cast_fp16_to_fp16, y = var_6030_to_fp16)[name = tensor("op_6031_cast_fp16")]; + tensor var_6036 = const()[name = tensor("op_6036"), val = tensor([1, 1, 1, 512])]; + tensor var_6037_cast_fp16 = reshape(shape = var_6036, x = var_6031_cast_fp16)[name = tensor("op_6037_cast_fp16")]; + tensor var_6038_to_fp16 = const()[name = tensor("op_6038_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_6039_cast_fp16 = mul(x = var_6037_cast_fp16, y = var_6038_to_fp16)[name = tensor("op_6039_cast_fp16")]; + tensor scores_cast_fp16 = add(x = scores_93_cast_fp16, y = var_6039_cast_fp16)[name = tensor("scores_cast_fp16")]; + tensor var_6042 = const()[name = tensor("op_6042"), val = tensor(-1)]; + tensor weights_cast_fp16 = softmax(axis = var_6042, x = scores_cast_fp16)[name = tensor("weights_cast_fp16")]; + tensor attn_transpose_x_0 = const()[name = tensor("attn_transpose_x_0"), val = tensor(false)]; + tensor attn_transpose_y_0 = const()[name = tensor("attn_transpose_y_0"), val = tensor(false)]; + tensor values_cast_fp16 = transpose(perm = var_6005, x = new_v_cache_cast_fp16)[name = tensor("transpose_25")]; + tensor attn_cast_fp16 = matmul(transpose_x = attn_transpose_x_0, transpose_y = attn_transpose_y_0, x = weights_cast_fp16, y = values_cast_fp16)[name = tensor("attn_cast_fp16")]; + tensor var_6050 = const()[name = tensor("op_6050"), val = tensor([0, 2, 1, 3])]; + tensor var_6055 = const()[name = tensor("op_6055"), val = tensor([1, 1, 1024])]; + tensor var_6051_cast_fp16 = transpose(perm = var_6050, x = attn_cast_fp16)[name = tensor("transpose_24")]; + tensor input_283_cast_fp16 = reshape(shape = var_6055, x = var_6051_cast_fp16)[name = tensor("input_283_cast_fp16")]; + tensor attn23_out_proj_weight_to_fp16 = const()[name = tensor("attn23_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(585393344)))]; + tensor linear_94_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn23_out_proj_weight_to_fp16, x = input_283_cast_fp16)[name = tensor("linear_94_cast_fp16")]; + tensor var_6061_to_fp16 = const()[name = tensor("op_6061_to_fp16"), val = tensor(0x1p+0)]; + tensor var_6062_cast_fp16 = add(x = position23_to_fp16, y = var_6061_to_fp16)[name = tensor("op_6062_cast_fp16")]; + tensor var_6062_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_6062_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_285_cast_fp16 = add(x = input_279_cast_fp16, y = linear_94_cast_fp16)[name = tensor("input_285_cast_fp16")]; + tensor input_287_axes_0 = const()[name = tensor("input_287_axes_0"), val = tensor([-1])]; + tensor norm23_2_weight_to_fp16 = const()[name = tensor("norm23_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(587490560)))]; + tensor norm23_2_bias_to_fp16 = const()[name = tensor("norm23_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(587492672)))]; + tensor var_6066_to_fp16 = const()[name = tensor("op_6066_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_287_cast_fp16 = layer_norm(axes = input_287_axes_0, beta = norm23_2_bias_to_fp16, epsilon = var_6066_to_fp16, gamma = norm23_2_weight_to_fp16, x = input_285_cast_fp16)[name = tensor("input_287_cast_fp16")]; + tensor linear23_1_weight_to_fp16 = const()[name = tensor("linear23_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(587494784)))]; + tensor linear_95_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear23_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_95_cast_fp16")]; + tensor input_289_mode_0 = const()[name = tensor("input_289_mode_0"), val = tensor("EXACT")]; + tensor input_289_cast_fp16 = gelu(mode = input_289_mode_0, x = linear_95_cast_fp16)[name = tensor("input_289_cast_fp16")]; + tensor linear23_2_weight_to_fp16 = const()[name = tensor("linear23_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(595883456)))]; + tensor linear_96_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear23_2_weight_to_fp16, x = input_289_cast_fp16)[name = tensor("linear_96_cast_fp16")]; + tensor input_291_cast_fp16 = add(x = input_285_cast_fp16, y = linear_96_cast_fp16)[name = tensor("input_291_cast_fp16")]; + tensor input_axes_0 = const()[name = tensor("input_axes_0"), val = tensor([-1])]; + tensor out_norm_weight_to_fp16 = const()[name = tensor("out_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(604272128)))]; + tensor out_norm_bias_to_fp16 = const()[name = tensor("out_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(604274240)))]; + tensor var_6083_to_fp16 = const()[name = tensor("op_6083_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_cast_fp16 = layer_norm(axes = input_axes_0, beta = out_norm_bias_to_fp16, epsilon = var_6083_to_fp16, gamma = out_norm_weight_to_fp16, x = input_291_cast_fp16)[name = tensor("input_cast_fp16")]; + tensor input_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("input_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor out_eos_weight_to_fp16 = const()[name = tensor("out_eos_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(604276352)))]; + tensor out_eos_bias_to_fp16 = const()[name = tensor("out_eos_bias_to_fp16"), val = tensor([-0x1.3ep-3])]; + tensor linear_97_cast_fp16 = linear(bias = out_eos_bias_to_fp16, weight = out_eos_weight_to_fp16, x = input_cast_fp16)[name = tensor("linear_97_cast_fp16")]; + tensor linear_97_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("linear_97_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor is_eos = cast(dtype = linear_97_cast_fp16_to_fp32_dtype_0, x = linear_97_cast_fp16)[name = tensor("cast_384")]; + tensor transformer_out = cast(dtype = input_cast_fp16_to_fp32_dtype_0, x = input_cast_fp16)[name = tensor("cast_385")]; + tensor new_position23 = cast(dtype = var_6062_cast_fp16_to_fp32_dtype_0, x = var_6062_cast_fp16)[name = tensor("cast_386")]; + tensor new_v_cache23 = cast(dtype = new_v_cache_cast_fp16_to_fp32_dtype_0, x = new_v_cache_cast_fp16)[name = tensor("cast_388")]; + tensor new_k_cache23 = cast(dtype = new_k_cache_cast_fp16_to_fp32_dtype_0, x = new_k_cache_cast_fp16)[name = tensor("cast_390")]; + tensor new_position22 = cast(dtype = var_5818_cast_fp16_to_fp32_dtype_0, x = var_5818_cast_fp16)[name = tensor("cast_394")]; + tensor new_v_cache22 = cast(dtype = new_v_cache_45_cast_fp16_to_fp32_dtype_0, x = new_v_cache_45_cast_fp16)[name = tensor("cast_396")]; + tensor new_k_cache22 = cast(dtype = new_k_cache_45_cast_fp16_to_fp32_dtype_0, x = new_k_cache_45_cast_fp16)[name = tensor("cast_398")]; + tensor new_position21 = cast(dtype = var_5574_cast_fp16_to_fp32_dtype_0, x = var_5574_cast_fp16)[name = tensor("cast_402")]; + tensor new_v_cache21 = cast(dtype = new_v_cache_43_cast_fp16_to_fp32_dtype_0, x = new_v_cache_43_cast_fp16)[name = tensor("cast_404")]; + tensor new_k_cache21 = cast(dtype = new_k_cache_43_cast_fp16_to_fp32_dtype_0, x = new_k_cache_43_cast_fp16)[name = tensor("cast_406")]; + tensor new_position20 = cast(dtype = var_5330_cast_fp16_to_fp32_dtype_0, x = var_5330_cast_fp16)[name = tensor("cast_410")]; + tensor new_v_cache20 = cast(dtype = new_v_cache_41_cast_fp16_to_fp32_dtype_0, x = new_v_cache_41_cast_fp16)[name = tensor("cast_412")]; + tensor new_k_cache20 = cast(dtype = new_k_cache_41_cast_fp16_to_fp32_dtype_0, x = new_k_cache_41_cast_fp16)[name = tensor("cast_414")]; + tensor new_position19 = cast(dtype = var_5086_cast_fp16_to_fp32_dtype_0, x = var_5086_cast_fp16)[name = tensor("cast_418")]; + tensor new_v_cache19 = cast(dtype = new_v_cache_39_cast_fp16_to_fp32_dtype_0, x = new_v_cache_39_cast_fp16)[name = tensor("cast_420")]; + tensor new_k_cache19 = cast(dtype = new_k_cache_39_cast_fp16_to_fp32_dtype_0, x = new_k_cache_39_cast_fp16)[name = tensor("cast_422")]; + tensor new_position18 = cast(dtype = var_4842_cast_fp16_to_fp32_dtype_0, x = var_4842_cast_fp16)[name = tensor("cast_426")]; + tensor new_v_cache18 = cast(dtype = new_v_cache_37_cast_fp16_to_fp32_dtype_0, x = new_v_cache_37_cast_fp16)[name = tensor("cast_428")]; + tensor new_k_cache18 = cast(dtype = new_k_cache_37_cast_fp16_to_fp32_dtype_0, x = new_k_cache_37_cast_fp16)[name = tensor("cast_430")]; + tensor new_position17 = cast(dtype = var_4598_cast_fp16_to_fp32_dtype_0, x = var_4598_cast_fp16)[name = tensor("cast_434")]; + tensor new_v_cache17 = cast(dtype = new_v_cache_35_cast_fp16_to_fp32_dtype_0, x = new_v_cache_35_cast_fp16)[name = tensor("cast_436")]; + tensor new_k_cache17 = cast(dtype = new_k_cache_35_cast_fp16_to_fp32_dtype_0, x = new_k_cache_35_cast_fp16)[name = tensor("cast_438")]; + tensor new_position16 = cast(dtype = var_4354_cast_fp16_to_fp32_dtype_0, x = var_4354_cast_fp16)[name = tensor("cast_442")]; + tensor new_v_cache16 = cast(dtype = new_v_cache_33_cast_fp16_to_fp32_dtype_0, x = new_v_cache_33_cast_fp16)[name = tensor("cast_444")]; + tensor new_k_cache16 = cast(dtype = new_k_cache_33_cast_fp16_to_fp32_dtype_0, x = new_k_cache_33_cast_fp16)[name = tensor("cast_446")]; + tensor new_position15 = cast(dtype = var_4110_cast_fp16_to_fp32_dtype_0, x = var_4110_cast_fp16)[name = tensor("cast_450")]; + tensor new_v_cache15 = cast(dtype = new_v_cache_31_cast_fp16_to_fp32_dtype_0, x = new_v_cache_31_cast_fp16)[name = tensor("cast_452")]; + tensor new_k_cache15 = cast(dtype = new_k_cache_31_cast_fp16_to_fp32_dtype_0, x = new_k_cache_31_cast_fp16)[name = tensor("cast_454")]; + tensor new_position14 = cast(dtype = var_3866_cast_fp16_to_fp32_dtype_0, x = var_3866_cast_fp16)[name = tensor("cast_458")]; + tensor new_v_cache14 = cast(dtype = new_v_cache_29_cast_fp16_to_fp32_dtype_0, x = new_v_cache_29_cast_fp16)[name = tensor("cast_460")]; + tensor new_k_cache14 = cast(dtype = new_k_cache_29_cast_fp16_to_fp32_dtype_0, x = new_k_cache_29_cast_fp16)[name = tensor("cast_462")]; + tensor new_position13 = cast(dtype = var_3622_cast_fp16_to_fp32_dtype_0, x = var_3622_cast_fp16)[name = tensor("cast_466")]; + tensor new_v_cache13 = cast(dtype = new_v_cache_27_cast_fp16_to_fp32_dtype_0, x = new_v_cache_27_cast_fp16)[name = tensor("cast_468")]; + tensor new_k_cache13 = cast(dtype = new_k_cache_27_cast_fp16_to_fp32_dtype_0, x = new_k_cache_27_cast_fp16)[name = tensor("cast_470")]; + tensor new_position12 = cast(dtype = var_3378_cast_fp16_to_fp32_dtype_0, x = var_3378_cast_fp16)[name = tensor("cast_474")]; + tensor new_v_cache12 = cast(dtype = new_v_cache_25_cast_fp16_to_fp32_dtype_0, x = new_v_cache_25_cast_fp16)[name = tensor("cast_476")]; + tensor new_k_cache12 = cast(dtype = new_k_cache_25_cast_fp16_to_fp32_dtype_0, x = new_k_cache_25_cast_fp16)[name = tensor("cast_478")]; + tensor new_position11 = cast(dtype = var_3134_cast_fp16_to_fp32_dtype_0, x = var_3134_cast_fp16)[name = tensor("cast_482")]; + tensor new_v_cache11 = cast(dtype = new_v_cache_23_cast_fp16_to_fp32_dtype_0, x = new_v_cache_23_cast_fp16)[name = tensor("cast_484")]; + tensor new_k_cache11 = cast(dtype = new_k_cache_23_cast_fp16_to_fp32_dtype_0, x = new_k_cache_23_cast_fp16)[name = tensor("cast_486")]; + tensor new_position10 = cast(dtype = var_2890_cast_fp16_to_fp32_dtype_0, x = var_2890_cast_fp16)[name = tensor("cast_490")]; + tensor new_v_cache10 = cast(dtype = new_v_cache_21_cast_fp16_to_fp32_dtype_0, x = new_v_cache_21_cast_fp16)[name = tensor("cast_492")]; + tensor new_k_cache10 = cast(dtype = new_k_cache_21_cast_fp16_to_fp32_dtype_0, x = new_k_cache_21_cast_fp16)[name = tensor("cast_494")]; + tensor new_position9 = cast(dtype = var_2646_cast_fp16_to_fp32_dtype_0, x = var_2646_cast_fp16)[name = tensor("cast_498")]; + tensor new_v_cache9 = cast(dtype = new_v_cache_19_cast_fp16_to_fp32_dtype_0, x = new_v_cache_19_cast_fp16)[name = tensor("cast_500")]; + tensor new_k_cache9 = cast(dtype = new_k_cache_19_cast_fp16_to_fp32_dtype_0, x = new_k_cache_19_cast_fp16)[name = tensor("cast_502")]; + tensor new_position8 = cast(dtype = var_2402_cast_fp16_to_fp32_dtype_0, x = var_2402_cast_fp16)[name = tensor("cast_506")]; + tensor new_v_cache8 = cast(dtype = new_v_cache_17_cast_fp16_to_fp32_dtype_0, x = new_v_cache_17_cast_fp16)[name = tensor("cast_508")]; + tensor new_k_cache8 = cast(dtype = new_k_cache_17_cast_fp16_to_fp32_dtype_0, x = new_k_cache_17_cast_fp16)[name = tensor("cast_510")]; + tensor new_position7 = cast(dtype = var_2158_cast_fp16_to_fp32_dtype_0, x = var_2158_cast_fp16)[name = tensor("cast_514")]; + tensor new_v_cache7 = cast(dtype = new_v_cache_15_cast_fp16_to_fp32_dtype_0, x = new_v_cache_15_cast_fp16)[name = tensor("cast_516")]; + tensor new_k_cache7 = cast(dtype = new_k_cache_15_cast_fp16_to_fp32_dtype_0, x = new_k_cache_15_cast_fp16)[name = tensor("cast_518")]; + tensor new_position6 = cast(dtype = var_1914_cast_fp16_to_fp32_dtype_0, x = var_1914_cast_fp16)[name = tensor("cast_522")]; + tensor new_v_cache6 = cast(dtype = new_v_cache_13_cast_fp16_to_fp32_dtype_0, x = new_v_cache_13_cast_fp16)[name = tensor("cast_524")]; + tensor new_k_cache6 = cast(dtype = new_k_cache_13_cast_fp16_to_fp32_dtype_0, x = new_k_cache_13_cast_fp16)[name = tensor("cast_526")]; + tensor new_position5 = cast(dtype = var_1670_cast_fp16_to_fp32_dtype_0, x = var_1670_cast_fp16)[name = tensor("cast_530")]; + tensor new_v_cache5 = cast(dtype = new_v_cache_11_cast_fp16_to_fp32_dtype_0, x = new_v_cache_11_cast_fp16)[name = tensor("cast_532")]; + tensor new_k_cache5 = cast(dtype = new_k_cache_11_cast_fp16_to_fp32_dtype_0, x = new_k_cache_11_cast_fp16)[name = tensor("cast_534")]; + tensor new_position4 = cast(dtype = var_1426_cast_fp16_to_fp32_dtype_0, x = var_1426_cast_fp16)[name = tensor("cast_538")]; + tensor new_v_cache4 = cast(dtype = new_v_cache_9_cast_fp16_to_fp32_dtype_0, x = new_v_cache_9_cast_fp16)[name = tensor("cast_540")]; + tensor new_k_cache4 = cast(dtype = new_k_cache_9_cast_fp16_to_fp32_dtype_0, x = new_k_cache_9_cast_fp16)[name = tensor("cast_542")]; + tensor new_position3 = cast(dtype = var_1182_cast_fp16_to_fp32_dtype_0, x = var_1182_cast_fp16)[name = tensor("cast_546")]; + tensor new_v_cache3 = cast(dtype = new_v_cache_7_cast_fp16_to_fp32_dtype_0, x = new_v_cache_7_cast_fp16)[name = tensor("cast_548")]; + tensor new_k_cache3 = cast(dtype = new_k_cache_7_cast_fp16_to_fp32_dtype_0, x = new_k_cache_7_cast_fp16)[name = tensor("cast_550")]; + tensor new_position2 = cast(dtype = var_938_cast_fp16_to_fp32_dtype_0, x = var_938_cast_fp16)[name = tensor("cast_554")]; + tensor new_v_cache2 = cast(dtype = new_v_cache_5_cast_fp16_to_fp32_dtype_0, x = new_v_cache_5_cast_fp16)[name = tensor("cast_556")]; + tensor new_k_cache2 = cast(dtype = new_k_cache_5_cast_fp16_to_fp32_dtype_0, x = new_k_cache_5_cast_fp16)[name = tensor("cast_558")]; + tensor new_position1 = cast(dtype = var_694_cast_fp16_to_fp32_dtype_0, x = var_694_cast_fp16)[name = tensor("cast_562")]; + tensor new_v_cache1 = cast(dtype = new_v_cache_3_cast_fp16_to_fp32_dtype_0, x = new_v_cache_3_cast_fp16)[name = tensor("cast_564")]; + tensor new_k_cache1 = cast(dtype = new_k_cache_3_cast_fp16_to_fp32_dtype_0, x = new_k_cache_3_cast_fp16)[name = tensor("cast_566")]; + tensor new_position0 = cast(dtype = var_450_cast_fp16_to_fp32_dtype_0, x = var_450_cast_fp16)[name = tensor("cast_570")]; + tensor new_v_cache0 = cast(dtype = new_v_cache_1_cast_fp16_to_fp32_dtype_0, x = new_v_cache_1_cast_fp16)[name = tensor("cast_572")]; + tensor new_k_cache0 = cast(dtype = new_k_cache_1_cast_fp16_to_fp32_dtype_0, x = new_k_cache_1_cast_fp16)[name = tensor("cast_574")]; + } -> (transformer_out, is_eos, new_k_cache0, new_v_cache0, new_position0, new_k_cache1, new_v_cache1, new_position1, new_k_cache2, new_v_cache2, new_position2, new_k_cache3, new_v_cache3, new_position3, new_k_cache4, new_v_cache4, new_position4, new_k_cache5, new_v_cache5, new_position5, new_k_cache6, new_v_cache6, new_position6, new_k_cache7, new_v_cache7, new_position7, new_k_cache8, new_v_cache8, new_position8, new_k_cache9, new_v_cache9, new_position9, new_k_cache10, new_v_cache10, new_position10, new_k_cache11, new_v_cache11, new_position11, new_k_cache12, new_v_cache12, new_position12, new_k_cache13, new_v_cache13, new_position13, new_k_cache14, new_v_cache14, new_position14, new_k_cache15, new_v_cache15, new_position15, new_k_cache16, new_v_cache16, new_position16, new_k_cache17, new_v_cache17, new_position17, new_k_cache18, new_v_cache18, new_position18, new_k_cache19, new_v_cache19, new_position19, new_k_cache20, new_v_cache20, new_position20, new_k_cache21, new_v_cache21, new_position21, new_k_cache22, new_v_cache22, new_position22, new_k_cache23, new_v_cache23, new_position23); +} \ No newline at end of file diff --git a/v2.1/german_24l/flowlm_step_ane.mlmodelc/weights/weight.bin b/v2.1/german_24l/flowlm_step_ane.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..48e818591d25b37ac9443c02a464cb8457f255e2 --- /dev/null +++ b/v2.1/german_24l/flowlm_step_ane.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c83b78f5871d4587b5c76b067255fa9029d6631c0ac64c9d02d937f02c5cd8ae +size 604278464 diff --git a/v2.1/german_24l/flowlm_step_ane.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/german_24l/flowlm_step_ane.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..7db35f9d836ca2a867654e04bd10de1b40457b91 --- /dev/null +++ b/v2.1/german_24l/flowlm_step_ane.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:552821e8ca858c4e54c561991e2d4e255d931ccd4329d8c605d73ff1d61b15f5 +size 545925 diff --git a/v2.1/german_24l/flowlm_step_ane.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/german_24l/flowlm_step_ane.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..48e818591d25b37ac9443c02a464cb8457f255e2 --- /dev/null +++ b/v2.1/german_24l/flowlm_step_ane.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c83b78f5871d4587b5c76b067255fa9029d6631c0ac64c9d02d937f02c5cd8ae +size 604278464 diff --git a/v2.1/german_24l/flowlm_step_ane.mlpackage/Manifest.json b/v2.1/german_24l/flowlm_step_ane.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..7d5d2a2aec2735640ade4dba734d3a329a09ce2b --- /dev/null +++ b/v2.1/german_24l/flowlm_step_ane.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "7E06AB87-4D13-4821-AF33-DB54A876642A": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "8A95EB89-85F0-40ED-AFEC-9906DDD14E3D": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "8A95EB89-85F0-40ED-AFEC-9906DDD14E3D" +} diff --git a/v2.1/italian/cond_prefill_ane.mlmodelc/analytics/coremldata.bin b/v2.1/italian/cond_prefill_ane.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..b99f56a7510ef5254a89de3afeeca3f0de3f5e8f --- /dev/null +++ b/v2.1/italian/cond_prefill_ane.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d6f3cf58fdeef96430089c8fb5b26d04da36aa074c43222edd4f7390cc7c8f1b +size 243 diff --git a/v2.1/italian/cond_prefill_ane.mlmodelc/coremldata.bin b/v2.1/italian/cond_prefill_ane.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..fe7217f1eb6ea3d8a1102b0a2abf49b17df8a63a --- /dev/null +++ b/v2.1/italian/cond_prefill_ane.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1046f9cd30e90a9e50f31f16dd0569eadf97ff7f7a0a6d3c5a5a47e55a2150ab +size 1388 diff --git a/v2.1/italian/cond_prefill_ane.mlmodelc/model.mil b/v2.1/italian/cond_prefill_ane.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..2aef44ccd7c895876ef314e46a742909df52fe28 --- /dev/null +++ b/v2.1/italian/cond_prefill_ane.mlmodelc/model.mil @@ -0,0 +1,804 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.9.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor conditioning, tensor k_cache0, tensor k_cache1, tensor k_cache2, tensor k_cache3, tensor k_cache4, tensor k_cache5, tensor position0, tensor position1, tensor position2, tensor position3, tensor position4, tensor position5, tensor v_cache0, tensor v_cache1, tensor v_cache2, tensor v_cache3, tensor v_cache4, tensor v_cache5, tensor valid_len) { + tensor token_indices_to_fp16 = const()[name = tensor("token_indices_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor position0_to_fp16_dtype_0 = const()[name = tensor("position0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position0_to_fp16 = cast(dtype = position0_to_fp16_dtype_0, x = position0)[name = tensor("cast_128")]; + tensor ts_cast_fp16 = add(x = token_indices_to_fp16, y = position0_to_fp16)[name = tensor("ts_cast_fp16")]; + tensor var_69 = const()[name = tensor("op_69"), val = tensor([256, 1, 1])]; + tensor var_70_cast_fp16 = reshape(shape = var_69, x = ts_cast_fp16)[name = tensor("op_70_cast_fp16")]; + tensor var_65_to_fp16 = const()[name = tensor("op_65_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(640)))]; + tensor angles_cast_fp16 = mul(x = var_65_to_fp16, y = var_70_cast_fp16)[name = tensor("angles_cast_fp16")]; + tensor rotr_cast_fp16 = cos(x = angles_cast_fp16)[name = tensor("rotr_cast_fp16")]; + tensor roti_cast_fp16 = sin(x = angles_cast_fp16)[name = tensor("roti_cast_fp16")]; + tensor _inversed_75_y_0_to_fp16 = const()[name = tensor("_inversed_75_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_75_cast_fp16 = mul(x = ts_cast_fp16, y = _inversed_75_y_0_to_fp16)[name = tensor("_inversed_75_cast_fp16")]; + tensor var_76_cast_fp16 = floor(x = _inversed_75_cast_fp16)[name = tensor("op_76_cast_fp16")]; + tensor var_77_to_fp16 = const()[name = tensor("op_77_to_fp16"), val = tensor(0x1p+9)]; + tensor var_78_cast_fp16 = mul(x = var_76_cast_fp16, y = var_77_to_fp16)[name = tensor("op_78_cast_fp16")]; + tensor write_idx_1_cast_fp16 = sub(x = ts_cast_fp16, y = var_78_cast_fp16)[name = tensor("write_idx_1_cast_fp16")]; + tensor valid_len_to_fp16_dtype_0 = const()[name = tensor("valid_len_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor valid_len_to_fp16 = cast(dtype = valid_len_to_fp16_dtype_0, x = valid_len)[name = tensor("cast_127")]; + tensor var_84_cast_fp16 = less(x = token_indices_to_fp16, y = valid_len_to_fp16)[name = tensor("op_84_cast_fp16")]; + tensor token_valid_to_fp16_dtype_0 = const()[name = tensor("token_valid_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_84_cast_fp16_to_fp16 = cast(dtype = token_valid_to_fp16_dtype_0, x = var_84_cast_fp16)[name = tensor("cast_126")]; + tensor var_90_cast_fp16 = mul(x = write_idx_1_cast_fp16, y = var_84_cast_fp16_to_fp16)[name = tensor("op_90_cast_fp16")]; + tensor var_91_to_fp16 = const()[name = tensor("op_91_to_fp16"), val = tensor(0x1p+0)]; + tensor var_93_cast_fp16 = sub(x = var_91_to_fp16, y = var_84_cast_fp16_to_fp16)[name = tensor("op_93_cast_fp16")]; + tensor var_94_to_fp16 = const()[name = tensor("op_94_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_95_cast_fp16 = mul(x = var_93_cast_fp16, y = var_94_to_fp16)[name = tensor("op_95_cast_fp16")]; + tensor write_idx_cast_fp16 = add(x = var_90_cast_fp16, y = var_95_cast_fp16)[name = tensor("write_idx_cast_fp16")]; + tensor var_100 = const()[name = tensor("op_100"), val = tensor([256, 1])]; + tensor var_101_cast_fp16 = reshape(shape = var_100, x = write_idx_cast_fp16)[name = tensor("op_101_cast_fp16")]; + tensor var_105_to_fp16 = const()[name = tensor("op_105_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(768)))]; + tensor var_106_cast_fp16 = equal(x = var_101_cast_fp16, y = var_105_to_fp16)[name = tensor("op_106_cast_fp16")]; + tensor var_116_axes_0 = const()[name = tensor("op_116_axes_0"), val = tensor([0])]; + tensor var_116_keep_dims_0 = const()[name = tensor("op_116_keep_dims_0"), val = tensor(false)]; + tensor assign_to_fp16_dtype_0 = const()[name = tensor("assign_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_106_cast_fp16_to_fp16 = cast(dtype = assign_to_fp16_dtype_0, x = var_106_cast_fp16)[name = tensor("cast_125")]; + tensor var_116_cast_fp16 = reduce_sum(axes = var_116_axes_0, keep_dims = var_116_keep_dims_0, x = var_106_cast_fp16_to_fp16)[name = tensor("op_116_cast_fp16")]; + tensor const_0_to_fp16 = const()[name = tensor("const_0_to_fp16"), val = tensor(-inf)]; + tensor var_118_to_fp16 = const()[name = tensor("op_118_to_fp16"), val = tensor(0x1p+0)]; + tensor clip_0_cast_fp16 = clip(alpha = const_0_to_fp16, beta = var_118_to_fp16, x = var_116_cast_fp16)[name = tensor("clip_0_cast_fp16")]; + tensor var_122 = const()[name = tensor("op_122"), val = tensor([1, 1])]; + tensor var_123_cast_fp16 = reshape(shape = var_122, x = position0_to_fp16)[name = tensor("op_123_cast_fp16")]; + tensor var_127_to_fp16 = const()[name = tensor("op_127_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1856)))]; + tensor q_positions_cast_fp16 = add(x = var_123_cast_fp16, y = var_127_to_fp16)[name = tensor("q_positions_cast_fp16")]; + tensor var_135_cast_fp16 = add(x = position0_to_fp16, y = valid_len_to_fp16)[name = tensor("op_135_cast_fp16")]; + tensor var_138 = const()[name = tensor("op_138"), val = tensor([1, 1])]; + tensor valid_end_cast_fp16 = reshape(shape = var_138, x = var_135_cast_fp16)[name = tensor("valid_end_cast_fp16")]; + tensor var_140_cast_fp16 = less(x = var_105_to_fp16, y = valid_end_cast_fp16)[name = tensor("op_140_cast_fp16")]; + tensor var_141_cast_fp16 = less_equal(x = var_105_to_fp16, y = q_positions_cast_fp16)[name = tensor("op_141_cast_fp16")]; + tensor var_142 = logical_and(x = var_140_cast_fp16, y = var_141_cast_fp16)[name = tensor("op_142")]; + tensor mask_to_fp16_dtype_0 = const()[name = tensor("mask_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_149_to_fp16 = const()[name = tensor("op_149_to_fp16"), val = tensor(0x1p+0)]; + tensor var_142_to_fp16 = cast(dtype = mask_to_fp16_dtype_0, x = var_142)[name = tensor("cast_124")]; + tensor var_150_cast_fp16 = sub(x = var_142_to_fp16, y = var_149_to_fp16)[name = tensor("op_150_cast_fp16")]; + tensor var_151_to_fp16 = const()[name = tensor("op_151_to_fp16"), val = tensor(0x1.388p+13)]; + tensor mask_bias_cast_fp16 = mul(x = var_150_cast_fp16, y = var_151_to_fp16)[name = tensor("mask_bias_cast_fp16")]; + tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; + tensor conditioning_to_fp16_dtype_0 = const()[name = tensor("conditioning_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor norm0_1_weight_to_fp16 = const()[name = tensor("norm0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2432)))]; + tensor norm0_1_bias_to_fp16 = const()[name = tensor("norm0_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4544)))]; + tensor var_154_to_fp16 = const()[name = tensor("op_154_to_fp16"), val = tensor(0x1.5p-17)]; + tensor conditioning_to_fp16 = cast(dtype = conditioning_to_fp16_dtype_0, x = conditioning)[name = tensor("cast_123")]; + tensor input_3_cast_fp16 = layer_norm(axes = input_3_axes_0, beta = norm0_1_bias_to_fp16, epsilon = var_154_to_fp16, gamma = norm0_1_weight_to_fp16, x = conditioning_to_fp16)[name = tensor("input_3_cast_fp16")]; + tensor attn0_in_proj_weight_to_fp16 = const()[name = tensor("attn0_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6656)))]; + tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6298176)))]; + tensor linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn0_in_proj_weight_to_fp16, x = input_3_cast_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor var_168 = const()[name = tensor("op_168"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_1_cast_fp16 = reshape(shape = var_168, x = linear_0_cast_fp16)[name = tensor("qkv_1_cast_fp16")]; + tensor x_1_begin_0 = const()[name = tensor("x_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_1_end_0 = const()[name = tensor("x_1_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_1_end_mask_0 = const()[name = tensor("x_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_1_squeeze_mask_0 = const()[name = tensor("x_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_1_cast_fp16 = slice_by_index(begin = x_1_begin_0, end = x_1_end_0, end_mask = x_1_end_mask_0, squeeze_mask = x_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("x_1_cast_fp16")]; + tensor x_3_begin_0 = const()[name = tensor("x_3_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_3_end_0 = const()[name = tensor("x_3_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_3_end_mask_0 = const()[name = tensor("x_3_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_3_squeeze_mask_0 = const()[name = tensor("x_3_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_3_cast_fp16 = slice_by_index(begin = x_3_begin_0, end = x_3_end_0, end_mask = x_3_end_mask_0, squeeze_mask = x_3_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("x_3_cast_fp16")]; + tensor v_1_begin_0 = const()[name = tensor("v_1_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_1_end_0 = const()[name = tensor("v_1_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_1_end_mask_0 = const()[name = tensor("v_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_1_squeeze_mask_0 = const()[name = tensor("v_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_1_cast_fp16 = slice_by_index(begin = v_1_begin_0, end = v_1_end_0, end_mask = v_1_end_mask_0, squeeze_mask = v_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("v_1_cast_fp16")]; + tensor var_213 = const()[name = tensor("op_213"), val = tensor([256, 16, 32, 2])]; + tensor pairs_1_cast_fp16 = reshape(shape = var_213, x = x_1_cast_fp16)[name = tensor("pairs_1_cast_fp16")]; + tensor xr_1_begin_0 = const()[name = tensor("xr_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_1_end_0 = const()[name = tensor("xr_1_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_1_end_mask_0 = const()[name = tensor("xr_1_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_1_squeeze_mask_0 = const()[name = tensor("xr_1_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_1_cast_fp16 = slice_by_index(begin = xr_1_begin_0, end = xr_1_end_0, end_mask = xr_1_end_mask_0, squeeze_mask = xr_1_squeeze_mask_0, x = pairs_1_cast_fp16)[name = tensor("xr_1_cast_fp16")]; + tensor xi_1_begin_0 = const()[name = tensor("xi_1_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_1_end_0 = const()[name = tensor("xi_1_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_1_end_mask_0 = const()[name = tensor("xi_1_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_1_squeeze_mask_0 = const()[name = tensor("xi_1_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_1_cast_fp16 = slice_by_index(begin = xi_1_begin_0, end = xi_1_end_0, end_mask = xi_1_end_mask_0, squeeze_mask = xi_1_squeeze_mask_0, x = pairs_1_cast_fp16)[name = tensor("xi_1_cast_fp16")]; + tensor var_221_cast_fp16 = mul(x = xr_1_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_221_cast_fp16")]; + tensor var_222_cast_fp16 = mul(x = xi_1_cast_fp16, y = roti_cast_fp16)[name = tensor("op_222_cast_fp16")]; + tensor xor__1_cast_fp16 = sub(x = var_221_cast_fp16, y = var_222_cast_fp16)[name = tensor("xor__1_cast_fp16")]; + tensor var_225_cast_fp16 = mul(x = xr_1_cast_fp16, y = roti_cast_fp16)[name = tensor("op_225_cast_fp16")]; + tensor var_226_cast_fp16 = mul(x = xi_1_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_226_cast_fp16")]; + tensor xoi_1_cast_fp16 = add(x = var_225_cast_fp16, y = var_226_cast_fp16)[name = tensor("xoi_1_cast_fp16")]; + tensor out_1_axis_0 = const()[name = tensor("out_1_axis_0"), val = tensor(-1)]; + tensor out_1_cast_fp16 = stack(axis = out_1_axis_0, values = (xor__1_cast_fp16, xoi_1_cast_fp16))[name = tensor("out_1_cast_fp16")]; + tensor var_236 = const()[name = tensor("op_236"), val = tensor([1, 256, 16, 64])]; + tensor q_1_cast_fp16 = reshape(shape = var_236, x = out_1_cast_fp16)[name = tensor("q_1_cast_fp16")]; + tensor var_242 = const()[name = tensor("op_242"), val = tensor([256, 16, 32, 2])]; + tensor pairs_3_cast_fp16 = reshape(shape = var_242, x = x_3_cast_fp16)[name = tensor("pairs_3_cast_fp16")]; + tensor xr_3_begin_0 = const()[name = tensor("xr_3_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_3_end_0 = const()[name = tensor("xr_3_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_3_end_mask_0 = const()[name = tensor("xr_3_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_3_squeeze_mask_0 = const()[name = tensor("xr_3_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_3_cast_fp16 = slice_by_index(begin = xr_3_begin_0, end = xr_3_end_0, end_mask = xr_3_end_mask_0, squeeze_mask = xr_3_squeeze_mask_0, x = pairs_3_cast_fp16)[name = tensor("xr_3_cast_fp16")]; + tensor xi_3_begin_0 = const()[name = tensor("xi_3_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_3_end_0 = const()[name = tensor("xi_3_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_3_end_mask_0 = const()[name = tensor("xi_3_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_3_squeeze_mask_0 = const()[name = tensor("xi_3_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_3_cast_fp16 = slice_by_index(begin = xi_3_begin_0, end = xi_3_end_0, end_mask = xi_3_end_mask_0, squeeze_mask = xi_3_squeeze_mask_0, x = pairs_3_cast_fp16)[name = tensor("xi_3_cast_fp16")]; + tensor var_250_cast_fp16 = mul(x = xr_3_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_250_cast_fp16")]; + tensor var_251_cast_fp16 = mul(x = xi_3_cast_fp16, y = roti_cast_fp16)[name = tensor("op_251_cast_fp16")]; + tensor xor__3_cast_fp16 = sub(x = var_250_cast_fp16, y = var_251_cast_fp16)[name = tensor("xor__3_cast_fp16")]; + tensor var_254_cast_fp16 = mul(x = xr_3_cast_fp16, y = roti_cast_fp16)[name = tensor("op_254_cast_fp16")]; + tensor var_255_cast_fp16 = mul(x = xi_3_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_255_cast_fp16")]; + tensor xoi_3_cast_fp16 = add(x = var_254_cast_fp16, y = var_255_cast_fp16)[name = tensor("xoi_3_cast_fp16")]; + tensor out_3_axis_0 = const()[name = tensor("out_3_axis_0"), val = tensor(-1)]; + tensor out_3_cast_fp16 = stack(axis = out_3_axis_0, values = (xor__3_cast_fp16, xoi_3_cast_fp16))[name = tensor("out_3_cast_fp16")]; + tensor var_272 = const()[name = tensor("op_272"), val = tensor([256, 1024])]; + tensor var_273_cast_fp16 = reshape(shape = var_272, x = out_3_cast_fp16)[name = tensor("op_273_cast_fp16")]; + tensor written_k_1_transpose_x_1 = const()[name = tensor("written_k_1_transpose_x_1"), val = tensor(true)]; + tensor written_k_1_transpose_y_1 = const()[name = tensor("written_k_1_transpose_y_1"), val = tensor(false)]; + tensor written_k_1_cast_fp16 = matmul(transpose_x = written_k_1_transpose_x_1, transpose_y = written_k_1_transpose_y_1, x = var_106_cast_fp16_to_fp16, y = var_273_cast_fp16)[name = tensor("written_k_1_cast_fp16")]; + tensor var_280 = const()[name = tensor("op_280"), val = tensor([256, 1024])]; + tensor var_281_cast_fp16 = reshape(shape = var_280, x = v_1_cast_fp16)[name = tensor("op_281_cast_fp16")]; + tensor written_v_1_transpose_x_1 = const()[name = tensor("written_v_1_transpose_x_1"), val = tensor(true)]; + tensor written_v_1_transpose_y_1 = const()[name = tensor("written_v_1_transpose_y_1"), val = tensor(false)]; + tensor written_v_1_cast_fp16 = matmul(transpose_x = written_v_1_transpose_x_1, transpose_y = written_v_1_transpose_y_1, x = var_106_cast_fp16_to_fp16, y = var_281_cast_fp16)[name = tensor("written_v_1_cast_fp16")]; + tensor var_283_to_fp16 = const()[name = tensor("op_283_to_fp16"), val = tensor(0x1p+0)]; + tensor var_285_cast_fp16 = sub(x = var_283_to_fp16, y = clip_0_cast_fp16)[name = tensor("op_285_cast_fp16")]; + tensor var_290 = const()[name = tensor("op_290"), val = tensor([1, 512, 1, 1])]; + tensor keep_1_cast_fp16 = reshape(shape = var_290, x = var_285_cast_fp16)[name = tensor("keep_1_cast_fp16")]; + tensor k_cache0_to_fp16_dtype_0 = const()[name = tensor("k_cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache0_to_fp16 = cast(dtype = k_cache0_to_fp16_dtype_0, x = k_cache0)[name = tensor("cast_122")]; + tensor var_292_cast_fp16 = mul(x = k_cache0_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_292_cast_fp16")]; + tensor var_297 = const()[name = tensor("op_297"), val = tensor([1, 512, 16, 64])]; + tensor var_298_cast_fp16 = reshape(shape = var_297, x = written_k_1_cast_fp16)[name = tensor("op_298_cast_fp16")]; + tensor new_k_cache_1_cast_fp16 = add(x = var_292_cast_fp16, y = var_298_cast_fp16)[name = tensor("new_k_cache_1_cast_fp16")]; + tensor new_k_cache_1_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_1_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache0_to_fp16_dtype_0 = const()[name = tensor("v_cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache0_to_fp16 = cast(dtype = v_cache0_to_fp16_dtype_0, x = v_cache0)[name = tensor("cast_120")]; + tensor var_301_cast_fp16 = mul(x = v_cache0_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_301_cast_fp16")]; + tensor var_306 = const()[name = tensor("op_306"), val = tensor([1, 512, 16, 64])]; + tensor var_307_cast_fp16 = reshape(shape = var_306, x = written_v_1_cast_fp16)[name = tensor("op_307_cast_fp16")]; + tensor new_v_cache_1_cast_fp16 = add(x = var_301_cast_fp16, y = var_307_cast_fp16)[name = tensor("new_v_cache_1_cast_fp16")]; + tensor new_v_cache_1_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_1_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_320 = const()[name = tensor("op_320"), val = tensor([0, 2, 1, 3])]; + tensor var_331_transpose_x_0 = const()[name = tensor("op_331_transpose_x_0"), val = tensor(false)]; + tensor var_331_transpose_y_0 = const()[name = tensor("op_331_transpose_y_0"), val = tensor(false)]; + tensor transpose_15_perm_0 = const()[name = tensor("transpose_15_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_16_perm_0 = const()[name = tensor("transpose_16_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_16 = transpose(perm = transpose_16_perm_0, x = new_k_cache_1_cast_fp16)[name = tensor("transpose_42")]; + tensor transpose_15 = transpose(perm = transpose_15_perm_0, x = q_1_cast_fp16)[name = tensor("transpose_43")]; + tensor var_331_cast_fp16 = matmul(transpose_x = var_331_transpose_x_0, transpose_y = var_331_transpose_y_0, x = transpose_15, y = transpose_16)[name = tensor("op_331_cast_fp16")]; + tensor var_332_to_fp16 = const()[name = tensor("op_332_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_1_cast_fp16 = mul(x = var_331_cast_fp16, y = var_332_to_fp16)[name = tensor("scores_1_cast_fp16")]; + tensor var_338 = const()[name = tensor("op_338"), val = tensor([1, 1, 256, 512])]; + tensor var_339_cast_fp16 = reshape(shape = var_338, x = mask_bias_cast_fp16)[name = tensor("op_339_cast_fp16")]; + tensor scores_3_cast_fp16 = add(x = scores_1_cast_fp16, y = var_339_cast_fp16)[name = tensor("scores_3_cast_fp16")]; + tensor var_342 = const()[name = tensor("op_342"), val = tensor(-1)]; + tensor weights_1_cast_fp16 = softmax(axis = var_342, x = scores_3_cast_fp16)[name = tensor("weights_1_cast_fp16")]; + tensor attn_1_transpose_x_0 = const()[name = tensor("attn_1_transpose_x_0"), val = tensor(false)]; + tensor attn_1_transpose_y_0 = const()[name = tensor("attn_1_transpose_y_0"), val = tensor(false)]; + tensor values_1_cast_fp16 = transpose(perm = var_320, x = new_v_cache_1_cast_fp16)[name = tensor("transpose_44")]; + tensor attn_1_cast_fp16 = matmul(transpose_x = attn_1_transpose_x_0, transpose_y = attn_1_transpose_y_0, x = weights_1_cast_fp16, y = values_1_cast_fp16)[name = tensor("attn_1_cast_fp16")]; + tensor var_350 = const()[name = tensor("op_350"), val = tensor([0, 2, 1, 3])]; + tensor var_355 = const()[name = tensor("op_355"), val = tensor([1, 256, 1024])]; + tensor var_351_cast_fp16 = transpose(perm = var_350, x = attn_1_cast_fp16)[name = tensor("transpose_41")]; + tensor input_5_cast_fp16 = reshape(shape = var_355, x = var_351_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor attn0_out_proj_weight_to_fp16 = const()[name = tensor("attn0_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6304384)))]; + tensor linear_1_bias_0_to_fp16 = const()[name = tensor("linear_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401600)))]; + tensor linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn0_out_proj_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("linear_1_cast_fp16")]; + tensor var_364_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_364_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_7_cast_fp16 = add(x = conditioning_to_fp16, y = linear_1_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor input_9_axes_0 = const()[name = tensor("input_9_axes_0"), val = tensor([-1])]; + tensor norm0_2_weight_to_fp16 = const()[name = tensor("norm0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8403712)))]; + tensor norm0_2_bias_to_fp16 = const()[name = tensor("norm0_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8405824)))]; + tensor var_368_to_fp16 = const()[name = tensor("op_368_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_9_cast_fp16 = layer_norm(axes = input_9_axes_0, beta = norm0_2_bias_to_fp16, epsilon = var_368_to_fp16, gamma = norm0_2_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor linear0_1_weight_to_fp16 = const()[name = tensor("linear0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8407936)))]; + tensor linear_2_bias_0_to_fp16 = const()[name = tensor("linear_2_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16796608)))]; + tensor linear_2_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear0_1_weight_to_fp16, x = input_9_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor input_11_mode_0 = const()[name = tensor("input_11_mode_0"), val = tensor("EXACT")]; + tensor input_11_cast_fp16 = gelu(mode = input_11_mode_0, x = linear_2_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor linear0_2_weight_to_fp16 = const()[name = tensor("linear0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16804864)))]; + tensor linear_3_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear0_2_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("linear_3_cast_fp16")]; + tensor input_13_cast_fp16 = add(x = input_7_cast_fp16, y = linear_3_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor input_15_axes_0 = const()[name = tensor("input_15_axes_0"), val = tensor([-1])]; + tensor norm1_1_weight_to_fp16 = const()[name = tensor("norm1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25193536)))]; + tensor norm1_1_bias_to_fp16 = const()[name = tensor("norm1_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25195648)))]; + tensor var_385_to_fp16 = const()[name = tensor("op_385_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_15_cast_fp16 = layer_norm(axes = input_15_axes_0, beta = norm1_1_bias_to_fp16, epsilon = var_385_to_fp16, gamma = norm1_1_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor attn1_in_proj_weight_to_fp16 = const()[name = tensor("attn1_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25197760)))]; + tensor linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn1_in_proj_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor var_399 = const()[name = tensor("op_399"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_3_cast_fp16 = reshape(shape = var_399, x = linear_4_cast_fp16)[name = tensor("qkv_3_cast_fp16")]; + tensor x_5_begin_0 = const()[name = tensor("x_5_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_5_end_0 = const()[name = tensor("x_5_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_5_end_mask_0 = const()[name = tensor("x_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_5_squeeze_mask_0 = const()[name = tensor("x_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_5_cast_fp16 = slice_by_index(begin = x_5_begin_0, end = x_5_end_0, end_mask = x_5_end_mask_0, squeeze_mask = x_5_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("x_5_cast_fp16")]; + tensor x_7_begin_0 = const()[name = tensor("x_7_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_7_end_0 = const()[name = tensor("x_7_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_7_end_mask_0 = const()[name = tensor("x_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_7_squeeze_mask_0 = const()[name = tensor("x_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_7_cast_fp16 = slice_by_index(begin = x_7_begin_0, end = x_7_end_0, end_mask = x_7_end_mask_0, squeeze_mask = x_7_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("x_7_cast_fp16")]; + tensor v_3_begin_0 = const()[name = tensor("v_3_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_3_end_0 = const()[name = tensor("v_3_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_3_end_mask_0 = const()[name = tensor("v_3_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_3_squeeze_mask_0 = const()[name = tensor("v_3_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_3_cast_fp16 = slice_by_index(begin = v_3_begin_0, end = v_3_end_0, end_mask = v_3_end_mask_0, squeeze_mask = v_3_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("v_3_cast_fp16")]; + tensor var_444 = const()[name = tensor("op_444"), val = tensor([256, 16, 32, 2])]; + tensor pairs_5_cast_fp16 = reshape(shape = var_444, x = x_5_cast_fp16)[name = tensor("pairs_5_cast_fp16")]; + tensor xr_5_begin_0 = const()[name = tensor("xr_5_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_5_end_0 = const()[name = tensor("xr_5_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_5_end_mask_0 = const()[name = tensor("xr_5_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_5_squeeze_mask_0 = const()[name = tensor("xr_5_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_5_cast_fp16 = slice_by_index(begin = xr_5_begin_0, end = xr_5_end_0, end_mask = xr_5_end_mask_0, squeeze_mask = xr_5_squeeze_mask_0, x = pairs_5_cast_fp16)[name = tensor("xr_5_cast_fp16")]; + tensor xi_5_begin_0 = const()[name = tensor("xi_5_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_5_end_0 = const()[name = tensor("xi_5_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_5_end_mask_0 = const()[name = tensor("xi_5_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_5_squeeze_mask_0 = const()[name = tensor("xi_5_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_5_cast_fp16 = slice_by_index(begin = xi_5_begin_0, end = xi_5_end_0, end_mask = xi_5_end_mask_0, squeeze_mask = xi_5_squeeze_mask_0, x = pairs_5_cast_fp16)[name = tensor("xi_5_cast_fp16")]; + tensor var_452_cast_fp16 = mul(x = xr_5_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_452_cast_fp16")]; + tensor var_453_cast_fp16 = mul(x = xi_5_cast_fp16, y = roti_cast_fp16)[name = tensor("op_453_cast_fp16")]; + tensor xor__5_cast_fp16 = sub(x = var_452_cast_fp16, y = var_453_cast_fp16)[name = tensor("xor__5_cast_fp16")]; + tensor var_456_cast_fp16 = mul(x = xr_5_cast_fp16, y = roti_cast_fp16)[name = tensor("op_456_cast_fp16")]; + tensor var_457_cast_fp16 = mul(x = xi_5_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_457_cast_fp16")]; + tensor xoi_5_cast_fp16 = add(x = var_456_cast_fp16, y = var_457_cast_fp16)[name = tensor("xoi_5_cast_fp16")]; + tensor out_5_axis_0 = const()[name = tensor("out_5_axis_0"), val = tensor(-1)]; + tensor out_5_cast_fp16 = stack(axis = out_5_axis_0, values = (xor__5_cast_fp16, xoi_5_cast_fp16))[name = tensor("out_5_cast_fp16")]; + tensor var_467 = const()[name = tensor("op_467"), val = tensor([1, 256, 16, 64])]; + tensor q_3_cast_fp16 = reshape(shape = var_467, x = out_5_cast_fp16)[name = tensor("q_3_cast_fp16")]; + tensor var_473 = const()[name = tensor("op_473"), val = tensor([256, 16, 32, 2])]; + tensor pairs_7_cast_fp16 = reshape(shape = var_473, x = x_7_cast_fp16)[name = tensor("pairs_7_cast_fp16")]; + tensor xr_7_begin_0 = const()[name = tensor("xr_7_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_7_end_0 = const()[name = tensor("xr_7_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_7_end_mask_0 = const()[name = tensor("xr_7_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_7_squeeze_mask_0 = const()[name = tensor("xr_7_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_7_cast_fp16 = slice_by_index(begin = xr_7_begin_0, end = xr_7_end_0, end_mask = xr_7_end_mask_0, squeeze_mask = xr_7_squeeze_mask_0, x = pairs_7_cast_fp16)[name = tensor("xr_7_cast_fp16")]; + tensor xi_7_begin_0 = const()[name = tensor("xi_7_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_7_end_0 = const()[name = tensor("xi_7_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_7_end_mask_0 = const()[name = tensor("xi_7_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_7_squeeze_mask_0 = const()[name = tensor("xi_7_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_7_cast_fp16 = slice_by_index(begin = xi_7_begin_0, end = xi_7_end_0, end_mask = xi_7_end_mask_0, squeeze_mask = xi_7_squeeze_mask_0, x = pairs_7_cast_fp16)[name = tensor("xi_7_cast_fp16")]; + tensor var_481_cast_fp16 = mul(x = xr_7_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_481_cast_fp16")]; + tensor var_482_cast_fp16 = mul(x = xi_7_cast_fp16, y = roti_cast_fp16)[name = tensor("op_482_cast_fp16")]; + tensor xor__7_cast_fp16 = sub(x = var_481_cast_fp16, y = var_482_cast_fp16)[name = tensor("xor__7_cast_fp16")]; + tensor var_485_cast_fp16 = mul(x = xr_7_cast_fp16, y = roti_cast_fp16)[name = tensor("op_485_cast_fp16")]; + tensor var_486_cast_fp16 = mul(x = xi_7_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_486_cast_fp16")]; + tensor xoi_7_cast_fp16 = add(x = var_485_cast_fp16, y = var_486_cast_fp16)[name = tensor("xoi_7_cast_fp16")]; + tensor out_7_axis_0 = const()[name = tensor("out_7_axis_0"), val = tensor(-1)]; + tensor out_7_cast_fp16 = stack(axis = out_7_axis_0, values = (xor__7_cast_fp16, xoi_7_cast_fp16))[name = tensor("out_7_cast_fp16")]; + tensor var_503 = const()[name = tensor("op_503"), val = tensor([256, 1024])]; + tensor var_504_cast_fp16 = reshape(shape = var_503, x = out_7_cast_fp16)[name = tensor("op_504_cast_fp16")]; + tensor written_k_3_transpose_x_1 = const()[name = tensor("written_k_3_transpose_x_1"), val = tensor(true)]; + tensor written_k_3_transpose_y_1 = const()[name = tensor("written_k_3_transpose_y_1"), val = tensor(false)]; + tensor written_k_3_cast_fp16 = matmul(transpose_x = written_k_3_transpose_x_1, transpose_y = written_k_3_transpose_y_1, x = var_106_cast_fp16_to_fp16, y = var_504_cast_fp16)[name = tensor("written_k_3_cast_fp16")]; + tensor var_511 = const()[name = tensor("op_511"), val = tensor([256, 1024])]; + tensor var_512_cast_fp16 = reshape(shape = var_511, x = v_3_cast_fp16)[name = tensor("op_512_cast_fp16")]; + tensor written_v_3_transpose_x_1 = const()[name = tensor("written_v_3_transpose_x_1"), val = tensor(true)]; + tensor written_v_3_transpose_y_1 = const()[name = tensor("written_v_3_transpose_y_1"), val = tensor(false)]; + tensor written_v_3_cast_fp16 = matmul(transpose_x = written_v_3_transpose_x_1, transpose_y = written_v_3_transpose_y_1, x = var_106_cast_fp16_to_fp16, y = var_512_cast_fp16)[name = tensor("written_v_3_cast_fp16")]; + tensor k_cache1_to_fp16_dtype_0 = const()[name = tensor("k_cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache1_to_fp16 = cast(dtype = k_cache1_to_fp16_dtype_0, x = k_cache1)[name = tensor("cast_117")]; + tensor var_523_cast_fp16 = mul(x = k_cache1_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_523_cast_fp16")]; + tensor var_528 = const()[name = tensor("op_528"), val = tensor([1, 512, 16, 64])]; + tensor var_529_cast_fp16 = reshape(shape = var_528, x = written_k_3_cast_fp16)[name = tensor("op_529_cast_fp16")]; + tensor new_k_cache_3_cast_fp16 = add(x = var_523_cast_fp16, y = var_529_cast_fp16)[name = tensor("new_k_cache_3_cast_fp16")]; + tensor new_k_cache_3_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_3_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache1_to_fp16_dtype_0 = const()[name = tensor("v_cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache1_to_fp16 = cast(dtype = v_cache1_to_fp16_dtype_0, x = v_cache1)[name = tensor("cast_115")]; + tensor var_532_cast_fp16 = mul(x = v_cache1_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_532_cast_fp16")]; + tensor var_537 = const()[name = tensor("op_537"), val = tensor([1, 512, 16, 64])]; + tensor var_538_cast_fp16 = reshape(shape = var_537, x = written_v_3_cast_fp16)[name = tensor("op_538_cast_fp16")]; + tensor new_v_cache_3_cast_fp16 = add(x = var_532_cast_fp16, y = var_538_cast_fp16)[name = tensor("new_v_cache_3_cast_fp16")]; + tensor new_v_cache_3_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_3_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_551 = const()[name = tensor("op_551"), val = tensor([0, 2, 1, 3])]; + tensor var_562_transpose_x_0 = const()[name = tensor("op_562_transpose_x_0"), val = tensor(false)]; + tensor var_562_transpose_y_0 = const()[name = tensor("op_562_transpose_y_0"), val = tensor(false)]; + tensor transpose_17_perm_0 = const()[name = tensor("transpose_17_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_18_perm_0 = const()[name = tensor("transpose_18_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_18 = transpose(perm = transpose_18_perm_0, x = new_k_cache_3_cast_fp16)[name = tensor("transpose_38")]; + tensor transpose_17 = transpose(perm = transpose_17_perm_0, x = q_3_cast_fp16)[name = tensor("transpose_39")]; + tensor var_562_cast_fp16 = matmul(transpose_x = var_562_transpose_x_0, transpose_y = var_562_transpose_y_0, x = transpose_17, y = transpose_18)[name = tensor("op_562_cast_fp16")]; + tensor var_563_to_fp16 = const()[name = tensor("op_563_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_5_cast_fp16 = mul(x = var_562_cast_fp16, y = var_563_to_fp16)[name = tensor("scores_5_cast_fp16")]; + tensor scores_7_cast_fp16 = add(x = scores_5_cast_fp16, y = var_339_cast_fp16)[name = tensor("scores_7_cast_fp16")]; + tensor var_573 = const()[name = tensor("op_573"), val = tensor(-1)]; + tensor weights_3_cast_fp16 = softmax(axis = var_573, x = scores_7_cast_fp16)[name = tensor("weights_3_cast_fp16")]; + tensor attn_3_transpose_x_0 = const()[name = tensor("attn_3_transpose_x_0"), val = tensor(false)]; + tensor attn_3_transpose_y_0 = const()[name = tensor("attn_3_transpose_y_0"), val = tensor(false)]; + tensor values_3_cast_fp16 = transpose(perm = var_551, x = new_v_cache_3_cast_fp16)[name = tensor("transpose_40")]; + tensor attn_3_cast_fp16 = matmul(transpose_x = attn_3_transpose_x_0, transpose_y = attn_3_transpose_y_0, x = weights_3_cast_fp16, y = values_3_cast_fp16)[name = tensor("attn_3_cast_fp16")]; + tensor var_581 = const()[name = tensor("op_581"), val = tensor([0, 2, 1, 3])]; + tensor var_586 = const()[name = tensor("op_586"), val = tensor([1, 256, 1024])]; + tensor var_582_cast_fp16 = transpose(perm = var_581, x = attn_3_cast_fp16)[name = tensor("transpose_37")]; + tensor input_17_cast_fp16 = reshape(shape = var_586, x = var_582_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor attn1_out_proj_weight_to_fp16 = const()[name = tensor("attn1_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31489280)))]; + tensor linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn1_out_proj_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor position1_to_fp16_dtype_0 = const()[name = tensor("position1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position1_to_fp16 = cast(dtype = position1_to_fp16_dtype_0, x = position1)[name = tensor("cast_113")]; + tensor var_595_cast_fp16 = add(x = position1_to_fp16, y = valid_len_to_fp16)[name = tensor("op_595_cast_fp16")]; + tensor var_595_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_595_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_19_cast_fp16 = add(x = input_13_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor input_21_axes_0 = const()[name = tensor("input_21_axes_0"), val = tensor([-1])]; + tensor norm1_2_weight_to_fp16 = const()[name = tensor("norm1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33586496)))]; + tensor norm1_2_bias_to_fp16 = const()[name = tensor("norm1_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33588608)))]; + tensor var_599_to_fp16 = const()[name = tensor("op_599_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_21_cast_fp16 = layer_norm(axes = input_21_axes_0, beta = norm1_2_bias_to_fp16, epsilon = var_599_to_fp16, gamma = norm1_2_weight_to_fp16, x = input_19_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor linear1_1_weight_to_fp16 = const()[name = tensor("linear1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33590720)))]; + tensor linear_6_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear1_1_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor input_23_mode_0 = const()[name = tensor("input_23_mode_0"), val = tensor("EXACT")]; + tensor input_23_cast_fp16 = gelu(mode = input_23_mode_0, x = linear_6_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor linear1_2_weight_to_fp16 = const()[name = tensor("linear1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41979392)))]; + tensor linear_7_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear1_2_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_25_cast_fp16 = add(x = input_19_cast_fp16, y = linear_7_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor input_27_axes_0 = const()[name = tensor("input_27_axes_0"), val = tensor([-1])]; + tensor norm2_1_weight_to_fp16 = const()[name = tensor("norm2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50368064)))]; + tensor norm2_1_bias_to_fp16 = const()[name = tensor("norm2_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50370176)))]; + tensor var_616_to_fp16 = const()[name = tensor("op_616_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = norm2_1_bias_to_fp16, epsilon = var_616_to_fp16, gamma = norm2_1_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor attn2_in_proj_weight_to_fp16 = const()[name = tensor("attn2_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50372288)))]; + tensor linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn2_in_proj_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor var_630 = const()[name = tensor("op_630"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_5_cast_fp16 = reshape(shape = var_630, x = linear_8_cast_fp16)[name = tensor("qkv_5_cast_fp16")]; + tensor x_9_begin_0 = const()[name = tensor("x_9_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_9_end_0 = const()[name = tensor("x_9_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_9_end_mask_0 = const()[name = tensor("x_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_9_squeeze_mask_0 = const()[name = tensor("x_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_9_cast_fp16 = slice_by_index(begin = x_9_begin_0, end = x_9_end_0, end_mask = x_9_end_mask_0, squeeze_mask = x_9_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("x_9_cast_fp16")]; + tensor x_11_begin_0 = const()[name = tensor("x_11_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_11_end_0 = const()[name = tensor("x_11_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_11_end_mask_0 = const()[name = tensor("x_11_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_11_squeeze_mask_0 = const()[name = tensor("x_11_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_11_cast_fp16 = slice_by_index(begin = x_11_begin_0, end = x_11_end_0, end_mask = x_11_end_mask_0, squeeze_mask = x_11_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("x_11_cast_fp16")]; + tensor v_5_begin_0 = const()[name = tensor("v_5_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_5_end_0 = const()[name = tensor("v_5_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_5_end_mask_0 = const()[name = tensor("v_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_5_squeeze_mask_0 = const()[name = tensor("v_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_5_cast_fp16 = slice_by_index(begin = v_5_begin_0, end = v_5_end_0, end_mask = v_5_end_mask_0, squeeze_mask = v_5_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("v_5_cast_fp16")]; + tensor var_675 = const()[name = tensor("op_675"), val = tensor([256, 16, 32, 2])]; + tensor pairs_9_cast_fp16 = reshape(shape = var_675, x = x_9_cast_fp16)[name = tensor("pairs_9_cast_fp16")]; + tensor xr_9_begin_0 = const()[name = tensor("xr_9_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_9_end_0 = const()[name = tensor("xr_9_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_9_end_mask_0 = const()[name = tensor("xr_9_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_9_squeeze_mask_0 = const()[name = tensor("xr_9_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_9_cast_fp16 = slice_by_index(begin = xr_9_begin_0, end = xr_9_end_0, end_mask = xr_9_end_mask_0, squeeze_mask = xr_9_squeeze_mask_0, x = pairs_9_cast_fp16)[name = tensor("xr_9_cast_fp16")]; + tensor xi_9_begin_0 = const()[name = tensor("xi_9_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_9_end_0 = const()[name = tensor("xi_9_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_9_end_mask_0 = const()[name = tensor("xi_9_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_9_squeeze_mask_0 = const()[name = tensor("xi_9_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_9_cast_fp16 = slice_by_index(begin = xi_9_begin_0, end = xi_9_end_0, end_mask = xi_9_end_mask_0, squeeze_mask = xi_9_squeeze_mask_0, x = pairs_9_cast_fp16)[name = tensor("xi_9_cast_fp16")]; + tensor var_683_cast_fp16 = mul(x = xr_9_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_683_cast_fp16")]; + tensor var_684_cast_fp16 = mul(x = xi_9_cast_fp16, y = roti_cast_fp16)[name = tensor("op_684_cast_fp16")]; + tensor xor__9_cast_fp16 = sub(x = var_683_cast_fp16, y = var_684_cast_fp16)[name = tensor("xor__9_cast_fp16")]; + tensor var_687_cast_fp16 = mul(x = xr_9_cast_fp16, y = roti_cast_fp16)[name = tensor("op_687_cast_fp16")]; + tensor var_688_cast_fp16 = mul(x = xi_9_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_688_cast_fp16")]; + tensor xoi_9_cast_fp16 = add(x = var_687_cast_fp16, y = var_688_cast_fp16)[name = tensor("xoi_9_cast_fp16")]; + tensor out_9_axis_0 = const()[name = tensor("out_9_axis_0"), val = tensor(-1)]; + tensor out_9_cast_fp16 = stack(axis = out_9_axis_0, values = (xor__9_cast_fp16, xoi_9_cast_fp16))[name = tensor("out_9_cast_fp16")]; + tensor var_698 = const()[name = tensor("op_698"), val = tensor([1, 256, 16, 64])]; + tensor q_5_cast_fp16 = reshape(shape = var_698, x = out_9_cast_fp16)[name = tensor("q_5_cast_fp16")]; + tensor var_704 = const()[name = tensor("op_704"), val = tensor([256, 16, 32, 2])]; + tensor pairs_11_cast_fp16 = reshape(shape = var_704, x = x_11_cast_fp16)[name = tensor("pairs_11_cast_fp16")]; + tensor xr_11_begin_0 = const()[name = tensor("xr_11_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_11_end_0 = const()[name = tensor("xr_11_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_11_end_mask_0 = const()[name = tensor("xr_11_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_11_squeeze_mask_0 = const()[name = tensor("xr_11_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_11_cast_fp16 = slice_by_index(begin = xr_11_begin_0, end = xr_11_end_0, end_mask = xr_11_end_mask_0, squeeze_mask = xr_11_squeeze_mask_0, x = pairs_11_cast_fp16)[name = tensor("xr_11_cast_fp16")]; + tensor xi_11_begin_0 = const()[name = tensor("xi_11_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_11_end_0 = const()[name = tensor("xi_11_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_11_end_mask_0 = const()[name = tensor("xi_11_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_11_squeeze_mask_0 = const()[name = tensor("xi_11_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_11_cast_fp16 = slice_by_index(begin = xi_11_begin_0, end = xi_11_end_0, end_mask = xi_11_end_mask_0, squeeze_mask = xi_11_squeeze_mask_0, x = pairs_11_cast_fp16)[name = tensor("xi_11_cast_fp16")]; + tensor var_712_cast_fp16 = mul(x = xr_11_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_712_cast_fp16")]; + tensor var_713_cast_fp16 = mul(x = xi_11_cast_fp16, y = roti_cast_fp16)[name = tensor("op_713_cast_fp16")]; + tensor xor__11_cast_fp16 = sub(x = var_712_cast_fp16, y = var_713_cast_fp16)[name = tensor("xor__11_cast_fp16")]; + tensor var_716_cast_fp16 = mul(x = xr_11_cast_fp16, y = roti_cast_fp16)[name = tensor("op_716_cast_fp16")]; + tensor var_717_cast_fp16 = mul(x = xi_11_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_717_cast_fp16")]; + tensor xoi_11_cast_fp16 = add(x = var_716_cast_fp16, y = var_717_cast_fp16)[name = tensor("xoi_11_cast_fp16")]; + tensor out_11_axis_0 = const()[name = tensor("out_11_axis_0"), val = tensor(-1)]; + tensor out_11_cast_fp16 = stack(axis = out_11_axis_0, values = (xor__11_cast_fp16, xoi_11_cast_fp16))[name = tensor("out_11_cast_fp16")]; + tensor var_734 = const()[name = tensor("op_734"), val = tensor([256, 1024])]; + tensor var_735_cast_fp16 = reshape(shape = var_734, x = out_11_cast_fp16)[name = tensor("op_735_cast_fp16")]; + tensor written_k_5_transpose_x_1 = const()[name = tensor("written_k_5_transpose_x_1"), val = tensor(true)]; + tensor written_k_5_transpose_y_1 = const()[name = tensor("written_k_5_transpose_y_1"), val = tensor(false)]; + tensor written_k_5_cast_fp16 = matmul(transpose_x = written_k_5_transpose_x_1, transpose_y = written_k_5_transpose_y_1, x = var_106_cast_fp16_to_fp16, y = var_735_cast_fp16)[name = tensor("written_k_5_cast_fp16")]; + tensor var_742 = const()[name = tensor("op_742"), val = tensor([256, 1024])]; + tensor var_743_cast_fp16 = reshape(shape = var_742, x = v_5_cast_fp16)[name = tensor("op_743_cast_fp16")]; + tensor written_v_5_transpose_x_1 = const()[name = tensor("written_v_5_transpose_x_1"), val = tensor(true)]; + tensor written_v_5_transpose_y_1 = const()[name = tensor("written_v_5_transpose_y_1"), val = tensor(false)]; + tensor written_v_5_cast_fp16 = matmul(transpose_x = written_v_5_transpose_x_1, transpose_y = written_v_5_transpose_y_1, x = var_106_cast_fp16_to_fp16, y = var_743_cast_fp16)[name = tensor("written_v_5_cast_fp16")]; + tensor k_cache2_to_fp16_dtype_0 = const()[name = tensor("k_cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache2_to_fp16 = cast(dtype = k_cache2_to_fp16_dtype_0, x = k_cache2)[name = tensor("cast_111")]; + tensor var_754_cast_fp16 = mul(x = k_cache2_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_754_cast_fp16")]; + tensor var_759 = const()[name = tensor("op_759"), val = tensor([1, 512, 16, 64])]; + tensor var_760_cast_fp16 = reshape(shape = var_759, x = written_k_5_cast_fp16)[name = tensor("op_760_cast_fp16")]; + tensor new_k_cache_5_cast_fp16 = add(x = var_754_cast_fp16, y = var_760_cast_fp16)[name = tensor("new_k_cache_5_cast_fp16")]; + tensor new_k_cache_5_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_5_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache2_to_fp16_dtype_0 = const()[name = tensor("v_cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache2_to_fp16 = cast(dtype = v_cache2_to_fp16_dtype_0, x = v_cache2)[name = tensor("cast_109")]; + tensor var_763_cast_fp16 = mul(x = v_cache2_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_763_cast_fp16")]; + tensor var_768 = const()[name = tensor("op_768"), val = tensor([1, 512, 16, 64])]; + tensor var_769_cast_fp16 = reshape(shape = var_768, x = written_v_5_cast_fp16)[name = tensor("op_769_cast_fp16")]; + tensor new_v_cache_5_cast_fp16 = add(x = var_763_cast_fp16, y = var_769_cast_fp16)[name = tensor("new_v_cache_5_cast_fp16")]; + tensor new_v_cache_5_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_5_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_782 = const()[name = tensor("op_782"), val = tensor([0, 2, 1, 3])]; + tensor var_793_transpose_x_0 = const()[name = tensor("op_793_transpose_x_0"), val = tensor(false)]; + tensor var_793_transpose_y_0 = const()[name = tensor("op_793_transpose_y_0"), val = tensor(false)]; + tensor transpose_19_perm_0 = const()[name = tensor("transpose_19_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_20_perm_0 = const()[name = tensor("transpose_20_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_20 = transpose(perm = transpose_20_perm_0, x = new_k_cache_5_cast_fp16)[name = tensor("transpose_34")]; + tensor transpose_19 = transpose(perm = transpose_19_perm_0, x = q_5_cast_fp16)[name = tensor("transpose_35")]; + tensor var_793_cast_fp16 = matmul(transpose_x = var_793_transpose_x_0, transpose_y = var_793_transpose_y_0, x = transpose_19, y = transpose_20)[name = tensor("op_793_cast_fp16")]; + tensor var_794_to_fp16 = const()[name = tensor("op_794_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_9_cast_fp16 = mul(x = var_793_cast_fp16, y = var_794_to_fp16)[name = tensor("scores_9_cast_fp16")]; + tensor scores_11_cast_fp16 = add(x = scores_9_cast_fp16, y = var_339_cast_fp16)[name = tensor("scores_11_cast_fp16")]; + tensor var_804 = const()[name = tensor("op_804"), val = tensor(-1)]; + tensor weights_5_cast_fp16 = softmax(axis = var_804, x = scores_11_cast_fp16)[name = tensor("weights_5_cast_fp16")]; + tensor attn_5_transpose_x_0 = const()[name = tensor("attn_5_transpose_x_0"), val = tensor(false)]; + tensor attn_5_transpose_y_0 = const()[name = tensor("attn_5_transpose_y_0"), val = tensor(false)]; + tensor values_5_cast_fp16 = transpose(perm = var_782, x = new_v_cache_5_cast_fp16)[name = tensor("transpose_36")]; + tensor attn_5_cast_fp16 = matmul(transpose_x = attn_5_transpose_x_0, transpose_y = attn_5_transpose_y_0, x = weights_5_cast_fp16, y = values_5_cast_fp16)[name = tensor("attn_5_cast_fp16")]; + tensor var_812 = const()[name = tensor("op_812"), val = tensor([0, 2, 1, 3])]; + tensor var_817 = const()[name = tensor("op_817"), val = tensor([1, 256, 1024])]; + tensor var_813_cast_fp16 = transpose(perm = var_812, x = attn_5_cast_fp16)[name = tensor("transpose_33")]; + tensor input_29_cast_fp16 = reshape(shape = var_817, x = var_813_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor attn2_out_proj_weight_to_fp16 = const()[name = tensor("attn2_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(56663808)))]; + tensor linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn2_out_proj_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor position2_to_fp16_dtype_0 = const()[name = tensor("position2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position2_to_fp16 = cast(dtype = position2_to_fp16_dtype_0, x = position2)[name = tensor("cast_107")]; + tensor var_826_cast_fp16 = add(x = position2_to_fp16, y = valid_len_to_fp16)[name = tensor("op_826_cast_fp16")]; + tensor var_826_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_826_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_31_cast_fp16 = add(x = input_25_cast_fp16, y = linear_9_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; + tensor norm2_2_weight_to_fp16 = const()[name = tensor("norm2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58761024)))]; + tensor norm2_2_bias_to_fp16 = const()[name = tensor("norm2_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58763136)))]; + tensor var_830_to_fp16 = const()[name = tensor("op_830_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_33_cast_fp16 = layer_norm(axes = input_33_axes_0, beta = norm2_2_bias_to_fp16, epsilon = var_830_to_fp16, gamma = norm2_2_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor linear2_1_weight_to_fp16 = const()[name = tensor("linear2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58765248)))]; + tensor linear_10_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear2_1_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor input_35_mode_0 = const()[name = tensor("input_35_mode_0"), val = tensor("EXACT")]; + tensor input_35_cast_fp16 = gelu(mode = input_35_mode_0, x = linear_10_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor linear2_2_weight_to_fp16 = const()[name = tensor("linear2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67153920)))]; + tensor linear_11_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear2_2_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor input_37_cast_fp16 = add(x = input_31_cast_fp16, y = linear_11_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor input_39_axes_0 = const()[name = tensor("input_39_axes_0"), val = tensor([-1])]; + tensor norm3_1_weight_to_fp16 = const()[name = tensor("norm3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75542592)))]; + tensor norm3_1_bias_to_fp16 = const()[name = tensor("norm3_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75544704)))]; + tensor var_847_to_fp16 = const()[name = tensor("op_847_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_39_cast_fp16 = layer_norm(axes = input_39_axes_0, beta = norm3_1_bias_to_fp16, epsilon = var_847_to_fp16, gamma = norm3_1_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor attn3_in_proj_weight_to_fp16 = const()[name = tensor("attn3_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75546816)))]; + tensor linear_12_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn3_in_proj_weight_to_fp16, x = input_39_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor var_861 = const()[name = tensor("op_861"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_7_cast_fp16 = reshape(shape = var_861, x = linear_12_cast_fp16)[name = tensor("qkv_7_cast_fp16")]; + tensor x_13_begin_0 = const()[name = tensor("x_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_13_end_0 = const()[name = tensor("x_13_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_13_end_mask_0 = const()[name = tensor("x_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_13_squeeze_mask_0 = const()[name = tensor("x_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_13_cast_fp16 = slice_by_index(begin = x_13_begin_0, end = x_13_end_0, end_mask = x_13_end_mask_0, squeeze_mask = x_13_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("x_13_cast_fp16")]; + tensor x_15_begin_0 = const()[name = tensor("x_15_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_15_end_0 = const()[name = tensor("x_15_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_15_end_mask_0 = const()[name = tensor("x_15_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_15_squeeze_mask_0 = const()[name = tensor("x_15_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_15_cast_fp16 = slice_by_index(begin = x_15_begin_0, end = x_15_end_0, end_mask = x_15_end_mask_0, squeeze_mask = x_15_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("x_15_cast_fp16")]; + tensor v_7_begin_0 = const()[name = tensor("v_7_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_7_end_0 = const()[name = tensor("v_7_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_7_end_mask_0 = const()[name = tensor("v_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_7_squeeze_mask_0 = const()[name = tensor("v_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_7_cast_fp16 = slice_by_index(begin = v_7_begin_0, end = v_7_end_0, end_mask = v_7_end_mask_0, squeeze_mask = v_7_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("v_7_cast_fp16")]; + tensor var_906 = const()[name = tensor("op_906"), val = tensor([256, 16, 32, 2])]; + tensor pairs_13_cast_fp16 = reshape(shape = var_906, x = x_13_cast_fp16)[name = tensor("pairs_13_cast_fp16")]; + tensor xr_13_begin_0 = const()[name = tensor("xr_13_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_13_end_0 = const()[name = tensor("xr_13_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_13_end_mask_0 = const()[name = tensor("xr_13_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_13_squeeze_mask_0 = const()[name = tensor("xr_13_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_13_cast_fp16 = slice_by_index(begin = xr_13_begin_0, end = xr_13_end_0, end_mask = xr_13_end_mask_0, squeeze_mask = xr_13_squeeze_mask_0, x = pairs_13_cast_fp16)[name = tensor("xr_13_cast_fp16")]; + tensor xi_13_begin_0 = const()[name = tensor("xi_13_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_13_end_0 = const()[name = tensor("xi_13_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_13_end_mask_0 = const()[name = tensor("xi_13_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_13_squeeze_mask_0 = const()[name = tensor("xi_13_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_13_cast_fp16 = slice_by_index(begin = xi_13_begin_0, end = xi_13_end_0, end_mask = xi_13_end_mask_0, squeeze_mask = xi_13_squeeze_mask_0, x = pairs_13_cast_fp16)[name = tensor("xi_13_cast_fp16")]; + tensor var_914_cast_fp16 = mul(x = xr_13_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_914_cast_fp16")]; + tensor var_915_cast_fp16 = mul(x = xi_13_cast_fp16, y = roti_cast_fp16)[name = tensor("op_915_cast_fp16")]; + tensor xor__13_cast_fp16 = sub(x = var_914_cast_fp16, y = var_915_cast_fp16)[name = tensor("xor__13_cast_fp16")]; + tensor var_918_cast_fp16 = mul(x = xr_13_cast_fp16, y = roti_cast_fp16)[name = tensor("op_918_cast_fp16")]; + tensor var_919_cast_fp16 = mul(x = xi_13_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_919_cast_fp16")]; + tensor xoi_13_cast_fp16 = add(x = var_918_cast_fp16, y = var_919_cast_fp16)[name = tensor("xoi_13_cast_fp16")]; + tensor out_13_axis_0 = const()[name = tensor("out_13_axis_0"), val = tensor(-1)]; + tensor out_13_cast_fp16 = stack(axis = out_13_axis_0, values = (xor__13_cast_fp16, xoi_13_cast_fp16))[name = tensor("out_13_cast_fp16")]; + tensor var_929 = const()[name = tensor("op_929"), val = tensor([1, 256, 16, 64])]; + tensor q_7_cast_fp16 = reshape(shape = var_929, x = out_13_cast_fp16)[name = tensor("q_7_cast_fp16")]; + tensor var_935 = const()[name = tensor("op_935"), val = tensor([256, 16, 32, 2])]; + tensor pairs_15_cast_fp16 = reshape(shape = var_935, x = x_15_cast_fp16)[name = tensor("pairs_15_cast_fp16")]; + tensor xr_15_begin_0 = const()[name = tensor("xr_15_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_15_end_0 = const()[name = tensor("xr_15_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_15_end_mask_0 = const()[name = tensor("xr_15_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_15_squeeze_mask_0 = const()[name = tensor("xr_15_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_15_cast_fp16 = slice_by_index(begin = xr_15_begin_0, end = xr_15_end_0, end_mask = xr_15_end_mask_0, squeeze_mask = xr_15_squeeze_mask_0, x = pairs_15_cast_fp16)[name = tensor("xr_15_cast_fp16")]; + tensor xi_15_begin_0 = const()[name = tensor("xi_15_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_15_end_0 = const()[name = tensor("xi_15_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_15_end_mask_0 = const()[name = tensor("xi_15_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_15_squeeze_mask_0 = const()[name = tensor("xi_15_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_15_cast_fp16 = slice_by_index(begin = xi_15_begin_0, end = xi_15_end_0, end_mask = xi_15_end_mask_0, squeeze_mask = xi_15_squeeze_mask_0, x = pairs_15_cast_fp16)[name = tensor("xi_15_cast_fp16")]; + tensor var_943_cast_fp16 = mul(x = xr_15_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_943_cast_fp16")]; + tensor var_944_cast_fp16 = mul(x = xi_15_cast_fp16, y = roti_cast_fp16)[name = tensor("op_944_cast_fp16")]; + tensor xor__15_cast_fp16 = sub(x = var_943_cast_fp16, y = var_944_cast_fp16)[name = tensor("xor__15_cast_fp16")]; + tensor var_947_cast_fp16 = mul(x = xr_15_cast_fp16, y = roti_cast_fp16)[name = tensor("op_947_cast_fp16")]; + tensor var_948_cast_fp16 = mul(x = xi_15_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_948_cast_fp16")]; + tensor xoi_15_cast_fp16 = add(x = var_947_cast_fp16, y = var_948_cast_fp16)[name = tensor("xoi_15_cast_fp16")]; + tensor out_15_axis_0 = const()[name = tensor("out_15_axis_0"), val = tensor(-1)]; + tensor out_15_cast_fp16 = stack(axis = out_15_axis_0, values = (xor__15_cast_fp16, xoi_15_cast_fp16))[name = tensor("out_15_cast_fp16")]; + tensor var_965 = const()[name = tensor("op_965"), val = tensor([256, 1024])]; + tensor var_966_cast_fp16 = reshape(shape = var_965, x = out_15_cast_fp16)[name = tensor("op_966_cast_fp16")]; + tensor written_k_7_transpose_x_1 = const()[name = tensor("written_k_7_transpose_x_1"), val = tensor(true)]; + tensor written_k_7_transpose_y_1 = const()[name = tensor("written_k_7_transpose_y_1"), val = tensor(false)]; + tensor written_k_7_cast_fp16 = matmul(transpose_x = written_k_7_transpose_x_1, transpose_y = written_k_7_transpose_y_1, x = var_106_cast_fp16_to_fp16, y = var_966_cast_fp16)[name = tensor("written_k_7_cast_fp16")]; + tensor var_973 = const()[name = tensor("op_973"), val = tensor([256, 1024])]; + tensor var_974_cast_fp16 = reshape(shape = var_973, x = v_7_cast_fp16)[name = tensor("op_974_cast_fp16")]; + tensor written_v_7_transpose_x_1 = const()[name = tensor("written_v_7_transpose_x_1"), val = tensor(true)]; + tensor written_v_7_transpose_y_1 = const()[name = tensor("written_v_7_transpose_y_1"), val = tensor(false)]; + tensor written_v_7_cast_fp16 = matmul(transpose_x = written_v_7_transpose_x_1, transpose_y = written_v_7_transpose_y_1, x = var_106_cast_fp16_to_fp16, y = var_974_cast_fp16)[name = tensor("written_v_7_cast_fp16")]; + tensor k_cache3_to_fp16_dtype_0 = const()[name = tensor("k_cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache3_to_fp16 = cast(dtype = k_cache3_to_fp16_dtype_0, x = k_cache3)[name = tensor("cast_105")]; + tensor var_985_cast_fp16 = mul(x = k_cache3_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_985_cast_fp16")]; + tensor var_990 = const()[name = tensor("op_990"), val = tensor([1, 512, 16, 64])]; + tensor var_991_cast_fp16 = reshape(shape = var_990, x = written_k_7_cast_fp16)[name = tensor("op_991_cast_fp16")]; + tensor new_k_cache_7_cast_fp16 = add(x = var_985_cast_fp16, y = var_991_cast_fp16)[name = tensor("new_k_cache_7_cast_fp16")]; + tensor new_k_cache_7_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_7_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache3_to_fp16_dtype_0 = const()[name = tensor("v_cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache3_to_fp16 = cast(dtype = v_cache3_to_fp16_dtype_0, x = v_cache3)[name = tensor("cast_103")]; + tensor var_994_cast_fp16 = mul(x = v_cache3_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_994_cast_fp16")]; + tensor var_999 = const()[name = tensor("op_999"), val = tensor([1, 512, 16, 64])]; + tensor var_1000_cast_fp16 = reshape(shape = var_999, x = written_v_7_cast_fp16)[name = tensor("op_1000_cast_fp16")]; + tensor new_v_cache_7_cast_fp16 = add(x = var_994_cast_fp16, y = var_1000_cast_fp16)[name = tensor("new_v_cache_7_cast_fp16")]; + tensor new_v_cache_7_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_7_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_1013 = const()[name = tensor("op_1013"), val = tensor([0, 2, 1, 3])]; + tensor var_1024_transpose_x_0 = const()[name = tensor("op_1024_transpose_x_0"), val = tensor(false)]; + tensor var_1024_transpose_y_0 = const()[name = tensor("op_1024_transpose_y_0"), val = tensor(false)]; + tensor transpose_21_perm_0 = const()[name = tensor("transpose_21_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_22_perm_0 = const()[name = tensor("transpose_22_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_22 = transpose(perm = transpose_22_perm_0, x = new_k_cache_7_cast_fp16)[name = tensor("transpose_30")]; + tensor transpose_21 = transpose(perm = transpose_21_perm_0, x = q_7_cast_fp16)[name = tensor("transpose_31")]; + tensor var_1024_cast_fp16 = matmul(transpose_x = var_1024_transpose_x_0, transpose_y = var_1024_transpose_y_0, x = transpose_21, y = transpose_22)[name = tensor("op_1024_cast_fp16")]; + tensor var_1025_to_fp16 = const()[name = tensor("op_1025_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_13_cast_fp16 = mul(x = var_1024_cast_fp16, y = var_1025_to_fp16)[name = tensor("scores_13_cast_fp16")]; + tensor scores_15_cast_fp16 = add(x = scores_13_cast_fp16, y = var_339_cast_fp16)[name = tensor("scores_15_cast_fp16")]; + tensor var_1035 = const()[name = tensor("op_1035"), val = tensor(-1)]; + tensor weights_7_cast_fp16 = softmax(axis = var_1035, x = scores_15_cast_fp16)[name = tensor("weights_7_cast_fp16")]; + tensor attn_7_transpose_x_0 = const()[name = tensor("attn_7_transpose_x_0"), val = tensor(false)]; + tensor attn_7_transpose_y_0 = const()[name = tensor("attn_7_transpose_y_0"), val = tensor(false)]; + tensor values_7_cast_fp16 = transpose(perm = var_1013, x = new_v_cache_7_cast_fp16)[name = tensor("transpose_32")]; + tensor attn_7_cast_fp16 = matmul(transpose_x = attn_7_transpose_x_0, transpose_y = attn_7_transpose_y_0, x = weights_7_cast_fp16, y = values_7_cast_fp16)[name = tensor("attn_7_cast_fp16")]; + tensor var_1043 = const()[name = tensor("op_1043"), val = tensor([0, 2, 1, 3])]; + tensor var_1048 = const()[name = tensor("op_1048"), val = tensor([1, 256, 1024])]; + tensor var_1044_cast_fp16 = transpose(perm = var_1043, x = attn_7_cast_fp16)[name = tensor("transpose_29")]; + tensor input_41_cast_fp16 = reshape(shape = var_1048, x = var_1044_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor attn3_out_proj_weight_to_fp16 = const()[name = tensor("attn3_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(81838336)))]; + tensor linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn3_out_proj_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor position3_to_fp16_dtype_0 = const()[name = tensor("position3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position3_to_fp16 = cast(dtype = position3_to_fp16_dtype_0, x = position3)[name = tensor("cast_101")]; + tensor var_1057_cast_fp16 = add(x = position3_to_fp16, y = valid_len_to_fp16)[name = tensor("op_1057_cast_fp16")]; + tensor var_1057_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1057_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_43_cast_fp16 = add(x = input_37_cast_fp16, y = linear_13_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; + tensor norm3_2_weight_to_fp16 = const()[name = tensor("norm3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(83935552)))]; + tensor norm3_2_bias_to_fp16 = const()[name = tensor("norm3_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(83937664)))]; + tensor var_1061_to_fp16 = const()[name = tensor("op_1061_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_45_cast_fp16 = layer_norm(axes = input_45_axes_0, beta = norm3_2_bias_to_fp16, epsilon = var_1061_to_fp16, gamma = norm3_2_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor linear3_1_weight_to_fp16 = const()[name = tensor("linear3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(83939776)))]; + tensor linear_14_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear3_1_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor input_47_mode_0 = const()[name = tensor("input_47_mode_0"), val = tensor("EXACT")]; + tensor input_47_cast_fp16 = gelu(mode = input_47_mode_0, x = linear_14_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor linear3_2_weight_to_fp16 = const()[name = tensor("linear3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(92328448)))]; + tensor linear_15_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear3_2_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor input_49_cast_fp16 = add(x = input_43_cast_fp16, y = linear_15_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor input_51_axes_0 = const()[name = tensor("input_51_axes_0"), val = tensor([-1])]; + tensor norm4_1_weight_to_fp16 = const()[name = tensor("norm4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100717120)))]; + tensor norm4_1_bias_to_fp16 = const()[name = tensor("norm4_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100719232)))]; + tensor var_1078_to_fp16 = const()[name = tensor("op_1078_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_51_cast_fp16 = layer_norm(axes = input_51_axes_0, beta = norm4_1_bias_to_fp16, epsilon = var_1078_to_fp16, gamma = norm4_1_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor attn4_in_proj_weight_to_fp16 = const()[name = tensor("attn4_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100721344)))]; + tensor linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn4_in_proj_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor var_1092 = const()[name = tensor("op_1092"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_9_cast_fp16 = reshape(shape = var_1092, x = linear_16_cast_fp16)[name = tensor("qkv_9_cast_fp16")]; + tensor x_17_begin_0 = const()[name = tensor("x_17_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_17_end_0 = const()[name = tensor("x_17_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_17_end_mask_0 = const()[name = tensor("x_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_17_squeeze_mask_0 = const()[name = tensor("x_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_17_cast_fp16 = slice_by_index(begin = x_17_begin_0, end = x_17_end_0, end_mask = x_17_end_mask_0, squeeze_mask = x_17_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("x_17_cast_fp16")]; + tensor x_19_begin_0 = const()[name = tensor("x_19_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_19_end_0 = const()[name = tensor("x_19_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_19_end_mask_0 = const()[name = tensor("x_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_19_squeeze_mask_0 = const()[name = tensor("x_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_19_cast_fp16 = slice_by_index(begin = x_19_begin_0, end = x_19_end_0, end_mask = x_19_end_mask_0, squeeze_mask = x_19_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("x_19_cast_fp16")]; + tensor v_9_begin_0 = const()[name = tensor("v_9_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_9_end_0 = const()[name = tensor("v_9_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_9_end_mask_0 = const()[name = tensor("v_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_9_squeeze_mask_0 = const()[name = tensor("v_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_9_cast_fp16 = slice_by_index(begin = v_9_begin_0, end = v_9_end_0, end_mask = v_9_end_mask_0, squeeze_mask = v_9_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("v_9_cast_fp16")]; + tensor var_1137 = const()[name = tensor("op_1137"), val = tensor([256, 16, 32, 2])]; + tensor pairs_17_cast_fp16 = reshape(shape = var_1137, x = x_17_cast_fp16)[name = tensor("pairs_17_cast_fp16")]; + tensor xr_17_begin_0 = const()[name = tensor("xr_17_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_17_end_0 = const()[name = tensor("xr_17_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_17_end_mask_0 = const()[name = tensor("xr_17_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_17_squeeze_mask_0 = const()[name = tensor("xr_17_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_17_cast_fp16 = slice_by_index(begin = xr_17_begin_0, end = xr_17_end_0, end_mask = xr_17_end_mask_0, squeeze_mask = xr_17_squeeze_mask_0, x = pairs_17_cast_fp16)[name = tensor("xr_17_cast_fp16")]; + tensor xi_17_begin_0 = const()[name = tensor("xi_17_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_17_end_0 = const()[name = tensor("xi_17_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_17_end_mask_0 = const()[name = tensor("xi_17_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_17_squeeze_mask_0 = const()[name = tensor("xi_17_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_17_cast_fp16 = slice_by_index(begin = xi_17_begin_0, end = xi_17_end_0, end_mask = xi_17_end_mask_0, squeeze_mask = xi_17_squeeze_mask_0, x = pairs_17_cast_fp16)[name = tensor("xi_17_cast_fp16")]; + tensor var_1145_cast_fp16 = mul(x = xr_17_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1145_cast_fp16")]; + tensor var_1146_cast_fp16 = mul(x = xi_17_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1146_cast_fp16")]; + tensor xor__17_cast_fp16 = sub(x = var_1145_cast_fp16, y = var_1146_cast_fp16)[name = tensor("xor__17_cast_fp16")]; + tensor var_1149_cast_fp16 = mul(x = xr_17_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1149_cast_fp16")]; + tensor var_1150_cast_fp16 = mul(x = xi_17_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1150_cast_fp16")]; + tensor xoi_17_cast_fp16 = add(x = var_1149_cast_fp16, y = var_1150_cast_fp16)[name = tensor("xoi_17_cast_fp16")]; + tensor out_17_axis_0 = const()[name = tensor("out_17_axis_0"), val = tensor(-1)]; + tensor out_17_cast_fp16 = stack(axis = out_17_axis_0, values = (xor__17_cast_fp16, xoi_17_cast_fp16))[name = tensor("out_17_cast_fp16")]; + tensor var_1160 = const()[name = tensor("op_1160"), val = tensor([1, 256, 16, 64])]; + tensor q_9_cast_fp16 = reshape(shape = var_1160, x = out_17_cast_fp16)[name = tensor("q_9_cast_fp16")]; + tensor var_1166 = const()[name = tensor("op_1166"), val = tensor([256, 16, 32, 2])]; + tensor pairs_19_cast_fp16 = reshape(shape = var_1166, x = x_19_cast_fp16)[name = tensor("pairs_19_cast_fp16")]; + tensor xr_19_begin_0 = const()[name = tensor("xr_19_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_19_end_0 = const()[name = tensor("xr_19_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_19_end_mask_0 = const()[name = tensor("xr_19_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_19_squeeze_mask_0 = const()[name = tensor("xr_19_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_19_cast_fp16 = slice_by_index(begin = xr_19_begin_0, end = xr_19_end_0, end_mask = xr_19_end_mask_0, squeeze_mask = xr_19_squeeze_mask_0, x = pairs_19_cast_fp16)[name = tensor("xr_19_cast_fp16")]; + tensor xi_19_begin_0 = const()[name = tensor("xi_19_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_19_end_0 = const()[name = tensor("xi_19_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_19_end_mask_0 = const()[name = tensor("xi_19_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_19_squeeze_mask_0 = const()[name = tensor("xi_19_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_19_cast_fp16 = slice_by_index(begin = xi_19_begin_0, end = xi_19_end_0, end_mask = xi_19_end_mask_0, squeeze_mask = xi_19_squeeze_mask_0, x = pairs_19_cast_fp16)[name = tensor("xi_19_cast_fp16")]; + tensor var_1174_cast_fp16 = mul(x = xr_19_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1174_cast_fp16")]; + tensor var_1175_cast_fp16 = mul(x = xi_19_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1175_cast_fp16")]; + tensor xor__19_cast_fp16 = sub(x = var_1174_cast_fp16, y = var_1175_cast_fp16)[name = tensor("xor__19_cast_fp16")]; + tensor var_1178_cast_fp16 = mul(x = xr_19_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1178_cast_fp16")]; + tensor var_1179_cast_fp16 = mul(x = xi_19_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1179_cast_fp16")]; + tensor xoi_19_cast_fp16 = add(x = var_1178_cast_fp16, y = var_1179_cast_fp16)[name = tensor("xoi_19_cast_fp16")]; + tensor out_19_axis_0 = const()[name = tensor("out_19_axis_0"), val = tensor(-1)]; + tensor out_19_cast_fp16 = stack(axis = out_19_axis_0, values = (xor__19_cast_fp16, xoi_19_cast_fp16))[name = tensor("out_19_cast_fp16")]; + tensor var_1196 = const()[name = tensor("op_1196"), val = tensor([256, 1024])]; + tensor var_1197_cast_fp16 = reshape(shape = var_1196, x = out_19_cast_fp16)[name = tensor("op_1197_cast_fp16")]; + tensor written_k_9_transpose_x_1 = const()[name = tensor("written_k_9_transpose_x_1"), val = tensor(true)]; + tensor written_k_9_transpose_y_1 = const()[name = tensor("written_k_9_transpose_y_1"), val = tensor(false)]; + tensor written_k_9_cast_fp16 = matmul(transpose_x = written_k_9_transpose_x_1, transpose_y = written_k_9_transpose_y_1, x = var_106_cast_fp16_to_fp16, y = var_1197_cast_fp16)[name = tensor("written_k_9_cast_fp16")]; + tensor var_1204 = const()[name = tensor("op_1204"), val = tensor([256, 1024])]; + tensor var_1205_cast_fp16 = reshape(shape = var_1204, x = v_9_cast_fp16)[name = tensor("op_1205_cast_fp16")]; + tensor written_v_9_transpose_x_1 = const()[name = tensor("written_v_9_transpose_x_1"), val = tensor(true)]; + tensor written_v_9_transpose_y_1 = const()[name = tensor("written_v_9_transpose_y_1"), val = tensor(false)]; + tensor written_v_9_cast_fp16 = matmul(transpose_x = written_v_9_transpose_x_1, transpose_y = written_v_9_transpose_y_1, x = var_106_cast_fp16_to_fp16, y = var_1205_cast_fp16)[name = tensor("written_v_9_cast_fp16")]; + tensor k_cache4_to_fp16_dtype_0 = const()[name = tensor("k_cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache4_to_fp16 = cast(dtype = k_cache4_to_fp16_dtype_0, x = k_cache4)[name = tensor("cast_99")]; + tensor var_1216_cast_fp16 = mul(x = k_cache4_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_1216_cast_fp16")]; + tensor var_1221 = const()[name = tensor("op_1221"), val = tensor([1, 512, 16, 64])]; + tensor var_1222_cast_fp16 = reshape(shape = var_1221, x = written_k_9_cast_fp16)[name = tensor("op_1222_cast_fp16")]; + tensor new_k_cache_9_cast_fp16 = add(x = var_1216_cast_fp16, y = var_1222_cast_fp16)[name = tensor("new_k_cache_9_cast_fp16")]; + tensor new_k_cache_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache4_to_fp16_dtype_0 = const()[name = tensor("v_cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache4_to_fp16 = cast(dtype = v_cache4_to_fp16_dtype_0, x = v_cache4)[name = tensor("cast_97")]; + tensor var_1225_cast_fp16 = mul(x = v_cache4_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_1225_cast_fp16")]; + tensor var_1230 = const()[name = tensor("op_1230"), val = tensor([1, 512, 16, 64])]; + tensor var_1231_cast_fp16 = reshape(shape = var_1230, x = written_v_9_cast_fp16)[name = tensor("op_1231_cast_fp16")]; + tensor new_v_cache_9_cast_fp16 = add(x = var_1225_cast_fp16, y = var_1231_cast_fp16)[name = tensor("new_v_cache_9_cast_fp16")]; + tensor new_v_cache_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_1244 = const()[name = tensor("op_1244"), val = tensor([0, 2, 1, 3])]; + tensor var_1255_transpose_x_0 = const()[name = tensor("op_1255_transpose_x_0"), val = tensor(false)]; + tensor var_1255_transpose_y_0 = const()[name = tensor("op_1255_transpose_y_0"), val = tensor(false)]; + tensor transpose_23_perm_0 = const()[name = tensor("transpose_23_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_24_perm_0 = const()[name = tensor("transpose_24_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_24 = transpose(perm = transpose_24_perm_0, x = new_k_cache_9_cast_fp16)[name = tensor("transpose_26")]; + tensor transpose_23 = transpose(perm = transpose_23_perm_0, x = q_9_cast_fp16)[name = tensor("transpose_27")]; + tensor var_1255_cast_fp16 = matmul(transpose_x = var_1255_transpose_x_0, transpose_y = var_1255_transpose_y_0, x = transpose_23, y = transpose_24)[name = tensor("op_1255_cast_fp16")]; + tensor var_1256_to_fp16 = const()[name = tensor("op_1256_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_17_cast_fp16 = mul(x = var_1255_cast_fp16, y = var_1256_to_fp16)[name = tensor("scores_17_cast_fp16")]; + tensor scores_19_cast_fp16 = add(x = scores_17_cast_fp16, y = var_339_cast_fp16)[name = tensor("scores_19_cast_fp16")]; + tensor var_1266 = const()[name = tensor("op_1266"), val = tensor(-1)]; + tensor weights_9_cast_fp16 = softmax(axis = var_1266, x = scores_19_cast_fp16)[name = tensor("weights_9_cast_fp16")]; + tensor attn_9_transpose_x_0 = const()[name = tensor("attn_9_transpose_x_0"), val = tensor(false)]; + tensor attn_9_transpose_y_0 = const()[name = tensor("attn_9_transpose_y_0"), val = tensor(false)]; + tensor values_9_cast_fp16 = transpose(perm = var_1244, x = new_v_cache_9_cast_fp16)[name = tensor("transpose_28")]; + tensor attn_9_cast_fp16 = matmul(transpose_x = attn_9_transpose_x_0, transpose_y = attn_9_transpose_y_0, x = weights_9_cast_fp16, y = values_9_cast_fp16)[name = tensor("attn_9_cast_fp16")]; + tensor var_1274 = const()[name = tensor("op_1274"), val = tensor([0, 2, 1, 3])]; + tensor var_1279 = const()[name = tensor("op_1279"), val = tensor([1, 256, 1024])]; + tensor var_1275_cast_fp16 = transpose(perm = var_1274, x = attn_9_cast_fp16)[name = tensor("transpose_25")]; + tensor input_53_cast_fp16 = reshape(shape = var_1279, x = var_1275_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor attn4_out_proj_weight_to_fp16 = const()[name = tensor("attn4_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(107012864)))]; + tensor linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn4_out_proj_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor position4_to_fp16_dtype_0 = const()[name = tensor("position4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position4_to_fp16 = cast(dtype = position4_to_fp16_dtype_0, x = position4)[name = tensor("cast_95")]; + tensor var_1288_cast_fp16 = add(x = position4_to_fp16, y = valid_len_to_fp16)[name = tensor("op_1288_cast_fp16")]; + tensor var_1288_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1288_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_55_cast_fp16 = add(x = input_49_cast_fp16, y = linear_17_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor input_57_axes_0 = const()[name = tensor("input_57_axes_0"), val = tensor([-1])]; + tensor norm4_2_weight_to_fp16 = const()[name = tensor("norm4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109110080)))]; + tensor norm4_2_bias_to_fp16 = const()[name = tensor("norm4_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109112192)))]; + tensor var_1292_to_fp16 = const()[name = tensor("op_1292_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_57_cast_fp16 = layer_norm(axes = input_57_axes_0, beta = norm4_2_bias_to_fp16, epsilon = var_1292_to_fp16, gamma = norm4_2_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor linear4_1_weight_to_fp16 = const()[name = tensor("linear4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109114304)))]; + tensor linear_18_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear4_1_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor input_59_mode_0 = const()[name = tensor("input_59_mode_0"), val = tensor("EXACT")]; + tensor input_59_cast_fp16 = gelu(mode = input_59_mode_0, x = linear_18_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor linear4_2_weight_to_fp16 = const()[name = tensor("linear4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(117502976)))]; + tensor linear_19_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear4_2_weight_to_fp16, x = input_59_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_61_cast_fp16 = add(x = input_55_cast_fp16, y = linear_19_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor input_63_axes_0 = const()[name = tensor("input_63_axes_0"), val = tensor([-1])]; + tensor norm5_1_weight_to_fp16 = const()[name = tensor("norm5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125891648)))]; + tensor norm5_1_bias_to_fp16 = const()[name = tensor("norm5_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125893760)))]; + tensor var_1309_to_fp16 = const()[name = tensor("op_1309_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_63_cast_fp16 = layer_norm(axes = input_63_axes_0, beta = norm5_1_bias_to_fp16, epsilon = var_1309_to_fp16, gamma = norm5_1_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor attn5_in_proj_weight_to_fp16 = const()[name = tensor("attn5_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125895872)))]; + tensor linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn5_in_proj_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor var_1323 = const()[name = tensor("op_1323"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_cast_fp16 = reshape(shape = var_1323, x = linear_20_cast_fp16)[name = tensor("qkv_cast_fp16")]; + tensor x_begin_0 = const()[name = tensor("x_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_end_0 = const()[name = tensor("x_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_end_mask_0 = const()[name = tensor("x_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_squeeze_mask_0 = const()[name = tensor("x_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_cast_fp16 = slice_by_index(begin = x_begin_0, end = x_end_0, end_mask = x_end_mask_0, squeeze_mask = x_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("x_cast_fp16")]; + tensor v_begin_0 = const()[name = tensor("v_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_end_0 = const()[name = tensor("v_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_end_mask_0 = const()[name = tensor("v_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_squeeze_mask_0 = const()[name = tensor("v_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_cast_fp16 = slice_by_index(begin = v_begin_0, end = v_end_0, end_mask = v_end_mask_0, squeeze_mask = v_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("v_cast_fp16")]; + tensor var_1355 = const()[name = tensor("op_1355"), val = tensor([256, 16, 32, 2])]; + tensor pairs_cast_fp16 = reshape(shape = var_1355, x = x_cast_fp16)[name = tensor("pairs_cast_fp16")]; + tensor xr_begin_0 = const()[name = tensor("xr_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_end_0 = const()[name = tensor("xr_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_end_mask_0 = const()[name = tensor("xr_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_squeeze_mask_0 = const()[name = tensor("xr_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_cast_fp16 = slice_by_index(begin = xr_begin_0, end = xr_end_0, end_mask = xr_end_mask_0, squeeze_mask = xr_squeeze_mask_0, x = pairs_cast_fp16)[name = tensor("xr_cast_fp16")]; + tensor xi_begin_0 = const()[name = tensor("xi_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_end_0 = const()[name = tensor("xi_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_end_mask_0 = const()[name = tensor("xi_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_squeeze_mask_0 = const()[name = tensor("xi_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_cast_fp16 = slice_by_index(begin = xi_begin_0, end = xi_end_0, end_mask = xi_end_mask_0, squeeze_mask = xi_squeeze_mask_0, x = pairs_cast_fp16)[name = tensor("xi_cast_fp16")]; + tensor var_1363_cast_fp16 = mul(x = xr_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1363_cast_fp16")]; + tensor var_1364_cast_fp16 = mul(x = xi_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1364_cast_fp16")]; + tensor xor__cast_fp16 = sub(x = var_1363_cast_fp16, y = var_1364_cast_fp16)[name = tensor("xor__cast_fp16")]; + tensor var_1367_cast_fp16 = mul(x = xr_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1367_cast_fp16")]; + tensor var_1368_cast_fp16 = mul(x = xi_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1368_cast_fp16")]; + tensor xoi_cast_fp16 = add(x = var_1367_cast_fp16, y = var_1368_cast_fp16)[name = tensor("xoi_cast_fp16")]; + tensor out_axis_0 = const()[name = tensor("out_axis_0"), val = tensor(-1)]; + tensor out_cast_fp16 = stack(axis = out_axis_0, values = (xor__cast_fp16, xoi_cast_fp16))[name = tensor("out_cast_fp16")]; + tensor var_1385 = const()[name = tensor("op_1385"), val = tensor([256, 1024])]; + tensor var_1386_cast_fp16 = reshape(shape = var_1385, x = out_cast_fp16)[name = tensor("op_1386_cast_fp16")]; + tensor written_k_transpose_x_1 = const()[name = tensor("written_k_transpose_x_1"), val = tensor(true)]; + tensor written_k_transpose_y_1 = const()[name = tensor("written_k_transpose_y_1"), val = tensor(false)]; + tensor written_k_cast_fp16 = matmul(transpose_x = written_k_transpose_x_1, transpose_y = written_k_transpose_y_1, x = var_106_cast_fp16_to_fp16, y = var_1386_cast_fp16)[name = tensor("written_k_cast_fp16")]; + tensor var_1393 = const()[name = tensor("op_1393"), val = tensor([256, 1024])]; + tensor var_1394_cast_fp16 = reshape(shape = var_1393, x = v_cast_fp16)[name = tensor("op_1394_cast_fp16")]; + tensor written_v_transpose_x_1 = const()[name = tensor("written_v_transpose_x_1"), val = tensor(true)]; + tensor written_v_transpose_y_1 = const()[name = tensor("written_v_transpose_y_1"), val = tensor(false)]; + tensor written_v_cast_fp16 = matmul(transpose_x = written_v_transpose_x_1, transpose_y = written_v_transpose_y_1, x = var_106_cast_fp16_to_fp16, y = var_1394_cast_fp16)[name = tensor("written_v_cast_fp16")]; + tensor k_cache5_to_fp16_dtype_0 = const()[name = tensor("k_cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache5_to_fp16 = cast(dtype = k_cache5_to_fp16_dtype_0, x = k_cache5)[name = tensor("cast_93")]; + tensor var_1405_cast_fp16 = mul(x = k_cache5_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_1405_cast_fp16")]; + tensor var_1410 = const()[name = tensor("op_1410"), val = tensor([1, 512, 16, 64])]; + tensor var_1411_cast_fp16 = reshape(shape = var_1410, x = written_k_cast_fp16)[name = tensor("op_1411_cast_fp16")]; + tensor new_k_cache_cast_fp16 = add(x = var_1405_cast_fp16, y = var_1411_cast_fp16)[name = tensor("new_k_cache_cast_fp16")]; + tensor new_k_cache_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache5_to_fp16_dtype_0 = const()[name = tensor("v_cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache5_to_fp16 = cast(dtype = v_cache5_to_fp16_dtype_0, x = v_cache5)[name = tensor("cast_91")]; + tensor var_1414_cast_fp16 = mul(x = v_cache5_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_1414_cast_fp16")]; + tensor var_1419 = const()[name = tensor("op_1419"), val = tensor([1, 512, 16, 64])]; + tensor var_1420_cast_fp16 = reshape(shape = var_1419, x = written_v_cast_fp16)[name = tensor("op_1420_cast_fp16")]; + tensor new_v_cache_cast_fp16 = add(x = var_1414_cast_fp16, y = var_1420_cast_fp16)[name = tensor("new_v_cache_cast_fp16")]; + tensor new_v_cache_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor position5_to_fp16_dtype_0 = const()[name = tensor("position5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position5_to_fp16 = cast(dtype = position5_to_fp16_dtype_0, x = position5)[name = tensor("cast_89")]; + tensor var_1427_cast_fp16 = add(x = position5_to_fp16, y = valid_len_to_fp16)[name = tensor("op_1427_cast_fp16")]; + tensor var_1427_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1427_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor new_position5 = cast(dtype = var_1427_cast_fp16_to_fp32_dtype_0, x = var_1427_cast_fp16)[name = tensor("cast_88")]; + tensor new_v_cache5 = cast(dtype = new_v_cache_cast_fp16_to_fp32_dtype_0, x = new_v_cache_cast_fp16)[name = tensor("cast_90")]; + tensor new_k_cache5 = cast(dtype = new_k_cache_cast_fp16_to_fp32_dtype_0, x = new_k_cache_cast_fp16)[name = tensor("cast_92")]; + tensor new_position4 = cast(dtype = var_1288_cast_fp16_to_fp32_dtype_0, x = var_1288_cast_fp16)[name = tensor("cast_94")]; + tensor new_v_cache4 = cast(dtype = new_v_cache_9_cast_fp16_to_fp32_dtype_0, x = new_v_cache_9_cast_fp16)[name = tensor("cast_96")]; + tensor new_k_cache4 = cast(dtype = new_k_cache_9_cast_fp16_to_fp32_dtype_0, x = new_k_cache_9_cast_fp16)[name = tensor("cast_98")]; + tensor new_position3 = cast(dtype = var_1057_cast_fp16_to_fp32_dtype_0, x = var_1057_cast_fp16)[name = tensor("cast_100")]; + tensor new_v_cache3 = cast(dtype = new_v_cache_7_cast_fp16_to_fp32_dtype_0, x = new_v_cache_7_cast_fp16)[name = tensor("cast_102")]; + tensor new_k_cache3 = cast(dtype = new_k_cache_7_cast_fp16_to_fp32_dtype_0, x = new_k_cache_7_cast_fp16)[name = tensor("cast_104")]; + tensor new_position2 = cast(dtype = var_826_cast_fp16_to_fp32_dtype_0, x = var_826_cast_fp16)[name = tensor("cast_106")]; + tensor new_v_cache2 = cast(dtype = new_v_cache_5_cast_fp16_to_fp32_dtype_0, x = new_v_cache_5_cast_fp16)[name = tensor("cast_108")]; + tensor new_k_cache2 = cast(dtype = new_k_cache_5_cast_fp16_to_fp32_dtype_0, x = new_k_cache_5_cast_fp16)[name = tensor("cast_110")]; + tensor new_position1 = cast(dtype = var_595_cast_fp16_to_fp32_dtype_0, x = var_595_cast_fp16)[name = tensor("cast_112")]; + tensor new_v_cache1 = cast(dtype = new_v_cache_3_cast_fp16_to_fp32_dtype_0, x = new_v_cache_3_cast_fp16)[name = tensor("cast_114")]; + tensor new_k_cache1 = cast(dtype = new_k_cache_3_cast_fp16_to_fp32_dtype_0, x = new_k_cache_3_cast_fp16)[name = tensor("cast_116")]; + tensor new_position0 = cast(dtype = var_364_cast_fp16_to_fp32_dtype_0, x = var_135_cast_fp16)[name = tensor("cast_118")]; + tensor new_v_cache0 = cast(dtype = new_v_cache_1_cast_fp16_to_fp32_dtype_0, x = new_v_cache_1_cast_fp16)[name = tensor("cast_119")]; + tensor new_k_cache0 = cast(dtype = new_k_cache_1_cast_fp16_to_fp32_dtype_0, x = new_k_cache_1_cast_fp16)[name = tensor("cast_121")]; + } -> (new_k_cache0, new_v_cache0, new_position0, new_k_cache1, new_v_cache1, new_position1, new_k_cache2, new_v_cache2, new_position2, new_k_cache3, new_v_cache3, new_position3, new_k_cache4, new_v_cache4, new_position4, new_k_cache5, new_v_cache5, new_position5); +} \ No newline at end of file diff --git a/v2.1/italian/cond_prefill_ane.mlmodelc/weights/weight.bin b/v2.1/italian/cond_prefill_ane.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..e9eb4cca08be50d90d5874e8bc0d9a7570c3a707 --- /dev/null +++ b/v2.1/italian/cond_prefill_ane.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:97bc6141ac29a3fe6122bd2eade45f5a1b5704e4d0fdb1bedf306305fa2bd00a +size 132187392 diff --git a/v2.1/italian/cond_prefill_ane.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/italian/cond_prefill_ane.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..5284df164607a1e02a9c55bd6d31147a368c3827 --- /dev/null +++ b/v2.1/italian/cond_prefill_ane.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eba20e8fd5c2deaf08a3e44b12f7f7d1b7a3a5c3cdb0f12ec09c91773728f9ef +size 119086 diff --git a/v2.1/italian/cond_prefill_ane.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/italian/cond_prefill_ane.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..e9eb4cca08be50d90d5874e8bc0d9a7570c3a707 --- /dev/null +++ b/v2.1/italian/cond_prefill_ane.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:97bc6141ac29a3fe6122bd2eade45f5a1b5704e4d0fdb1bedf306305fa2bd00a +size 132187392 diff --git a/v2.1/italian/cond_prefill_ane.mlpackage/Manifest.json b/v2.1/italian/cond_prefill_ane.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..5795286c25d19db7bc0c58f7d925fdbf206b1649 --- /dev/null +++ b/v2.1/italian/cond_prefill_ane.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "0CB3A8C9-3AEC-4E28-A313-FCC3FF683AE3": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "567FB4E6-8B41-4369-9ECB-F2F87E90FE94": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "567FB4E6-8B41-4369-9ECB-F2F87E90FE94" +} diff --git a/v2.1/italian/flowlm_step_ane.mlmodelc/analytics/coremldata.bin b/v2.1/italian/flowlm_step_ane.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..bfa3f44095167f50f13e108f84ca47d9cc53eb33 --- /dev/null +++ b/v2.1/italian/flowlm_step_ane.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9c9fbd985ced14e2dfd3152953d154b44a394e2db8e7616ff04adb4bb64d5d03 +size 243 diff --git a/v2.1/italian/flowlm_step_ane.mlmodelc/coremldata.bin b/v2.1/italian/flowlm_step_ane.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..a5c8a29fc28ba866705d54f9840a976de0a28ace --- /dev/null +++ b/v2.1/italian/flowlm_step_ane.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9174d028cface1bd7426a912a570dabcc6bf4f592dc3123f332604e2201ab6a6 +size 1414 diff --git a/v2.1/italian/flowlm_step_ane.mlmodelc/model.mil b/v2.1/italian/flowlm_step_ane.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..ce93c157c3ffb774d53632de6fc1633c4dd81499 --- /dev/null +++ b/v2.1/italian/flowlm_step_ane.mlmodelc/model.mil @@ -0,0 +1,945 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.9.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor k_cache0, tensor k_cache1, tensor k_cache2, tensor k_cache3, tensor k_cache4, tensor k_cache5, tensor position0, tensor position1, tensor position2, tensor position3, tensor position4, tensor position5, tensor sequence, tensor v_cache0, tensor v_cache1, tensor v_cache2, tensor v_cache3, tensor v_cache4, tensor v_cache5) { + tensor sequence_to_fp16_dtype_0 = const()[name = tensor("sequence_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor input_linear_weight_to_fp16 = const()[name = tensor("input_linear_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(65664)))]; + tensor sequence_to_fp16 = cast(dtype = sequence_to_fp16_dtype_0, x = sequence)[name = tensor("cast_146")]; + tensor linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = input_linear_weight_to_fp16, x = sequence_to_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; + tensor norm0_1_weight_to_fp16 = const()[name = tensor("norm0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67776)))]; + tensor norm0_1_bias_to_fp16 = const()[name = tensor("norm0_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(69888)))]; + tensor var_65_to_fp16 = const()[name = tensor("op_65_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_5_cast_fp16 = layer_norm(axes = input_5_axes_0, beta = norm0_1_bias_to_fp16, epsilon = var_65_to_fp16, gamma = norm0_1_weight_to_fp16, x = linear_0_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor attn0_in_proj_weight_to_fp16 = const()[name = tensor("attn0_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(72000)))]; + tensor linear_1_bias_0_to_fp16 = const()[name = tensor("linear_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6363520)))]; + tensor linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn0_in_proj_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("linear_1_cast_fp16")]; + tensor var_78 = const()[name = tensor("op_78"), val = tensor([1, 3, 16, 64])]; + tensor qkv_1_cast_fp16 = reshape(shape = var_78, x = linear_1_cast_fp16)[name = tensor("qkv_1_cast_fp16")]; + tensor x_1_begin_0 = const()[name = tensor("x_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_1_end_0 = const()[name = tensor("x_1_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_1_end_mask_0 = const()[name = tensor("x_1_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_1_squeeze_mask_0 = const()[name = tensor("x_1_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_1_cast_fp16 = slice_by_index(begin = x_1_begin_0, end = x_1_end_0, end_mask = x_1_end_mask_0, squeeze_mask = x_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("x_1_cast_fp16")]; + tensor x_3_begin_0 = const()[name = tensor("x_3_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_3_end_0 = const()[name = tensor("x_3_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_3_end_mask_0 = const()[name = tensor("x_3_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_3_squeeze_mask_0 = const()[name = tensor("x_3_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_3_cast_fp16 = slice_by_index(begin = x_3_begin_0, end = x_3_end_0, end_mask = x_3_end_mask_0, squeeze_mask = x_3_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("x_3_cast_fp16")]; + tensor v_1_begin_0 = const()[name = tensor("v_1_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_1_end_0 = const()[name = tensor("v_1_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_1_end_mask_0 = const()[name = tensor("v_1_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_1_squeeze_mask_0 = const()[name = tensor("v_1_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_1_cast_fp16 = slice_by_index(begin = v_1_begin_0, end = v_1_end_0, end_mask = v_1_end_mask_0, squeeze_mask = v_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("v_1_cast_fp16")]; + tensor var_107 = const()[name = tensor("op_107"), val = tensor([1, 1, 1])]; + tensor position0_to_fp16_dtype_0 = const()[name = tensor("position0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position0_to_fp16 = cast(dtype = position0_to_fp16_dtype_0, x = position0)[name = tensor("cast_145")]; + tensor var_108_cast_fp16 = reshape(shape = var_107, x = position0_to_fp16)[name = tensor("op_108_cast_fp16")]; + tensor rope_freqs_to_fp16 = const()[name = tensor("rope_freqs_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6369728)))]; + tensor angles_1_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_108_cast_fp16)[name = tensor("angles_1_cast_fp16")]; + tensor rotr_1_cast_fp16 = cos(x = angles_1_cast_fp16)[name = tensor("rotr_1_cast_fp16")]; + tensor roti_1_cast_fp16 = sin(x = angles_1_cast_fp16)[name = tensor("roti_1_cast_fp16")]; + tensor var_116 = const()[name = tensor("op_116"), val = tensor([1, 16, 32, 2])]; + tensor pairs_1_cast_fp16 = reshape(shape = var_116, x = x_1_cast_fp16)[name = tensor("pairs_1_cast_fp16")]; + tensor xr_1_begin_0 = const()[name = tensor("xr_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_1_end_0 = const()[name = tensor("xr_1_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_1_end_mask_0 = const()[name = tensor("xr_1_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_1_squeeze_mask_0 = const()[name = tensor("xr_1_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_1_cast_fp16 = slice_by_index(begin = xr_1_begin_0, end = xr_1_end_0, end_mask = xr_1_end_mask_0, squeeze_mask = xr_1_squeeze_mask_0, x = pairs_1_cast_fp16)[name = tensor("xr_1_cast_fp16")]; + tensor xi_1_begin_0 = const()[name = tensor("xi_1_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_1_end_0 = const()[name = tensor("xi_1_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_1_end_mask_0 = const()[name = tensor("xi_1_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_1_squeeze_mask_0 = const()[name = tensor("xi_1_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_1_cast_fp16 = slice_by_index(begin = xi_1_begin_0, end = xi_1_end_0, end_mask = xi_1_end_mask_0, squeeze_mask = xi_1_squeeze_mask_0, x = pairs_1_cast_fp16)[name = tensor("xi_1_cast_fp16")]; + tensor var_124_cast_fp16 = mul(x = xr_1_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_124_cast_fp16")]; + tensor var_125_cast_fp16 = mul(x = xi_1_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_125_cast_fp16")]; + tensor xor__1_cast_fp16 = sub(x = var_124_cast_fp16, y = var_125_cast_fp16)[name = tensor("xor__1_cast_fp16")]; + tensor var_128_cast_fp16 = mul(x = xr_1_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_128_cast_fp16")]; + tensor var_129_cast_fp16 = mul(x = xi_1_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_129_cast_fp16")]; + tensor xoi_1_cast_fp16 = add(x = var_128_cast_fp16, y = var_129_cast_fp16)[name = tensor("xoi_1_cast_fp16")]; + tensor out_1_axis_0 = const()[name = tensor("out_1_axis_0"), val = tensor(-1)]; + tensor out_1_cast_fp16 = stack(axis = out_1_axis_0, values = (xor__1_cast_fp16, xoi_1_cast_fp16))[name = tensor("out_1_cast_fp16")]; + tensor var_152 = const()[name = tensor("op_152"), val = tensor([1, 16, 32, 2])]; + tensor pairs_3_cast_fp16 = reshape(shape = var_152, x = x_3_cast_fp16)[name = tensor("pairs_3_cast_fp16")]; + tensor xr_3_begin_0 = const()[name = tensor("xr_3_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_3_end_0 = const()[name = tensor("xr_3_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_3_end_mask_0 = const()[name = tensor("xr_3_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_3_squeeze_mask_0 = const()[name = tensor("xr_3_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_3_cast_fp16 = slice_by_index(begin = xr_3_begin_0, end = xr_3_end_0, end_mask = xr_3_end_mask_0, squeeze_mask = xr_3_squeeze_mask_0, x = pairs_3_cast_fp16)[name = tensor("xr_3_cast_fp16")]; + tensor xi_3_begin_0 = const()[name = tensor("xi_3_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_3_end_0 = const()[name = tensor("xi_3_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_3_end_mask_0 = const()[name = tensor("xi_3_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_3_squeeze_mask_0 = const()[name = tensor("xi_3_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_3_cast_fp16 = slice_by_index(begin = xi_3_begin_0, end = xi_3_end_0, end_mask = xi_3_end_mask_0, squeeze_mask = xi_3_squeeze_mask_0, x = pairs_3_cast_fp16)[name = tensor("xi_3_cast_fp16")]; + tensor var_160_cast_fp16 = mul(x = xr_3_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_160_cast_fp16")]; + tensor var_161_cast_fp16 = mul(x = xi_3_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_161_cast_fp16")]; + tensor xor__3_cast_fp16 = sub(x = var_160_cast_fp16, y = var_161_cast_fp16)[name = tensor("xor__3_cast_fp16")]; + tensor var_164_cast_fp16 = mul(x = xr_3_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_164_cast_fp16")]; + tensor var_165_cast_fp16 = mul(x = xi_3_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_165_cast_fp16")]; + tensor xoi_3_cast_fp16 = add(x = var_164_cast_fp16, y = var_165_cast_fp16)[name = tensor("xoi_3_cast_fp16")]; + tensor out_3_axis_0 = const()[name = tensor("out_3_axis_0"), val = tensor(-1)]; + tensor out_3_cast_fp16 = stack(axis = out_3_axis_0, values = (xor__3_cast_fp16, xoi_3_cast_fp16))[name = tensor("out_3_cast_fp16")]; + tensor _inversed_177_y_0_to_fp16 = const()[name = tensor("_inversed_177_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_177_cast_fp16 = mul(x = position0_to_fp16, y = _inversed_177_y_0_to_fp16)[name = tensor("_inversed_177_cast_fp16")]; + tensor var_178_cast_fp16 = floor(x = _inversed_177_cast_fp16)[name = tensor("op_178_cast_fp16")]; + tensor var_179_to_fp16 = const()[name = tensor("op_179_to_fp16"), val = tensor(0x1p+9)]; + tensor var_180_cast_fp16 = mul(x = var_178_cast_fp16, y = var_179_to_fp16)[name = tensor("op_180_cast_fp16")]; + tensor write_idx_1_cast_fp16 = sub(x = position0_to_fp16, y = var_180_cast_fp16)[name = tensor("write_idx_1_cast_fp16")]; + tensor slot_indices_to_fp16 = const()[name = tensor("slot_indices_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6369856)))]; + tensor var_183_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_1_cast_fp16)[name = tensor("op_183_cast_fp16")]; + tensor var_193 = const()[name = tensor("op_193"), val = tensor([1, 512, 1, 1])]; + tensor onehot_1_to_fp16_dtype_0 = const()[name = tensor("onehot_1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_183_cast_fp16_to_fp16 = cast(dtype = onehot_1_to_fp16_dtype_0, x = var_183_cast_fp16)[name = tensor("cast_144")]; + tensor onehot_3_cast_fp16 = reshape(shape = var_193, x = var_183_cast_fp16_to_fp16)[name = tensor("onehot_3_cast_fp16")]; + tensor var_195_to_fp16 = const()[name = tensor("op_195_to_fp16"), val = tensor(0x1p+0)]; + tensor var_197_cast_fp16 = sub(x = var_195_to_fp16, y = onehot_3_cast_fp16)[name = tensor("op_197_cast_fp16")]; + tensor k_cache0_to_fp16_dtype_0 = const()[name = tensor("k_cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache0_to_fp16 = cast(dtype = k_cache0_to_fp16_dtype_0, x = k_cache0)[name = tensor("cast_143")]; + tensor var_198_cast_fp16 = mul(x = k_cache0_to_fp16, y = var_197_cast_fp16)[name = tensor("op_198_cast_fp16")]; + tensor var_203 = const()[name = tensor("op_203"), val = tensor([1, 1, 16, 64])]; + tensor var_204_cast_fp16 = reshape(shape = var_203, x = out_3_cast_fp16)[name = tensor("op_204_cast_fp16")]; + tensor var_205_cast_fp16 = mul(x = var_204_cast_fp16, y = onehot_3_cast_fp16)[name = tensor("op_205_cast_fp16")]; + tensor new_k_cache_1_cast_fp16 = add(x = var_198_cast_fp16, y = var_205_cast_fp16)[name = tensor("new_k_cache_1_cast_fp16")]; + tensor new_k_cache_1_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_1_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache0_to_fp16_dtype_0 = const()[name = tensor("v_cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache0_to_fp16 = cast(dtype = v_cache0_to_fp16_dtype_0, x = v_cache0)[name = tensor("cast_141")]; + tensor var_211_cast_fp16 = mul(x = v_cache0_to_fp16, y = var_197_cast_fp16)[name = tensor("op_211_cast_fp16")]; + tensor var_216 = const()[name = tensor("op_216"), val = tensor([1, 1, 16, 64])]; + tensor var_217_cast_fp16 = reshape(shape = var_216, x = v_1_cast_fp16)[name = tensor("op_217_cast_fp16")]; + tensor var_218_cast_fp16 = mul(x = var_217_cast_fp16, y = onehot_3_cast_fp16)[name = tensor("op_218_cast_fp16")]; + tensor new_v_cache_1_cast_fp16 = add(x = var_211_cast_fp16, y = var_218_cast_fp16)[name = tensor("new_v_cache_1_cast_fp16")]; + tensor new_v_cache_1_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_1_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_0_perm_0 = const()[name = tensor("transpose_0_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_231 = const()[name = tensor("op_231"), val = tensor([0, 2, 1, 3])]; + tensor var_237 = const()[name = tensor("op_237"), val = tensor([1, 16, 1, 64])]; + tensor q4_1_cast_fp16 = reshape(shape = var_237, x = out_1_cast_fp16)[name = tensor("q4_1_cast_fp16")]; + tensor var_242_transpose_x_0 = const()[name = tensor("op_242_transpose_x_0"), val = tensor(false)]; + tensor var_242_transpose_y_0 = const()[name = tensor("op_242_transpose_y_0"), val = tensor(false)]; + tensor transpose_0_cast_fp16 = transpose(perm = transpose_0_perm_0, x = new_k_cache_1_cast_fp16)[name = tensor("transpose_23")]; + tensor var_242_cast_fp16 = matmul(transpose_x = var_242_transpose_x_0, transpose_y = var_242_transpose_y_0, x = q4_1_cast_fp16, y = transpose_0_cast_fp16)[name = tensor("op_242_cast_fp16")]; + tensor var_243_to_fp16 = const()[name = tensor("op_243_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_1_cast_fp16 = mul(x = var_242_cast_fp16, y = var_243_to_fp16)[name = tensor("scores_1_cast_fp16")]; + tensor var_247 = const()[name = tensor("op_247"), val = tensor([1, 1])]; + tensor var_248_cast_fp16 = reshape(shape = var_247, x = position0_to_fp16)[name = tensor("op_248_cast_fp16")]; + tensor var_249_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_248_cast_fp16)[name = tensor("op_249_cast_fp16")]; + tensor mask_1_to_fp16_dtype_0 = const()[name = tensor("mask_1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_256_to_fp16 = const()[name = tensor("op_256_to_fp16"), val = tensor(0x1p+0)]; + tensor var_249_cast_fp16_to_fp16 = cast(dtype = mask_1_to_fp16_dtype_0, x = var_249_cast_fp16)[name = tensor("cast_139")]; + tensor var_257_cast_fp16 = sub(x = var_249_cast_fp16_to_fp16, y = var_256_to_fp16)[name = tensor("op_257_cast_fp16")]; + tensor var_262 = const()[name = tensor("op_262"), val = tensor([1, 1, 1, 512])]; + tensor var_263_cast_fp16 = reshape(shape = var_262, x = var_257_cast_fp16)[name = tensor("op_263_cast_fp16")]; + tensor var_264_to_fp16 = const()[name = tensor("op_264_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_265_cast_fp16 = mul(x = var_263_cast_fp16, y = var_264_to_fp16)[name = tensor("op_265_cast_fp16")]; + tensor scores_3_cast_fp16 = add(x = scores_1_cast_fp16, y = var_265_cast_fp16)[name = tensor("scores_3_cast_fp16")]; + tensor var_268 = const()[name = tensor("op_268"), val = tensor(-1)]; + tensor weights_1_cast_fp16 = softmax(axis = var_268, x = scores_3_cast_fp16)[name = tensor("weights_1_cast_fp16")]; + tensor attn_1_transpose_x_0 = const()[name = tensor("attn_1_transpose_x_0"), val = tensor(false)]; + tensor attn_1_transpose_y_0 = const()[name = tensor("attn_1_transpose_y_0"), val = tensor(false)]; + tensor values_1_cast_fp16 = transpose(perm = var_231, x = new_v_cache_1_cast_fp16)[name = tensor("transpose_22")]; + tensor attn_1_cast_fp16 = matmul(transpose_x = attn_1_transpose_x_0, transpose_y = attn_1_transpose_y_0, x = weights_1_cast_fp16, y = values_1_cast_fp16)[name = tensor("attn_1_cast_fp16")]; + tensor var_276 = const()[name = tensor("op_276"), val = tensor([0, 2, 1, 3])]; + tensor var_281 = const()[name = tensor("op_281"), val = tensor([1, 1, 1024])]; + tensor var_277_cast_fp16 = transpose(perm = var_276, x = attn_1_cast_fp16)[name = tensor("transpose_21")]; + tensor input_7_cast_fp16 = reshape(shape = var_281, x = var_277_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor attn0_out_proj_weight_to_fp16 = const()[name = tensor("attn0_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6370944)))]; + tensor linear_2_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn0_out_proj_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor var_287_to_fp16 = const()[name = tensor("op_287_to_fp16"), val = tensor(0x1p+0)]; + tensor var_288_cast_fp16 = add(x = position0_to_fp16, y = var_287_to_fp16)[name = tensor("op_288_cast_fp16")]; + tensor var_288_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_288_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_9_cast_fp16 = add(x = linear_0_cast_fp16, y = linear_2_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor input_11_axes_0 = const()[name = tensor("input_11_axes_0"), val = tensor([-1])]; + tensor norm0_2_weight_to_fp16 = const()[name = tensor("norm0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8468160)))]; + tensor norm0_2_bias_to_fp16 = const()[name = tensor("norm0_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8470272)))]; + tensor var_292_to_fp16 = const()[name = tensor("op_292_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_11_cast_fp16 = layer_norm(axes = input_11_axes_0, beta = norm0_2_bias_to_fp16, epsilon = var_292_to_fp16, gamma = norm0_2_weight_to_fp16, x = input_9_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor linear0_1_weight_to_fp16 = const()[name = tensor("linear0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8472384)))]; + tensor linear_3_bias_0_to_fp16 = const()[name = tensor("linear_3_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16861056)))]; + tensor linear_3_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear0_1_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("linear_3_cast_fp16")]; + tensor input_13_mode_0 = const()[name = tensor("input_13_mode_0"), val = tensor("EXACT")]; + tensor input_13_cast_fp16 = gelu(mode = input_13_mode_0, x = linear_3_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor linear0_2_weight_to_fp16 = const()[name = tensor("linear0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16869312)))]; + tensor linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear0_2_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor input_15_cast_fp16 = add(x = input_9_cast_fp16, y = linear_4_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor input_17_axes_0 = const()[name = tensor("input_17_axes_0"), val = tensor([-1])]; + tensor norm1_1_weight_to_fp16 = const()[name = tensor("norm1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25257984)))]; + tensor norm1_1_bias_to_fp16 = const()[name = tensor("norm1_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25260096)))]; + tensor var_309_to_fp16 = const()[name = tensor("op_309_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_17_cast_fp16 = layer_norm(axes = input_17_axes_0, beta = norm1_1_bias_to_fp16, epsilon = var_309_to_fp16, gamma = norm1_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor attn1_in_proj_weight_to_fp16 = const()[name = tensor("attn1_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25262208)))]; + tensor linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn1_in_proj_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor var_322 = const()[name = tensor("op_322"), val = tensor([1, 3, 16, 64])]; + tensor qkv_3_cast_fp16 = reshape(shape = var_322, x = linear_5_cast_fp16)[name = tensor("qkv_3_cast_fp16")]; + tensor x_5_begin_0 = const()[name = tensor("x_5_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_5_end_0 = const()[name = tensor("x_5_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_5_end_mask_0 = const()[name = tensor("x_5_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_5_squeeze_mask_0 = const()[name = tensor("x_5_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_5_cast_fp16 = slice_by_index(begin = x_5_begin_0, end = x_5_end_0, end_mask = x_5_end_mask_0, squeeze_mask = x_5_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("x_5_cast_fp16")]; + tensor x_7_begin_0 = const()[name = tensor("x_7_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_7_end_0 = const()[name = tensor("x_7_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_7_end_mask_0 = const()[name = tensor("x_7_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_7_squeeze_mask_0 = const()[name = tensor("x_7_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_7_cast_fp16 = slice_by_index(begin = x_7_begin_0, end = x_7_end_0, end_mask = x_7_end_mask_0, squeeze_mask = x_7_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("x_7_cast_fp16")]; + tensor v_3_begin_0 = const()[name = tensor("v_3_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_3_end_0 = const()[name = tensor("v_3_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_3_end_mask_0 = const()[name = tensor("v_3_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_3_squeeze_mask_0 = const()[name = tensor("v_3_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_3_cast_fp16 = slice_by_index(begin = v_3_begin_0, end = v_3_end_0, end_mask = v_3_end_mask_0, squeeze_mask = v_3_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("v_3_cast_fp16")]; + tensor var_351 = const()[name = tensor("op_351"), val = tensor([1, 1, 1])]; + tensor position1_to_fp16_dtype_0 = const()[name = tensor("position1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position1_to_fp16 = cast(dtype = position1_to_fp16_dtype_0, x = position1)[name = tensor("cast_137")]; + tensor var_352_cast_fp16 = reshape(shape = var_351, x = position1_to_fp16)[name = tensor("op_352_cast_fp16")]; + tensor angles_5_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_352_cast_fp16)[name = tensor("angles_5_cast_fp16")]; + tensor rotr_5_cast_fp16 = cos(x = angles_5_cast_fp16)[name = tensor("rotr_5_cast_fp16")]; + tensor roti_5_cast_fp16 = sin(x = angles_5_cast_fp16)[name = tensor("roti_5_cast_fp16")]; + tensor var_360 = const()[name = tensor("op_360"), val = tensor([1, 16, 32, 2])]; + tensor pairs_5_cast_fp16 = reshape(shape = var_360, x = x_5_cast_fp16)[name = tensor("pairs_5_cast_fp16")]; + tensor xr_5_begin_0 = const()[name = tensor("xr_5_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_5_end_0 = const()[name = tensor("xr_5_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_5_end_mask_0 = const()[name = tensor("xr_5_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_5_squeeze_mask_0 = const()[name = tensor("xr_5_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_5_cast_fp16 = slice_by_index(begin = xr_5_begin_0, end = xr_5_end_0, end_mask = xr_5_end_mask_0, squeeze_mask = xr_5_squeeze_mask_0, x = pairs_5_cast_fp16)[name = tensor("xr_5_cast_fp16")]; + tensor xi_5_begin_0 = const()[name = tensor("xi_5_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_5_end_0 = const()[name = tensor("xi_5_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_5_end_mask_0 = const()[name = tensor("xi_5_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_5_squeeze_mask_0 = const()[name = tensor("xi_5_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_5_cast_fp16 = slice_by_index(begin = xi_5_begin_0, end = xi_5_end_0, end_mask = xi_5_end_mask_0, squeeze_mask = xi_5_squeeze_mask_0, x = pairs_5_cast_fp16)[name = tensor("xi_5_cast_fp16")]; + tensor var_368_cast_fp16 = mul(x = xr_5_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_368_cast_fp16")]; + tensor var_369_cast_fp16 = mul(x = xi_5_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_369_cast_fp16")]; + tensor xor__5_cast_fp16 = sub(x = var_368_cast_fp16, y = var_369_cast_fp16)[name = tensor("xor__5_cast_fp16")]; + tensor var_372_cast_fp16 = mul(x = xr_5_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_372_cast_fp16")]; + tensor var_373_cast_fp16 = mul(x = xi_5_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_373_cast_fp16")]; + tensor xoi_5_cast_fp16 = add(x = var_372_cast_fp16, y = var_373_cast_fp16)[name = tensor("xoi_5_cast_fp16")]; + tensor out_5_axis_0 = const()[name = tensor("out_5_axis_0"), val = tensor(-1)]; + tensor out_5_cast_fp16 = stack(axis = out_5_axis_0, values = (xor__5_cast_fp16, xoi_5_cast_fp16))[name = tensor("out_5_cast_fp16")]; + tensor var_396 = const()[name = tensor("op_396"), val = tensor([1, 16, 32, 2])]; + tensor pairs_7_cast_fp16 = reshape(shape = var_396, x = x_7_cast_fp16)[name = tensor("pairs_7_cast_fp16")]; + tensor xr_7_begin_0 = const()[name = tensor("xr_7_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_7_end_0 = const()[name = tensor("xr_7_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_7_end_mask_0 = const()[name = tensor("xr_7_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_7_squeeze_mask_0 = const()[name = tensor("xr_7_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_7_cast_fp16 = slice_by_index(begin = xr_7_begin_0, end = xr_7_end_0, end_mask = xr_7_end_mask_0, squeeze_mask = xr_7_squeeze_mask_0, x = pairs_7_cast_fp16)[name = tensor("xr_7_cast_fp16")]; + tensor xi_7_begin_0 = const()[name = tensor("xi_7_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_7_end_0 = const()[name = tensor("xi_7_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_7_end_mask_0 = const()[name = tensor("xi_7_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_7_squeeze_mask_0 = const()[name = tensor("xi_7_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_7_cast_fp16 = slice_by_index(begin = xi_7_begin_0, end = xi_7_end_0, end_mask = xi_7_end_mask_0, squeeze_mask = xi_7_squeeze_mask_0, x = pairs_7_cast_fp16)[name = tensor("xi_7_cast_fp16")]; + tensor var_404_cast_fp16 = mul(x = xr_7_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_404_cast_fp16")]; + tensor var_405_cast_fp16 = mul(x = xi_7_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_405_cast_fp16")]; + tensor xor__7_cast_fp16 = sub(x = var_404_cast_fp16, y = var_405_cast_fp16)[name = tensor("xor__7_cast_fp16")]; + tensor var_408_cast_fp16 = mul(x = xr_7_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_408_cast_fp16")]; + tensor var_409_cast_fp16 = mul(x = xi_7_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_409_cast_fp16")]; + tensor xoi_7_cast_fp16 = add(x = var_408_cast_fp16, y = var_409_cast_fp16)[name = tensor("xoi_7_cast_fp16")]; + tensor out_7_axis_0 = const()[name = tensor("out_7_axis_0"), val = tensor(-1)]; + tensor out_7_cast_fp16 = stack(axis = out_7_axis_0, values = (xor__7_cast_fp16, xoi_7_cast_fp16))[name = tensor("out_7_cast_fp16")]; + tensor _inversed_421_y_0_to_fp16 = const()[name = tensor("_inversed_421_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_421_cast_fp16 = mul(x = position1_to_fp16, y = _inversed_421_y_0_to_fp16)[name = tensor("_inversed_421_cast_fp16")]; + tensor var_422_cast_fp16 = floor(x = _inversed_421_cast_fp16)[name = tensor("op_422_cast_fp16")]; + tensor var_423_to_fp16 = const()[name = tensor("op_423_to_fp16"), val = tensor(0x1p+9)]; + tensor var_424_cast_fp16 = mul(x = var_422_cast_fp16, y = var_423_to_fp16)[name = tensor("op_424_cast_fp16")]; + tensor write_idx_3_cast_fp16 = sub(x = position1_to_fp16, y = var_424_cast_fp16)[name = tensor("write_idx_3_cast_fp16")]; + tensor var_427_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_3_cast_fp16)[name = tensor("op_427_cast_fp16")]; + tensor var_437 = const()[name = tensor("op_437"), val = tensor([1, 512, 1, 1])]; + tensor onehot_5_to_fp16_dtype_0 = const()[name = tensor("onehot_5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_427_cast_fp16_to_fp16 = cast(dtype = onehot_5_to_fp16_dtype_0, x = var_427_cast_fp16)[name = tensor("cast_136")]; + tensor onehot_7_cast_fp16 = reshape(shape = var_437, x = var_427_cast_fp16_to_fp16)[name = tensor("onehot_7_cast_fp16")]; + tensor var_439_to_fp16 = const()[name = tensor("op_439_to_fp16"), val = tensor(0x1p+0)]; + tensor var_441_cast_fp16 = sub(x = var_439_to_fp16, y = onehot_7_cast_fp16)[name = tensor("op_441_cast_fp16")]; + tensor k_cache1_to_fp16_dtype_0 = const()[name = tensor("k_cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache1_to_fp16 = cast(dtype = k_cache1_to_fp16_dtype_0, x = k_cache1)[name = tensor("cast_135")]; + tensor var_442_cast_fp16 = mul(x = k_cache1_to_fp16, y = var_441_cast_fp16)[name = tensor("op_442_cast_fp16")]; + tensor var_447 = const()[name = tensor("op_447"), val = tensor([1, 1, 16, 64])]; + tensor var_448_cast_fp16 = reshape(shape = var_447, x = out_7_cast_fp16)[name = tensor("op_448_cast_fp16")]; + tensor var_449_cast_fp16 = mul(x = var_448_cast_fp16, y = onehot_7_cast_fp16)[name = tensor("op_449_cast_fp16")]; + tensor new_k_cache_3_cast_fp16 = add(x = var_442_cast_fp16, y = var_449_cast_fp16)[name = tensor("new_k_cache_3_cast_fp16")]; + tensor new_k_cache_3_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_3_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache1_to_fp16_dtype_0 = const()[name = tensor("v_cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache1_to_fp16 = cast(dtype = v_cache1_to_fp16_dtype_0, x = v_cache1)[name = tensor("cast_133")]; + tensor var_455_cast_fp16 = mul(x = v_cache1_to_fp16, y = var_441_cast_fp16)[name = tensor("op_455_cast_fp16")]; + tensor var_460 = const()[name = tensor("op_460"), val = tensor([1, 1, 16, 64])]; + tensor var_461_cast_fp16 = reshape(shape = var_460, x = v_3_cast_fp16)[name = tensor("op_461_cast_fp16")]; + tensor var_462_cast_fp16 = mul(x = var_461_cast_fp16, y = onehot_7_cast_fp16)[name = tensor("op_462_cast_fp16")]; + tensor new_v_cache_3_cast_fp16 = add(x = var_455_cast_fp16, y = var_462_cast_fp16)[name = tensor("new_v_cache_3_cast_fp16")]; + tensor new_v_cache_3_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_3_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_1_perm_0 = const()[name = tensor("transpose_1_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_475 = const()[name = tensor("op_475"), val = tensor([0, 2, 1, 3])]; + tensor var_481 = const()[name = tensor("op_481"), val = tensor([1, 16, 1, 64])]; + tensor q4_3_cast_fp16 = reshape(shape = var_481, x = out_5_cast_fp16)[name = tensor("q4_3_cast_fp16")]; + tensor var_486_transpose_x_0 = const()[name = tensor("op_486_transpose_x_0"), val = tensor(false)]; + tensor var_486_transpose_y_0 = const()[name = tensor("op_486_transpose_y_0"), val = tensor(false)]; + tensor transpose_1_cast_fp16 = transpose(perm = transpose_1_perm_0, x = new_k_cache_3_cast_fp16)[name = tensor("transpose_20")]; + tensor var_486_cast_fp16 = matmul(transpose_x = var_486_transpose_x_0, transpose_y = var_486_transpose_y_0, x = q4_3_cast_fp16, y = transpose_1_cast_fp16)[name = tensor("op_486_cast_fp16")]; + tensor var_487_to_fp16 = const()[name = tensor("op_487_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_5_cast_fp16 = mul(x = var_486_cast_fp16, y = var_487_to_fp16)[name = tensor("scores_5_cast_fp16")]; + tensor var_491 = const()[name = tensor("op_491"), val = tensor([1, 1])]; + tensor var_492_cast_fp16 = reshape(shape = var_491, x = position1_to_fp16)[name = tensor("op_492_cast_fp16")]; + tensor var_493_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_492_cast_fp16)[name = tensor("op_493_cast_fp16")]; + tensor mask_3_to_fp16_dtype_0 = const()[name = tensor("mask_3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_500_to_fp16 = const()[name = tensor("op_500_to_fp16"), val = tensor(0x1p+0)]; + tensor var_493_cast_fp16_to_fp16 = cast(dtype = mask_3_to_fp16_dtype_0, x = var_493_cast_fp16)[name = tensor("cast_131")]; + tensor var_501_cast_fp16 = sub(x = var_493_cast_fp16_to_fp16, y = var_500_to_fp16)[name = tensor("op_501_cast_fp16")]; + tensor var_506 = const()[name = tensor("op_506"), val = tensor([1, 1, 1, 512])]; + tensor var_507_cast_fp16 = reshape(shape = var_506, x = var_501_cast_fp16)[name = tensor("op_507_cast_fp16")]; + tensor var_508_to_fp16 = const()[name = tensor("op_508_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_509_cast_fp16 = mul(x = var_507_cast_fp16, y = var_508_to_fp16)[name = tensor("op_509_cast_fp16")]; + tensor scores_7_cast_fp16 = add(x = scores_5_cast_fp16, y = var_509_cast_fp16)[name = tensor("scores_7_cast_fp16")]; + tensor var_512 = const()[name = tensor("op_512"), val = tensor(-1)]; + tensor weights_3_cast_fp16 = softmax(axis = var_512, x = scores_7_cast_fp16)[name = tensor("weights_3_cast_fp16")]; + tensor attn_3_transpose_x_0 = const()[name = tensor("attn_3_transpose_x_0"), val = tensor(false)]; + tensor attn_3_transpose_y_0 = const()[name = tensor("attn_3_transpose_y_0"), val = tensor(false)]; + tensor values_3_cast_fp16 = transpose(perm = var_475, x = new_v_cache_3_cast_fp16)[name = tensor("transpose_19")]; + tensor attn_3_cast_fp16 = matmul(transpose_x = attn_3_transpose_x_0, transpose_y = attn_3_transpose_y_0, x = weights_3_cast_fp16, y = values_3_cast_fp16)[name = tensor("attn_3_cast_fp16")]; + tensor var_520 = const()[name = tensor("op_520"), val = tensor([0, 2, 1, 3])]; + tensor var_525 = const()[name = tensor("op_525"), val = tensor([1, 1, 1024])]; + tensor var_521_cast_fp16 = transpose(perm = var_520, x = attn_3_cast_fp16)[name = tensor("transpose_18")]; + tensor input_19_cast_fp16 = reshape(shape = var_525, x = var_521_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor attn1_out_proj_weight_to_fp16 = const()[name = tensor("attn1_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31553728)))]; + tensor linear_6_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn1_out_proj_weight_to_fp16, x = input_19_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor var_531_to_fp16 = const()[name = tensor("op_531_to_fp16"), val = tensor(0x1p+0)]; + tensor var_532_cast_fp16 = add(x = position1_to_fp16, y = var_531_to_fp16)[name = tensor("op_532_cast_fp16")]; + tensor var_532_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_532_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_21_cast_fp16 = add(x = input_15_cast_fp16, y = linear_6_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor input_23_axes_0 = const()[name = tensor("input_23_axes_0"), val = tensor([-1])]; + tensor norm1_2_weight_to_fp16 = const()[name = tensor("norm1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33650944)))]; + tensor norm1_2_bias_to_fp16 = const()[name = tensor("norm1_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33653056)))]; + tensor var_536_to_fp16 = const()[name = tensor("op_536_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_23_cast_fp16 = layer_norm(axes = input_23_axes_0, beta = norm1_2_bias_to_fp16, epsilon = var_536_to_fp16, gamma = norm1_2_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor linear1_1_weight_to_fp16 = const()[name = tensor("linear1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33655168)))]; + tensor linear_7_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear1_1_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_25_mode_0 = const()[name = tensor("input_25_mode_0"), val = tensor("EXACT")]; + tensor input_25_cast_fp16 = gelu(mode = input_25_mode_0, x = linear_7_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor linear1_2_weight_to_fp16 = const()[name = tensor("linear1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42043840)))]; + tensor linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear1_2_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor input_27_cast_fp16 = add(x = input_21_cast_fp16, y = linear_8_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor input_29_axes_0 = const()[name = tensor("input_29_axes_0"), val = tensor([-1])]; + tensor norm2_1_weight_to_fp16 = const()[name = tensor("norm2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50432512)))]; + tensor norm2_1_bias_to_fp16 = const()[name = tensor("norm2_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50434624)))]; + tensor var_553_to_fp16 = const()[name = tensor("op_553_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_29_cast_fp16 = layer_norm(axes = input_29_axes_0, beta = norm2_1_bias_to_fp16, epsilon = var_553_to_fp16, gamma = norm2_1_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor attn2_in_proj_weight_to_fp16 = const()[name = tensor("attn2_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50436736)))]; + tensor linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn2_in_proj_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor var_566 = const()[name = tensor("op_566"), val = tensor([1, 3, 16, 64])]; + tensor qkv_5_cast_fp16 = reshape(shape = var_566, x = linear_9_cast_fp16)[name = tensor("qkv_5_cast_fp16")]; + tensor x_9_begin_0 = const()[name = tensor("x_9_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_9_end_0 = const()[name = tensor("x_9_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_9_end_mask_0 = const()[name = tensor("x_9_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_9_squeeze_mask_0 = const()[name = tensor("x_9_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_9_cast_fp16 = slice_by_index(begin = x_9_begin_0, end = x_9_end_0, end_mask = x_9_end_mask_0, squeeze_mask = x_9_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("x_9_cast_fp16")]; + tensor x_11_begin_0 = const()[name = tensor("x_11_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_11_end_0 = const()[name = tensor("x_11_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_11_end_mask_0 = const()[name = tensor("x_11_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_11_squeeze_mask_0 = const()[name = tensor("x_11_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_11_cast_fp16 = slice_by_index(begin = x_11_begin_0, end = x_11_end_0, end_mask = x_11_end_mask_0, squeeze_mask = x_11_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("x_11_cast_fp16")]; + tensor v_5_begin_0 = const()[name = tensor("v_5_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_5_end_0 = const()[name = tensor("v_5_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_5_end_mask_0 = const()[name = tensor("v_5_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_5_squeeze_mask_0 = const()[name = tensor("v_5_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_5_cast_fp16 = slice_by_index(begin = v_5_begin_0, end = v_5_end_0, end_mask = v_5_end_mask_0, squeeze_mask = v_5_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("v_5_cast_fp16")]; + tensor var_595 = const()[name = tensor("op_595"), val = tensor([1, 1, 1])]; + tensor position2_to_fp16_dtype_0 = const()[name = tensor("position2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position2_to_fp16 = cast(dtype = position2_to_fp16_dtype_0, x = position2)[name = tensor("cast_129")]; + tensor var_596_cast_fp16 = reshape(shape = var_595, x = position2_to_fp16)[name = tensor("op_596_cast_fp16")]; + tensor angles_9_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_596_cast_fp16)[name = tensor("angles_9_cast_fp16")]; + tensor rotr_9_cast_fp16 = cos(x = angles_9_cast_fp16)[name = tensor("rotr_9_cast_fp16")]; + tensor roti_9_cast_fp16 = sin(x = angles_9_cast_fp16)[name = tensor("roti_9_cast_fp16")]; + tensor var_604 = const()[name = tensor("op_604"), val = tensor([1, 16, 32, 2])]; + tensor pairs_9_cast_fp16 = reshape(shape = var_604, x = x_9_cast_fp16)[name = tensor("pairs_9_cast_fp16")]; + tensor xr_9_begin_0 = const()[name = tensor("xr_9_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_9_end_0 = const()[name = tensor("xr_9_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_9_end_mask_0 = const()[name = tensor("xr_9_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_9_squeeze_mask_0 = const()[name = tensor("xr_9_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_9_cast_fp16 = slice_by_index(begin = xr_9_begin_0, end = xr_9_end_0, end_mask = xr_9_end_mask_0, squeeze_mask = xr_9_squeeze_mask_0, x = pairs_9_cast_fp16)[name = tensor("xr_9_cast_fp16")]; + tensor xi_9_begin_0 = const()[name = tensor("xi_9_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_9_end_0 = const()[name = tensor("xi_9_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_9_end_mask_0 = const()[name = tensor("xi_9_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_9_squeeze_mask_0 = const()[name = tensor("xi_9_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_9_cast_fp16 = slice_by_index(begin = xi_9_begin_0, end = xi_9_end_0, end_mask = xi_9_end_mask_0, squeeze_mask = xi_9_squeeze_mask_0, x = pairs_9_cast_fp16)[name = tensor("xi_9_cast_fp16")]; + tensor var_612_cast_fp16 = mul(x = xr_9_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_612_cast_fp16")]; + tensor var_613_cast_fp16 = mul(x = xi_9_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_613_cast_fp16")]; + tensor xor__9_cast_fp16 = sub(x = var_612_cast_fp16, y = var_613_cast_fp16)[name = tensor("xor__9_cast_fp16")]; + tensor var_616_cast_fp16 = mul(x = xr_9_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_616_cast_fp16")]; + tensor var_617_cast_fp16 = mul(x = xi_9_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_617_cast_fp16")]; + tensor xoi_9_cast_fp16 = add(x = var_616_cast_fp16, y = var_617_cast_fp16)[name = tensor("xoi_9_cast_fp16")]; + tensor out_9_axis_0 = const()[name = tensor("out_9_axis_0"), val = tensor(-1)]; + tensor out_9_cast_fp16 = stack(axis = out_9_axis_0, values = (xor__9_cast_fp16, xoi_9_cast_fp16))[name = tensor("out_9_cast_fp16")]; + tensor var_640 = const()[name = tensor("op_640"), val = tensor([1, 16, 32, 2])]; + tensor pairs_11_cast_fp16 = reshape(shape = var_640, x = x_11_cast_fp16)[name = tensor("pairs_11_cast_fp16")]; + tensor xr_11_begin_0 = const()[name = tensor("xr_11_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_11_end_0 = const()[name = tensor("xr_11_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_11_end_mask_0 = const()[name = tensor("xr_11_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_11_squeeze_mask_0 = const()[name = tensor("xr_11_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_11_cast_fp16 = slice_by_index(begin = xr_11_begin_0, end = xr_11_end_0, end_mask = xr_11_end_mask_0, squeeze_mask = xr_11_squeeze_mask_0, x = pairs_11_cast_fp16)[name = tensor("xr_11_cast_fp16")]; + tensor xi_11_begin_0 = const()[name = tensor("xi_11_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_11_end_0 = const()[name = tensor("xi_11_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_11_end_mask_0 = const()[name = tensor("xi_11_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_11_squeeze_mask_0 = const()[name = tensor("xi_11_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_11_cast_fp16 = slice_by_index(begin = xi_11_begin_0, end = xi_11_end_0, end_mask = xi_11_end_mask_0, squeeze_mask = xi_11_squeeze_mask_0, x = pairs_11_cast_fp16)[name = tensor("xi_11_cast_fp16")]; + tensor var_648_cast_fp16 = mul(x = xr_11_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_648_cast_fp16")]; + tensor var_649_cast_fp16 = mul(x = xi_11_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_649_cast_fp16")]; + tensor xor__11_cast_fp16 = sub(x = var_648_cast_fp16, y = var_649_cast_fp16)[name = tensor("xor__11_cast_fp16")]; + tensor var_652_cast_fp16 = mul(x = xr_11_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_652_cast_fp16")]; + tensor var_653_cast_fp16 = mul(x = xi_11_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_653_cast_fp16")]; + tensor xoi_11_cast_fp16 = add(x = var_652_cast_fp16, y = var_653_cast_fp16)[name = tensor("xoi_11_cast_fp16")]; + tensor out_11_axis_0 = const()[name = tensor("out_11_axis_0"), val = tensor(-1)]; + tensor out_11_cast_fp16 = stack(axis = out_11_axis_0, values = (xor__11_cast_fp16, xoi_11_cast_fp16))[name = tensor("out_11_cast_fp16")]; + tensor _inversed_665_y_0_to_fp16 = const()[name = tensor("_inversed_665_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_665_cast_fp16 = mul(x = position2_to_fp16, y = _inversed_665_y_0_to_fp16)[name = tensor("_inversed_665_cast_fp16")]; + tensor var_666_cast_fp16 = floor(x = _inversed_665_cast_fp16)[name = tensor("op_666_cast_fp16")]; + tensor var_667_to_fp16 = const()[name = tensor("op_667_to_fp16"), val = tensor(0x1p+9)]; + tensor var_668_cast_fp16 = mul(x = var_666_cast_fp16, y = var_667_to_fp16)[name = tensor("op_668_cast_fp16")]; + tensor write_idx_5_cast_fp16 = sub(x = position2_to_fp16, y = var_668_cast_fp16)[name = tensor("write_idx_5_cast_fp16")]; + tensor var_671_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_5_cast_fp16)[name = tensor("op_671_cast_fp16")]; + tensor var_681 = const()[name = tensor("op_681"), val = tensor([1, 512, 1, 1])]; + tensor onehot_9_to_fp16_dtype_0 = const()[name = tensor("onehot_9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_671_cast_fp16_to_fp16 = cast(dtype = onehot_9_to_fp16_dtype_0, x = var_671_cast_fp16)[name = tensor("cast_128")]; + tensor onehot_11_cast_fp16 = reshape(shape = var_681, x = var_671_cast_fp16_to_fp16)[name = tensor("onehot_11_cast_fp16")]; + tensor var_683_to_fp16 = const()[name = tensor("op_683_to_fp16"), val = tensor(0x1p+0)]; + tensor var_685_cast_fp16 = sub(x = var_683_to_fp16, y = onehot_11_cast_fp16)[name = tensor("op_685_cast_fp16")]; + tensor k_cache2_to_fp16_dtype_0 = const()[name = tensor("k_cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache2_to_fp16 = cast(dtype = k_cache2_to_fp16_dtype_0, x = k_cache2)[name = tensor("cast_127")]; + tensor var_686_cast_fp16 = mul(x = k_cache2_to_fp16, y = var_685_cast_fp16)[name = tensor("op_686_cast_fp16")]; + tensor var_691 = const()[name = tensor("op_691"), val = tensor([1, 1, 16, 64])]; + tensor var_692_cast_fp16 = reshape(shape = var_691, x = out_11_cast_fp16)[name = tensor("op_692_cast_fp16")]; + tensor var_693_cast_fp16 = mul(x = var_692_cast_fp16, y = onehot_11_cast_fp16)[name = tensor("op_693_cast_fp16")]; + tensor new_k_cache_5_cast_fp16 = add(x = var_686_cast_fp16, y = var_693_cast_fp16)[name = tensor("new_k_cache_5_cast_fp16")]; + tensor new_k_cache_5_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_5_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache2_to_fp16_dtype_0 = const()[name = tensor("v_cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache2_to_fp16 = cast(dtype = v_cache2_to_fp16_dtype_0, x = v_cache2)[name = tensor("cast_125")]; + tensor var_699_cast_fp16 = mul(x = v_cache2_to_fp16, y = var_685_cast_fp16)[name = tensor("op_699_cast_fp16")]; + tensor var_704 = const()[name = tensor("op_704"), val = tensor([1, 1, 16, 64])]; + tensor var_705_cast_fp16 = reshape(shape = var_704, x = v_5_cast_fp16)[name = tensor("op_705_cast_fp16")]; + tensor var_706_cast_fp16 = mul(x = var_705_cast_fp16, y = onehot_11_cast_fp16)[name = tensor("op_706_cast_fp16")]; + tensor new_v_cache_5_cast_fp16 = add(x = var_699_cast_fp16, y = var_706_cast_fp16)[name = tensor("new_v_cache_5_cast_fp16")]; + tensor new_v_cache_5_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_5_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_2_perm_0 = const()[name = tensor("transpose_2_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_719 = const()[name = tensor("op_719"), val = tensor([0, 2, 1, 3])]; + tensor var_725 = const()[name = tensor("op_725"), val = tensor([1, 16, 1, 64])]; + tensor q4_5_cast_fp16 = reshape(shape = var_725, x = out_9_cast_fp16)[name = tensor("q4_5_cast_fp16")]; + tensor var_730_transpose_x_0 = const()[name = tensor("op_730_transpose_x_0"), val = tensor(false)]; + tensor var_730_transpose_y_0 = const()[name = tensor("op_730_transpose_y_0"), val = tensor(false)]; + tensor transpose_2_cast_fp16 = transpose(perm = transpose_2_perm_0, x = new_k_cache_5_cast_fp16)[name = tensor("transpose_17")]; + tensor var_730_cast_fp16 = matmul(transpose_x = var_730_transpose_x_0, transpose_y = var_730_transpose_y_0, x = q4_5_cast_fp16, y = transpose_2_cast_fp16)[name = tensor("op_730_cast_fp16")]; + tensor var_731_to_fp16 = const()[name = tensor("op_731_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_9_cast_fp16 = mul(x = var_730_cast_fp16, y = var_731_to_fp16)[name = tensor("scores_9_cast_fp16")]; + tensor var_735 = const()[name = tensor("op_735"), val = tensor([1, 1])]; + tensor var_736_cast_fp16 = reshape(shape = var_735, x = position2_to_fp16)[name = tensor("op_736_cast_fp16")]; + tensor var_737_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_736_cast_fp16)[name = tensor("op_737_cast_fp16")]; + tensor mask_5_to_fp16_dtype_0 = const()[name = tensor("mask_5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_744_to_fp16 = const()[name = tensor("op_744_to_fp16"), val = tensor(0x1p+0)]; + tensor var_737_cast_fp16_to_fp16 = cast(dtype = mask_5_to_fp16_dtype_0, x = var_737_cast_fp16)[name = tensor("cast_123")]; + tensor var_745_cast_fp16 = sub(x = var_737_cast_fp16_to_fp16, y = var_744_to_fp16)[name = tensor("op_745_cast_fp16")]; + tensor var_750 = const()[name = tensor("op_750"), val = tensor([1, 1, 1, 512])]; + tensor var_751_cast_fp16 = reshape(shape = var_750, x = var_745_cast_fp16)[name = tensor("op_751_cast_fp16")]; + tensor var_752_to_fp16 = const()[name = tensor("op_752_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_753_cast_fp16 = mul(x = var_751_cast_fp16, y = var_752_to_fp16)[name = tensor("op_753_cast_fp16")]; + tensor scores_11_cast_fp16 = add(x = scores_9_cast_fp16, y = var_753_cast_fp16)[name = tensor("scores_11_cast_fp16")]; + tensor var_756 = const()[name = tensor("op_756"), val = tensor(-1)]; + tensor weights_5_cast_fp16 = softmax(axis = var_756, x = scores_11_cast_fp16)[name = tensor("weights_5_cast_fp16")]; + tensor attn_5_transpose_x_0 = const()[name = tensor("attn_5_transpose_x_0"), val = tensor(false)]; + tensor attn_5_transpose_y_0 = const()[name = tensor("attn_5_transpose_y_0"), val = tensor(false)]; + tensor values_5_cast_fp16 = transpose(perm = var_719, x = new_v_cache_5_cast_fp16)[name = tensor("transpose_16")]; + tensor attn_5_cast_fp16 = matmul(transpose_x = attn_5_transpose_x_0, transpose_y = attn_5_transpose_y_0, x = weights_5_cast_fp16, y = values_5_cast_fp16)[name = tensor("attn_5_cast_fp16")]; + tensor var_764 = const()[name = tensor("op_764"), val = tensor([0, 2, 1, 3])]; + tensor var_769 = const()[name = tensor("op_769"), val = tensor([1, 1, 1024])]; + tensor var_765_cast_fp16 = transpose(perm = var_764, x = attn_5_cast_fp16)[name = tensor("transpose_15")]; + tensor input_31_cast_fp16 = reshape(shape = var_769, x = var_765_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor attn2_out_proj_weight_to_fp16 = const()[name = tensor("attn2_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(56728256)))]; + tensor linear_10_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn2_out_proj_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor var_775_to_fp16 = const()[name = tensor("op_775_to_fp16"), val = tensor(0x1p+0)]; + tensor var_776_cast_fp16 = add(x = position2_to_fp16, y = var_775_to_fp16)[name = tensor("op_776_cast_fp16")]; + tensor var_776_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_776_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_33_cast_fp16 = add(x = input_27_cast_fp16, y = linear_10_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor input_35_axes_0 = const()[name = tensor("input_35_axes_0"), val = tensor([-1])]; + tensor norm2_2_weight_to_fp16 = const()[name = tensor("norm2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58825472)))]; + tensor norm2_2_bias_to_fp16 = const()[name = tensor("norm2_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58827584)))]; + tensor var_780_to_fp16 = const()[name = tensor("op_780_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_35_cast_fp16 = layer_norm(axes = input_35_axes_0, beta = norm2_2_bias_to_fp16, epsilon = var_780_to_fp16, gamma = norm2_2_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor linear2_1_weight_to_fp16 = const()[name = tensor("linear2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58829696)))]; + tensor linear_11_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear2_1_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor input_37_mode_0 = const()[name = tensor("input_37_mode_0"), val = tensor("EXACT")]; + tensor input_37_cast_fp16 = gelu(mode = input_37_mode_0, x = linear_11_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor linear2_2_weight_to_fp16 = const()[name = tensor("linear2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67218368)))]; + tensor linear_12_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear2_2_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor input_39_cast_fp16 = add(x = input_33_cast_fp16, y = linear_12_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor input_41_axes_0 = const()[name = tensor("input_41_axes_0"), val = tensor([-1])]; + tensor norm3_1_weight_to_fp16 = const()[name = tensor("norm3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75607040)))]; + tensor norm3_1_bias_to_fp16 = const()[name = tensor("norm3_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75609152)))]; + tensor var_797_to_fp16 = const()[name = tensor("op_797_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_41_cast_fp16 = layer_norm(axes = input_41_axes_0, beta = norm3_1_bias_to_fp16, epsilon = var_797_to_fp16, gamma = norm3_1_weight_to_fp16, x = input_39_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor attn3_in_proj_weight_to_fp16 = const()[name = tensor("attn3_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75611264)))]; + tensor linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn3_in_proj_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor var_810 = const()[name = tensor("op_810"), val = tensor([1, 3, 16, 64])]; + tensor qkv_7_cast_fp16 = reshape(shape = var_810, x = linear_13_cast_fp16)[name = tensor("qkv_7_cast_fp16")]; + tensor x_13_begin_0 = const()[name = tensor("x_13_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_13_end_0 = const()[name = tensor("x_13_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_13_end_mask_0 = const()[name = tensor("x_13_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_13_squeeze_mask_0 = const()[name = tensor("x_13_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_13_cast_fp16 = slice_by_index(begin = x_13_begin_0, end = x_13_end_0, end_mask = x_13_end_mask_0, squeeze_mask = x_13_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("x_13_cast_fp16")]; + tensor x_15_begin_0 = const()[name = tensor("x_15_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_15_end_0 = const()[name = tensor("x_15_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_15_end_mask_0 = const()[name = tensor("x_15_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_15_squeeze_mask_0 = const()[name = tensor("x_15_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_15_cast_fp16 = slice_by_index(begin = x_15_begin_0, end = x_15_end_0, end_mask = x_15_end_mask_0, squeeze_mask = x_15_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("x_15_cast_fp16")]; + tensor v_7_begin_0 = const()[name = tensor("v_7_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_7_end_0 = const()[name = tensor("v_7_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_7_end_mask_0 = const()[name = tensor("v_7_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_7_squeeze_mask_0 = const()[name = tensor("v_7_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_7_cast_fp16 = slice_by_index(begin = v_7_begin_0, end = v_7_end_0, end_mask = v_7_end_mask_0, squeeze_mask = v_7_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("v_7_cast_fp16")]; + tensor var_839 = const()[name = tensor("op_839"), val = tensor([1, 1, 1])]; + tensor position3_to_fp16_dtype_0 = const()[name = tensor("position3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position3_to_fp16 = cast(dtype = position3_to_fp16_dtype_0, x = position3)[name = tensor("cast_121")]; + tensor var_840_cast_fp16 = reshape(shape = var_839, x = position3_to_fp16)[name = tensor("op_840_cast_fp16")]; + tensor angles_13_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_840_cast_fp16)[name = tensor("angles_13_cast_fp16")]; + tensor rotr_13_cast_fp16 = cos(x = angles_13_cast_fp16)[name = tensor("rotr_13_cast_fp16")]; + tensor roti_13_cast_fp16 = sin(x = angles_13_cast_fp16)[name = tensor("roti_13_cast_fp16")]; + tensor var_848 = const()[name = tensor("op_848"), val = tensor([1, 16, 32, 2])]; + tensor pairs_13_cast_fp16 = reshape(shape = var_848, x = x_13_cast_fp16)[name = tensor("pairs_13_cast_fp16")]; + tensor xr_13_begin_0 = const()[name = tensor("xr_13_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_13_end_0 = const()[name = tensor("xr_13_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_13_end_mask_0 = const()[name = tensor("xr_13_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_13_squeeze_mask_0 = const()[name = tensor("xr_13_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_13_cast_fp16 = slice_by_index(begin = xr_13_begin_0, end = xr_13_end_0, end_mask = xr_13_end_mask_0, squeeze_mask = xr_13_squeeze_mask_0, x = pairs_13_cast_fp16)[name = tensor("xr_13_cast_fp16")]; + tensor xi_13_begin_0 = const()[name = tensor("xi_13_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_13_end_0 = const()[name = tensor("xi_13_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_13_end_mask_0 = const()[name = tensor("xi_13_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_13_squeeze_mask_0 = const()[name = tensor("xi_13_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_13_cast_fp16 = slice_by_index(begin = xi_13_begin_0, end = xi_13_end_0, end_mask = xi_13_end_mask_0, squeeze_mask = xi_13_squeeze_mask_0, x = pairs_13_cast_fp16)[name = tensor("xi_13_cast_fp16")]; + tensor var_856_cast_fp16 = mul(x = xr_13_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_856_cast_fp16")]; + tensor var_857_cast_fp16 = mul(x = xi_13_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_857_cast_fp16")]; + tensor xor__13_cast_fp16 = sub(x = var_856_cast_fp16, y = var_857_cast_fp16)[name = tensor("xor__13_cast_fp16")]; + tensor var_860_cast_fp16 = mul(x = xr_13_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_860_cast_fp16")]; + tensor var_861_cast_fp16 = mul(x = xi_13_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_861_cast_fp16")]; + tensor xoi_13_cast_fp16 = add(x = var_860_cast_fp16, y = var_861_cast_fp16)[name = tensor("xoi_13_cast_fp16")]; + tensor out_13_axis_0 = const()[name = tensor("out_13_axis_0"), val = tensor(-1)]; + tensor out_13_cast_fp16 = stack(axis = out_13_axis_0, values = (xor__13_cast_fp16, xoi_13_cast_fp16))[name = tensor("out_13_cast_fp16")]; + tensor var_884 = const()[name = tensor("op_884"), val = tensor([1, 16, 32, 2])]; + tensor pairs_15_cast_fp16 = reshape(shape = var_884, x = x_15_cast_fp16)[name = tensor("pairs_15_cast_fp16")]; + tensor xr_15_begin_0 = const()[name = tensor("xr_15_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_15_end_0 = const()[name = tensor("xr_15_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_15_end_mask_0 = const()[name = tensor("xr_15_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_15_squeeze_mask_0 = const()[name = tensor("xr_15_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_15_cast_fp16 = slice_by_index(begin = xr_15_begin_0, end = xr_15_end_0, end_mask = xr_15_end_mask_0, squeeze_mask = xr_15_squeeze_mask_0, x = pairs_15_cast_fp16)[name = tensor("xr_15_cast_fp16")]; + tensor xi_15_begin_0 = const()[name = tensor("xi_15_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_15_end_0 = const()[name = tensor("xi_15_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_15_end_mask_0 = const()[name = tensor("xi_15_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_15_squeeze_mask_0 = const()[name = tensor("xi_15_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_15_cast_fp16 = slice_by_index(begin = xi_15_begin_0, end = xi_15_end_0, end_mask = xi_15_end_mask_0, squeeze_mask = xi_15_squeeze_mask_0, x = pairs_15_cast_fp16)[name = tensor("xi_15_cast_fp16")]; + tensor var_892_cast_fp16 = mul(x = xr_15_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_892_cast_fp16")]; + tensor var_893_cast_fp16 = mul(x = xi_15_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_893_cast_fp16")]; + tensor xor__15_cast_fp16 = sub(x = var_892_cast_fp16, y = var_893_cast_fp16)[name = tensor("xor__15_cast_fp16")]; + tensor var_896_cast_fp16 = mul(x = xr_15_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_896_cast_fp16")]; + tensor var_897_cast_fp16 = mul(x = xi_15_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_897_cast_fp16")]; + tensor xoi_15_cast_fp16 = add(x = var_896_cast_fp16, y = var_897_cast_fp16)[name = tensor("xoi_15_cast_fp16")]; + tensor out_15_axis_0 = const()[name = tensor("out_15_axis_0"), val = tensor(-1)]; + tensor out_15_cast_fp16 = stack(axis = out_15_axis_0, values = (xor__15_cast_fp16, xoi_15_cast_fp16))[name = tensor("out_15_cast_fp16")]; + tensor _inversed_909_y_0_to_fp16 = const()[name = tensor("_inversed_909_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_909_cast_fp16 = mul(x = position3_to_fp16, y = _inversed_909_y_0_to_fp16)[name = tensor("_inversed_909_cast_fp16")]; + tensor var_910_cast_fp16 = floor(x = _inversed_909_cast_fp16)[name = tensor("op_910_cast_fp16")]; + tensor var_911_to_fp16 = const()[name = tensor("op_911_to_fp16"), val = tensor(0x1p+9)]; + tensor var_912_cast_fp16 = mul(x = var_910_cast_fp16, y = var_911_to_fp16)[name = tensor("op_912_cast_fp16")]; + tensor write_idx_7_cast_fp16 = sub(x = position3_to_fp16, y = var_912_cast_fp16)[name = tensor("write_idx_7_cast_fp16")]; + tensor var_915_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_7_cast_fp16)[name = tensor("op_915_cast_fp16")]; + tensor var_925 = const()[name = tensor("op_925"), val = tensor([1, 512, 1, 1])]; + tensor onehot_13_to_fp16_dtype_0 = const()[name = tensor("onehot_13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_915_cast_fp16_to_fp16 = cast(dtype = onehot_13_to_fp16_dtype_0, x = var_915_cast_fp16)[name = tensor("cast_120")]; + tensor onehot_15_cast_fp16 = reshape(shape = var_925, x = var_915_cast_fp16_to_fp16)[name = tensor("onehot_15_cast_fp16")]; + tensor var_927_to_fp16 = const()[name = tensor("op_927_to_fp16"), val = tensor(0x1p+0)]; + tensor var_929_cast_fp16 = sub(x = var_927_to_fp16, y = onehot_15_cast_fp16)[name = tensor("op_929_cast_fp16")]; + tensor k_cache3_to_fp16_dtype_0 = const()[name = tensor("k_cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache3_to_fp16 = cast(dtype = k_cache3_to_fp16_dtype_0, x = k_cache3)[name = tensor("cast_119")]; + tensor var_930_cast_fp16 = mul(x = k_cache3_to_fp16, y = var_929_cast_fp16)[name = tensor("op_930_cast_fp16")]; + tensor var_935 = const()[name = tensor("op_935"), val = tensor([1, 1, 16, 64])]; + tensor var_936_cast_fp16 = reshape(shape = var_935, x = out_15_cast_fp16)[name = tensor("op_936_cast_fp16")]; + tensor var_937_cast_fp16 = mul(x = var_936_cast_fp16, y = onehot_15_cast_fp16)[name = tensor("op_937_cast_fp16")]; + tensor new_k_cache_7_cast_fp16 = add(x = var_930_cast_fp16, y = var_937_cast_fp16)[name = tensor("new_k_cache_7_cast_fp16")]; + tensor new_k_cache_7_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_7_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache3_to_fp16_dtype_0 = const()[name = tensor("v_cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache3_to_fp16 = cast(dtype = v_cache3_to_fp16_dtype_0, x = v_cache3)[name = tensor("cast_117")]; + tensor var_943_cast_fp16 = mul(x = v_cache3_to_fp16, y = var_929_cast_fp16)[name = tensor("op_943_cast_fp16")]; + tensor var_948 = const()[name = tensor("op_948"), val = tensor([1, 1, 16, 64])]; + tensor var_949_cast_fp16 = reshape(shape = var_948, x = v_7_cast_fp16)[name = tensor("op_949_cast_fp16")]; + tensor var_950_cast_fp16 = mul(x = var_949_cast_fp16, y = onehot_15_cast_fp16)[name = tensor("op_950_cast_fp16")]; + tensor new_v_cache_7_cast_fp16 = add(x = var_943_cast_fp16, y = var_950_cast_fp16)[name = tensor("new_v_cache_7_cast_fp16")]; + tensor new_v_cache_7_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_7_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_3_perm_0 = const()[name = tensor("transpose_3_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_963 = const()[name = tensor("op_963"), val = tensor([0, 2, 1, 3])]; + tensor var_969 = const()[name = tensor("op_969"), val = tensor([1, 16, 1, 64])]; + tensor q4_7_cast_fp16 = reshape(shape = var_969, x = out_13_cast_fp16)[name = tensor("q4_7_cast_fp16")]; + tensor var_974_transpose_x_0 = const()[name = tensor("op_974_transpose_x_0"), val = tensor(false)]; + tensor var_974_transpose_y_0 = const()[name = tensor("op_974_transpose_y_0"), val = tensor(false)]; + tensor transpose_3_cast_fp16 = transpose(perm = transpose_3_perm_0, x = new_k_cache_7_cast_fp16)[name = tensor("transpose_14")]; + tensor var_974_cast_fp16 = matmul(transpose_x = var_974_transpose_x_0, transpose_y = var_974_transpose_y_0, x = q4_7_cast_fp16, y = transpose_3_cast_fp16)[name = tensor("op_974_cast_fp16")]; + tensor var_975_to_fp16 = const()[name = tensor("op_975_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_13_cast_fp16 = mul(x = var_974_cast_fp16, y = var_975_to_fp16)[name = tensor("scores_13_cast_fp16")]; + tensor var_979 = const()[name = tensor("op_979"), val = tensor([1, 1])]; + tensor var_980_cast_fp16 = reshape(shape = var_979, x = position3_to_fp16)[name = tensor("op_980_cast_fp16")]; + tensor var_981_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_980_cast_fp16)[name = tensor("op_981_cast_fp16")]; + tensor mask_7_to_fp16_dtype_0 = const()[name = tensor("mask_7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_988_to_fp16 = const()[name = tensor("op_988_to_fp16"), val = tensor(0x1p+0)]; + tensor var_981_cast_fp16_to_fp16 = cast(dtype = mask_7_to_fp16_dtype_0, x = var_981_cast_fp16)[name = tensor("cast_115")]; + tensor var_989_cast_fp16 = sub(x = var_981_cast_fp16_to_fp16, y = var_988_to_fp16)[name = tensor("op_989_cast_fp16")]; + tensor var_994 = const()[name = tensor("op_994"), val = tensor([1, 1, 1, 512])]; + tensor var_995_cast_fp16 = reshape(shape = var_994, x = var_989_cast_fp16)[name = tensor("op_995_cast_fp16")]; + tensor var_996_to_fp16 = const()[name = tensor("op_996_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_997_cast_fp16 = mul(x = var_995_cast_fp16, y = var_996_to_fp16)[name = tensor("op_997_cast_fp16")]; + tensor scores_15_cast_fp16 = add(x = scores_13_cast_fp16, y = var_997_cast_fp16)[name = tensor("scores_15_cast_fp16")]; + tensor var_1000 = const()[name = tensor("op_1000"), val = tensor(-1)]; + tensor weights_7_cast_fp16 = softmax(axis = var_1000, x = scores_15_cast_fp16)[name = tensor("weights_7_cast_fp16")]; + tensor attn_7_transpose_x_0 = const()[name = tensor("attn_7_transpose_x_0"), val = tensor(false)]; + tensor attn_7_transpose_y_0 = const()[name = tensor("attn_7_transpose_y_0"), val = tensor(false)]; + tensor values_7_cast_fp16 = transpose(perm = var_963, x = new_v_cache_7_cast_fp16)[name = tensor("transpose_13")]; + tensor attn_7_cast_fp16 = matmul(transpose_x = attn_7_transpose_x_0, transpose_y = attn_7_transpose_y_0, x = weights_7_cast_fp16, y = values_7_cast_fp16)[name = tensor("attn_7_cast_fp16")]; + tensor var_1008 = const()[name = tensor("op_1008"), val = tensor([0, 2, 1, 3])]; + tensor var_1013 = const()[name = tensor("op_1013"), val = tensor([1, 1, 1024])]; + tensor var_1009_cast_fp16 = transpose(perm = var_1008, x = attn_7_cast_fp16)[name = tensor("transpose_12")]; + tensor input_43_cast_fp16 = reshape(shape = var_1013, x = var_1009_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor attn3_out_proj_weight_to_fp16 = const()[name = tensor("attn3_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(81902784)))]; + tensor linear_14_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn3_out_proj_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor var_1019_to_fp16 = const()[name = tensor("op_1019_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1020_cast_fp16 = add(x = position3_to_fp16, y = var_1019_to_fp16)[name = tensor("op_1020_cast_fp16")]; + tensor var_1020_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1020_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_45_cast_fp16 = add(x = input_39_cast_fp16, y = linear_14_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor input_47_axes_0 = const()[name = tensor("input_47_axes_0"), val = tensor([-1])]; + tensor norm3_2_weight_to_fp16 = const()[name = tensor("norm3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84000000)))]; + tensor norm3_2_bias_to_fp16 = const()[name = tensor("norm3_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84002112)))]; + tensor var_1024_to_fp16 = const()[name = tensor("op_1024_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_47_cast_fp16 = layer_norm(axes = input_47_axes_0, beta = norm3_2_bias_to_fp16, epsilon = var_1024_to_fp16, gamma = norm3_2_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor linear3_1_weight_to_fp16 = const()[name = tensor("linear3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84004224)))]; + tensor linear_15_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear3_1_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor input_49_mode_0 = const()[name = tensor("input_49_mode_0"), val = tensor("EXACT")]; + tensor input_49_cast_fp16 = gelu(mode = input_49_mode_0, x = linear_15_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor linear3_2_weight_to_fp16 = const()[name = tensor("linear3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(92392896)))]; + tensor linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear3_2_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor input_51_cast_fp16 = add(x = input_45_cast_fp16, y = linear_16_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor input_53_axes_0 = const()[name = tensor("input_53_axes_0"), val = tensor([-1])]; + tensor norm4_1_weight_to_fp16 = const()[name = tensor("norm4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100781568)))]; + tensor norm4_1_bias_to_fp16 = const()[name = tensor("norm4_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100783680)))]; + tensor var_1041_to_fp16 = const()[name = tensor("op_1041_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_53_cast_fp16 = layer_norm(axes = input_53_axes_0, beta = norm4_1_bias_to_fp16, epsilon = var_1041_to_fp16, gamma = norm4_1_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor attn4_in_proj_weight_to_fp16 = const()[name = tensor("attn4_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100785792)))]; + tensor linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn4_in_proj_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor var_1054 = const()[name = tensor("op_1054"), val = tensor([1, 3, 16, 64])]; + tensor qkv_9_cast_fp16 = reshape(shape = var_1054, x = linear_17_cast_fp16)[name = tensor("qkv_9_cast_fp16")]; + tensor x_17_begin_0 = const()[name = tensor("x_17_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_17_end_0 = const()[name = tensor("x_17_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_17_end_mask_0 = const()[name = tensor("x_17_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_17_squeeze_mask_0 = const()[name = tensor("x_17_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_17_cast_fp16 = slice_by_index(begin = x_17_begin_0, end = x_17_end_0, end_mask = x_17_end_mask_0, squeeze_mask = x_17_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("x_17_cast_fp16")]; + tensor x_19_begin_0 = const()[name = tensor("x_19_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_19_end_0 = const()[name = tensor("x_19_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_19_end_mask_0 = const()[name = tensor("x_19_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_19_squeeze_mask_0 = const()[name = tensor("x_19_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_19_cast_fp16 = slice_by_index(begin = x_19_begin_0, end = x_19_end_0, end_mask = x_19_end_mask_0, squeeze_mask = x_19_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("x_19_cast_fp16")]; + tensor v_9_begin_0 = const()[name = tensor("v_9_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_9_end_0 = const()[name = tensor("v_9_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_9_end_mask_0 = const()[name = tensor("v_9_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_9_squeeze_mask_0 = const()[name = tensor("v_9_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_9_cast_fp16 = slice_by_index(begin = v_9_begin_0, end = v_9_end_0, end_mask = v_9_end_mask_0, squeeze_mask = v_9_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("v_9_cast_fp16")]; + tensor var_1083 = const()[name = tensor("op_1083"), val = tensor([1, 1, 1])]; + tensor position4_to_fp16_dtype_0 = const()[name = tensor("position4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position4_to_fp16 = cast(dtype = position4_to_fp16_dtype_0, x = position4)[name = tensor("cast_113")]; + tensor var_1084_cast_fp16 = reshape(shape = var_1083, x = position4_to_fp16)[name = tensor("op_1084_cast_fp16")]; + tensor angles_17_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_1084_cast_fp16)[name = tensor("angles_17_cast_fp16")]; + tensor rotr_17_cast_fp16 = cos(x = angles_17_cast_fp16)[name = tensor("rotr_17_cast_fp16")]; + tensor roti_17_cast_fp16 = sin(x = angles_17_cast_fp16)[name = tensor("roti_17_cast_fp16")]; + tensor var_1092 = const()[name = tensor("op_1092"), val = tensor([1, 16, 32, 2])]; + tensor pairs_17_cast_fp16 = reshape(shape = var_1092, x = x_17_cast_fp16)[name = tensor("pairs_17_cast_fp16")]; + tensor xr_17_begin_0 = const()[name = tensor("xr_17_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_17_end_0 = const()[name = tensor("xr_17_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_17_end_mask_0 = const()[name = tensor("xr_17_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_17_squeeze_mask_0 = const()[name = tensor("xr_17_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_17_cast_fp16 = slice_by_index(begin = xr_17_begin_0, end = xr_17_end_0, end_mask = xr_17_end_mask_0, squeeze_mask = xr_17_squeeze_mask_0, x = pairs_17_cast_fp16)[name = tensor("xr_17_cast_fp16")]; + tensor xi_17_begin_0 = const()[name = tensor("xi_17_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_17_end_0 = const()[name = tensor("xi_17_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_17_end_mask_0 = const()[name = tensor("xi_17_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_17_squeeze_mask_0 = const()[name = tensor("xi_17_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_17_cast_fp16 = slice_by_index(begin = xi_17_begin_0, end = xi_17_end_0, end_mask = xi_17_end_mask_0, squeeze_mask = xi_17_squeeze_mask_0, x = pairs_17_cast_fp16)[name = tensor("xi_17_cast_fp16")]; + tensor var_1100_cast_fp16 = mul(x = xr_17_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_1100_cast_fp16")]; + tensor var_1101_cast_fp16 = mul(x = xi_17_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_1101_cast_fp16")]; + tensor xor__17_cast_fp16 = sub(x = var_1100_cast_fp16, y = var_1101_cast_fp16)[name = tensor("xor__17_cast_fp16")]; + tensor var_1104_cast_fp16 = mul(x = xr_17_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_1104_cast_fp16")]; + tensor var_1105_cast_fp16 = mul(x = xi_17_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_1105_cast_fp16")]; + tensor xoi_17_cast_fp16 = add(x = var_1104_cast_fp16, y = var_1105_cast_fp16)[name = tensor("xoi_17_cast_fp16")]; + tensor out_17_axis_0 = const()[name = tensor("out_17_axis_0"), val = tensor(-1)]; + tensor out_17_cast_fp16 = stack(axis = out_17_axis_0, values = (xor__17_cast_fp16, xoi_17_cast_fp16))[name = tensor("out_17_cast_fp16")]; + tensor var_1128 = const()[name = tensor("op_1128"), val = tensor([1, 16, 32, 2])]; + tensor pairs_19_cast_fp16 = reshape(shape = var_1128, x = x_19_cast_fp16)[name = tensor("pairs_19_cast_fp16")]; + tensor xr_19_begin_0 = const()[name = tensor("xr_19_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_19_end_0 = const()[name = tensor("xr_19_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_19_end_mask_0 = const()[name = tensor("xr_19_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_19_squeeze_mask_0 = const()[name = tensor("xr_19_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_19_cast_fp16 = slice_by_index(begin = xr_19_begin_0, end = xr_19_end_0, end_mask = xr_19_end_mask_0, squeeze_mask = xr_19_squeeze_mask_0, x = pairs_19_cast_fp16)[name = tensor("xr_19_cast_fp16")]; + tensor xi_19_begin_0 = const()[name = tensor("xi_19_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_19_end_0 = const()[name = tensor("xi_19_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_19_end_mask_0 = const()[name = tensor("xi_19_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_19_squeeze_mask_0 = const()[name = tensor("xi_19_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_19_cast_fp16 = slice_by_index(begin = xi_19_begin_0, end = xi_19_end_0, end_mask = xi_19_end_mask_0, squeeze_mask = xi_19_squeeze_mask_0, x = pairs_19_cast_fp16)[name = tensor("xi_19_cast_fp16")]; + tensor var_1136_cast_fp16 = mul(x = xr_19_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_1136_cast_fp16")]; + tensor var_1137_cast_fp16 = mul(x = xi_19_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_1137_cast_fp16")]; + tensor xor__19_cast_fp16 = sub(x = var_1136_cast_fp16, y = var_1137_cast_fp16)[name = tensor("xor__19_cast_fp16")]; + tensor var_1140_cast_fp16 = mul(x = xr_19_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_1140_cast_fp16")]; + tensor var_1141_cast_fp16 = mul(x = xi_19_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_1141_cast_fp16")]; + tensor xoi_19_cast_fp16 = add(x = var_1140_cast_fp16, y = var_1141_cast_fp16)[name = tensor("xoi_19_cast_fp16")]; + tensor out_19_axis_0 = const()[name = tensor("out_19_axis_0"), val = tensor(-1)]; + tensor out_19_cast_fp16 = stack(axis = out_19_axis_0, values = (xor__19_cast_fp16, xoi_19_cast_fp16))[name = tensor("out_19_cast_fp16")]; + tensor _inversed_1153_y_0_to_fp16 = const()[name = tensor("_inversed_1153_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1153_cast_fp16 = mul(x = position4_to_fp16, y = _inversed_1153_y_0_to_fp16)[name = tensor("_inversed_1153_cast_fp16")]; + tensor var_1154_cast_fp16 = floor(x = _inversed_1153_cast_fp16)[name = tensor("op_1154_cast_fp16")]; + tensor var_1155_to_fp16 = const()[name = tensor("op_1155_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1156_cast_fp16 = mul(x = var_1154_cast_fp16, y = var_1155_to_fp16)[name = tensor("op_1156_cast_fp16")]; + tensor write_idx_9_cast_fp16 = sub(x = position4_to_fp16, y = var_1156_cast_fp16)[name = tensor("write_idx_9_cast_fp16")]; + tensor var_1159_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_9_cast_fp16)[name = tensor("op_1159_cast_fp16")]; + tensor var_1169 = const()[name = tensor("op_1169"), val = tensor([1, 512, 1, 1])]; + tensor onehot_17_to_fp16_dtype_0 = const()[name = tensor("onehot_17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_1159_cast_fp16_to_fp16 = cast(dtype = onehot_17_to_fp16_dtype_0, x = var_1159_cast_fp16)[name = tensor("cast_112")]; + tensor onehot_19_cast_fp16 = reshape(shape = var_1169, x = var_1159_cast_fp16_to_fp16)[name = tensor("onehot_19_cast_fp16")]; + tensor var_1171_to_fp16 = const()[name = tensor("op_1171_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1173_cast_fp16 = sub(x = var_1171_to_fp16, y = onehot_19_cast_fp16)[name = tensor("op_1173_cast_fp16")]; + tensor k_cache4_to_fp16_dtype_0 = const()[name = tensor("k_cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache4_to_fp16 = cast(dtype = k_cache4_to_fp16_dtype_0, x = k_cache4)[name = tensor("cast_111")]; + tensor var_1174_cast_fp16 = mul(x = k_cache4_to_fp16, y = var_1173_cast_fp16)[name = tensor("op_1174_cast_fp16")]; + tensor var_1179 = const()[name = tensor("op_1179"), val = tensor([1, 1, 16, 64])]; + tensor var_1180_cast_fp16 = reshape(shape = var_1179, x = out_19_cast_fp16)[name = tensor("op_1180_cast_fp16")]; + tensor var_1181_cast_fp16 = mul(x = var_1180_cast_fp16, y = onehot_19_cast_fp16)[name = tensor("op_1181_cast_fp16")]; + tensor new_k_cache_9_cast_fp16 = add(x = var_1174_cast_fp16, y = var_1181_cast_fp16)[name = tensor("new_k_cache_9_cast_fp16")]; + tensor new_k_cache_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache4_to_fp16_dtype_0 = const()[name = tensor("v_cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache4_to_fp16 = cast(dtype = v_cache4_to_fp16_dtype_0, x = v_cache4)[name = tensor("cast_109")]; + tensor var_1187_cast_fp16 = mul(x = v_cache4_to_fp16, y = var_1173_cast_fp16)[name = tensor("op_1187_cast_fp16")]; + tensor var_1192 = const()[name = tensor("op_1192"), val = tensor([1, 1, 16, 64])]; + tensor var_1193_cast_fp16 = reshape(shape = var_1192, x = v_9_cast_fp16)[name = tensor("op_1193_cast_fp16")]; + tensor var_1194_cast_fp16 = mul(x = var_1193_cast_fp16, y = onehot_19_cast_fp16)[name = tensor("op_1194_cast_fp16")]; + tensor new_v_cache_9_cast_fp16 = add(x = var_1187_cast_fp16, y = var_1194_cast_fp16)[name = tensor("new_v_cache_9_cast_fp16")]; + tensor new_v_cache_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_4_perm_0 = const()[name = tensor("transpose_4_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_1207 = const()[name = tensor("op_1207"), val = tensor([0, 2, 1, 3])]; + tensor var_1213 = const()[name = tensor("op_1213"), val = tensor([1, 16, 1, 64])]; + tensor q4_9_cast_fp16 = reshape(shape = var_1213, x = out_17_cast_fp16)[name = tensor("q4_9_cast_fp16")]; + tensor var_1218_transpose_x_0 = const()[name = tensor("op_1218_transpose_x_0"), val = tensor(false)]; + tensor var_1218_transpose_y_0 = const()[name = tensor("op_1218_transpose_y_0"), val = tensor(false)]; + tensor transpose_4_cast_fp16 = transpose(perm = transpose_4_perm_0, x = new_k_cache_9_cast_fp16)[name = tensor("transpose_11")]; + tensor var_1218_cast_fp16 = matmul(transpose_x = var_1218_transpose_x_0, transpose_y = var_1218_transpose_y_0, x = q4_9_cast_fp16, y = transpose_4_cast_fp16)[name = tensor("op_1218_cast_fp16")]; + tensor var_1219_to_fp16 = const()[name = tensor("op_1219_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_17_cast_fp16 = mul(x = var_1218_cast_fp16, y = var_1219_to_fp16)[name = tensor("scores_17_cast_fp16")]; + tensor var_1223 = const()[name = tensor("op_1223"), val = tensor([1, 1])]; + tensor var_1224_cast_fp16 = reshape(shape = var_1223, x = position4_to_fp16)[name = tensor("op_1224_cast_fp16")]; + tensor var_1225_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_1224_cast_fp16)[name = tensor("op_1225_cast_fp16")]; + tensor mask_9_to_fp16_dtype_0 = const()[name = tensor("mask_9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_1232_to_fp16 = const()[name = tensor("op_1232_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1225_cast_fp16_to_fp16 = cast(dtype = mask_9_to_fp16_dtype_0, x = var_1225_cast_fp16)[name = tensor("cast_107")]; + tensor var_1233_cast_fp16 = sub(x = var_1225_cast_fp16_to_fp16, y = var_1232_to_fp16)[name = tensor("op_1233_cast_fp16")]; + tensor var_1238 = const()[name = tensor("op_1238"), val = tensor([1, 1, 1, 512])]; + tensor var_1239_cast_fp16 = reshape(shape = var_1238, x = var_1233_cast_fp16)[name = tensor("op_1239_cast_fp16")]; + tensor var_1240_to_fp16 = const()[name = tensor("op_1240_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_1241_cast_fp16 = mul(x = var_1239_cast_fp16, y = var_1240_to_fp16)[name = tensor("op_1241_cast_fp16")]; + tensor scores_19_cast_fp16 = add(x = scores_17_cast_fp16, y = var_1241_cast_fp16)[name = tensor("scores_19_cast_fp16")]; + tensor var_1244 = const()[name = tensor("op_1244"), val = tensor(-1)]; + tensor weights_9_cast_fp16 = softmax(axis = var_1244, x = scores_19_cast_fp16)[name = tensor("weights_9_cast_fp16")]; + tensor attn_9_transpose_x_0 = const()[name = tensor("attn_9_transpose_x_0"), val = tensor(false)]; + tensor attn_9_transpose_y_0 = const()[name = tensor("attn_9_transpose_y_0"), val = tensor(false)]; + tensor values_9_cast_fp16 = transpose(perm = var_1207, x = new_v_cache_9_cast_fp16)[name = tensor("transpose_10")]; + tensor attn_9_cast_fp16 = matmul(transpose_x = attn_9_transpose_x_0, transpose_y = attn_9_transpose_y_0, x = weights_9_cast_fp16, y = values_9_cast_fp16)[name = tensor("attn_9_cast_fp16")]; + tensor var_1252 = const()[name = tensor("op_1252"), val = tensor([0, 2, 1, 3])]; + tensor var_1257 = const()[name = tensor("op_1257"), val = tensor([1, 1, 1024])]; + tensor var_1253_cast_fp16 = transpose(perm = var_1252, x = attn_9_cast_fp16)[name = tensor("transpose_9")]; + tensor input_55_cast_fp16 = reshape(shape = var_1257, x = var_1253_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor attn4_out_proj_weight_to_fp16 = const()[name = tensor("attn4_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(107077312)))]; + tensor linear_18_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn4_out_proj_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor var_1263_to_fp16 = const()[name = tensor("op_1263_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1264_cast_fp16 = add(x = position4_to_fp16, y = var_1263_to_fp16)[name = tensor("op_1264_cast_fp16")]; + tensor var_1264_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1264_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_57_cast_fp16 = add(x = input_51_cast_fp16, y = linear_18_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor input_59_axes_0 = const()[name = tensor("input_59_axes_0"), val = tensor([-1])]; + tensor norm4_2_weight_to_fp16 = const()[name = tensor("norm4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109174528)))]; + tensor norm4_2_bias_to_fp16 = const()[name = tensor("norm4_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109176640)))]; + tensor var_1268_to_fp16 = const()[name = tensor("op_1268_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_59_cast_fp16 = layer_norm(axes = input_59_axes_0, beta = norm4_2_bias_to_fp16, epsilon = var_1268_to_fp16, gamma = norm4_2_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor linear4_1_weight_to_fp16 = const()[name = tensor("linear4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109178752)))]; + tensor linear_19_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear4_1_weight_to_fp16, x = input_59_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_61_mode_0 = const()[name = tensor("input_61_mode_0"), val = tensor("EXACT")]; + tensor input_61_cast_fp16 = gelu(mode = input_61_mode_0, x = linear_19_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor linear4_2_weight_to_fp16 = const()[name = tensor("linear4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(117567424)))]; + tensor linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear4_2_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor input_63_cast_fp16 = add(x = input_57_cast_fp16, y = linear_20_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor input_65_axes_0 = const()[name = tensor("input_65_axes_0"), val = tensor([-1])]; + tensor norm5_1_weight_to_fp16 = const()[name = tensor("norm5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125956096)))]; + tensor norm5_1_bias_to_fp16 = const()[name = tensor("norm5_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125958208)))]; + tensor var_1285_to_fp16 = const()[name = tensor("op_1285_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_65_cast_fp16 = layer_norm(axes = input_65_axes_0, beta = norm5_1_bias_to_fp16, epsilon = var_1285_to_fp16, gamma = norm5_1_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("input_65_cast_fp16")]; + tensor attn5_in_proj_weight_to_fp16 = const()[name = tensor("attn5_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125960320)))]; + tensor linear_21_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn5_in_proj_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("linear_21_cast_fp16")]; + tensor var_1298 = const()[name = tensor("op_1298"), val = tensor([1, 3, 16, 64])]; + tensor qkv_cast_fp16 = reshape(shape = var_1298, x = linear_21_cast_fp16)[name = tensor("qkv_cast_fp16")]; + tensor x_21_begin_0 = const()[name = tensor("x_21_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_21_end_0 = const()[name = tensor("x_21_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_21_end_mask_0 = const()[name = tensor("x_21_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_21_squeeze_mask_0 = const()[name = tensor("x_21_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_21_cast_fp16 = slice_by_index(begin = x_21_begin_0, end = x_21_end_0, end_mask = x_21_end_mask_0, squeeze_mask = x_21_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("x_21_cast_fp16")]; + tensor x_begin_0 = const()[name = tensor("x_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_end_0 = const()[name = tensor("x_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_end_mask_0 = const()[name = tensor("x_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_squeeze_mask_0 = const()[name = tensor("x_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_cast_fp16 = slice_by_index(begin = x_begin_0, end = x_end_0, end_mask = x_end_mask_0, squeeze_mask = x_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("x_cast_fp16")]; + tensor v_begin_0 = const()[name = tensor("v_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_end_0 = const()[name = tensor("v_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_end_mask_0 = const()[name = tensor("v_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_squeeze_mask_0 = const()[name = tensor("v_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_cast_fp16 = slice_by_index(begin = v_begin_0, end = v_end_0, end_mask = v_end_mask_0, squeeze_mask = v_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("v_cast_fp16")]; + tensor var_1327 = const()[name = tensor("op_1327"), val = tensor([1, 1, 1])]; + tensor position5_to_fp16_dtype_0 = const()[name = tensor("position5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position5_to_fp16 = cast(dtype = position5_to_fp16_dtype_0, x = position5)[name = tensor("cast_105")]; + tensor var_1328_cast_fp16 = reshape(shape = var_1327, x = position5_to_fp16)[name = tensor("op_1328_cast_fp16")]; + tensor angles_21_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_1328_cast_fp16)[name = tensor("angles_21_cast_fp16")]; + tensor rotr_21_cast_fp16 = cos(x = angles_21_cast_fp16)[name = tensor("rotr_21_cast_fp16")]; + tensor roti_21_cast_fp16 = sin(x = angles_21_cast_fp16)[name = tensor("roti_21_cast_fp16")]; + tensor var_1336 = const()[name = tensor("op_1336"), val = tensor([1, 16, 32, 2])]; + tensor pairs_21_cast_fp16 = reshape(shape = var_1336, x = x_21_cast_fp16)[name = tensor("pairs_21_cast_fp16")]; + tensor xr_21_begin_0 = const()[name = tensor("xr_21_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_21_end_0 = const()[name = tensor("xr_21_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_21_end_mask_0 = const()[name = tensor("xr_21_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_21_squeeze_mask_0 = const()[name = tensor("xr_21_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_21_cast_fp16 = slice_by_index(begin = xr_21_begin_0, end = xr_21_end_0, end_mask = xr_21_end_mask_0, squeeze_mask = xr_21_squeeze_mask_0, x = pairs_21_cast_fp16)[name = tensor("xr_21_cast_fp16")]; + tensor xi_21_begin_0 = const()[name = tensor("xi_21_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_21_end_0 = const()[name = tensor("xi_21_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_21_end_mask_0 = const()[name = tensor("xi_21_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_21_squeeze_mask_0 = const()[name = tensor("xi_21_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_21_cast_fp16 = slice_by_index(begin = xi_21_begin_0, end = xi_21_end_0, end_mask = xi_21_end_mask_0, squeeze_mask = xi_21_squeeze_mask_0, x = pairs_21_cast_fp16)[name = tensor("xi_21_cast_fp16")]; + tensor var_1344_cast_fp16 = mul(x = xr_21_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_1344_cast_fp16")]; + tensor var_1345_cast_fp16 = mul(x = xi_21_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_1345_cast_fp16")]; + tensor xor__21_cast_fp16 = sub(x = var_1344_cast_fp16, y = var_1345_cast_fp16)[name = tensor("xor__21_cast_fp16")]; + tensor var_1348_cast_fp16 = mul(x = xr_21_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_1348_cast_fp16")]; + tensor var_1349_cast_fp16 = mul(x = xi_21_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_1349_cast_fp16")]; + tensor xoi_21_cast_fp16 = add(x = var_1348_cast_fp16, y = var_1349_cast_fp16)[name = tensor("xoi_21_cast_fp16")]; + tensor out_21_axis_0 = const()[name = tensor("out_21_axis_0"), val = tensor(-1)]; + tensor out_21_cast_fp16 = stack(axis = out_21_axis_0, values = (xor__21_cast_fp16, xoi_21_cast_fp16))[name = tensor("out_21_cast_fp16")]; + tensor var_1372 = const()[name = tensor("op_1372"), val = tensor([1, 16, 32, 2])]; + tensor pairs_cast_fp16 = reshape(shape = var_1372, x = x_cast_fp16)[name = tensor("pairs_cast_fp16")]; + tensor xr_begin_0 = const()[name = tensor("xr_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_end_0 = const()[name = tensor("xr_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_end_mask_0 = const()[name = tensor("xr_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_squeeze_mask_0 = const()[name = tensor("xr_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_cast_fp16 = slice_by_index(begin = xr_begin_0, end = xr_end_0, end_mask = xr_end_mask_0, squeeze_mask = xr_squeeze_mask_0, x = pairs_cast_fp16)[name = tensor("xr_cast_fp16")]; + tensor xi_begin_0 = const()[name = tensor("xi_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_end_0 = const()[name = tensor("xi_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_end_mask_0 = const()[name = tensor("xi_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_squeeze_mask_0 = const()[name = tensor("xi_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_cast_fp16 = slice_by_index(begin = xi_begin_0, end = xi_end_0, end_mask = xi_end_mask_0, squeeze_mask = xi_squeeze_mask_0, x = pairs_cast_fp16)[name = tensor("xi_cast_fp16")]; + tensor var_1380_cast_fp16 = mul(x = xr_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_1380_cast_fp16")]; + tensor var_1381_cast_fp16 = mul(x = xi_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_1381_cast_fp16")]; + tensor xor__cast_fp16 = sub(x = var_1380_cast_fp16, y = var_1381_cast_fp16)[name = tensor("xor__cast_fp16")]; + tensor var_1384_cast_fp16 = mul(x = xr_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_1384_cast_fp16")]; + tensor var_1385_cast_fp16 = mul(x = xi_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_1385_cast_fp16")]; + tensor xoi_cast_fp16 = add(x = var_1384_cast_fp16, y = var_1385_cast_fp16)[name = tensor("xoi_cast_fp16")]; + tensor out_axis_0 = const()[name = tensor("out_axis_0"), val = tensor(-1)]; + tensor out_cast_fp16 = stack(axis = out_axis_0, values = (xor__cast_fp16, xoi_cast_fp16))[name = tensor("out_cast_fp16")]; + tensor _inversed_1397_y_0_to_fp16 = const()[name = tensor("_inversed_1397_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1397_cast_fp16 = mul(x = position5_to_fp16, y = _inversed_1397_y_0_to_fp16)[name = tensor("_inversed_1397_cast_fp16")]; + tensor var_1398_cast_fp16 = floor(x = _inversed_1397_cast_fp16)[name = tensor("op_1398_cast_fp16")]; + tensor var_1399_to_fp16 = const()[name = tensor("op_1399_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1400_cast_fp16 = mul(x = var_1398_cast_fp16, y = var_1399_to_fp16)[name = tensor("op_1400_cast_fp16")]; + tensor write_idx_cast_fp16 = sub(x = position5_to_fp16, y = var_1400_cast_fp16)[name = tensor("write_idx_cast_fp16")]; + tensor var_1403_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_cast_fp16)[name = tensor("op_1403_cast_fp16")]; + tensor var_1413 = const()[name = tensor("op_1413"), val = tensor([1, 512, 1, 1])]; + tensor onehot_21_to_fp16_dtype_0 = const()[name = tensor("onehot_21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_1403_cast_fp16_to_fp16 = cast(dtype = onehot_21_to_fp16_dtype_0, x = var_1403_cast_fp16)[name = tensor("cast_104")]; + tensor onehot_cast_fp16 = reshape(shape = var_1413, x = var_1403_cast_fp16_to_fp16)[name = tensor("onehot_cast_fp16")]; + tensor var_1415_to_fp16 = const()[name = tensor("op_1415_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1417_cast_fp16 = sub(x = var_1415_to_fp16, y = onehot_cast_fp16)[name = tensor("op_1417_cast_fp16")]; + tensor k_cache5_to_fp16_dtype_0 = const()[name = tensor("k_cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache5_to_fp16 = cast(dtype = k_cache5_to_fp16_dtype_0, x = k_cache5)[name = tensor("cast_103")]; + tensor var_1418_cast_fp16 = mul(x = k_cache5_to_fp16, y = var_1417_cast_fp16)[name = tensor("op_1418_cast_fp16")]; + tensor var_1423 = const()[name = tensor("op_1423"), val = tensor([1, 1, 16, 64])]; + tensor var_1424_cast_fp16 = reshape(shape = var_1423, x = out_cast_fp16)[name = tensor("op_1424_cast_fp16")]; + tensor var_1425_cast_fp16 = mul(x = var_1424_cast_fp16, y = onehot_cast_fp16)[name = tensor("op_1425_cast_fp16")]; + tensor new_k_cache_cast_fp16 = add(x = var_1418_cast_fp16, y = var_1425_cast_fp16)[name = tensor("new_k_cache_cast_fp16")]; + tensor new_k_cache_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache5_to_fp16_dtype_0 = const()[name = tensor("v_cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache5_to_fp16 = cast(dtype = v_cache5_to_fp16_dtype_0, x = v_cache5)[name = tensor("cast_101")]; + tensor var_1431_cast_fp16 = mul(x = v_cache5_to_fp16, y = var_1417_cast_fp16)[name = tensor("op_1431_cast_fp16")]; + tensor var_1436 = const()[name = tensor("op_1436"), val = tensor([1, 1, 16, 64])]; + tensor var_1437_cast_fp16 = reshape(shape = var_1436, x = v_cast_fp16)[name = tensor("op_1437_cast_fp16")]; + tensor var_1438_cast_fp16 = mul(x = var_1437_cast_fp16, y = onehot_cast_fp16)[name = tensor("op_1438_cast_fp16")]; + tensor new_v_cache_cast_fp16 = add(x = var_1431_cast_fp16, y = var_1438_cast_fp16)[name = tensor("new_v_cache_cast_fp16")]; + tensor new_v_cache_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_5_perm_0 = const()[name = tensor("transpose_5_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_1451 = const()[name = tensor("op_1451"), val = tensor([0, 2, 1, 3])]; + tensor var_1457 = const()[name = tensor("op_1457"), val = tensor([1, 16, 1, 64])]; + tensor q4_cast_fp16 = reshape(shape = var_1457, x = out_21_cast_fp16)[name = tensor("q4_cast_fp16")]; + tensor var_1462_transpose_x_0 = const()[name = tensor("op_1462_transpose_x_0"), val = tensor(false)]; + tensor var_1462_transpose_y_0 = const()[name = tensor("op_1462_transpose_y_0"), val = tensor(false)]; + tensor transpose_5_cast_fp16 = transpose(perm = transpose_5_perm_0, x = new_k_cache_cast_fp16)[name = tensor("transpose_8")]; + tensor var_1462_cast_fp16 = matmul(transpose_x = var_1462_transpose_x_0, transpose_y = var_1462_transpose_y_0, x = q4_cast_fp16, y = transpose_5_cast_fp16)[name = tensor("op_1462_cast_fp16")]; + tensor var_1463_to_fp16 = const()[name = tensor("op_1463_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_21_cast_fp16 = mul(x = var_1462_cast_fp16, y = var_1463_to_fp16)[name = tensor("scores_21_cast_fp16")]; + tensor var_1467 = const()[name = tensor("op_1467"), val = tensor([1, 1])]; + tensor var_1468_cast_fp16 = reshape(shape = var_1467, x = position5_to_fp16)[name = tensor("op_1468_cast_fp16")]; + tensor var_1469_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_1468_cast_fp16)[name = tensor("op_1469_cast_fp16")]; + tensor mask_to_fp16_dtype_0 = const()[name = tensor("mask_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_1476_to_fp16 = const()[name = tensor("op_1476_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1469_cast_fp16_to_fp16 = cast(dtype = mask_to_fp16_dtype_0, x = var_1469_cast_fp16)[name = tensor("cast_99")]; + tensor var_1477_cast_fp16 = sub(x = var_1469_cast_fp16_to_fp16, y = var_1476_to_fp16)[name = tensor("op_1477_cast_fp16")]; + tensor var_1482 = const()[name = tensor("op_1482"), val = tensor([1, 1, 1, 512])]; + tensor var_1483_cast_fp16 = reshape(shape = var_1482, x = var_1477_cast_fp16)[name = tensor("op_1483_cast_fp16")]; + tensor var_1484_to_fp16 = const()[name = tensor("op_1484_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_1485_cast_fp16 = mul(x = var_1483_cast_fp16, y = var_1484_to_fp16)[name = tensor("op_1485_cast_fp16")]; + tensor scores_cast_fp16 = add(x = scores_21_cast_fp16, y = var_1485_cast_fp16)[name = tensor("scores_cast_fp16")]; + tensor var_1488 = const()[name = tensor("op_1488"), val = tensor(-1)]; + tensor weights_cast_fp16 = softmax(axis = var_1488, x = scores_cast_fp16)[name = tensor("weights_cast_fp16")]; + tensor attn_transpose_x_0 = const()[name = tensor("attn_transpose_x_0"), val = tensor(false)]; + tensor attn_transpose_y_0 = const()[name = tensor("attn_transpose_y_0"), val = tensor(false)]; + tensor values_cast_fp16 = transpose(perm = var_1451, x = new_v_cache_cast_fp16)[name = tensor("transpose_7")]; + tensor attn_cast_fp16 = matmul(transpose_x = attn_transpose_x_0, transpose_y = attn_transpose_y_0, x = weights_cast_fp16, y = values_cast_fp16)[name = tensor("attn_cast_fp16")]; + tensor var_1496 = const()[name = tensor("op_1496"), val = tensor([0, 2, 1, 3])]; + tensor var_1501 = const()[name = tensor("op_1501"), val = tensor([1, 1, 1024])]; + tensor var_1497_cast_fp16 = transpose(perm = var_1496, x = attn_cast_fp16)[name = tensor("transpose_6")]; + tensor input_67_cast_fp16 = reshape(shape = var_1501, x = var_1497_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor attn5_out_proj_weight_to_fp16 = const()[name = tensor("attn5_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132251840)))]; + tensor linear_22_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn5_out_proj_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("linear_22_cast_fp16")]; + tensor var_1507_to_fp16 = const()[name = tensor("op_1507_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1508_cast_fp16 = add(x = position5_to_fp16, y = var_1507_to_fp16)[name = tensor("op_1508_cast_fp16")]; + tensor var_1508_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1508_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_69_cast_fp16 = add(x = input_63_cast_fp16, y = linear_22_cast_fp16)[name = tensor("input_69_cast_fp16")]; + tensor input_71_axes_0 = const()[name = tensor("input_71_axes_0"), val = tensor([-1])]; + tensor norm5_2_weight_to_fp16 = const()[name = tensor("norm5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134349056)))]; + tensor norm5_2_bias_to_fp16 = const()[name = tensor("norm5_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134351168)))]; + tensor var_1512_to_fp16 = const()[name = tensor("op_1512_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_71_cast_fp16 = layer_norm(axes = input_71_axes_0, beta = norm5_2_bias_to_fp16, epsilon = var_1512_to_fp16, gamma = norm5_2_weight_to_fp16, x = input_69_cast_fp16)[name = tensor("input_71_cast_fp16")]; + tensor linear5_1_weight_to_fp16 = const()[name = tensor("linear5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134353280)))]; + tensor linear_23_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear5_1_weight_to_fp16, x = input_71_cast_fp16)[name = tensor("linear_23_cast_fp16")]; + tensor input_73_mode_0 = const()[name = tensor("input_73_mode_0"), val = tensor("EXACT")]; + tensor input_73_cast_fp16 = gelu(mode = input_73_mode_0, x = linear_23_cast_fp16)[name = tensor("input_73_cast_fp16")]; + tensor linear5_2_weight_to_fp16 = const()[name = tensor("linear5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142741952)))]; + tensor linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear5_2_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("linear_24_cast_fp16")]; + tensor input_75_cast_fp16 = add(x = input_69_cast_fp16, y = linear_24_cast_fp16)[name = tensor("input_75_cast_fp16")]; + tensor input_axes_0 = const()[name = tensor("input_axes_0"), val = tensor([-1])]; + tensor out_norm_weight_to_fp16 = const()[name = tensor("out_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151130624)))]; + tensor out_norm_bias_to_fp16 = const()[name = tensor("out_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151132736)))]; + tensor var_1529_to_fp16 = const()[name = tensor("op_1529_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_cast_fp16 = layer_norm(axes = input_axes_0, beta = out_norm_bias_to_fp16, epsilon = var_1529_to_fp16, gamma = out_norm_weight_to_fp16, x = input_75_cast_fp16)[name = tensor("input_cast_fp16")]; + tensor input_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("input_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor out_eos_weight_to_fp16 = const()[name = tensor("out_eos_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151134848)))]; + tensor out_eos_bias_to_fp16 = const()[name = tensor("out_eos_bias_to_fp16"), val = tensor([-0x1.64p-3])]; + tensor linear_25_cast_fp16 = linear(bias = out_eos_bias_to_fp16, weight = out_eos_weight_to_fp16, x = input_cast_fp16)[name = tensor("linear_25_cast_fp16")]; + tensor linear_25_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("linear_25_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor is_eos = cast(dtype = linear_25_cast_fp16_to_fp32_dtype_0, x = linear_25_cast_fp16)[name = tensor("cast_96")]; + tensor transformer_out = cast(dtype = input_cast_fp16_to_fp32_dtype_0, x = input_cast_fp16)[name = tensor("cast_97")]; + tensor new_position5 = cast(dtype = var_1508_cast_fp16_to_fp32_dtype_0, x = var_1508_cast_fp16)[name = tensor("cast_98")]; + tensor new_v_cache5 = cast(dtype = new_v_cache_cast_fp16_to_fp32_dtype_0, x = new_v_cache_cast_fp16)[name = tensor("cast_100")]; + tensor new_k_cache5 = cast(dtype = new_k_cache_cast_fp16_to_fp32_dtype_0, x = new_k_cache_cast_fp16)[name = tensor("cast_102")]; + tensor new_position4 = cast(dtype = var_1264_cast_fp16_to_fp32_dtype_0, x = var_1264_cast_fp16)[name = tensor("cast_106")]; + tensor new_v_cache4 = cast(dtype = new_v_cache_9_cast_fp16_to_fp32_dtype_0, x = new_v_cache_9_cast_fp16)[name = tensor("cast_108")]; + tensor new_k_cache4 = cast(dtype = new_k_cache_9_cast_fp16_to_fp32_dtype_0, x = new_k_cache_9_cast_fp16)[name = tensor("cast_110")]; + tensor new_position3 = cast(dtype = var_1020_cast_fp16_to_fp32_dtype_0, x = var_1020_cast_fp16)[name = tensor("cast_114")]; + tensor new_v_cache3 = cast(dtype = new_v_cache_7_cast_fp16_to_fp32_dtype_0, x = new_v_cache_7_cast_fp16)[name = tensor("cast_116")]; + tensor new_k_cache3 = cast(dtype = new_k_cache_7_cast_fp16_to_fp32_dtype_0, x = new_k_cache_7_cast_fp16)[name = tensor("cast_118")]; + tensor new_position2 = cast(dtype = var_776_cast_fp16_to_fp32_dtype_0, x = var_776_cast_fp16)[name = tensor("cast_122")]; + tensor new_v_cache2 = cast(dtype = new_v_cache_5_cast_fp16_to_fp32_dtype_0, x = new_v_cache_5_cast_fp16)[name = tensor("cast_124")]; + tensor new_k_cache2 = cast(dtype = new_k_cache_5_cast_fp16_to_fp32_dtype_0, x = new_k_cache_5_cast_fp16)[name = tensor("cast_126")]; + tensor new_position1 = cast(dtype = var_532_cast_fp16_to_fp32_dtype_0, x = var_532_cast_fp16)[name = tensor("cast_130")]; + tensor new_v_cache1 = cast(dtype = new_v_cache_3_cast_fp16_to_fp32_dtype_0, x = new_v_cache_3_cast_fp16)[name = tensor("cast_132")]; + tensor new_k_cache1 = cast(dtype = new_k_cache_3_cast_fp16_to_fp32_dtype_0, x = new_k_cache_3_cast_fp16)[name = tensor("cast_134")]; + tensor new_position0 = cast(dtype = var_288_cast_fp16_to_fp32_dtype_0, x = var_288_cast_fp16)[name = tensor("cast_138")]; + tensor new_v_cache0 = cast(dtype = new_v_cache_1_cast_fp16_to_fp32_dtype_0, x = new_v_cache_1_cast_fp16)[name = tensor("cast_140")]; + tensor new_k_cache0 = cast(dtype = new_k_cache_1_cast_fp16_to_fp32_dtype_0, x = new_k_cache_1_cast_fp16)[name = tensor("cast_142")]; + } -> (transformer_out, is_eos, new_k_cache0, new_v_cache0, new_position0, new_k_cache1, new_v_cache1, new_position1, new_k_cache2, new_v_cache2, new_position2, new_k_cache3, new_v_cache3, new_position3, new_k_cache4, new_v_cache4, new_position4, new_k_cache5, new_v_cache5, new_position5); +} \ No newline at end of file diff --git a/v2.1/italian/flowlm_step_ane.mlmodelc/weights/weight.bin b/v2.1/italian/flowlm_step_ane.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..67ec228c0c622a1a05355b250521e40f05b496a8 --- /dev/null +++ b/v2.1/italian/flowlm_step_ane.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1279ebb95e1f3a1b14a3d8bec0b1fc593843fdc5b08fb6f22afc34457be2f340 +size 151136960 diff --git a/v2.1/italian/flowlm_step_ane.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/italian/flowlm_step_ane.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..f42e9a6c90f3dbe3f352c96bb01333d4a2c9b27d --- /dev/null +++ b/v2.1/italian/flowlm_step_ane.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:96d45427d7d02a9ba3c4f0a23555e76fbc7932e5f26bec40eccfe1a949850f16 +size 137731 diff --git a/v2.1/italian/flowlm_step_ane.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/italian/flowlm_step_ane.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..67ec228c0c622a1a05355b250521e40f05b496a8 --- /dev/null +++ b/v2.1/italian/flowlm_step_ane.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1279ebb95e1f3a1b14a3d8bec0b1fc593843fdc5b08fb6f22afc34457be2f340 +size 151136960 diff --git a/v2.1/italian/flowlm_step_ane.mlpackage/Manifest.json b/v2.1/italian/flowlm_step_ane.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..8e7db22530610f9fdb5cb3f792ea22562bcfa024 --- /dev/null +++ b/v2.1/italian/flowlm_step_ane.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "D48C09DD-6B22-47F5-A187-EC46E4E79D0E": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "ECC45704-76EE-4047-AD3D-984691D893B8": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "D48C09DD-6B22-47F5-A187-EC46E4E79D0E" +} diff --git a/v2.1/italian_24l/cond_prefill_ane.mlmodelc/analytics/coremldata.bin b/v2.1/italian_24l/cond_prefill_ane.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..8f40e5584cee2fcf14ef7b252ea9cefdb1192b2d --- /dev/null +++ b/v2.1/italian_24l/cond_prefill_ane.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c8406fc88cb3582c666b91557aa1f5a84d1ffa7058a608fbe40401ccb22c9a39 +size 243 diff --git a/v2.1/italian_24l/cond_prefill_ane.mlmodelc/coremldata.bin b/v2.1/italian_24l/cond_prefill_ane.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..8361121093fcaa1c999a1507ebdadb587f65ed49 --- /dev/null +++ b/v2.1/italian_24l/cond_prefill_ane.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3fdd745c516cc3d146416eee9aff93bf981070807ab6fa3510d742f4eb6b5e78 +size 4496 diff --git a/v2.1/italian_24l/cond_prefill_ane.mlmodelc/model.mil b/v2.1/italian_24l/cond_prefill_ane.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..f102b6e147b28d002e028bb659c273f9e9c7f4c9 --- /dev/null +++ b/v2.1/italian_24l/cond_prefill_ane.mlmodelc/model.mil @@ -0,0 +1,3198 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.9.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor conditioning, tensor k_cache0, tensor k_cache1, tensor k_cache10, tensor k_cache11, tensor k_cache12, tensor k_cache13, tensor k_cache14, tensor k_cache15, tensor k_cache16, tensor k_cache17, tensor k_cache18, tensor k_cache19, tensor k_cache2, tensor k_cache20, tensor k_cache21, tensor k_cache22, tensor k_cache23, tensor k_cache3, tensor k_cache4, tensor k_cache5, tensor k_cache6, tensor k_cache7, tensor k_cache8, tensor k_cache9, tensor position0, tensor position1, tensor position10, tensor position11, tensor position12, tensor position13, tensor position14, tensor position15, tensor position16, tensor position17, tensor position18, tensor position19, tensor position2, tensor position20, tensor position21, tensor position22, tensor position23, tensor position3, tensor position4, tensor position5, tensor position6, tensor position7, tensor position8, tensor position9, tensor v_cache0, tensor v_cache1, tensor v_cache10, tensor v_cache11, tensor v_cache12, tensor v_cache13, tensor v_cache14, tensor v_cache15, tensor v_cache16, tensor v_cache17, tensor v_cache18, tensor v_cache19, tensor v_cache2, tensor v_cache20, tensor v_cache21, tensor v_cache22, tensor v_cache23, tensor v_cache3, tensor v_cache4, tensor v_cache5, tensor v_cache6, tensor v_cache7, tensor v_cache8, tensor v_cache9, tensor valid_len) { + tensor token_indices_to_fp16 = const()[name = tensor("token_indices_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor position0_to_fp16_dtype_0 = const()[name = tensor("position0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position0_to_fp16 = cast(dtype = position0_to_fp16_dtype_0, x = position0)[name = tensor("cast_470")]; + tensor ts_cast_fp16 = add(x = token_indices_to_fp16, y = position0_to_fp16)[name = tensor("ts_cast_fp16")]; + tensor var_231 = const()[name = tensor("op_231"), val = tensor([256, 1, 1])]; + tensor var_232_cast_fp16 = reshape(shape = var_231, x = ts_cast_fp16)[name = tensor("op_232_cast_fp16")]; + tensor var_227_to_fp16 = const()[name = tensor("op_227_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(640)))]; + tensor angles_cast_fp16 = mul(x = var_227_to_fp16, y = var_232_cast_fp16)[name = tensor("angles_cast_fp16")]; + tensor rotr_cast_fp16 = cos(x = angles_cast_fp16)[name = tensor("rotr_cast_fp16")]; + tensor roti_cast_fp16 = sin(x = angles_cast_fp16)[name = tensor("roti_cast_fp16")]; + tensor _inversed_237_y_0_to_fp16 = const()[name = tensor("_inversed_237_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_237_cast_fp16 = mul(x = ts_cast_fp16, y = _inversed_237_y_0_to_fp16)[name = tensor("_inversed_237_cast_fp16")]; + tensor var_238_cast_fp16 = floor(x = _inversed_237_cast_fp16)[name = tensor("op_238_cast_fp16")]; + tensor var_239_to_fp16 = const()[name = tensor("op_239_to_fp16"), val = tensor(0x1p+9)]; + tensor var_240_cast_fp16 = mul(x = var_238_cast_fp16, y = var_239_to_fp16)[name = tensor("op_240_cast_fp16")]; + tensor write_idx_1_cast_fp16 = sub(x = ts_cast_fp16, y = var_240_cast_fp16)[name = tensor("write_idx_1_cast_fp16")]; + tensor valid_len_to_fp16_dtype_0 = const()[name = tensor("valid_len_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor valid_len_to_fp16 = cast(dtype = valid_len_to_fp16_dtype_0, x = valid_len)[name = tensor("cast_469")]; + tensor var_246_cast_fp16 = less(x = token_indices_to_fp16, y = valid_len_to_fp16)[name = tensor("op_246_cast_fp16")]; + tensor token_valid_to_fp16_dtype_0 = const()[name = tensor("token_valid_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_246_cast_fp16_to_fp16 = cast(dtype = token_valid_to_fp16_dtype_0, x = var_246_cast_fp16)[name = tensor("cast_468")]; + tensor var_252_cast_fp16 = mul(x = write_idx_1_cast_fp16, y = var_246_cast_fp16_to_fp16)[name = tensor("op_252_cast_fp16")]; + tensor var_253_to_fp16 = const()[name = tensor("op_253_to_fp16"), val = tensor(0x1p+0)]; + tensor var_255_cast_fp16 = sub(x = var_253_to_fp16, y = var_246_cast_fp16_to_fp16)[name = tensor("op_255_cast_fp16")]; + tensor var_256_to_fp16 = const()[name = tensor("op_256_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_257_cast_fp16 = mul(x = var_255_cast_fp16, y = var_256_to_fp16)[name = tensor("op_257_cast_fp16")]; + tensor write_idx_cast_fp16 = add(x = var_252_cast_fp16, y = var_257_cast_fp16)[name = tensor("write_idx_cast_fp16")]; + tensor var_262 = const()[name = tensor("op_262"), val = tensor([256, 1])]; + tensor var_263_cast_fp16 = reshape(shape = var_262, x = write_idx_cast_fp16)[name = tensor("op_263_cast_fp16")]; + tensor var_267_to_fp16 = const()[name = tensor("op_267_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(768)))]; + tensor var_268_cast_fp16 = equal(x = var_263_cast_fp16, y = var_267_to_fp16)[name = tensor("op_268_cast_fp16")]; + tensor var_278_axes_0 = const()[name = tensor("op_278_axes_0"), val = tensor([0])]; + tensor var_278_keep_dims_0 = const()[name = tensor("op_278_keep_dims_0"), val = tensor(false)]; + tensor assign_to_fp16_dtype_0 = const()[name = tensor("assign_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_268_cast_fp16_to_fp16 = cast(dtype = assign_to_fp16_dtype_0, x = var_268_cast_fp16)[name = tensor("cast_467")]; + tensor var_278_cast_fp16 = reduce_sum(axes = var_278_axes_0, keep_dims = var_278_keep_dims_0, x = var_268_cast_fp16_to_fp16)[name = tensor("op_278_cast_fp16")]; + tensor const_0_to_fp16 = const()[name = tensor("const_0_to_fp16"), val = tensor(-inf)]; + tensor var_280_to_fp16 = const()[name = tensor("op_280_to_fp16"), val = tensor(0x1p+0)]; + tensor clip_0_cast_fp16 = clip(alpha = const_0_to_fp16, beta = var_280_to_fp16, x = var_278_cast_fp16)[name = tensor("clip_0_cast_fp16")]; + tensor var_284 = const()[name = tensor("op_284"), val = tensor([1, 1])]; + tensor var_285_cast_fp16 = reshape(shape = var_284, x = position0_to_fp16)[name = tensor("op_285_cast_fp16")]; + tensor var_289_to_fp16 = const()[name = tensor("op_289_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1856)))]; + tensor q_positions_cast_fp16 = add(x = var_285_cast_fp16, y = var_289_to_fp16)[name = tensor("q_positions_cast_fp16")]; + tensor var_297_cast_fp16 = add(x = position0_to_fp16, y = valid_len_to_fp16)[name = tensor("op_297_cast_fp16")]; + tensor var_300 = const()[name = tensor("op_300"), val = tensor([1, 1])]; + tensor valid_end_cast_fp16 = reshape(shape = var_300, x = var_297_cast_fp16)[name = tensor("valid_end_cast_fp16")]; + tensor var_302_cast_fp16 = less(x = var_267_to_fp16, y = valid_end_cast_fp16)[name = tensor("op_302_cast_fp16")]; + tensor var_303_cast_fp16 = less_equal(x = var_267_to_fp16, y = q_positions_cast_fp16)[name = tensor("op_303_cast_fp16")]; + tensor var_304 = logical_and(x = var_302_cast_fp16, y = var_303_cast_fp16)[name = tensor("op_304")]; + tensor mask_to_fp16_dtype_0 = const()[name = tensor("mask_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_311_to_fp16 = const()[name = tensor("op_311_to_fp16"), val = tensor(0x1p+0)]; + tensor var_304_to_fp16 = cast(dtype = mask_to_fp16_dtype_0, x = var_304)[name = tensor("cast_466")]; + tensor var_312_cast_fp16 = sub(x = var_304_to_fp16, y = var_311_to_fp16)[name = tensor("op_312_cast_fp16")]; + tensor var_313_to_fp16 = const()[name = tensor("op_313_to_fp16"), val = tensor(0x1.388p+13)]; + tensor mask_bias_cast_fp16 = mul(x = var_312_cast_fp16, y = var_313_to_fp16)[name = tensor("mask_bias_cast_fp16")]; + tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; + tensor conditioning_to_fp16_dtype_0 = const()[name = tensor("conditioning_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor norm0_1_weight_to_fp16 = const()[name = tensor("norm0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2432)))]; + tensor norm0_1_bias_to_fp16 = const()[name = tensor("norm0_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4544)))]; + tensor var_316_to_fp16 = const()[name = tensor("op_316_to_fp16"), val = tensor(0x1.5p-17)]; + tensor conditioning_to_fp16 = cast(dtype = conditioning_to_fp16_dtype_0, x = conditioning)[name = tensor("cast_465")]; + tensor input_3_cast_fp16 = layer_norm(axes = input_3_axes_0, beta = norm0_1_bias_to_fp16, epsilon = var_316_to_fp16, gamma = norm0_1_weight_to_fp16, x = conditioning_to_fp16)[name = tensor("input_3_cast_fp16")]; + tensor attn0_in_proj_weight_to_fp16 = const()[name = tensor("attn0_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6656)))]; + tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6298176)))]; + tensor linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn0_in_proj_weight_to_fp16, x = input_3_cast_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor var_330 = const()[name = tensor("op_330"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_1_cast_fp16 = reshape(shape = var_330, x = linear_0_cast_fp16)[name = tensor("qkv_1_cast_fp16")]; + tensor x_1_begin_0 = const()[name = tensor("x_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_1_end_0 = const()[name = tensor("x_1_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_1_end_mask_0 = const()[name = tensor("x_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_1_squeeze_mask_0 = const()[name = tensor("x_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_1_cast_fp16 = slice_by_index(begin = x_1_begin_0, end = x_1_end_0, end_mask = x_1_end_mask_0, squeeze_mask = x_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("x_1_cast_fp16")]; + tensor x_3_begin_0 = const()[name = tensor("x_3_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_3_end_0 = const()[name = tensor("x_3_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_3_end_mask_0 = const()[name = tensor("x_3_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_3_squeeze_mask_0 = const()[name = tensor("x_3_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_3_cast_fp16 = slice_by_index(begin = x_3_begin_0, end = x_3_end_0, end_mask = x_3_end_mask_0, squeeze_mask = x_3_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("x_3_cast_fp16")]; + tensor v_1_begin_0 = const()[name = tensor("v_1_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_1_end_0 = const()[name = tensor("v_1_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_1_end_mask_0 = const()[name = tensor("v_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_1_squeeze_mask_0 = const()[name = tensor("v_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_1_cast_fp16 = slice_by_index(begin = v_1_begin_0, end = v_1_end_0, end_mask = v_1_end_mask_0, squeeze_mask = v_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("v_1_cast_fp16")]; + tensor var_375 = const()[name = tensor("op_375"), val = tensor([256, 16, 32, 2])]; + tensor pairs_1_cast_fp16 = reshape(shape = var_375, x = x_1_cast_fp16)[name = tensor("pairs_1_cast_fp16")]; + tensor xr_1_begin_0 = const()[name = tensor("xr_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_1_end_0 = const()[name = tensor("xr_1_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_1_end_mask_0 = const()[name = tensor("xr_1_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_1_squeeze_mask_0 = const()[name = tensor("xr_1_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_1_cast_fp16 = slice_by_index(begin = xr_1_begin_0, end = xr_1_end_0, end_mask = xr_1_end_mask_0, squeeze_mask = xr_1_squeeze_mask_0, x = pairs_1_cast_fp16)[name = tensor("xr_1_cast_fp16")]; + tensor xi_1_begin_0 = const()[name = tensor("xi_1_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_1_end_0 = const()[name = tensor("xi_1_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_1_end_mask_0 = const()[name = tensor("xi_1_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_1_squeeze_mask_0 = const()[name = tensor("xi_1_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_1_cast_fp16 = slice_by_index(begin = xi_1_begin_0, end = xi_1_end_0, end_mask = xi_1_end_mask_0, squeeze_mask = xi_1_squeeze_mask_0, x = pairs_1_cast_fp16)[name = tensor("xi_1_cast_fp16")]; + tensor var_383_cast_fp16 = mul(x = xr_1_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_383_cast_fp16")]; + tensor var_384_cast_fp16 = mul(x = xi_1_cast_fp16, y = roti_cast_fp16)[name = tensor("op_384_cast_fp16")]; + tensor xor__1_cast_fp16 = sub(x = var_383_cast_fp16, y = var_384_cast_fp16)[name = tensor("xor__1_cast_fp16")]; + tensor var_387_cast_fp16 = mul(x = xr_1_cast_fp16, y = roti_cast_fp16)[name = tensor("op_387_cast_fp16")]; + tensor var_388_cast_fp16 = mul(x = xi_1_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_388_cast_fp16")]; + tensor xoi_1_cast_fp16 = add(x = var_387_cast_fp16, y = var_388_cast_fp16)[name = tensor("xoi_1_cast_fp16")]; + tensor out_1_axis_0 = const()[name = tensor("out_1_axis_0"), val = tensor(-1)]; + tensor out_1_cast_fp16 = stack(axis = out_1_axis_0, values = (xor__1_cast_fp16, xoi_1_cast_fp16))[name = tensor("out_1_cast_fp16")]; + tensor var_398 = const()[name = tensor("op_398"), val = tensor([1, 256, 16, 64])]; + tensor q_1_cast_fp16 = reshape(shape = var_398, x = out_1_cast_fp16)[name = tensor("q_1_cast_fp16")]; + tensor var_404 = const()[name = tensor("op_404"), val = tensor([256, 16, 32, 2])]; + tensor pairs_3_cast_fp16 = reshape(shape = var_404, x = x_3_cast_fp16)[name = tensor("pairs_3_cast_fp16")]; + tensor xr_3_begin_0 = const()[name = tensor("xr_3_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_3_end_0 = const()[name = tensor("xr_3_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_3_end_mask_0 = const()[name = tensor("xr_3_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_3_squeeze_mask_0 = const()[name = tensor("xr_3_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_3_cast_fp16 = slice_by_index(begin = xr_3_begin_0, end = xr_3_end_0, end_mask = xr_3_end_mask_0, squeeze_mask = xr_3_squeeze_mask_0, x = pairs_3_cast_fp16)[name = tensor("xr_3_cast_fp16")]; + tensor xi_3_begin_0 = const()[name = tensor("xi_3_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_3_end_0 = const()[name = tensor("xi_3_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_3_end_mask_0 = const()[name = tensor("xi_3_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_3_squeeze_mask_0 = const()[name = tensor("xi_3_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_3_cast_fp16 = slice_by_index(begin = xi_3_begin_0, end = xi_3_end_0, end_mask = xi_3_end_mask_0, squeeze_mask = xi_3_squeeze_mask_0, x = pairs_3_cast_fp16)[name = tensor("xi_3_cast_fp16")]; + tensor var_412_cast_fp16 = mul(x = xr_3_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_412_cast_fp16")]; + tensor var_413_cast_fp16 = mul(x = xi_3_cast_fp16, y = roti_cast_fp16)[name = tensor("op_413_cast_fp16")]; + tensor xor__3_cast_fp16 = sub(x = var_412_cast_fp16, y = var_413_cast_fp16)[name = tensor("xor__3_cast_fp16")]; + tensor var_416_cast_fp16 = mul(x = xr_3_cast_fp16, y = roti_cast_fp16)[name = tensor("op_416_cast_fp16")]; + tensor var_417_cast_fp16 = mul(x = xi_3_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_417_cast_fp16")]; + tensor xoi_3_cast_fp16 = add(x = var_416_cast_fp16, y = var_417_cast_fp16)[name = tensor("xoi_3_cast_fp16")]; + tensor out_3_axis_0 = const()[name = tensor("out_3_axis_0"), val = tensor(-1)]; + tensor out_3_cast_fp16 = stack(axis = out_3_axis_0, values = (xor__3_cast_fp16, xoi_3_cast_fp16))[name = tensor("out_3_cast_fp16")]; + tensor var_434 = const()[name = tensor("op_434"), val = tensor([256, 1024])]; + tensor var_435_cast_fp16 = reshape(shape = var_434, x = out_3_cast_fp16)[name = tensor("op_435_cast_fp16")]; + tensor written_k_1_transpose_x_1 = const()[name = tensor("written_k_1_transpose_x_1"), val = tensor(true)]; + tensor written_k_1_transpose_y_1 = const()[name = tensor("written_k_1_transpose_y_1"), val = tensor(false)]; + tensor written_k_1_cast_fp16 = matmul(transpose_x = written_k_1_transpose_x_1, transpose_y = written_k_1_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_435_cast_fp16)[name = tensor("written_k_1_cast_fp16")]; + tensor var_442 = const()[name = tensor("op_442"), val = tensor([256, 1024])]; + tensor var_443_cast_fp16 = reshape(shape = var_442, x = v_1_cast_fp16)[name = tensor("op_443_cast_fp16")]; + tensor written_v_1_transpose_x_1 = const()[name = tensor("written_v_1_transpose_x_1"), val = tensor(true)]; + tensor written_v_1_transpose_y_1 = const()[name = tensor("written_v_1_transpose_y_1"), val = tensor(false)]; + tensor written_v_1_cast_fp16 = matmul(transpose_x = written_v_1_transpose_x_1, transpose_y = written_v_1_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_443_cast_fp16)[name = tensor("written_v_1_cast_fp16")]; + tensor var_445_to_fp16 = const()[name = tensor("op_445_to_fp16"), val = tensor(0x1p+0)]; + tensor var_447_cast_fp16 = sub(x = var_445_to_fp16, y = clip_0_cast_fp16)[name = tensor("op_447_cast_fp16")]; + tensor var_452 = const()[name = tensor("op_452"), val = tensor([1, 512, 1, 1])]; + tensor keep_1_cast_fp16 = reshape(shape = var_452, x = var_447_cast_fp16)[name = tensor("keep_1_cast_fp16")]; + tensor k_cache0_to_fp16_dtype_0 = const()[name = tensor("k_cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache0_to_fp16 = cast(dtype = k_cache0_to_fp16_dtype_0, x = k_cache0)[name = tensor("cast_464")]; + tensor var_454_cast_fp16 = mul(x = k_cache0_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_454_cast_fp16")]; + tensor var_459 = const()[name = tensor("op_459"), val = tensor([1, 512, 16, 64])]; + tensor var_460_cast_fp16 = reshape(shape = var_459, x = written_k_1_cast_fp16)[name = tensor("op_460_cast_fp16")]; + tensor new_k_cache_1_cast_fp16 = add(x = var_454_cast_fp16, y = var_460_cast_fp16)[name = tensor("new_k_cache_1_cast_fp16")]; + tensor new_k_cache_1_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_1_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache0_to_fp16_dtype_0 = const()[name = tensor("v_cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache0_to_fp16 = cast(dtype = v_cache0_to_fp16_dtype_0, x = v_cache0)[name = tensor("cast_462")]; + tensor var_463_cast_fp16 = mul(x = v_cache0_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_463_cast_fp16")]; + tensor var_468 = const()[name = tensor("op_468"), val = tensor([1, 512, 16, 64])]; + tensor var_469_cast_fp16 = reshape(shape = var_468, x = written_v_1_cast_fp16)[name = tensor("op_469_cast_fp16")]; + tensor new_v_cache_1_cast_fp16 = add(x = var_463_cast_fp16, y = var_469_cast_fp16)[name = tensor("new_v_cache_1_cast_fp16")]; + tensor new_v_cache_1_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_1_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_482 = const()[name = tensor("op_482"), val = tensor([0, 2, 1, 3])]; + tensor var_493_transpose_x_0 = const()[name = tensor("op_493_transpose_x_0"), val = tensor(false)]; + tensor var_493_transpose_y_0 = const()[name = tensor("op_493_transpose_y_0"), val = tensor(false)]; + tensor transpose_69_perm_0 = const()[name = tensor("transpose_69_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_70_perm_0 = const()[name = tensor("transpose_70_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_70 = transpose(perm = transpose_70_perm_0, x = new_k_cache_1_cast_fp16)[name = tensor("transpose_204")]; + tensor transpose_69 = transpose(perm = transpose_69_perm_0, x = q_1_cast_fp16)[name = tensor("transpose_205")]; + tensor var_493_cast_fp16 = matmul(transpose_x = var_493_transpose_x_0, transpose_y = var_493_transpose_y_0, x = transpose_69, y = transpose_70)[name = tensor("op_493_cast_fp16")]; + tensor var_494_to_fp16 = const()[name = tensor("op_494_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_1_cast_fp16 = mul(x = var_493_cast_fp16, y = var_494_to_fp16)[name = tensor("scores_1_cast_fp16")]; + tensor var_500 = const()[name = tensor("op_500"), val = tensor([1, 1, 256, 512])]; + tensor var_501_cast_fp16 = reshape(shape = var_500, x = mask_bias_cast_fp16)[name = tensor("op_501_cast_fp16")]; + tensor scores_3_cast_fp16 = add(x = scores_1_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_3_cast_fp16")]; + tensor var_504 = const()[name = tensor("op_504"), val = tensor(-1)]; + tensor weights_1_cast_fp16 = softmax(axis = var_504, x = scores_3_cast_fp16)[name = tensor("weights_1_cast_fp16")]; + tensor attn_1_transpose_x_0 = const()[name = tensor("attn_1_transpose_x_0"), val = tensor(false)]; + tensor attn_1_transpose_y_0 = const()[name = tensor("attn_1_transpose_y_0"), val = tensor(false)]; + tensor values_1_cast_fp16 = transpose(perm = var_482, x = new_v_cache_1_cast_fp16)[name = tensor("transpose_206")]; + tensor attn_1_cast_fp16 = matmul(transpose_x = attn_1_transpose_x_0, transpose_y = attn_1_transpose_y_0, x = weights_1_cast_fp16, y = values_1_cast_fp16)[name = tensor("attn_1_cast_fp16")]; + tensor var_512 = const()[name = tensor("op_512"), val = tensor([0, 2, 1, 3])]; + tensor var_517 = const()[name = tensor("op_517"), val = tensor([1, 256, 1024])]; + tensor var_513_cast_fp16 = transpose(perm = var_512, x = attn_1_cast_fp16)[name = tensor("transpose_203")]; + tensor input_5_cast_fp16 = reshape(shape = var_517, x = var_513_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor attn0_out_proj_weight_to_fp16 = const()[name = tensor("attn0_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6304384)))]; + tensor linear_1_bias_0_to_fp16 = const()[name = tensor("linear_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401600)))]; + tensor linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn0_out_proj_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("linear_1_cast_fp16")]; + tensor var_526_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_526_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_7_cast_fp16 = add(x = conditioning_to_fp16, y = linear_1_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor input_9_axes_0 = const()[name = tensor("input_9_axes_0"), val = tensor([-1])]; + tensor norm0_2_weight_to_fp16 = const()[name = tensor("norm0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8403712)))]; + tensor norm0_2_bias_to_fp16 = const()[name = tensor("norm0_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8405824)))]; + tensor var_530_to_fp16 = const()[name = tensor("op_530_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_9_cast_fp16 = layer_norm(axes = input_9_axes_0, beta = norm0_2_bias_to_fp16, epsilon = var_530_to_fp16, gamma = norm0_2_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor linear0_1_weight_to_fp16 = const()[name = tensor("linear0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8407936)))]; + tensor linear_2_bias_0_to_fp16 = const()[name = tensor("linear_2_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16796608)))]; + tensor linear_2_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear0_1_weight_to_fp16, x = input_9_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor input_11_mode_0 = const()[name = tensor("input_11_mode_0"), val = tensor("EXACT")]; + tensor input_11_cast_fp16 = gelu(mode = input_11_mode_0, x = linear_2_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor linear0_2_weight_to_fp16 = const()[name = tensor("linear0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16804864)))]; + tensor linear_3_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear0_2_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("linear_3_cast_fp16")]; + tensor input_13_cast_fp16 = add(x = input_7_cast_fp16, y = linear_3_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor input_15_axes_0 = const()[name = tensor("input_15_axes_0"), val = tensor([-1])]; + tensor norm1_1_weight_to_fp16 = const()[name = tensor("norm1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25193536)))]; + tensor norm1_1_bias_to_fp16 = const()[name = tensor("norm1_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25195648)))]; + tensor var_547_to_fp16 = const()[name = tensor("op_547_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_15_cast_fp16 = layer_norm(axes = input_15_axes_0, beta = norm1_1_bias_to_fp16, epsilon = var_547_to_fp16, gamma = norm1_1_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor attn1_in_proj_weight_to_fp16 = const()[name = tensor("attn1_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25197760)))]; + tensor linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn1_in_proj_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor var_561 = const()[name = tensor("op_561"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_3_cast_fp16 = reshape(shape = var_561, x = linear_4_cast_fp16)[name = tensor("qkv_3_cast_fp16")]; + tensor x_5_begin_0 = const()[name = tensor("x_5_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_5_end_0 = const()[name = tensor("x_5_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_5_end_mask_0 = const()[name = tensor("x_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_5_squeeze_mask_0 = const()[name = tensor("x_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_5_cast_fp16 = slice_by_index(begin = x_5_begin_0, end = x_5_end_0, end_mask = x_5_end_mask_0, squeeze_mask = x_5_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("x_5_cast_fp16")]; + tensor x_7_begin_0 = const()[name = tensor("x_7_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_7_end_0 = const()[name = tensor("x_7_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_7_end_mask_0 = const()[name = tensor("x_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_7_squeeze_mask_0 = const()[name = tensor("x_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_7_cast_fp16 = slice_by_index(begin = x_7_begin_0, end = x_7_end_0, end_mask = x_7_end_mask_0, squeeze_mask = x_7_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("x_7_cast_fp16")]; + tensor v_3_begin_0 = const()[name = tensor("v_3_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_3_end_0 = const()[name = tensor("v_3_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_3_end_mask_0 = const()[name = tensor("v_3_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_3_squeeze_mask_0 = const()[name = tensor("v_3_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_3_cast_fp16 = slice_by_index(begin = v_3_begin_0, end = v_3_end_0, end_mask = v_3_end_mask_0, squeeze_mask = v_3_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("v_3_cast_fp16")]; + tensor var_606 = const()[name = tensor("op_606"), val = tensor([256, 16, 32, 2])]; + tensor pairs_5_cast_fp16 = reshape(shape = var_606, x = x_5_cast_fp16)[name = tensor("pairs_5_cast_fp16")]; + tensor xr_5_begin_0 = const()[name = tensor("xr_5_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_5_end_0 = const()[name = tensor("xr_5_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_5_end_mask_0 = const()[name = tensor("xr_5_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_5_squeeze_mask_0 = const()[name = tensor("xr_5_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_5_cast_fp16 = slice_by_index(begin = xr_5_begin_0, end = xr_5_end_0, end_mask = xr_5_end_mask_0, squeeze_mask = xr_5_squeeze_mask_0, x = pairs_5_cast_fp16)[name = tensor("xr_5_cast_fp16")]; + tensor xi_5_begin_0 = const()[name = tensor("xi_5_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_5_end_0 = const()[name = tensor("xi_5_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_5_end_mask_0 = const()[name = tensor("xi_5_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_5_squeeze_mask_0 = const()[name = tensor("xi_5_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_5_cast_fp16 = slice_by_index(begin = xi_5_begin_0, end = xi_5_end_0, end_mask = xi_5_end_mask_0, squeeze_mask = xi_5_squeeze_mask_0, x = pairs_5_cast_fp16)[name = tensor("xi_5_cast_fp16")]; + tensor var_614_cast_fp16 = mul(x = xr_5_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_614_cast_fp16")]; + tensor var_615_cast_fp16 = mul(x = xi_5_cast_fp16, y = roti_cast_fp16)[name = tensor("op_615_cast_fp16")]; + tensor xor__5_cast_fp16 = sub(x = var_614_cast_fp16, y = var_615_cast_fp16)[name = tensor("xor__5_cast_fp16")]; + tensor var_618_cast_fp16 = mul(x = xr_5_cast_fp16, y = roti_cast_fp16)[name = tensor("op_618_cast_fp16")]; + tensor var_619_cast_fp16 = mul(x = xi_5_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_619_cast_fp16")]; + tensor xoi_5_cast_fp16 = add(x = var_618_cast_fp16, y = var_619_cast_fp16)[name = tensor("xoi_5_cast_fp16")]; + tensor out_5_axis_0 = const()[name = tensor("out_5_axis_0"), val = tensor(-1)]; + tensor out_5_cast_fp16 = stack(axis = out_5_axis_0, values = (xor__5_cast_fp16, xoi_5_cast_fp16))[name = tensor("out_5_cast_fp16")]; + tensor var_629 = const()[name = tensor("op_629"), val = tensor([1, 256, 16, 64])]; + tensor q_3_cast_fp16 = reshape(shape = var_629, x = out_5_cast_fp16)[name = tensor("q_3_cast_fp16")]; + tensor var_635 = const()[name = tensor("op_635"), val = tensor([256, 16, 32, 2])]; + tensor pairs_7_cast_fp16 = reshape(shape = var_635, x = x_7_cast_fp16)[name = tensor("pairs_7_cast_fp16")]; + tensor xr_7_begin_0 = const()[name = tensor("xr_7_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_7_end_0 = const()[name = tensor("xr_7_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_7_end_mask_0 = const()[name = tensor("xr_7_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_7_squeeze_mask_0 = const()[name = tensor("xr_7_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_7_cast_fp16 = slice_by_index(begin = xr_7_begin_0, end = xr_7_end_0, end_mask = xr_7_end_mask_0, squeeze_mask = xr_7_squeeze_mask_0, x = pairs_7_cast_fp16)[name = tensor("xr_7_cast_fp16")]; + tensor xi_7_begin_0 = const()[name = tensor("xi_7_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_7_end_0 = const()[name = tensor("xi_7_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_7_end_mask_0 = const()[name = tensor("xi_7_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_7_squeeze_mask_0 = const()[name = tensor("xi_7_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_7_cast_fp16 = slice_by_index(begin = xi_7_begin_0, end = xi_7_end_0, end_mask = xi_7_end_mask_0, squeeze_mask = xi_7_squeeze_mask_0, x = pairs_7_cast_fp16)[name = tensor("xi_7_cast_fp16")]; + tensor var_643_cast_fp16 = mul(x = xr_7_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_643_cast_fp16")]; + tensor var_644_cast_fp16 = mul(x = xi_7_cast_fp16, y = roti_cast_fp16)[name = tensor("op_644_cast_fp16")]; + tensor xor__7_cast_fp16 = sub(x = var_643_cast_fp16, y = var_644_cast_fp16)[name = tensor("xor__7_cast_fp16")]; + tensor var_647_cast_fp16 = mul(x = xr_7_cast_fp16, y = roti_cast_fp16)[name = tensor("op_647_cast_fp16")]; + tensor var_648_cast_fp16 = mul(x = xi_7_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_648_cast_fp16")]; + tensor xoi_7_cast_fp16 = add(x = var_647_cast_fp16, y = var_648_cast_fp16)[name = tensor("xoi_7_cast_fp16")]; + tensor out_7_axis_0 = const()[name = tensor("out_7_axis_0"), val = tensor(-1)]; + tensor out_7_cast_fp16 = stack(axis = out_7_axis_0, values = (xor__7_cast_fp16, xoi_7_cast_fp16))[name = tensor("out_7_cast_fp16")]; + tensor var_665 = const()[name = tensor("op_665"), val = tensor([256, 1024])]; + tensor var_666_cast_fp16 = reshape(shape = var_665, x = out_7_cast_fp16)[name = tensor("op_666_cast_fp16")]; + tensor written_k_3_transpose_x_1 = const()[name = tensor("written_k_3_transpose_x_1"), val = tensor(true)]; + tensor written_k_3_transpose_y_1 = const()[name = tensor("written_k_3_transpose_y_1"), val = tensor(false)]; + tensor written_k_3_cast_fp16 = matmul(transpose_x = written_k_3_transpose_x_1, transpose_y = written_k_3_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_666_cast_fp16)[name = tensor("written_k_3_cast_fp16")]; + tensor var_673 = const()[name = tensor("op_673"), val = tensor([256, 1024])]; + tensor var_674_cast_fp16 = reshape(shape = var_673, x = v_3_cast_fp16)[name = tensor("op_674_cast_fp16")]; + tensor written_v_3_transpose_x_1 = const()[name = tensor("written_v_3_transpose_x_1"), val = tensor(true)]; + tensor written_v_3_transpose_y_1 = const()[name = tensor("written_v_3_transpose_y_1"), val = tensor(false)]; + tensor written_v_3_cast_fp16 = matmul(transpose_x = written_v_3_transpose_x_1, transpose_y = written_v_3_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_674_cast_fp16)[name = tensor("written_v_3_cast_fp16")]; + tensor k_cache1_to_fp16_dtype_0 = const()[name = tensor("k_cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache1_to_fp16 = cast(dtype = k_cache1_to_fp16_dtype_0, x = k_cache1)[name = tensor("cast_459")]; + tensor var_685_cast_fp16 = mul(x = k_cache1_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_685_cast_fp16")]; + tensor var_690 = const()[name = tensor("op_690"), val = tensor([1, 512, 16, 64])]; + tensor var_691_cast_fp16 = reshape(shape = var_690, x = written_k_3_cast_fp16)[name = tensor("op_691_cast_fp16")]; + tensor new_k_cache_3_cast_fp16 = add(x = var_685_cast_fp16, y = var_691_cast_fp16)[name = tensor("new_k_cache_3_cast_fp16")]; + tensor new_k_cache_3_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_3_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache1_to_fp16_dtype_0 = const()[name = tensor("v_cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache1_to_fp16 = cast(dtype = v_cache1_to_fp16_dtype_0, x = v_cache1)[name = tensor("cast_457")]; + tensor var_694_cast_fp16 = mul(x = v_cache1_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_694_cast_fp16")]; + tensor var_699 = const()[name = tensor("op_699"), val = tensor([1, 512, 16, 64])]; + tensor var_700_cast_fp16 = reshape(shape = var_699, x = written_v_3_cast_fp16)[name = tensor("op_700_cast_fp16")]; + tensor new_v_cache_3_cast_fp16 = add(x = var_694_cast_fp16, y = var_700_cast_fp16)[name = tensor("new_v_cache_3_cast_fp16")]; + tensor new_v_cache_3_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_3_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_713 = const()[name = tensor("op_713"), val = tensor([0, 2, 1, 3])]; + tensor var_724_transpose_x_0 = const()[name = tensor("op_724_transpose_x_0"), val = tensor(false)]; + tensor var_724_transpose_y_0 = const()[name = tensor("op_724_transpose_y_0"), val = tensor(false)]; + tensor transpose_71_perm_0 = const()[name = tensor("transpose_71_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_72_perm_0 = const()[name = tensor("transpose_72_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_72 = transpose(perm = transpose_72_perm_0, x = new_k_cache_3_cast_fp16)[name = tensor("transpose_200")]; + tensor transpose_71 = transpose(perm = transpose_71_perm_0, x = q_3_cast_fp16)[name = tensor("transpose_201")]; + tensor var_724_cast_fp16 = matmul(transpose_x = var_724_transpose_x_0, transpose_y = var_724_transpose_y_0, x = transpose_71, y = transpose_72)[name = tensor("op_724_cast_fp16")]; + tensor var_725_to_fp16 = const()[name = tensor("op_725_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_5_cast_fp16 = mul(x = var_724_cast_fp16, y = var_725_to_fp16)[name = tensor("scores_5_cast_fp16")]; + tensor scores_7_cast_fp16 = add(x = scores_5_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_7_cast_fp16")]; + tensor var_735 = const()[name = tensor("op_735"), val = tensor(-1)]; + tensor weights_3_cast_fp16 = softmax(axis = var_735, x = scores_7_cast_fp16)[name = tensor("weights_3_cast_fp16")]; + tensor attn_3_transpose_x_0 = const()[name = tensor("attn_3_transpose_x_0"), val = tensor(false)]; + tensor attn_3_transpose_y_0 = const()[name = tensor("attn_3_transpose_y_0"), val = tensor(false)]; + tensor values_3_cast_fp16 = transpose(perm = var_713, x = new_v_cache_3_cast_fp16)[name = tensor("transpose_202")]; + tensor attn_3_cast_fp16 = matmul(transpose_x = attn_3_transpose_x_0, transpose_y = attn_3_transpose_y_0, x = weights_3_cast_fp16, y = values_3_cast_fp16)[name = tensor("attn_3_cast_fp16")]; + tensor var_743 = const()[name = tensor("op_743"), val = tensor([0, 2, 1, 3])]; + tensor var_748 = const()[name = tensor("op_748"), val = tensor([1, 256, 1024])]; + tensor var_744_cast_fp16 = transpose(perm = var_743, x = attn_3_cast_fp16)[name = tensor("transpose_199")]; + tensor input_17_cast_fp16 = reshape(shape = var_748, x = var_744_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor attn1_out_proj_weight_to_fp16 = const()[name = tensor("attn1_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31489280)))]; + tensor linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn1_out_proj_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor position1_to_fp16_dtype_0 = const()[name = tensor("position1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position1_to_fp16 = cast(dtype = position1_to_fp16_dtype_0, x = position1)[name = tensor("cast_455")]; + tensor var_757_cast_fp16 = add(x = position1_to_fp16, y = valid_len_to_fp16)[name = tensor("op_757_cast_fp16")]; + tensor var_757_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_757_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_19_cast_fp16 = add(x = input_13_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor input_21_axes_0 = const()[name = tensor("input_21_axes_0"), val = tensor([-1])]; + tensor norm1_2_weight_to_fp16 = const()[name = tensor("norm1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33586496)))]; + tensor norm1_2_bias_to_fp16 = const()[name = tensor("norm1_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33588608)))]; + tensor var_761_to_fp16 = const()[name = tensor("op_761_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_21_cast_fp16 = layer_norm(axes = input_21_axes_0, beta = norm1_2_bias_to_fp16, epsilon = var_761_to_fp16, gamma = norm1_2_weight_to_fp16, x = input_19_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor linear1_1_weight_to_fp16 = const()[name = tensor("linear1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33590720)))]; + tensor linear_6_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear1_1_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor input_23_mode_0 = const()[name = tensor("input_23_mode_0"), val = tensor("EXACT")]; + tensor input_23_cast_fp16 = gelu(mode = input_23_mode_0, x = linear_6_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor linear1_2_weight_to_fp16 = const()[name = tensor("linear1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41979392)))]; + tensor linear_7_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear1_2_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_25_cast_fp16 = add(x = input_19_cast_fp16, y = linear_7_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor input_27_axes_0 = const()[name = tensor("input_27_axes_0"), val = tensor([-1])]; + tensor norm2_1_weight_to_fp16 = const()[name = tensor("norm2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50368064)))]; + tensor norm2_1_bias_to_fp16 = const()[name = tensor("norm2_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50370176)))]; + tensor var_778_to_fp16 = const()[name = tensor("op_778_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = norm2_1_bias_to_fp16, epsilon = var_778_to_fp16, gamma = norm2_1_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor attn2_in_proj_weight_to_fp16 = const()[name = tensor("attn2_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50372288)))]; + tensor linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn2_in_proj_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor var_792 = const()[name = tensor("op_792"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_5_cast_fp16 = reshape(shape = var_792, x = linear_8_cast_fp16)[name = tensor("qkv_5_cast_fp16")]; + tensor x_9_begin_0 = const()[name = tensor("x_9_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_9_end_0 = const()[name = tensor("x_9_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_9_end_mask_0 = const()[name = tensor("x_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_9_squeeze_mask_0 = const()[name = tensor("x_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_9_cast_fp16 = slice_by_index(begin = x_9_begin_0, end = x_9_end_0, end_mask = x_9_end_mask_0, squeeze_mask = x_9_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("x_9_cast_fp16")]; + tensor x_11_begin_0 = const()[name = tensor("x_11_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_11_end_0 = const()[name = tensor("x_11_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_11_end_mask_0 = const()[name = tensor("x_11_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_11_squeeze_mask_0 = const()[name = tensor("x_11_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_11_cast_fp16 = slice_by_index(begin = x_11_begin_0, end = x_11_end_0, end_mask = x_11_end_mask_0, squeeze_mask = x_11_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("x_11_cast_fp16")]; + tensor v_5_begin_0 = const()[name = tensor("v_5_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_5_end_0 = const()[name = tensor("v_5_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_5_end_mask_0 = const()[name = tensor("v_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_5_squeeze_mask_0 = const()[name = tensor("v_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_5_cast_fp16 = slice_by_index(begin = v_5_begin_0, end = v_5_end_0, end_mask = v_5_end_mask_0, squeeze_mask = v_5_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("v_5_cast_fp16")]; + tensor var_837 = const()[name = tensor("op_837"), val = tensor([256, 16, 32, 2])]; + tensor pairs_9_cast_fp16 = reshape(shape = var_837, x = x_9_cast_fp16)[name = tensor("pairs_9_cast_fp16")]; + tensor xr_9_begin_0 = const()[name = tensor("xr_9_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_9_end_0 = const()[name = tensor("xr_9_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_9_end_mask_0 = const()[name = tensor("xr_9_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_9_squeeze_mask_0 = const()[name = tensor("xr_9_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_9_cast_fp16 = slice_by_index(begin = xr_9_begin_0, end = xr_9_end_0, end_mask = xr_9_end_mask_0, squeeze_mask = xr_9_squeeze_mask_0, x = pairs_9_cast_fp16)[name = tensor("xr_9_cast_fp16")]; + tensor xi_9_begin_0 = const()[name = tensor("xi_9_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_9_end_0 = const()[name = tensor("xi_9_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_9_end_mask_0 = const()[name = tensor("xi_9_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_9_squeeze_mask_0 = const()[name = tensor("xi_9_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_9_cast_fp16 = slice_by_index(begin = xi_9_begin_0, end = xi_9_end_0, end_mask = xi_9_end_mask_0, squeeze_mask = xi_9_squeeze_mask_0, x = pairs_9_cast_fp16)[name = tensor("xi_9_cast_fp16")]; + tensor var_845_cast_fp16 = mul(x = xr_9_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_845_cast_fp16")]; + tensor var_846_cast_fp16 = mul(x = xi_9_cast_fp16, y = roti_cast_fp16)[name = tensor("op_846_cast_fp16")]; + tensor xor__9_cast_fp16 = sub(x = var_845_cast_fp16, y = var_846_cast_fp16)[name = tensor("xor__9_cast_fp16")]; + tensor var_849_cast_fp16 = mul(x = xr_9_cast_fp16, y = roti_cast_fp16)[name = tensor("op_849_cast_fp16")]; + tensor var_850_cast_fp16 = mul(x = xi_9_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_850_cast_fp16")]; + tensor xoi_9_cast_fp16 = add(x = var_849_cast_fp16, y = var_850_cast_fp16)[name = tensor("xoi_9_cast_fp16")]; + tensor out_9_axis_0 = const()[name = tensor("out_9_axis_0"), val = tensor(-1)]; + tensor out_9_cast_fp16 = stack(axis = out_9_axis_0, values = (xor__9_cast_fp16, xoi_9_cast_fp16))[name = tensor("out_9_cast_fp16")]; + tensor var_860 = const()[name = tensor("op_860"), val = tensor([1, 256, 16, 64])]; + tensor q_5_cast_fp16 = reshape(shape = var_860, x = out_9_cast_fp16)[name = tensor("q_5_cast_fp16")]; + tensor var_866 = const()[name = tensor("op_866"), val = tensor([256, 16, 32, 2])]; + tensor pairs_11_cast_fp16 = reshape(shape = var_866, x = x_11_cast_fp16)[name = tensor("pairs_11_cast_fp16")]; + tensor xr_11_begin_0 = const()[name = tensor("xr_11_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_11_end_0 = const()[name = tensor("xr_11_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_11_end_mask_0 = const()[name = tensor("xr_11_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_11_squeeze_mask_0 = const()[name = tensor("xr_11_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_11_cast_fp16 = slice_by_index(begin = xr_11_begin_0, end = xr_11_end_0, end_mask = xr_11_end_mask_0, squeeze_mask = xr_11_squeeze_mask_0, x = pairs_11_cast_fp16)[name = tensor("xr_11_cast_fp16")]; + tensor xi_11_begin_0 = const()[name = tensor("xi_11_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_11_end_0 = const()[name = tensor("xi_11_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_11_end_mask_0 = const()[name = tensor("xi_11_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_11_squeeze_mask_0 = const()[name = tensor("xi_11_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_11_cast_fp16 = slice_by_index(begin = xi_11_begin_0, end = xi_11_end_0, end_mask = xi_11_end_mask_0, squeeze_mask = xi_11_squeeze_mask_0, x = pairs_11_cast_fp16)[name = tensor("xi_11_cast_fp16")]; + tensor var_874_cast_fp16 = mul(x = xr_11_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_874_cast_fp16")]; + tensor var_875_cast_fp16 = mul(x = xi_11_cast_fp16, y = roti_cast_fp16)[name = tensor("op_875_cast_fp16")]; + tensor xor__11_cast_fp16 = sub(x = var_874_cast_fp16, y = var_875_cast_fp16)[name = tensor("xor__11_cast_fp16")]; + tensor var_878_cast_fp16 = mul(x = xr_11_cast_fp16, y = roti_cast_fp16)[name = tensor("op_878_cast_fp16")]; + tensor var_879_cast_fp16 = mul(x = xi_11_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_879_cast_fp16")]; + tensor xoi_11_cast_fp16 = add(x = var_878_cast_fp16, y = var_879_cast_fp16)[name = tensor("xoi_11_cast_fp16")]; + tensor out_11_axis_0 = const()[name = tensor("out_11_axis_0"), val = tensor(-1)]; + tensor out_11_cast_fp16 = stack(axis = out_11_axis_0, values = (xor__11_cast_fp16, xoi_11_cast_fp16))[name = tensor("out_11_cast_fp16")]; + tensor var_896 = const()[name = tensor("op_896"), val = tensor([256, 1024])]; + tensor var_897_cast_fp16 = reshape(shape = var_896, x = out_11_cast_fp16)[name = tensor("op_897_cast_fp16")]; + tensor written_k_5_transpose_x_1 = const()[name = tensor("written_k_5_transpose_x_1"), val = tensor(true)]; + tensor written_k_5_transpose_y_1 = const()[name = tensor("written_k_5_transpose_y_1"), val = tensor(false)]; + tensor written_k_5_cast_fp16 = matmul(transpose_x = written_k_5_transpose_x_1, transpose_y = written_k_5_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_897_cast_fp16)[name = tensor("written_k_5_cast_fp16")]; + tensor var_904 = const()[name = tensor("op_904"), val = tensor([256, 1024])]; + tensor var_905_cast_fp16 = reshape(shape = var_904, x = v_5_cast_fp16)[name = tensor("op_905_cast_fp16")]; + tensor written_v_5_transpose_x_1 = const()[name = tensor("written_v_5_transpose_x_1"), val = tensor(true)]; + tensor written_v_5_transpose_y_1 = const()[name = tensor("written_v_5_transpose_y_1"), val = tensor(false)]; + tensor written_v_5_cast_fp16 = matmul(transpose_x = written_v_5_transpose_x_1, transpose_y = written_v_5_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_905_cast_fp16)[name = tensor("written_v_5_cast_fp16")]; + tensor k_cache2_to_fp16_dtype_0 = const()[name = tensor("k_cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache2_to_fp16 = cast(dtype = k_cache2_to_fp16_dtype_0, x = k_cache2)[name = tensor("cast_453")]; + tensor var_916_cast_fp16 = mul(x = k_cache2_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_916_cast_fp16")]; + tensor var_921 = const()[name = tensor("op_921"), val = tensor([1, 512, 16, 64])]; + tensor var_922_cast_fp16 = reshape(shape = var_921, x = written_k_5_cast_fp16)[name = tensor("op_922_cast_fp16")]; + tensor new_k_cache_5_cast_fp16 = add(x = var_916_cast_fp16, y = var_922_cast_fp16)[name = tensor("new_k_cache_5_cast_fp16")]; + tensor new_k_cache_5_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_5_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache2_to_fp16_dtype_0 = const()[name = tensor("v_cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache2_to_fp16 = cast(dtype = v_cache2_to_fp16_dtype_0, x = v_cache2)[name = tensor("cast_451")]; + tensor var_925_cast_fp16 = mul(x = v_cache2_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_925_cast_fp16")]; + tensor var_930 = const()[name = tensor("op_930"), val = tensor([1, 512, 16, 64])]; + tensor var_931_cast_fp16 = reshape(shape = var_930, x = written_v_5_cast_fp16)[name = tensor("op_931_cast_fp16")]; + tensor new_v_cache_5_cast_fp16 = add(x = var_925_cast_fp16, y = var_931_cast_fp16)[name = tensor("new_v_cache_5_cast_fp16")]; + tensor new_v_cache_5_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_5_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_944 = const()[name = tensor("op_944"), val = tensor([0, 2, 1, 3])]; + tensor var_955_transpose_x_0 = const()[name = tensor("op_955_transpose_x_0"), val = tensor(false)]; + tensor var_955_transpose_y_0 = const()[name = tensor("op_955_transpose_y_0"), val = tensor(false)]; + tensor transpose_73_perm_0 = const()[name = tensor("transpose_73_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_74_perm_0 = const()[name = tensor("transpose_74_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_74 = transpose(perm = transpose_74_perm_0, x = new_k_cache_5_cast_fp16)[name = tensor("transpose_196")]; + tensor transpose_73 = transpose(perm = transpose_73_perm_0, x = q_5_cast_fp16)[name = tensor("transpose_197")]; + tensor var_955_cast_fp16 = matmul(transpose_x = var_955_transpose_x_0, transpose_y = var_955_transpose_y_0, x = transpose_73, y = transpose_74)[name = tensor("op_955_cast_fp16")]; + tensor var_956_to_fp16 = const()[name = tensor("op_956_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_9_cast_fp16 = mul(x = var_955_cast_fp16, y = var_956_to_fp16)[name = tensor("scores_9_cast_fp16")]; + tensor scores_11_cast_fp16 = add(x = scores_9_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_11_cast_fp16")]; + tensor var_966 = const()[name = tensor("op_966"), val = tensor(-1)]; + tensor weights_5_cast_fp16 = softmax(axis = var_966, x = scores_11_cast_fp16)[name = tensor("weights_5_cast_fp16")]; + tensor attn_5_transpose_x_0 = const()[name = tensor("attn_5_transpose_x_0"), val = tensor(false)]; + tensor attn_5_transpose_y_0 = const()[name = tensor("attn_5_transpose_y_0"), val = tensor(false)]; + tensor values_5_cast_fp16 = transpose(perm = var_944, x = new_v_cache_5_cast_fp16)[name = tensor("transpose_198")]; + tensor attn_5_cast_fp16 = matmul(transpose_x = attn_5_transpose_x_0, transpose_y = attn_5_transpose_y_0, x = weights_5_cast_fp16, y = values_5_cast_fp16)[name = tensor("attn_5_cast_fp16")]; + tensor var_974 = const()[name = tensor("op_974"), val = tensor([0, 2, 1, 3])]; + tensor var_979 = const()[name = tensor("op_979"), val = tensor([1, 256, 1024])]; + tensor var_975_cast_fp16 = transpose(perm = var_974, x = attn_5_cast_fp16)[name = tensor("transpose_195")]; + tensor input_29_cast_fp16 = reshape(shape = var_979, x = var_975_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor attn2_out_proj_weight_to_fp16 = const()[name = tensor("attn2_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(56663808)))]; + tensor linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn2_out_proj_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor position2_to_fp16_dtype_0 = const()[name = tensor("position2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position2_to_fp16 = cast(dtype = position2_to_fp16_dtype_0, x = position2)[name = tensor("cast_449")]; + tensor var_988_cast_fp16 = add(x = position2_to_fp16, y = valid_len_to_fp16)[name = tensor("op_988_cast_fp16")]; + tensor var_988_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_988_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_31_cast_fp16 = add(x = input_25_cast_fp16, y = linear_9_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; + tensor norm2_2_weight_to_fp16 = const()[name = tensor("norm2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58761024)))]; + tensor norm2_2_bias_to_fp16 = const()[name = tensor("norm2_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58763136)))]; + tensor var_992_to_fp16 = const()[name = tensor("op_992_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_33_cast_fp16 = layer_norm(axes = input_33_axes_0, beta = norm2_2_bias_to_fp16, epsilon = var_992_to_fp16, gamma = norm2_2_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor linear2_1_weight_to_fp16 = const()[name = tensor("linear2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58765248)))]; + tensor linear_10_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear2_1_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor input_35_mode_0 = const()[name = tensor("input_35_mode_0"), val = tensor("EXACT")]; + tensor input_35_cast_fp16 = gelu(mode = input_35_mode_0, x = linear_10_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor linear2_2_weight_to_fp16 = const()[name = tensor("linear2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67153920)))]; + tensor linear_11_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear2_2_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor input_37_cast_fp16 = add(x = input_31_cast_fp16, y = linear_11_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor input_39_axes_0 = const()[name = tensor("input_39_axes_0"), val = tensor([-1])]; + tensor norm3_1_weight_to_fp16 = const()[name = tensor("norm3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75542592)))]; + tensor norm3_1_bias_to_fp16 = const()[name = tensor("norm3_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75544704)))]; + tensor var_1009_to_fp16 = const()[name = tensor("op_1009_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_39_cast_fp16 = layer_norm(axes = input_39_axes_0, beta = norm3_1_bias_to_fp16, epsilon = var_1009_to_fp16, gamma = norm3_1_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor attn3_in_proj_weight_to_fp16 = const()[name = tensor("attn3_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75546816)))]; + tensor linear_12_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn3_in_proj_weight_to_fp16, x = input_39_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor var_1023 = const()[name = tensor("op_1023"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_7_cast_fp16 = reshape(shape = var_1023, x = linear_12_cast_fp16)[name = tensor("qkv_7_cast_fp16")]; + tensor x_13_begin_0 = const()[name = tensor("x_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_13_end_0 = const()[name = tensor("x_13_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_13_end_mask_0 = const()[name = tensor("x_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_13_squeeze_mask_0 = const()[name = tensor("x_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_13_cast_fp16 = slice_by_index(begin = x_13_begin_0, end = x_13_end_0, end_mask = x_13_end_mask_0, squeeze_mask = x_13_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("x_13_cast_fp16")]; + tensor x_15_begin_0 = const()[name = tensor("x_15_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_15_end_0 = const()[name = tensor("x_15_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_15_end_mask_0 = const()[name = tensor("x_15_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_15_squeeze_mask_0 = const()[name = tensor("x_15_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_15_cast_fp16 = slice_by_index(begin = x_15_begin_0, end = x_15_end_0, end_mask = x_15_end_mask_0, squeeze_mask = x_15_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("x_15_cast_fp16")]; + tensor v_7_begin_0 = const()[name = tensor("v_7_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_7_end_0 = const()[name = tensor("v_7_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_7_end_mask_0 = const()[name = tensor("v_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_7_squeeze_mask_0 = const()[name = tensor("v_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_7_cast_fp16 = slice_by_index(begin = v_7_begin_0, end = v_7_end_0, end_mask = v_7_end_mask_0, squeeze_mask = v_7_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("v_7_cast_fp16")]; + tensor var_1068 = const()[name = tensor("op_1068"), val = tensor([256, 16, 32, 2])]; + tensor pairs_13_cast_fp16 = reshape(shape = var_1068, x = x_13_cast_fp16)[name = tensor("pairs_13_cast_fp16")]; + tensor xr_13_begin_0 = const()[name = tensor("xr_13_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_13_end_0 = const()[name = tensor("xr_13_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_13_end_mask_0 = const()[name = tensor("xr_13_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_13_squeeze_mask_0 = const()[name = tensor("xr_13_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_13_cast_fp16 = slice_by_index(begin = xr_13_begin_0, end = xr_13_end_0, end_mask = xr_13_end_mask_0, squeeze_mask = xr_13_squeeze_mask_0, x = pairs_13_cast_fp16)[name = tensor("xr_13_cast_fp16")]; + tensor xi_13_begin_0 = const()[name = tensor("xi_13_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_13_end_0 = const()[name = tensor("xi_13_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_13_end_mask_0 = const()[name = tensor("xi_13_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_13_squeeze_mask_0 = const()[name = tensor("xi_13_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_13_cast_fp16 = slice_by_index(begin = xi_13_begin_0, end = xi_13_end_0, end_mask = xi_13_end_mask_0, squeeze_mask = xi_13_squeeze_mask_0, x = pairs_13_cast_fp16)[name = tensor("xi_13_cast_fp16")]; + tensor var_1076_cast_fp16 = mul(x = xr_13_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1076_cast_fp16")]; + tensor var_1077_cast_fp16 = mul(x = xi_13_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1077_cast_fp16")]; + tensor xor__13_cast_fp16 = sub(x = var_1076_cast_fp16, y = var_1077_cast_fp16)[name = tensor("xor__13_cast_fp16")]; + tensor var_1080_cast_fp16 = mul(x = xr_13_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1080_cast_fp16")]; + tensor var_1081_cast_fp16 = mul(x = xi_13_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1081_cast_fp16")]; + tensor xoi_13_cast_fp16 = add(x = var_1080_cast_fp16, y = var_1081_cast_fp16)[name = tensor("xoi_13_cast_fp16")]; + tensor out_13_axis_0 = const()[name = tensor("out_13_axis_0"), val = tensor(-1)]; + tensor out_13_cast_fp16 = stack(axis = out_13_axis_0, values = (xor__13_cast_fp16, xoi_13_cast_fp16))[name = tensor("out_13_cast_fp16")]; + tensor var_1091 = const()[name = tensor("op_1091"), val = tensor([1, 256, 16, 64])]; + tensor q_7_cast_fp16 = reshape(shape = var_1091, x = out_13_cast_fp16)[name = tensor("q_7_cast_fp16")]; + tensor var_1097 = const()[name = tensor("op_1097"), val = tensor([256, 16, 32, 2])]; + tensor pairs_15_cast_fp16 = reshape(shape = var_1097, x = x_15_cast_fp16)[name = tensor("pairs_15_cast_fp16")]; + tensor xr_15_begin_0 = const()[name = tensor("xr_15_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_15_end_0 = const()[name = tensor("xr_15_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_15_end_mask_0 = const()[name = tensor("xr_15_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_15_squeeze_mask_0 = const()[name = tensor("xr_15_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_15_cast_fp16 = slice_by_index(begin = xr_15_begin_0, end = xr_15_end_0, end_mask = xr_15_end_mask_0, squeeze_mask = xr_15_squeeze_mask_0, x = pairs_15_cast_fp16)[name = tensor("xr_15_cast_fp16")]; + tensor xi_15_begin_0 = const()[name = tensor("xi_15_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_15_end_0 = const()[name = tensor("xi_15_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_15_end_mask_0 = const()[name = tensor("xi_15_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_15_squeeze_mask_0 = const()[name = tensor("xi_15_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_15_cast_fp16 = slice_by_index(begin = xi_15_begin_0, end = xi_15_end_0, end_mask = xi_15_end_mask_0, squeeze_mask = xi_15_squeeze_mask_0, x = pairs_15_cast_fp16)[name = tensor("xi_15_cast_fp16")]; + tensor var_1105_cast_fp16 = mul(x = xr_15_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1105_cast_fp16")]; + tensor var_1106_cast_fp16 = mul(x = xi_15_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1106_cast_fp16")]; + tensor xor__15_cast_fp16 = sub(x = var_1105_cast_fp16, y = var_1106_cast_fp16)[name = tensor("xor__15_cast_fp16")]; + tensor var_1109_cast_fp16 = mul(x = xr_15_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1109_cast_fp16")]; + tensor var_1110_cast_fp16 = mul(x = xi_15_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1110_cast_fp16")]; + tensor xoi_15_cast_fp16 = add(x = var_1109_cast_fp16, y = var_1110_cast_fp16)[name = tensor("xoi_15_cast_fp16")]; + tensor out_15_axis_0 = const()[name = tensor("out_15_axis_0"), val = tensor(-1)]; + tensor out_15_cast_fp16 = stack(axis = out_15_axis_0, values = (xor__15_cast_fp16, xoi_15_cast_fp16))[name = tensor("out_15_cast_fp16")]; + tensor var_1127 = const()[name = tensor("op_1127"), val = tensor([256, 1024])]; + tensor var_1128_cast_fp16 = reshape(shape = var_1127, x = out_15_cast_fp16)[name = tensor("op_1128_cast_fp16")]; + tensor written_k_7_transpose_x_1 = const()[name = tensor("written_k_7_transpose_x_1"), val = tensor(true)]; + tensor written_k_7_transpose_y_1 = const()[name = tensor("written_k_7_transpose_y_1"), val = tensor(false)]; + tensor written_k_7_cast_fp16 = matmul(transpose_x = written_k_7_transpose_x_1, transpose_y = written_k_7_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_1128_cast_fp16)[name = tensor("written_k_7_cast_fp16")]; + tensor var_1135 = const()[name = tensor("op_1135"), val = tensor([256, 1024])]; + tensor var_1136_cast_fp16 = reshape(shape = var_1135, x = v_7_cast_fp16)[name = tensor("op_1136_cast_fp16")]; + tensor written_v_7_transpose_x_1 = const()[name = tensor("written_v_7_transpose_x_1"), val = tensor(true)]; + tensor written_v_7_transpose_y_1 = const()[name = tensor("written_v_7_transpose_y_1"), val = tensor(false)]; + tensor written_v_7_cast_fp16 = matmul(transpose_x = written_v_7_transpose_x_1, transpose_y = written_v_7_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_1136_cast_fp16)[name = tensor("written_v_7_cast_fp16")]; + tensor k_cache3_to_fp16_dtype_0 = const()[name = tensor("k_cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache3_to_fp16 = cast(dtype = k_cache3_to_fp16_dtype_0, x = k_cache3)[name = tensor("cast_447")]; + tensor var_1147_cast_fp16 = mul(x = k_cache3_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_1147_cast_fp16")]; + tensor var_1152 = const()[name = tensor("op_1152"), val = tensor([1, 512, 16, 64])]; + tensor var_1153_cast_fp16 = reshape(shape = var_1152, x = written_k_7_cast_fp16)[name = tensor("op_1153_cast_fp16")]; + tensor new_k_cache_7_cast_fp16 = add(x = var_1147_cast_fp16, y = var_1153_cast_fp16)[name = tensor("new_k_cache_7_cast_fp16")]; + tensor new_k_cache_7_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_7_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache3_to_fp16_dtype_0 = const()[name = tensor("v_cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache3_to_fp16 = cast(dtype = v_cache3_to_fp16_dtype_0, x = v_cache3)[name = tensor("cast_445")]; + tensor var_1156_cast_fp16 = mul(x = v_cache3_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_1156_cast_fp16")]; + tensor var_1161 = const()[name = tensor("op_1161"), val = tensor([1, 512, 16, 64])]; + tensor var_1162_cast_fp16 = reshape(shape = var_1161, x = written_v_7_cast_fp16)[name = tensor("op_1162_cast_fp16")]; + tensor new_v_cache_7_cast_fp16 = add(x = var_1156_cast_fp16, y = var_1162_cast_fp16)[name = tensor("new_v_cache_7_cast_fp16")]; + tensor new_v_cache_7_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_7_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_1175 = const()[name = tensor("op_1175"), val = tensor([0, 2, 1, 3])]; + tensor var_1186_transpose_x_0 = const()[name = tensor("op_1186_transpose_x_0"), val = tensor(false)]; + tensor var_1186_transpose_y_0 = const()[name = tensor("op_1186_transpose_y_0"), val = tensor(false)]; + tensor transpose_75_perm_0 = const()[name = tensor("transpose_75_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_76_perm_0 = const()[name = tensor("transpose_76_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_76 = transpose(perm = transpose_76_perm_0, x = new_k_cache_7_cast_fp16)[name = tensor("transpose_192")]; + tensor transpose_75 = transpose(perm = transpose_75_perm_0, x = q_7_cast_fp16)[name = tensor("transpose_193")]; + tensor var_1186_cast_fp16 = matmul(transpose_x = var_1186_transpose_x_0, transpose_y = var_1186_transpose_y_0, x = transpose_75, y = transpose_76)[name = tensor("op_1186_cast_fp16")]; + tensor var_1187_to_fp16 = const()[name = tensor("op_1187_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_13_cast_fp16 = mul(x = var_1186_cast_fp16, y = var_1187_to_fp16)[name = tensor("scores_13_cast_fp16")]; + tensor scores_15_cast_fp16 = add(x = scores_13_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_15_cast_fp16")]; + tensor var_1197 = const()[name = tensor("op_1197"), val = tensor(-1)]; + tensor weights_7_cast_fp16 = softmax(axis = var_1197, x = scores_15_cast_fp16)[name = tensor("weights_7_cast_fp16")]; + tensor attn_7_transpose_x_0 = const()[name = tensor("attn_7_transpose_x_0"), val = tensor(false)]; + tensor attn_7_transpose_y_0 = const()[name = tensor("attn_7_transpose_y_0"), val = tensor(false)]; + tensor values_7_cast_fp16 = transpose(perm = var_1175, x = new_v_cache_7_cast_fp16)[name = tensor("transpose_194")]; + tensor attn_7_cast_fp16 = matmul(transpose_x = attn_7_transpose_x_0, transpose_y = attn_7_transpose_y_0, x = weights_7_cast_fp16, y = values_7_cast_fp16)[name = tensor("attn_7_cast_fp16")]; + tensor var_1205 = const()[name = tensor("op_1205"), val = tensor([0, 2, 1, 3])]; + tensor var_1210 = const()[name = tensor("op_1210"), val = tensor([1, 256, 1024])]; + tensor var_1206_cast_fp16 = transpose(perm = var_1205, x = attn_7_cast_fp16)[name = tensor("transpose_191")]; + tensor input_41_cast_fp16 = reshape(shape = var_1210, x = var_1206_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor attn3_out_proj_weight_to_fp16 = const()[name = tensor("attn3_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(81838336)))]; + tensor linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn3_out_proj_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor position3_to_fp16_dtype_0 = const()[name = tensor("position3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position3_to_fp16 = cast(dtype = position3_to_fp16_dtype_0, x = position3)[name = tensor("cast_443")]; + tensor var_1219_cast_fp16 = add(x = position3_to_fp16, y = valid_len_to_fp16)[name = tensor("op_1219_cast_fp16")]; + tensor var_1219_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1219_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_43_cast_fp16 = add(x = input_37_cast_fp16, y = linear_13_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; + tensor norm3_2_weight_to_fp16 = const()[name = tensor("norm3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(83935552)))]; + tensor norm3_2_bias_to_fp16 = const()[name = tensor("norm3_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(83937664)))]; + tensor var_1223_to_fp16 = const()[name = tensor("op_1223_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_45_cast_fp16 = layer_norm(axes = input_45_axes_0, beta = norm3_2_bias_to_fp16, epsilon = var_1223_to_fp16, gamma = norm3_2_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor linear3_1_weight_to_fp16 = const()[name = tensor("linear3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(83939776)))]; + tensor linear_14_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear3_1_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor input_47_mode_0 = const()[name = tensor("input_47_mode_0"), val = tensor("EXACT")]; + tensor input_47_cast_fp16 = gelu(mode = input_47_mode_0, x = linear_14_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor linear3_2_weight_to_fp16 = const()[name = tensor("linear3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(92328448)))]; + tensor linear_15_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear3_2_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor input_49_cast_fp16 = add(x = input_43_cast_fp16, y = linear_15_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor input_51_axes_0 = const()[name = tensor("input_51_axes_0"), val = tensor([-1])]; + tensor norm4_1_weight_to_fp16 = const()[name = tensor("norm4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100717120)))]; + tensor norm4_1_bias_to_fp16 = const()[name = tensor("norm4_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100719232)))]; + tensor var_1240_to_fp16 = const()[name = tensor("op_1240_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_51_cast_fp16 = layer_norm(axes = input_51_axes_0, beta = norm4_1_bias_to_fp16, epsilon = var_1240_to_fp16, gamma = norm4_1_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor attn4_in_proj_weight_to_fp16 = const()[name = tensor("attn4_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100721344)))]; + tensor linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn4_in_proj_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor var_1254 = const()[name = tensor("op_1254"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_9_cast_fp16 = reshape(shape = var_1254, x = linear_16_cast_fp16)[name = tensor("qkv_9_cast_fp16")]; + tensor x_17_begin_0 = const()[name = tensor("x_17_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_17_end_0 = const()[name = tensor("x_17_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_17_end_mask_0 = const()[name = tensor("x_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_17_squeeze_mask_0 = const()[name = tensor("x_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_17_cast_fp16 = slice_by_index(begin = x_17_begin_0, end = x_17_end_0, end_mask = x_17_end_mask_0, squeeze_mask = x_17_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("x_17_cast_fp16")]; + tensor x_19_begin_0 = const()[name = tensor("x_19_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_19_end_0 = const()[name = tensor("x_19_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_19_end_mask_0 = const()[name = tensor("x_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_19_squeeze_mask_0 = const()[name = tensor("x_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_19_cast_fp16 = slice_by_index(begin = x_19_begin_0, end = x_19_end_0, end_mask = x_19_end_mask_0, squeeze_mask = x_19_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("x_19_cast_fp16")]; + tensor v_9_begin_0 = const()[name = tensor("v_9_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_9_end_0 = const()[name = tensor("v_9_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_9_end_mask_0 = const()[name = tensor("v_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_9_squeeze_mask_0 = const()[name = tensor("v_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_9_cast_fp16 = slice_by_index(begin = v_9_begin_0, end = v_9_end_0, end_mask = v_9_end_mask_0, squeeze_mask = v_9_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("v_9_cast_fp16")]; + tensor var_1299 = const()[name = tensor("op_1299"), val = tensor([256, 16, 32, 2])]; + tensor pairs_17_cast_fp16 = reshape(shape = var_1299, x = x_17_cast_fp16)[name = tensor("pairs_17_cast_fp16")]; + tensor xr_17_begin_0 = const()[name = tensor("xr_17_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_17_end_0 = const()[name = tensor("xr_17_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_17_end_mask_0 = const()[name = tensor("xr_17_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_17_squeeze_mask_0 = const()[name = tensor("xr_17_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_17_cast_fp16 = slice_by_index(begin = xr_17_begin_0, end = xr_17_end_0, end_mask = xr_17_end_mask_0, squeeze_mask = xr_17_squeeze_mask_0, x = pairs_17_cast_fp16)[name = tensor("xr_17_cast_fp16")]; + tensor xi_17_begin_0 = const()[name = tensor("xi_17_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_17_end_0 = const()[name = tensor("xi_17_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_17_end_mask_0 = const()[name = tensor("xi_17_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_17_squeeze_mask_0 = const()[name = tensor("xi_17_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_17_cast_fp16 = slice_by_index(begin = xi_17_begin_0, end = xi_17_end_0, end_mask = xi_17_end_mask_0, squeeze_mask = xi_17_squeeze_mask_0, x = pairs_17_cast_fp16)[name = tensor("xi_17_cast_fp16")]; + tensor var_1307_cast_fp16 = mul(x = xr_17_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1307_cast_fp16")]; + tensor var_1308_cast_fp16 = mul(x = xi_17_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1308_cast_fp16")]; + tensor xor__17_cast_fp16 = sub(x = var_1307_cast_fp16, y = var_1308_cast_fp16)[name = tensor("xor__17_cast_fp16")]; + tensor var_1311_cast_fp16 = mul(x = xr_17_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1311_cast_fp16")]; + tensor var_1312_cast_fp16 = mul(x = xi_17_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1312_cast_fp16")]; + tensor xoi_17_cast_fp16 = add(x = var_1311_cast_fp16, y = var_1312_cast_fp16)[name = tensor("xoi_17_cast_fp16")]; + tensor out_17_axis_0 = const()[name = tensor("out_17_axis_0"), val = tensor(-1)]; + tensor out_17_cast_fp16 = stack(axis = out_17_axis_0, values = (xor__17_cast_fp16, xoi_17_cast_fp16))[name = tensor("out_17_cast_fp16")]; + tensor var_1322 = const()[name = tensor("op_1322"), val = tensor([1, 256, 16, 64])]; + tensor q_9_cast_fp16 = reshape(shape = var_1322, x = out_17_cast_fp16)[name = tensor("q_9_cast_fp16")]; + tensor var_1328 = const()[name = tensor("op_1328"), val = tensor([256, 16, 32, 2])]; + tensor pairs_19_cast_fp16 = reshape(shape = var_1328, x = x_19_cast_fp16)[name = tensor("pairs_19_cast_fp16")]; + tensor xr_19_begin_0 = const()[name = tensor("xr_19_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_19_end_0 = const()[name = tensor("xr_19_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_19_end_mask_0 = const()[name = tensor("xr_19_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_19_squeeze_mask_0 = const()[name = tensor("xr_19_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_19_cast_fp16 = slice_by_index(begin = xr_19_begin_0, end = xr_19_end_0, end_mask = xr_19_end_mask_0, squeeze_mask = xr_19_squeeze_mask_0, x = pairs_19_cast_fp16)[name = tensor("xr_19_cast_fp16")]; + tensor xi_19_begin_0 = const()[name = tensor("xi_19_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_19_end_0 = const()[name = tensor("xi_19_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_19_end_mask_0 = const()[name = tensor("xi_19_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_19_squeeze_mask_0 = const()[name = tensor("xi_19_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_19_cast_fp16 = slice_by_index(begin = xi_19_begin_0, end = xi_19_end_0, end_mask = xi_19_end_mask_0, squeeze_mask = xi_19_squeeze_mask_0, x = pairs_19_cast_fp16)[name = tensor("xi_19_cast_fp16")]; + tensor var_1336_cast_fp16 = mul(x = xr_19_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1336_cast_fp16")]; + tensor var_1337_cast_fp16 = mul(x = xi_19_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1337_cast_fp16")]; + tensor xor__19_cast_fp16 = sub(x = var_1336_cast_fp16, y = var_1337_cast_fp16)[name = tensor("xor__19_cast_fp16")]; + tensor var_1340_cast_fp16 = mul(x = xr_19_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1340_cast_fp16")]; + tensor var_1341_cast_fp16 = mul(x = xi_19_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1341_cast_fp16")]; + tensor xoi_19_cast_fp16 = add(x = var_1340_cast_fp16, y = var_1341_cast_fp16)[name = tensor("xoi_19_cast_fp16")]; + tensor out_19_axis_0 = const()[name = tensor("out_19_axis_0"), val = tensor(-1)]; + tensor out_19_cast_fp16 = stack(axis = out_19_axis_0, values = (xor__19_cast_fp16, xoi_19_cast_fp16))[name = tensor("out_19_cast_fp16")]; + tensor var_1358 = const()[name = tensor("op_1358"), val = tensor([256, 1024])]; + tensor var_1359_cast_fp16 = reshape(shape = var_1358, x = out_19_cast_fp16)[name = tensor("op_1359_cast_fp16")]; + tensor written_k_9_transpose_x_1 = const()[name = tensor("written_k_9_transpose_x_1"), val = tensor(true)]; + tensor written_k_9_transpose_y_1 = const()[name = tensor("written_k_9_transpose_y_1"), val = tensor(false)]; + tensor written_k_9_cast_fp16 = matmul(transpose_x = written_k_9_transpose_x_1, transpose_y = written_k_9_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_1359_cast_fp16)[name = tensor("written_k_9_cast_fp16")]; + tensor var_1366 = const()[name = tensor("op_1366"), val = tensor([256, 1024])]; + tensor var_1367_cast_fp16 = reshape(shape = var_1366, x = v_9_cast_fp16)[name = tensor("op_1367_cast_fp16")]; + tensor written_v_9_transpose_x_1 = const()[name = tensor("written_v_9_transpose_x_1"), val = tensor(true)]; + tensor written_v_9_transpose_y_1 = const()[name = tensor("written_v_9_transpose_y_1"), val = tensor(false)]; + tensor written_v_9_cast_fp16 = matmul(transpose_x = written_v_9_transpose_x_1, transpose_y = written_v_9_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_1367_cast_fp16)[name = tensor("written_v_9_cast_fp16")]; + tensor k_cache4_to_fp16_dtype_0 = const()[name = tensor("k_cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache4_to_fp16 = cast(dtype = k_cache4_to_fp16_dtype_0, x = k_cache4)[name = tensor("cast_441")]; + tensor var_1378_cast_fp16 = mul(x = k_cache4_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_1378_cast_fp16")]; + tensor var_1383 = const()[name = tensor("op_1383"), val = tensor([1, 512, 16, 64])]; + tensor var_1384_cast_fp16 = reshape(shape = var_1383, x = written_k_9_cast_fp16)[name = tensor("op_1384_cast_fp16")]; + tensor new_k_cache_9_cast_fp16 = add(x = var_1378_cast_fp16, y = var_1384_cast_fp16)[name = tensor("new_k_cache_9_cast_fp16")]; + tensor new_k_cache_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache4_to_fp16_dtype_0 = const()[name = tensor("v_cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache4_to_fp16 = cast(dtype = v_cache4_to_fp16_dtype_0, x = v_cache4)[name = tensor("cast_439")]; + tensor var_1387_cast_fp16 = mul(x = v_cache4_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_1387_cast_fp16")]; + tensor var_1392 = const()[name = tensor("op_1392"), val = tensor([1, 512, 16, 64])]; + tensor var_1393_cast_fp16 = reshape(shape = var_1392, x = written_v_9_cast_fp16)[name = tensor("op_1393_cast_fp16")]; + tensor new_v_cache_9_cast_fp16 = add(x = var_1387_cast_fp16, y = var_1393_cast_fp16)[name = tensor("new_v_cache_9_cast_fp16")]; + tensor new_v_cache_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_1406 = const()[name = tensor("op_1406"), val = tensor([0, 2, 1, 3])]; + tensor var_1417_transpose_x_0 = const()[name = tensor("op_1417_transpose_x_0"), val = tensor(false)]; + tensor var_1417_transpose_y_0 = const()[name = tensor("op_1417_transpose_y_0"), val = tensor(false)]; + tensor transpose_77_perm_0 = const()[name = tensor("transpose_77_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_78_perm_0 = const()[name = tensor("transpose_78_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_78 = transpose(perm = transpose_78_perm_0, x = new_k_cache_9_cast_fp16)[name = tensor("transpose_188")]; + tensor transpose_77 = transpose(perm = transpose_77_perm_0, x = q_9_cast_fp16)[name = tensor("transpose_189")]; + tensor var_1417_cast_fp16 = matmul(transpose_x = var_1417_transpose_x_0, transpose_y = var_1417_transpose_y_0, x = transpose_77, y = transpose_78)[name = tensor("op_1417_cast_fp16")]; + tensor var_1418_to_fp16 = const()[name = tensor("op_1418_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_17_cast_fp16 = mul(x = var_1417_cast_fp16, y = var_1418_to_fp16)[name = tensor("scores_17_cast_fp16")]; + tensor scores_19_cast_fp16 = add(x = scores_17_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_19_cast_fp16")]; + tensor var_1428 = const()[name = tensor("op_1428"), val = tensor(-1)]; + tensor weights_9_cast_fp16 = softmax(axis = var_1428, x = scores_19_cast_fp16)[name = tensor("weights_9_cast_fp16")]; + tensor attn_9_transpose_x_0 = const()[name = tensor("attn_9_transpose_x_0"), val = tensor(false)]; + tensor attn_9_transpose_y_0 = const()[name = tensor("attn_9_transpose_y_0"), val = tensor(false)]; + tensor values_9_cast_fp16 = transpose(perm = var_1406, x = new_v_cache_9_cast_fp16)[name = tensor("transpose_190")]; + tensor attn_9_cast_fp16 = matmul(transpose_x = attn_9_transpose_x_0, transpose_y = attn_9_transpose_y_0, x = weights_9_cast_fp16, y = values_9_cast_fp16)[name = tensor("attn_9_cast_fp16")]; + tensor var_1436 = const()[name = tensor("op_1436"), val = tensor([0, 2, 1, 3])]; + tensor var_1441 = const()[name = tensor("op_1441"), val = tensor([1, 256, 1024])]; + tensor var_1437_cast_fp16 = transpose(perm = var_1436, x = attn_9_cast_fp16)[name = tensor("transpose_187")]; + tensor input_53_cast_fp16 = reshape(shape = var_1441, x = var_1437_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor attn4_out_proj_weight_to_fp16 = const()[name = tensor("attn4_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(107012864)))]; + tensor linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn4_out_proj_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor position4_to_fp16_dtype_0 = const()[name = tensor("position4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position4_to_fp16 = cast(dtype = position4_to_fp16_dtype_0, x = position4)[name = tensor("cast_437")]; + tensor var_1450_cast_fp16 = add(x = position4_to_fp16, y = valid_len_to_fp16)[name = tensor("op_1450_cast_fp16")]; + tensor var_1450_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1450_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_55_cast_fp16 = add(x = input_49_cast_fp16, y = linear_17_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor input_57_axes_0 = const()[name = tensor("input_57_axes_0"), val = tensor([-1])]; + tensor norm4_2_weight_to_fp16 = const()[name = tensor("norm4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109110080)))]; + tensor norm4_2_bias_to_fp16 = const()[name = tensor("norm4_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109112192)))]; + tensor var_1454_to_fp16 = const()[name = tensor("op_1454_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_57_cast_fp16 = layer_norm(axes = input_57_axes_0, beta = norm4_2_bias_to_fp16, epsilon = var_1454_to_fp16, gamma = norm4_2_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor linear4_1_weight_to_fp16 = const()[name = tensor("linear4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109114304)))]; + tensor linear_18_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear4_1_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor input_59_mode_0 = const()[name = tensor("input_59_mode_0"), val = tensor("EXACT")]; + tensor input_59_cast_fp16 = gelu(mode = input_59_mode_0, x = linear_18_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor linear4_2_weight_to_fp16 = const()[name = tensor("linear4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(117502976)))]; + tensor linear_19_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear4_2_weight_to_fp16, x = input_59_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_61_cast_fp16 = add(x = input_55_cast_fp16, y = linear_19_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor input_63_axes_0 = const()[name = tensor("input_63_axes_0"), val = tensor([-1])]; + tensor norm5_1_weight_to_fp16 = const()[name = tensor("norm5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125891648)))]; + tensor norm5_1_bias_to_fp16 = const()[name = tensor("norm5_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125893760)))]; + tensor var_1471_to_fp16 = const()[name = tensor("op_1471_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_63_cast_fp16 = layer_norm(axes = input_63_axes_0, beta = norm5_1_bias_to_fp16, epsilon = var_1471_to_fp16, gamma = norm5_1_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor attn5_in_proj_weight_to_fp16 = const()[name = tensor("attn5_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125895872)))]; + tensor linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn5_in_proj_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor var_1485 = const()[name = tensor("op_1485"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_11_cast_fp16 = reshape(shape = var_1485, x = linear_20_cast_fp16)[name = tensor("qkv_11_cast_fp16")]; + tensor x_21_begin_0 = const()[name = tensor("x_21_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_21_end_0 = const()[name = tensor("x_21_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_21_end_mask_0 = const()[name = tensor("x_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_21_squeeze_mask_0 = const()[name = tensor("x_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_21_cast_fp16 = slice_by_index(begin = x_21_begin_0, end = x_21_end_0, end_mask = x_21_end_mask_0, squeeze_mask = x_21_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("x_21_cast_fp16")]; + tensor x_23_begin_0 = const()[name = tensor("x_23_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_23_end_0 = const()[name = tensor("x_23_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_23_end_mask_0 = const()[name = tensor("x_23_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_23_squeeze_mask_0 = const()[name = tensor("x_23_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_23_cast_fp16 = slice_by_index(begin = x_23_begin_0, end = x_23_end_0, end_mask = x_23_end_mask_0, squeeze_mask = x_23_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("x_23_cast_fp16")]; + tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_11_cast_fp16 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("v_11_cast_fp16")]; + tensor var_1530 = const()[name = tensor("op_1530"), val = tensor([256, 16, 32, 2])]; + tensor pairs_21_cast_fp16 = reshape(shape = var_1530, x = x_21_cast_fp16)[name = tensor("pairs_21_cast_fp16")]; + tensor xr_21_begin_0 = const()[name = tensor("xr_21_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_21_end_0 = const()[name = tensor("xr_21_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_21_end_mask_0 = const()[name = tensor("xr_21_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_21_squeeze_mask_0 = const()[name = tensor("xr_21_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_21_cast_fp16 = slice_by_index(begin = xr_21_begin_0, end = xr_21_end_0, end_mask = xr_21_end_mask_0, squeeze_mask = xr_21_squeeze_mask_0, x = pairs_21_cast_fp16)[name = tensor("xr_21_cast_fp16")]; + tensor xi_21_begin_0 = const()[name = tensor("xi_21_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_21_end_0 = const()[name = tensor("xi_21_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_21_end_mask_0 = const()[name = tensor("xi_21_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_21_squeeze_mask_0 = const()[name = tensor("xi_21_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_21_cast_fp16 = slice_by_index(begin = xi_21_begin_0, end = xi_21_end_0, end_mask = xi_21_end_mask_0, squeeze_mask = xi_21_squeeze_mask_0, x = pairs_21_cast_fp16)[name = tensor("xi_21_cast_fp16")]; + tensor var_1538_cast_fp16 = mul(x = xr_21_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1538_cast_fp16")]; + tensor var_1539_cast_fp16 = mul(x = xi_21_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1539_cast_fp16")]; + tensor xor__21_cast_fp16 = sub(x = var_1538_cast_fp16, y = var_1539_cast_fp16)[name = tensor("xor__21_cast_fp16")]; + tensor var_1542_cast_fp16 = mul(x = xr_21_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1542_cast_fp16")]; + tensor var_1543_cast_fp16 = mul(x = xi_21_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1543_cast_fp16")]; + tensor xoi_21_cast_fp16 = add(x = var_1542_cast_fp16, y = var_1543_cast_fp16)[name = tensor("xoi_21_cast_fp16")]; + tensor out_21_axis_0 = const()[name = tensor("out_21_axis_0"), val = tensor(-1)]; + tensor out_21_cast_fp16 = stack(axis = out_21_axis_0, values = (xor__21_cast_fp16, xoi_21_cast_fp16))[name = tensor("out_21_cast_fp16")]; + tensor var_1553 = const()[name = tensor("op_1553"), val = tensor([1, 256, 16, 64])]; + tensor q_11_cast_fp16 = reshape(shape = var_1553, x = out_21_cast_fp16)[name = tensor("q_11_cast_fp16")]; + tensor var_1559 = const()[name = tensor("op_1559"), val = tensor([256, 16, 32, 2])]; + tensor pairs_23_cast_fp16 = reshape(shape = var_1559, x = x_23_cast_fp16)[name = tensor("pairs_23_cast_fp16")]; + tensor xr_23_begin_0 = const()[name = tensor("xr_23_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_23_end_0 = const()[name = tensor("xr_23_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_23_end_mask_0 = const()[name = tensor("xr_23_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_23_squeeze_mask_0 = const()[name = tensor("xr_23_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_23_cast_fp16 = slice_by_index(begin = xr_23_begin_0, end = xr_23_end_0, end_mask = xr_23_end_mask_0, squeeze_mask = xr_23_squeeze_mask_0, x = pairs_23_cast_fp16)[name = tensor("xr_23_cast_fp16")]; + tensor xi_23_begin_0 = const()[name = tensor("xi_23_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_23_end_0 = const()[name = tensor("xi_23_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_23_end_mask_0 = const()[name = tensor("xi_23_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_23_squeeze_mask_0 = const()[name = tensor("xi_23_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_23_cast_fp16 = slice_by_index(begin = xi_23_begin_0, end = xi_23_end_0, end_mask = xi_23_end_mask_0, squeeze_mask = xi_23_squeeze_mask_0, x = pairs_23_cast_fp16)[name = tensor("xi_23_cast_fp16")]; + tensor var_1567_cast_fp16 = mul(x = xr_23_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1567_cast_fp16")]; + tensor var_1568_cast_fp16 = mul(x = xi_23_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1568_cast_fp16")]; + tensor xor__23_cast_fp16 = sub(x = var_1567_cast_fp16, y = var_1568_cast_fp16)[name = tensor("xor__23_cast_fp16")]; + tensor var_1571_cast_fp16 = mul(x = xr_23_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1571_cast_fp16")]; + tensor var_1572_cast_fp16 = mul(x = xi_23_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1572_cast_fp16")]; + tensor xoi_23_cast_fp16 = add(x = var_1571_cast_fp16, y = var_1572_cast_fp16)[name = tensor("xoi_23_cast_fp16")]; + tensor out_23_axis_0 = const()[name = tensor("out_23_axis_0"), val = tensor(-1)]; + tensor out_23_cast_fp16 = stack(axis = out_23_axis_0, values = (xor__23_cast_fp16, xoi_23_cast_fp16))[name = tensor("out_23_cast_fp16")]; + tensor var_1589 = const()[name = tensor("op_1589"), val = tensor([256, 1024])]; + tensor var_1590_cast_fp16 = reshape(shape = var_1589, x = out_23_cast_fp16)[name = tensor("op_1590_cast_fp16")]; + tensor written_k_11_transpose_x_1 = const()[name = tensor("written_k_11_transpose_x_1"), val = tensor(true)]; + tensor written_k_11_transpose_y_1 = const()[name = tensor("written_k_11_transpose_y_1"), val = tensor(false)]; + tensor written_k_11_cast_fp16 = matmul(transpose_x = written_k_11_transpose_x_1, transpose_y = written_k_11_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_1590_cast_fp16)[name = tensor("written_k_11_cast_fp16")]; + tensor var_1597 = const()[name = tensor("op_1597"), val = tensor([256, 1024])]; + tensor var_1598_cast_fp16 = reshape(shape = var_1597, x = v_11_cast_fp16)[name = tensor("op_1598_cast_fp16")]; + tensor written_v_11_transpose_x_1 = const()[name = tensor("written_v_11_transpose_x_1"), val = tensor(true)]; + tensor written_v_11_transpose_y_1 = const()[name = tensor("written_v_11_transpose_y_1"), val = tensor(false)]; + tensor written_v_11_cast_fp16 = matmul(transpose_x = written_v_11_transpose_x_1, transpose_y = written_v_11_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_1598_cast_fp16)[name = tensor("written_v_11_cast_fp16")]; + tensor k_cache5_to_fp16_dtype_0 = const()[name = tensor("k_cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache5_to_fp16 = cast(dtype = k_cache5_to_fp16_dtype_0, x = k_cache5)[name = tensor("cast_435")]; + tensor var_1609_cast_fp16 = mul(x = k_cache5_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_1609_cast_fp16")]; + tensor var_1614 = const()[name = tensor("op_1614"), val = tensor([1, 512, 16, 64])]; + tensor var_1615_cast_fp16 = reshape(shape = var_1614, x = written_k_11_cast_fp16)[name = tensor("op_1615_cast_fp16")]; + tensor new_k_cache_11_cast_fp16 = add(x = var_1609_cast_fp16, y = var_1615_cast_fp16)[name = tensor("new_k_cache_11_cast_fp16")]; + tensor new_k_cache_11_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_11_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache5_to_fp16_dtype_0 = const()[name = tensor("v_cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache5_to_fp16 = cast(dtype = v_cache5_to_fp16_dtype_0, x = v_cache5)[name = tensor("cast_433")]; + tensor var_1618_cast_fp16 = mul(x = v_cache5_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_1618_cast_fp16")]; + tensor var_1623 = const()[name = tensor("op_1623"), val = tensor([1, 512, 16, 64])]; + tensor var_1624_cast_fp16 = reshape(shape = var_1623, x = written_v_11_cast_fp16)[name = tensor("op_1624_cast_fp16")]; + tensor new_v_cache_11_cast_fp16 = add(x = var_1618_cast_fp16, y = var_1624_cast_fp16)[name = tensor("new_v_cache_11_cast_fp16")]; + tensor new_v_cache_11_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_11_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_1637 = const()[name = tensor("op_1637"), val = tensor([0, 2, 1, 3])]; + tensor var_1648_transpose_x_0 = const()[name = tensor("op_1648_transpose_x_0"), val = tensor(false)]; + tensor var_1648_transpose_y_0 = const()[name = tensor("op_1648_transpose_y_0"), val = tensor(false)]; + tensor transpose_79_perm_0 = const()[name = tensor("transpose_79_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_80_perm_0 = const()[name = tensor("transpose_80_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_80 = transpose(perm = transpose_80_perm_0, x = new_k_cache_11_cast_fp16)[name = tensor("transpose_184")]; + tensor transpose_79 = transpose(perm = transpose_79_perm_0, x = q_11_cast_fp16)[name = tensor("transpose_185")]; + tensor var_1648_cast_fp16 = matmul(transpose_x = var_1648_transpose_x_0, transpose_y = var_1648_transpose_y_0, x = transpose_79, y = transpose_80)[name = tensor("op_1648_cast_fp16")]; + tensor var_1649_to_fp16 = const()[name = tensor("op_1649_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_21_cast_fp16 = mul(x = var_1648_cast_fp16, y = var_1649_to_fp16)[name = tensor("scores_21_cast_fp16")]; + tensor scores_23_cast_fp16 = add(x = scores_21_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_23_cast_fp16")]; + tensor var_1659 = const()[name = tensor("op_1659"), val = tensor(-1)]; + tensor weights_11_cast_fp16 = softmax(axis = var_1659, x = scores_23_cast_fp16)[name = tensor("weights_11_cast_fp16")]; + tensor attn_11_transpose_x_0 = const()[name = tensor("attn_11_transpose_x_0"), val = tensor(false)]; + tensor attn_11_transpose_y_0 = const()[name = tensor("attn_11_transpose_y_0"), val = tensor(false)]; + tensor values_11_cast_fp16 = transpose(perm = var_1637, x = new_v_cache_11_cast_fp16)[name = tensor("transpose_186")]; + tensor attn_11_cast_fp16 = matmul(transpose_x = attn_11_transpose_x_0, transpose_y = attn_11_transpose_y_0, x = weights_11_cast_fp16, y = values_11_cast_fp16)[name = tensor("attn_11_cast_fp16")]; + tensor var_1667 = const()[name = tensor("op_1667"), val = tensor([0, 2, 1, 3])]; + tensor var_1672 = const()[name = tensor("op_1672"), val = tensor([1, 256, 1024])]; + tensor var_1668_cast_fp16 = transpose(perm = var_1667, x = attn_11_cast_fp16)[name = tensor("transpose_183")]; + tensor input_65_cast_fp16 = reshape(shape = var_1672, x = var_1668_cast_fp16)[name = tensor("input_65_cast_fp16")]; + tensor attn5_out_proj_weight_to_fp16 = const()[name = tensor("attn5_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132187392)))]; + tensor linear_21_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn5_out_proj_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("linear_21_cast_fp16")]; + tensor position5_to_fp16_dtype_0 = const()[name = tensor("position5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position5_to_fp16 = cast(dtype = position5_to_fp16_dtype_0, x = position5)[name = tensor("cast_431")]; + tensor var_1681_cast_fp16 = add(x = position5_to_fp16, y = valid_len_to_fp16)[name = tensor("op_1681_cast_fp16")]; + tensor var_1681_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1681_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_67_cast_fp16 = add(x = input_61_cast_fp16, y = linear_21_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor input_69_axes_0 = const()[name = tensor("input_69_axes_0"), val = tensor([-1])]; + tensor norm5_2_weight_to_fp16 = const()[name = tensor("norm5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134284608)))]; + tensor norm5_2_bias_to_fp16 = const()[name = tensor("norm5_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134286720)))]; + tensor var_1685_to_fp16 = const()[name = tensor("op_1685_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_69_cast_fp16 = layer_norm(axes = input_69_axes_0, beta = norm5_2_bias_to_fp16, epsilon = var_1685_to_fp16, gamma = norm5_2_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("input_69_cast_fp16")]; + tensor linear5_1_weight_to_fp16 = const()[name = tensor("linear5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134288832)))]; + tensor linear_22_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear5_1_weight_to_fp16, x = input_69_cast_fp16)[name = tensor("linear_22_cast_fp16")]; + tensor input_71_mode_0 = const()[name = tensor("input_71_mode_0"), val = tensor("EXACT")]; + tensor input_71_cast_fp16 = gelu(mode = input_71_mode_0, x = linear_22_cast_fp16)[name = tensor("input_71_cast_fp16")]; + tensor linear5_2_weight_to_fp16 = const()[name = tensor("linear5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142677504)))]; + tensor linear_23_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear5_2_weight_to_fp16, x = input_71_cast_fp16)[name = tensor("linear_23_cast_fp16")]; + tensor input_73_cast_fp16 = add(x = input_67_cast_fp16, y = linear_23_cast_fp16)[name = tensor("input_73_cast_fp16")]; + tensor input_75_axes_0 = const()[name = tensor("input_75_axes_0"), val = tensor([-1])]; + tensor norm6_1_weight_to_fp16 = const()[name = tensor("norm6_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151066176)))]; + tensor norm6_1_bias_to_fp16 = const()[name = tensor("norm6_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151068288)))]; + tensor var_1702_to_fp16 = const()[name = tensor("op_1702_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_75_cast_fp16 = layer_norm(axes = input_75_axes_0, beta = norm6_1_bias_to_fp16, epsilon = var_1702_to_fp16, gamma = norm6_1_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("input_75_cast_fp16")]; + tensor attn6_in_proj_weight_to_fp16 = const()[name = tensor("attn6_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151070400)))]; + tensor linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn6_in_proj_weight_to_fp16, x = input_75_cast_fp16)[name = tensor("linear_24_cast_fp16")]; + tensor var_1716 = const()[name = tensor("op_1716"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_13_cast_fp16 = reshape(shape = var_1716, x = linear_24_cast_fp16)[name = tensor("qkv_13_cast_fp16")]; + tensor x_25_begin_0 = const()[name = tensor("x_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_25_end_0 = const()[name = tensor("x_25_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_25_end_mask_0 = const()[name = tensor("x_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_25_squeeze_mask_0 = const()[name = tensor("x_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_25_cast_fp16 = slice_by_index(begin = x_25_begin_0, end = x_25_end_0, end_mask = x_25_end_mask_0, squeeze_mask = x_25_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("x_25_cast_fp16")]; + tensor x_27_begin_0 = const()[name = tensor("x_27_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_27_end_0 = const()[name = tensor("x_27_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_27_end_mask_0 = const()[name = tensor("x_27_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_27_squeeze_mask_0 = const()[name = tensor("x_27_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_27_cast_fp16 = slice_by_index(begin = x_27_begin_0, end = x_27_end_0, end_mask = x_27_end_mask_0, squeeze_mask = x_27_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("x_27_cast_fp16")]; + tensor v_13_begin_0 = const()[name = tensor("v_13_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_13_end_0 = const()[name = tensor("v_13_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_13_end_mask_0 = const()[name = tensor("v_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_13_squeeze_mask_0 = const()[name = tensor("v_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_13_cast_fp16 = slice_by_index(begin = v_13_begin_0, end = v_13_end_0, end_mask = v_13_end_mask_0, squeeze_mask = v_13_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("v_13_cast_fp16")]; + tensor var_1761 = const()[name = tensor("op_1761"), val = tensor([256, 16, 32, 2])]; + tensor pairs_25_cast_fp16 = reshape(shape = var_1761, x = x_25_cast_fp16)[name = tensor("pairs_25_cast_fp16")]; + tensor xr_25_begin_0 = const()[name = tensor("xr_25_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_25_end_0 = const()[name = tensor("xr_25_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_25_end_mask_0 = const()[name = tensor("xr_25_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_25_squeeze_mask_0 = const()[name = tensor("xr_25_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_25_cast_fp16 = slice_by_index(begin = xr_25_begin_0, end = xr_25_end_0, end_mask = xr_25_end_mask_0, squeeze_mask = xr_25_squeeze_mask_0, x = pairs_25_cast_fp16)[name = tensor("xr_25_cast_fp16")]; + tensor xi_25_begin_0 = const()[name = tensor("xi_25_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_25_end_0 = const()[name = tensor("xi_25_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_25_end_mask_0 = const()[name = tensor("xi_25_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_25_squeeze_mask_0 = const()[name = tensor("xi_25_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_25_cast_fp16 = slice_by_index(begin = xi_25_begin_0, end = xi_25_end_0, end_mask = xi_25_end_mask_0, squeeze_mask = xi_25_squeeze_mask_0, x = pairs_25_cast_fp16)[name = tensor("xi_25_cast_fp16")]; + tensor var_1769_cast_fp16 = mul(x = xr_25_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1769_cast_fp16")]; + tensor var_1770_cast_fp16 = mul(x = xi_25_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1770_cast_fp16")]; + tensor xor__25_cast_fp16 = sub(x = var_1769_cast_fp16, y = var_1770_cast_fp16)[name = tensor("xor__25_cast_fp16")]; + tensor var_1773_cast_fp16 = mul(x = xr_25_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1773_cast_fp16")]; + tensor var_1774_cast_fp16 = mul(x = xi_25_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1774_cast_fp16")]; + tensor xoi_25_cast_fp16 = add(x = var_1773_cast_fp16, y = var_1774_cast_fp16)[name = tensor("xoi_25_cast_fp16")]; + tensor out_25_axis_0 = const()[name = tensor("out_25_axis_0"), val = tensor(-1)]; + tensor out_25_cast_fp16 = stack(axis = out_25_axis_0, values = (xor__25_cast_fp16, xoi_25_cast_fp16))[name = tensor("out_25_cast_fp16")]; + tensor var_1784 = const()[name = tensor("op_1784"), val = tensor([1, 256, 16, 64])]; + tensor q_13_cast_fp16 = reshape(shape = var_1784, x = out_25_cast_fp16)[name = tensor("q_13_cast_fp16")]; + tensor var_1790 = const()[name = tensor("op_1790"), val = tensor([256, 16, 32, 2])]; + tensor pairs_27_cast_fp16 = reshape(shape = var_1790, x = x_27_cast_fp16)[name = tensor("pairs_27_cast_fp16")]; + tensor xr_27_begin_0 = const()[name = tensor("xr_27_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_27_end_0 = const()[name = tensor("xr_27_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_27_end_mask_0 = const()[name = tensor("xr_27_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_27_squeeze_mask_0 = const()[name = tensor("xr_27_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_27_cast_fp16 = slice_by_index(begin = xr_27_begin_0, end = xr_27_end_0, end_mask = xr_27_end_mask_0, squeeze_mask = xr_27_squeeze_mask_0, x = pairs_27_cast_fp16)[name = tensor("xr_27_cast_fp16")]; + tensor xi_27_begin_0 = const()[name = tensor("xi_27_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_27_end_0 = const()[name = tensor("xi_27_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_27_end_mask_0 = const()[name = tensor("xi_27_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_27_squeeze_mask_0 = const()[name = tensor("xi_27_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_27_cast_fp16 = slice_by_index(begin = xi_27_begin_0, end = xi_27_end_0, end_mask = xi_27_end_mask_0, squeeze_mask = xi_27_squeeze_mask_0, x = pairs_27_cast_fp16)[name = tensor("xi_27_cast_fp16")]; + tensor var_1798_cast_fp16 = mul(x = xr_27_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1798_cast_fp16")]; + tensor var_1799_cast_fp16 = mul(x = xi_27_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1799_cast_fp16")]; + tensor xor__27_cast_fp16 = sub(x = var_1798_cast_fp16, y = var_1799_cast_fp16)[name = tensor("xor__27_cast_fp16")]; + tensor var_1802_cast_fp16 = mul(x = xr_27_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1802_cast_fp16")]; + tensor var_1803_cast_fp16 = mul(x = xi_27_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1803_cast_fp16")]; + tensor xoi_27_cast_fp16 = add(x = var_1802_cast_fp16, y = var_1803_cast_fp16)[name = tensor("xoi_27_cast_fp16")]; + tensor out_27_axis_0 = const()[name = tensor("out_27_axis_0"), val = tensor(-1)]; + tensor out_27_cast_fp16 = stack(axis = out_27_axis_0, values = (xor__27_cast_fp16, xoi_27_cast_fp16))[name = tensor("out_27_cast_fp16")]; + tensor var_1820 = const()[name = tensor("op_1820"), val = tensor([256, 1024])]; + tensor var_1821_cast_fp16 = reshape(shape = var_1820, x = out_27_cast_fp16)[name = tensor("op_1821_cast_fp16")]; + tensor written_k_13_transpose_x_1 = const()[name = tensor("written_k_13_transpose_x_1"), val = tensor(true)]; + tensor written_k_13_transpose_y_1 = const()[name = tensor("written_k_13_transpose_y_1"), val = tensor(false)]; + tensor written_k_13_cast_fp16 = matmul(transpose_x = written_k_13_transpose_x_1, transpose_y = written_k_13_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_1821_cast_fp16)[name = tensor("written_k_13_cast_fp16")]; + tensor var_1828 = const()[name = tensor("op_1828"), val = tensor([256, 1024])]; + tensor var_1829_cast_fp16 = reshape(shape = var_1828, x = v_13_cast_fp16)[name = tensor("op_1829_cast_fp16")]; + tensor written_v_13_transpose_x_1 = const()[name = tensor("written_v_13_transpose_x_1"), val = tensor(true)]; + tensor written_v_13_transpose_y_1 = const()[name = tensor("written_v_13_transpose_y_1"), val = tensor(false)]; + tensor written_v_13_cast_fp16 = matmul(transpose_x = written_v_13_transpose_x_1, transpose_y = written_v_13_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_1829_cast_fp16)[name = tensor("written_v_13_cast_fp16")]; + tensor k_cache6_to_fp16_dtype_0 = const()[name = tensor("k_cache6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache6_to_fp16 = cast(dtype = k_cache6_to_fp16_dtype_0, x = k_cache6)[name = tensor("cast_429")]; + tensor var_1840_cast_fp16 = mul(x = k_cache6_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_1840_cast_fp16")]; + tensor var_1845 = const()[name = tensor("op_1845"), val = tensor([1, 512, 16, 64])]; + tensor var_1846_cast_fp16 = reshape(shape = var_1845, x = written_k_13_cast_fp16)[name = tensor("op_1846_cast_fp16")]; + tensor new_k_cache_13_cast_fp16 = add(x = var_1840_cast_fp16, y = var_1846_cast_fp16)[name = tensor("new_k_cache_13_cast_fp16")]; + tensor new_k_cache_13_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_13_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache6_to_fp16_dtype_0 = const()[name = tensor("v_cache6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache6_to_fp16 = cast(dtype = v_cache6_to_fp16_dtype_0, x = v_cache6)[name = tensor("cast_427")]; + tensor var_1849_cast_fp16 = mul(x = v_cache6_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_1849_cast_fp16")]; + tensor var_1854 = const()[name = tensor("op_1854"), val = tensor([1, 512, 16, 64])]; + tensor var_1855_cast_fp16 = reshape(shape = var_1854, x = written_v_13_cast_fp16)[name = tensor("op_1855_cast_fp16")]; + tensor new_v_cache_13_cast_fp16 = add(x = var_1849_cast_fp16, y = var_1855_cast_fp16)[name = tensor("new_v_cache_13_cast_fp16")]; + tensor new_v_cache_13_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_13_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_1868 = const()[name = tensor("op_1868"), val = tensor([0, 2, 1, 3])]; + tensor var_1879_transpose_x_0 = const()[name = tensor("op_1879_transpose_x_0"), val = tensor(false)]; + tensor var_1879_transpose_y_0 = const()[name = tensor("op_1879_transpose_y_0"), val = tensor(false)]; + tensor transpose_81_perm_0 = const()[name = tensor("transpose_81_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_82_perm_0 = const()[name = tensor("transpose_82_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_82 = transpose(perm = transpose_82_perm_0, x = new_k_cache_13_cast_fp16)[name = tensor("transpose_180")]; + tensor transpose_81 = transpose(perm = transpose_81_perm_0, x = q_13_cast_fp16)[name = tensor("transpose_181")]; + tensor var_1879_cast_fp16 = matmul(transpose_x = var_1879_transpose_x_0, transpose_y = var_1879_transpose_y_0, x = transpose_81, y = transpose_82)[name = tensor("op_1879_cast_fp16")]; + tensor var_1880_to_fp16 = const()[name = tensor("op_1880_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_25_cast_fp16 = mul(x = var_1879_cast_fp16, y = var_1880_to_fp16)[name = tensor("scores_25_cast_fp16")]; + tensor scores_27_cast_fp16 = add(x = scores_25_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_27_cast_fp16")]; + tensor var_1890 = const()[name = tensor("op_1890"), val = tensor(-1)]; + tensor weights_13_cast_fp16 = softmax(axis = var_1890, x = scores_27_cast_fp16)[name = tensor("weights_13_cast_fp16")]; + tensor attn_13_transpose_x_0 = const()[name = tensor("attn_13_transpose_x_0"), val = tensor(false)]; + tensor attn_13_transpose_y_0 = const()[name = tensor("attn_13_transpose_y_0"), val = tensor(false)]; + tensor values_13_cast_fp16 = transpose(perm = var_1868, x = new_v_cache_13_cast_fp16)[name = tensor("transpose_182")]; + tensor attn_13_cast_fp16 = matmul(transpose_x = attn_13_transpose_x_0, transpose_y = attn_13_transpose_y_0, x = weights_13_cast_fp16, y = values_13_cast_fp16)[name = tensor("attn_13_cast_fp16")]; + tensor var_1898 = const()[name = tensor("op_1898"), val = tensor([0, 2, 1, 3])]; + tensor var_1903 = const()[name = tensor("op_1903"), val = tensor([1, 256, 1024])]; + tensor var_1899_cast_fp16 = transpose(perm = var_1898, x = attn_13_cast_fp16)[name = tensor("transpose_179")]; + tensor input_77_cast_fp16 = reshape(shape = var_1903, x = var_1899_cast_fp16)[name = tensor("input_77_cast_fp16")]; + tensor attn6_out_proj_weight_to_fp16 = const()[name = tensor("attn6_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(157361920)))]; + tensor linear_25_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn6_out_proj_weight_to_fp16, x = input_77_cast_fp16)[name = tensor("linear_25_cast_fp16")]; + tensor position6_to_fp16_dtype_0 = const()[name = tensor("position6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position6_to_fp16 = cast(dtype = position6_to_fp16_dtype_0, x = position6)[name = tensor("cast_425")]; + tensor var_1912_cast_fp16 = add(x = position6_to_fp16, y = valid_len_to_fp16)[name = tensor("op_1912_cast_fp16")]; + tensor var_1912_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1912_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_79_cast_fp16 = add(x = input_73_cast_fp16, y = linear_25_cast_fp16)[name = tensor("input_79_cast_fp16")]; + tensor input_81_axes_0 = const()[name = tensor("input_81_axes_0"), val = tensor([-1])]; + tensor norm6_2_weight_to_fp16 = const()[name = tensor("norm6_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159459136)))]; + tensor norm6_2_bias_to_fp16 = const()[name = tensor("norm6_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159461248)))]; + tensor var_1916_to_fp16 = const()[name = tensor("op_1916_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_81_cast_fp16 = layer_norm(axes = input_81_axes_0, beta = norm6_2_bias_to_fp16, epsilon = var_1916_to_fp16, gamma = norm6_2_weight_to_fp16, x = input_79_cast_fp16)[name = tensor("input_81_cast_fp16")]; + tensor linear6_1_weight_to_fp16 = const()[name = tensor("linear6_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159463360)))]; + tensor linear_26_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear6_1_weight_to_fp16, x = input_81_cast_fp16)[name = tensor("linear_26_cast_fp16")]; + tensor input_83_mode_0 = const()[name = tensor("input_83_mode_0"), val = tensor("EXACT")]; + tensor input_83_cast_fp16 = gelu(mode = input_83_mode_0, x = linear_26_cast_fp16)[name = tensor("input_83_cast_fp16")]; + tensor linear6_2_weight_to_fp16 = const()[name = tensor("linear6_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(167852032)))]; + tensor linear_27_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear6_2_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_27_cast_fp16")]; + tensor input_85_cast_fp16 = add(x = input_79_cast_fp16, y = linear_27_cast_fp16)[name = tensor("input_85_cast_fp16")]; + tensor input_87_axes_0 = const()[name = tensor("input_87_axes_0"), val = tensor([-1])]; + tensor norm7_1_weight_to_fp16 = const()[name = tensor("norm7_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176240704)))]; + tensor norm7_1_bias_to_fp16 = const()[name = tensor("norm7_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176242816)))]; + tensor var_1933_to_fp16 = const()[name = tensor("op_1933_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_87_cast_fp16 = layer_norm(axes = input_87_axes_0, beta = norm7_1_bias_to_fp16, epsilon = var_1933_to_fp16, gamma = norm7_1_weight_to_fp16, x = input_85_cast_fp16)[name = tensor("input_87_cast_fp16")]; + tensor attn7_in_proj_weight_to_fp16 = const()[name = tensor("attn7_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176244928)))]; + tensor linear_28_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn7_in_proj_weight_to_fp16, x = input_87_cast_fp16)[name = tensor("linear_28_cast_fp16")]; + tensor var_1947 = const()[name = tensor("op_1947"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_15_cast_fp16 = reshape(shape = var_1947, x = linear_28_cast_fp16)[name = tensor("qkv_15_cast_fp16")]; + tensor x_29_begin_0 = const()[name = tensor("x_29_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_29_end_0 = const()[name = tensor("x_29_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_29_end_mask_0 = const()[name = tensor("x_29_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_29_squeeze_mask_0 = const()[name = tensor("x_29_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_29_cast_fp16 = slice_by_index(begin = x_29_begin_0, end = x_29_end_0, end_mask = x_29_end_mask_0, squeeze_mask = x_29_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("x_29_cast_fp16")]; + tensor x_31_begin_0 = const()[name = tensor("x_31_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_31_end_0 = const()[name = tensor("x_31_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_31_end_mask_0 = const()[name = tensor("x_31_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_31_squeeze_mask_0 = const()[name = tensor("x_31_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_31_cast_fp16 = slice_by_index(begin = x_31_begin_0, end = x_31_end_0, end_mask = x_31_end_mask_0, squeeze_mask = x_31_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("x_31_cast_fp16")]; + tensor v_15_begin_0 = const()[name = tensor("v_15_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_15_end_0 = const()[name = tensor("v_15_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_15_end_mask_0 = const()[name = tensor("v_15_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_15_squeeze_mask_0 = const()[name = tensor("v_15_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_15_cast_fp16 = slice_by_index(begin = v_15_begin_0, end = v_15_end_0, end_mask = v_15_end_mask_0, squeeze_mask = v_15_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("v_15_cast_fp16")]; + tensor var_1992 = const()[name = tensor("op_1992"), val = tensor([256, 16, 32, 2])]; + tensor pairs_29_cast_fp16 = reshape(shape = var_1992, x = x_29_cast_fp16)[name = tensor("pairs_29_cast_fp16")]; + tensor xr_29_begin_0 = const()[name = tensor("xr_29_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_29_end_0 = const()[name = tensor("xr_29_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_29_end_mask_0 = const()[name = tensor("xr_29_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_29_squeeze_mask_0 = const()[name = tensor("xr_29_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_29_cast_fp16 = slice_by_index(begin = xr_29_begin_0, end = xr_29_end_0, end_mask = xr_29_end_mask_0, squeeze_mask = xr_29_squeeze_mask_0, x = pairs_29_cast_fp16)[name = tensor("xr_29_cast_fp16")]; + tensor xi_29_begin_0 = const()[name = tensor("xi_29_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_29_end_0 = const()[name = tensor("xi_29_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_29_end_mask_0 = const()[name = tensor("xi_29_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_29_squeeze_mask_0 = const()[name = tensor("xi_29_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_29_cast_fp16 = slice_by_index(begin = xi_29_begin_0, end = xi_29_end_0, end_mask = xi_29_end_mask_0, squeeze_mask = xi_29_squeeze_mask_0, x = pairs_29_cast_fp16)[name = tensor("xi_29_cast_fp16")]; + tensor var_2000_cast_fp16 = mul(x = xr_29_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2000_cast_fp16")]; + tensor var_2001_cast_fp16 = mul(x = xi_29_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2001_cast_fp16")]; + tensor xor__29_cast_fp16 = sub(x = var_2000_cast_fp16, y = var_2001_cast_fp16)[name = tensor("xor__29_cast_fp16")]; + tensor var_2004_cast_fp16 = mul(x = xr_29_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2004_cast_fp16")]; + tensor var_2005_cast_fp16 = mul(x = xi_29_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2005_cast_fp16")]; + tensor xoi_29_cast_fp16 = add(x = var_2004_cast_fp16, y = var_2005_cast_fp16)[name = tensor("xoi_29_cast_fp16")]; + tensor out_29_axis_0 = const()[name = tensor("out_29_axis_0"), val = tensor(-1)]; + tensor out_29_cast_fp16 = stack(axis = out_29_axis_0, values = (xor__29_cast_fp16, xoi_29_cast_fp16))[name = tensor("out_29_cast_fp16")]; + tensor var_2015 = const()[name = tensor("op_2015"), val = tensor([1, 256, 16, 64])]; + tensor q_15_cast_fp16 = reshape(shape = var_2015, x = out_29_cast_fp16)[name = tensor("q_15_cast_fp16")]; + tensor var_2021 = const()[name = tensor("op_2021"), val = tensor([256, 16, 32, 2])]; + tensor pairs_31_cast_fp16 = reshape(shape = var_2021, x = x_31_cast_fp16)[name = tensor("pairs_31_cast_fp16")]; + tensor xr_31_begin_0 = const()[name = tensor("xr_31_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_31_end_0 = const()[name = tensor("xr_31_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_31_end_mask_0 = const()[name = tensor("xr_31_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_31_squeeze_mask_0 = const()[name = tensor("xr_31_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_31_cast_fp16 = slice_by_index(begin = xr_31_begin_0, end = xr_31_end_0, end_mask = xr_31_end_mask_0, squeeze_mask = xr_31_squeeze_mask_0, x = pairs_31_cast_fp16)[name = tensor("xr_31_cast_fp16")]; + tensor xi_31_begin_0 = const()[name = tensor("xi_31_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_31_end_0 = const()[name = tensor("xi_31_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_31_end_mask_0 = const()[name = tensor("xi_31_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_31_squeeze_mask_0 = const()[name = tensor("xi_31_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_31_cast_fp16 = slice_by_index(begin = xi_31_begin_0, end = xi_31_end_0, end_mask = xi_31_end_mask_0, squeeze_mask = xi_31_squeeze_mask_0, x = pairs_31_cast_fp16)[name = tensor("xi_31_cast_fp16")]; + tensor var_2029_cast_fp16 = mul(x = xr_31_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2029_cast_fp16")]; + tensor var_2030_cast_fp16 = mul(x = xi_31_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2030_cast_fp16")]; + tensor xor__31_cast_fp16 = sub(x = var_2029_cast_fp16, y = var_2030_cast_fp16)[name = tensor("xor__31_cast_fp16")]; + tensor var_2033_cast_fp16 = mul(x = xr_31_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2033_cast_fp16")]; + tensor var_2034_cast_fp16 = mul(x = xi_31_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2034_cast_fp16")]; + tensor xoi_31_cast_fp16 = add(x = var_2033_cast_fp16, y = var_2034_cast_fp16)[name = tensor("xoi_31_cast_fp16")]; + tensor out_31_axis_0 = const()[name = tensor("out_31_axis_0"), val = tensor(-1)]; + tensor out_31_cast_fp16 = stack(axis = out_31_axis_0, values = (xor__31_cast_fp16, xoi_31_cast_fp16))[name = tensor("out_31_cast_fp16")]; + tensor var_2051 = const()[name = tensor("op_2051"), val = tensor([256, 1024])]; + tensor var_2052_cast_fp16 = reshape(shape = var_2051, x = out_31_cast_fp16)[name = tensor("op_2052_cast_fp16")]; + tensor written_k_15_transpose_x_1 = const()[name = tensor("written_k_15_transpose_x_1"), val = tensor(true)]; + tensor written_k_15_transpose_y_1 = const()[name = tensor("written_k_15_transpose_y_1"), val = tensor(false)]; + tensor written_k_15_cast_fp16 = matmul(transpose_x = written_k_15_transpose_x_1, transpose_y = written_k_15_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_2052_cast_fp16)[name = tensor("written_k_15_cast_fp16")]; + tensor var_2059 = const()[name = tensor("op_2059"), val = tensor([256, 1024])]; + tensor var_2060_cast_fp16 = reshape(shape = var_2059, x = v_15_cast_fp16)[name = tensor("op_2060_cast_fp16")]; + tensor written_v_15_transpose_x_1 = const()[name = tensor("written_v_15_transpose_x_1"), val = tensor(true)]; + tensor written_v_15_transpose_y_1 = const()[name = tensor("written_v_15_transpose_y_1"), val = tensor(false)]; + tensor written_v_15_cast_fp16 = matmul(transpose_x = written_v_15_transpose_x_1, transpose_y = written_v_15_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_2060_cast_fp16)[name = tensor("written_v_15_cast_fp16")]; + tensor k_cache7_to_fp16_dtype_0 = const()[name = tensor("k_cache7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache7_to_fp16 = cast(dtype = k_cache7_to_fp16_dtype_0, x = k_cache7)[name = tensor("cast_423")]; + tensor var_2071_cast_fp16 = mul(x = k_cache7_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_2071_cast_fp16")]; + tensor var_2076 = const()[name = tensor("op_2076"), val = tensor([1, 512, 16, 64])]; + tensor var_2077_cast_fp16 = reshape(shape = var_2076, x = written_k_15_cast_fp16)[name = tensor("op_2077_cast_fp16")]; + tensor new_k_cache_15_cast_fp16 = add(x = var_2071_cast_fp16, y = var_2077_cast_fp16)[name = tensor("new_k_cache_15_cast_fp16")]; + tensor new_k_cache_15_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_15_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache7_to_fp16_dtype_0 = const()[name = tensor("v_cache7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache7_to_fp16 = cast(dtype = v_cache7_to_fp16_dtype_0, x = v_cache7)[name = tensor("cast_421")]; + tensor var_2080_cast_fp16 = mul(x = v_cache7_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_2080_cast_fp16")]; + tensor var_2085 = const()[name = tensor("op_2085"), val = tensor([1, 512, 16, 64])]; + tensor var_2086_cast_fp16 = reshape(shape = var_2085, x = written_v_15_cast_fp16)[name = tensor("op_2086_cast_fp16")]; + tensor new_v_cache_15_cast_fp16 = add(x = var_2080_cast_fp16, y = var_2086_cast_fp16)[name = tensor("new_v_cache_15_cast_fp16")]; + tensor new_v_cache_15_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_15_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_2099 = const()[name = tensor("op_2099"), val = tensor([0, 2, 1, 3])]; + tensor var_2110_transpose_x_0 = const()[name = tensor("op_2110_transpose_x_0"), val = tensor(false)]; + tensor var_2110_transpose_y_0 = const()[name = tensor("op_2110_transpose_y_0"), val = tensor(false)]; + tensor transpose_83_perm_0 = const()[name = tensor("transpose_83_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_84_perm_0 = const()[name = tensor("transpose_84_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_84 = transpose(perm = transpose_84_perm_0, x = new_k_cache_15_cast_fp16)[name = tensor("transpose_176")]; + tensor transpose_83 = transpose(perm = transpose_83_perm_0, x = q_15_cast_fp16)[name = tensor("transpose_177")]; + tensor var_2110_cast_fp16 = matmul(transpose_x = var_2110_transpose_x_0, transpose_y = var_2110_transpose_y_0, x = transpose_83, y = transpose_84)[name = tensor("op_2110_cast_fp16")]; + tensor var_2111_to_fp16 = const()[name = tensor("op_2111_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_29_cast_fp16 = mul(x = var_2110_cast_fp16, y = var_2111_to_fp16)[name = tensor("scores_29_cast_fp16")]; + tensor scores_31_cast_fp16 = add(x = scores_29_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_31_cast_fp16")]; + tensor var_2121 = const()[name = tensor("op_2121"), val = tensor(-1)]; + tensor weights_15_cast_fp16 = softmax(axis = var_2121, x = scores_31_cast_fp16)[name = tensor("weights_15_cast_fp16")]; + tensor attn_15_transpose_x_0 = const()[name = tensor("attn_15_transpose_x_0"), val = tensor(false)]; + tensor attn_15_transpose_y_0 = const()[name = tensor("attn_15_transpose_y_0"), val = tensor(false)]; + tensor values_15_cast_fp16 = transpose(perm = var_2099, x = new_v_cache_15_cast_fp16)[name = tensor("transpose_178")]; + tensor attn_15_cast_fp16 = matmul(transpose_x = attn_15_transpose_x_0, transpose_y = attn_15_transpose_y_0, x = weights_15_cast_fp16, y = values_15_cast_fp16)[name = tensor("attn_15_cast_fp16")]; + tensor var_2129 = const()[name = tensor("op_2129"), val = tensor([0, 2, 1, 3])]; + tensor var_2134 = const()[name = tensor("op_2134"), val = tensor([1, 256, 1024])]; + tensor var_2130_cast_fp16 = transpose(perm = var_2129, x = attn_15_cast_fp16)[name = tensor("transpose_175")]; + tensor input_89_cast_fp16 = reshape(shape = var_2134, x = var_2130_cast_fp16)[name = tensor("input_89_cast_fp16")]; + tensor attn7_out_proj_weight_to_fp16 = const()[name = tensor("attn7_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(182536448)))]; + tensor linear_29_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn7_out_proj_weight_to_fp16, x = input_89_cast_fp16)[name = tensor("linear_29_cast_fp16")]; + tensor position7_to_fp16_dtype_0 = const()[name = tensor("position7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position7_to_fp16 = cast(dtype = position7_to_fp16_dtype_0, x = position7)[name = tensor("cast_419")]; + tensor var_2143_cast_fp16 = add(x = position7_to_fp16, y = valid_len_to_fp16)[name = tensor("op_2143_cast_fp16")]; + tensor var_2143_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2143_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_91_cast_fp16 = add(x = input_85_cast_fp16, y = linear_29_cast_fp16)[name = tensor("input_91_cast_fp16")]; + tensor input_93_axes_0 = const()[name = tensor("input_93_axes_0"), val = tensor([-1])]; + tensor norm7_2_weight_to_fp16 = const()[name = tensor("norm7_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184633664)))]; + tensor norm7_2_bias_to_fp16 = const()[name = tensor("norm7_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184635776)))]; + tensor var_2147_to_fp16 = const()[name = tensor("op_2147_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_93_cast_fp16 = layer_norm(axes = input_93_axes_0, beta = norm7_2_bias_to_fp16, epsilon = var_2147_to_fp16, gamma = norm7_2_weight_to_fp16, x = input_91_cast_fp16)[name = tensor("input_93_cast_fp16")]; + tensor linear7_1_weight_to_fp16 = const()[name = tensor("linear7_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184637888)))]; + tensor linear_30_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear7_1_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("linear_30_cast_fp16")]; + tensor input_95_mode_0 = const()[name = tensor("input_95_mode_0"), val = tensor("EXACT")]; + tensor input_95_cast_fp16 = gelu(mode = input_95_mode_0, x = linear_30_cast_fp16)[name = tensor("input_95_cast_fp16")]; + tensor linear7_2_weight_to_fp16 = const()[name = tensor("linear7_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(193026560)))]; + tensor linear_31_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear7_2_weight_to_fp16, x = input_95_cast_fp16)[name = tensor("linear_31_cast_fp16")]; + tensor input_97_cast_fp16 = add(x = input_91_cast_fp16, y = linear_31_cast_fp16)[name = tensor("input_97_cast_fp16")]; + tensor input_99_axes_0 = const()[name = tensor("input_99_axes_0"), val = tensor([-1])]; + tensor norm8_1_weight_to_fp16 = const()[name = tensor("norm8_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201415232)))]; + tensor norm8_1_bias_to_fp16 = const()[name = tensor("norm8_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201417344)))]; + tensor var_2164_to_fp16 = const()[name = tensor("op_2164_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_99_cast_fp16 = layer_norm(axes = input_99_axes_0, beta = norm8_1_bias_to_fp16, epsilon = var_2164_to_fp16, gamma = norm8_1_weight_to_fp16, x = input_97_cast_fp16)[name = tensor("input_99_cast_fp16")]; + tensor attn8_in_proj_weight_to_fp16 = const()[name = tensor("attn8_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201419456)))]; + tensor linear_32_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn8_in_proj_weight_to_fp16, x = input_99_cast_fp16)[name = tensor("linear_32_cast_fp16")]; + tensor var_2178 = const()[name = tensor("op_2178"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_17_cast_fp16 = reshape(shape = var_2178, x = linear_32_cast_fp16)[name = tensor("qkv_17_cast_fp16")]; + tensor x_33_begin_0 = const()[name = tensor("x_33_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_33_end_0 = const()[name = tensor("x_33_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_33_end_mask_0 = const()[name = tensor("x_33_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_33_squeeze_mask_0 = const()[name = tensor("x_33_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_33_cast_fp16 = slice_by_index(begin = x_33_begin_0, end = x_33_end_0, end_mask = x_33_end_mask_0, squeeze_mask = x_33_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("x_33_cast_fp16")]; + tensor x_35_begin_0 = const()[name = tensor("x_35_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_35_end_0 = const()[name = tensor("x_35_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_35_end_mask_0 = const()[name = tensor("x_35_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_35_squeeze_mask_0 = const()[name = tensor("x_35_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_35_cast_fp16 = slice_by_index(begin = x_35_begin_0, end = x_35_end_0, end_mask = x_35_end_mask_0, squeeze_mask = x_35_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("x_35_cast_fp16")]; + tensor v_17_begin_0 = const()[name = tensor("v_17_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_17_end_0 = const()[name = tensor("v_17_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_17_end_mask_0 = const()[name = tensor("v_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_17_squeeze_mask_0 = const()[name = tensor("v_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_17_cast_fp16 = slice_by_index(begin = v_17_begin_0, end = v_17_end_0, end_mask = v_17_end_mask_0, squeeze_mask = v_17_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("v_17_cast_fp16")]; + tensor var_2223 = const()[name = tensor("op_2223"), val = tensor([256, 16, 32, 2])]; + tensor pairs_33_cast_fp16 = reshape(shape = var_2223, x = x_33_cast_fp16)[name = tensor("pairs_33_cast_fp16")]; + tensor xr_33_begin_0 = const()[name = tensor("xr_33_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_33_end_0 = const()[name = tensor("xr_33_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_33_end_mask_0 = const()[name = tensor("xr_33_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_33_squeeze_mask_0 = const()[name = tensor("xr_33_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_33_cast_fp16 = slice_by_index(begin = xr_33_begin_0, end = xr_33_end_0, end_mask = xr_33_end_mask_0, squeeze_mask = xr_33_squeeze_mask_0, x = pairs_33_cast_fp16)[name = tensor("xr_33_cast_fp16")]; + tensor xi_33_begin_0 = const()[name = tensor("xi_33_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_33_end_0 = const()[name = tensor("xi_33_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_33_end_mask_0 = const()[name = tensor("xi_33_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_33_squeeze_mask_0 = const()[name = tensor("xi_33_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_33_cast_fp16 = slice_by_index(begin = xi_33_begin_0, end = xi_33_end_0, end_mask = xi_33_end_mask_0, squeeze_mask = xi_33_squeeze_mask_0, x = pairs_33_cast_fp16)[name = tensor("xi_33_cast_fp16")]; + tensor var_2231_cast_fp16 = mul(x = xr_33_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2231_cast_fp16")]; + tensor var_2232_cast_fp16 = mul(x = xi_33_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2232_cast_fp16")]; + tensor xor__33_cast_fp16 = sub(x = var_2231_cast_fp16, y = var_2232_cast_fp16)[name = tensor("xor__33_cast_fp16")]; + tensor var_2235_cast_fp16 = mul(x = xr_33_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2235_cast_fp16")]; + tensor var_2236_cast_fp16 = mul(x = xi_33_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2236_cast_fp16")]; + tensor xoi_33_cast_fp16 = add(x = var_2235_cast_fp16, y = var_2236_cast_fp16)[name = tensor("xoi_33_cast_fp16")]; + tensor out_33_axis_0 = const()[name = tensor("out_33_axis_0"), val = tensor(-1)]; + tensor out_33_cast_fp16 = stack(axis = out_33_axis_0, values = (xor__33_cast_fp16, xoi_33_cast_fp16))[name = tensor("out_33_cast_fp16")]; + tensor var_2246 = const()[name = tensor("op_2246"), val = tensor([1, 256, 16, 64])]; + tensor q_17_cast_fp16 = reshape(shape = var_2246, x = out_33_cast_fp16)[name = tensor("q_17_cast_fp16")]; + tensor var_2252 = const()[name = tensor("op_2252"), val = tensor([256, 16, 32, 2])]; + tensor pairs_35_cast_fp16 = reshape(shape = var_2252, x = x_35_cast_fp16)[name = tensor("pairs_35_cast_fp16")]; + tensor xr_35_begin_0 = const()[name = tensor("xr_35_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_35_end_0 = const()[name = tensor("xr_35_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_35_end_mask_0 = const()[name = tensor("xr_35_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_35_squeeze_mask_0 = const()[name = tensor("xr_35_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_35_cast_fp16 = slice_by_index(begin = xr_35_begin_0, end = xr_35_end_0, end_mask = xr_35_end_mask_0, squeeze_mask = xr_35_squeeze_mask_0, x = pairs_35_cast_fp16)[name = tensor("xr_35_cast_fp16")]; + tensor xi_35_begin_0 = const()[name = tensor("xi_35_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_35_end_0 = const()[name = tensor("xi_35_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_35_end_mask_0 = const()[name = tensor("xi_35_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_35_squeeze_mask_0 = const()[name = tensor("xi_35_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_35_cast_fp16 = slice_by_index(begin = xi_35_begin_0, end = xi_35_end_0, end_mask = xi_35_end_mask_0, squeeze_mask = xi_35_squeeze_mask_0, x = pairs_35_cast_fp16)[name = tensor("xi_35_cast_fp16")]; + tensor var_2260_cast_fp16 = mul(x = xr_35_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2260_cast_fp16")]; + tensor var_2261_cast_fp16 = mul(x = xi_35_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2261_cast_fp16")]; + tensor xor__35_cast_fp16 = sub(x = var_2260_cast_fp16, y = var_2261_cast_fp16)[name = tensor("xor__35_cast_fp16")]; + tensor var_2264_cast_fp16 = mul(x = xr_35_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2264_cast_fp16")]; + tensor var_2265_cast_fp16 = mul(x = xi_35_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2265_cast_fp16")]; + tensor xoi_35_cast_fp16 = add(x = var_2264_cast_fp16, y = var_2265_cast_fp16)[name = tensor("xoi_35_cast_fp16")]; + tensor out_35_axis_0 = const()[name = tensor("out_35_axis_0"), val = tensor(-1)]; + tensor out_35_cast_fp16 = stack(axis = out_35_axis_0, values = (xor__35_cast_fp16, xoi_35_cast_fp16))[name = tensor("out_35_cast_fp16")]; + tensor var_2282 = const()[name = tensor("op_2282"), val = tensor([256, 1024])]; + tensor var_2283_cast_fp16 = reshape(shape = var_2282, x = out_35_cast_fp16)[name = tensor("op_2283_cast_fp16")]; + tensor written_k_17_transpose_x_1 = const()[name = tensor("written_k_17_transpose_x_1"), val = tensor(true)]; + tensor written_k_17_transpose_y_1 = const()[name = tensor("written_k_17_transpose_y_1"), val = tensor(false)]; + tensor written_k_17_cast_fp16 = matmul(transpose_x = written_k_17_transpose_x_1, transpose_y = written_k_17_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_2283_cast_fp16)[name = tensor("written_k_17_cast_fp16")]; + tensor var_2290 = const()[name = tensor("op_2290"), val = tensor([256, 1024])]; + tensor var_2291_cast_fp16 = reshape(shape = var_2290, x = v_17_cast_fp16)[name = tensor("op_2291_cast_fp16")]; + tensor written_v_17_transpose_x_1 = const()[name = tensor("written_v_17_transpose_x_1"), val = tensor(true)]; + tensor written_v_17_transpose_y_1 = const()[name = tensor("written_v_17_transpose_y_1"), val = tensor(false)]; + tensor written_v_17_cast_fp16 = matmul(transpose_x = written_v_17_transpose_x_1, transpose_y = written_v_17_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_2291_cast_fp16)[name = tensor("written_v_17_cast_fp16")]; + tensor k_cache8_to_fp16_dtype_0 = const()[name = tensor("k_cache8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache8_to_fp16 = cast(dtype = k_cache8_to_fp16_dtype_0, x = k_cache8)[name = tensor("cast_417")]; + tensor var_2302_cast_fp16 = mul(x = k_cache8_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_2302_cast_fp16")]; + tensor var_2307 = const()[name = tensor("op_2307"), val = tensor([1, 512, 16, 64])]; + tensor var_2308_cast_fp16 = reshape(shape = var_2307, x = written_k_17_cast_fp16)[name = tensor("op_2308_cast_fp16")]; + tensor new_k_cache_17_cast_fp16 = add(x = var_2302_cast_fp16, y = var_2308_cast_fp16)[name = tensor("new_k_cache_17_cast_fp16")]; + tensor new_k_cache_17_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_17_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache8_to_fp16_dtype_0 = const()[name = tensor("v_cache8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache8_to_fp16 = cast(dtype = v_cache8_to_fp16_dtype_0, x = v_cache8)[name = tensor("cast_415")]; + tensor var_2311_cast_fp16 = mul(x = v_cache8_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_2311_cast_fp16")]; + tensor var_2316 = const()[name = tensor("op_2316"), val = tensor([1, 512, 16, 64])]; + tensor var_2317_cast_fp16 = reshape(shape = var_2316, x = written_v_17_cast_fp16)[name = tensor("op_2317_cast_fp16")]; + tensor new_v_cache_17_cast_fp16 = add(x = var_2311_cast_fp16, y = var_2317_cast_fp16)[name = tensor("new_v_cache_17_cast_fp16")]; + tensor new_v_cache_17_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_17_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_2330 = const()[name = tensor("op_2330"), val = tensor([0, 2, 1, 3])]; + tensor var_2341_transpose_x_0 = const()[name = tensor("op_2341_transpose_x_0"), val = tensor(false)]; + tensor var_2341_transpose_y_0 = const()[name = tensor("op_2341_transpose_y_0"), val = tensor(false)]; + tensor transpose_85_perm_0 = const()[name = tensor("transpose_85_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_86_perm_0 = const()[name = tensor("transpose_86_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_86 = transpose(perm = transpose_86_perm_0, x = new_k_cache_17_cast_fp16)[name = tensor("transpose_172")]; + tensor transpose_85 = transpose(perm = transpose_85_perm_0, x = q_17_cast_fp16)[name = tensor("transpose_173")]; + tensor var_2341_cast_fp16 = matmul(transpose_x = var_2341_transpose_x_0, transpose_y = var_2341_transpose_y_0, x = transpose_85, y = transpose_86)[name = tensor("op_2341_cast_fp16")]; + tensor var_2342_to_fp16 = const()[name = tensor("op_2342_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_33_cast_fp16 = mul(x = var_2341_cast_fp16, y = var_2342_to_fp16)[name = tensor("scores_33_cast_fp16")]; + tensor scores_35_cast_fp16 = add(x = scores_33_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_35_cast_fp16")]; + tensor var_2352 = const()[name = tensor("op_2352"), val = tensor(-1)]; + tensor weights_17_cast_fp16 = softmax(axis = var_2352, x = scores_35_cast_fp16)[name = tensor("weights_17_cast_fp16")]; + tensor attn_17_transpose_x_0 = const()[name = tensor("attn_17_transpose_x_0"), val = tensor(false)]; + tensor attn_17_transpose_y_0 = const()[name = tensor("attn_17_transpose_y_0"), val = tensor(false)]; + tensor values_17_cast_fp16 = transpose(perm = var_2330, x = new_v_cache_17_cast_fp16)[name = tensor("transpose_174")]; + tensor attn_17_cast_fp16 = matmul(transpose_x = attn_17_transpose_x_0, transpose_y = attn_17_transpose_y_0, x = weights_17_cast_fp16, y = values_17_cast_fp16)[name = tensor("attn_17_cast_fp16")]; + tensor var_2360 = const()[name = tensor("op_2360"), val = tensor([0, 2, 1, 3])]; + tensor var_2365 = const()[name = tensor("op_2365"), val = tensor([1, 256, 1024])]; + tensor var_2361_cast_fp16 = transpose(perm = var_2360, x = attn_17_cast_fp16)[name = tensor("transpose_171")]; + tensor input_101_cast_fp16 = reshape(shape = var_2365, x = var_2361_cast_fp16)[name = tensor("input_101_cast_fp16")]; + tensor attn8_out_proj_weight_to_fp16 = const()[name = tensor("attn8_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(207710976)))]; + tensor linear_33_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn8_out_proj_weight_to_fp16, x = input_101_cast_fp16)[name = tensor("linear_33_cast_fp16")]; + tensor position8_to_fp16_dtype_0 = const()[name = tensor("position8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position8_to_fp16 = cast(dtype = position8_to_fp16_dtype_0, x = position8)[name = tensor("cast_413")]; + tensor var_2374_cast_fp16 = add(x = position8_to_fp16, y = valid_len_to_fp16)[name = tensor("op_2374_cast_fp16")]; + tensor var_2374_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2374_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_103_cast_fp16 = add(x = input_97_cast_fp16, y = linear_33_cast_fp16)[name = tensor("input_103_cast_fp16")]; + tensor input_105_axes_0 = const()[name = tensor("input_105_axes_0"), val = tensor([-1])]; + tensor norm8_2_weight_to_fp16 = const()[name = tensor("norm8_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209808192)))]; + tensor norm8_2_bias_to_fp16 = const()[name = tensor("norm8_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209810304)))]; + tensor var_2378_to_fp16 = const()[name = tensor("op_2378_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_105_cast_fp16 = layer_norm(axes = input_105_axes_0, beta = norm8_2_bias_to_fp16, epsilon = var_2378_to_fp16, gamma = norm8_2_weight_to_fp16, x = input_103_cast_fp16)[name = tensor("input_105_cast_fp16")]; + tensor linear8_1_weight_to_fp16 = const()[name = tensor("linear8_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209812416)))]; + tensor linear_34_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear8_1_weight_to_fp16, x = input_105_cast_fp16)[name = tensor("linear_34_cast_fp16")]; + tensor input_107_mode_0 = const()[name = tensor("input_107_mode_0"), val = tensor("EXACT")]; + tensor input_107_cast_fp16 = gelu(mode = input_107_mode_0, x = linear_34_cast_fp16)[name = tensor("input_107_cast_fp16")]; + tensor linear8_2_weight_to_fp16 = const()[name = tensor("linear8_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(218201088)))]; + tensor linear_35_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear8_2_weight_to_fp16, x = input_107_cast_fp16)[name = tensor("linear_35_cast_fp16")]; + tensor input_109_cast_fp16 = add(x = input_103_cast_fp16, y = linear_35_cast_fp16)[name = tensor("input_109_cast_fp16")]; + tensor input_111_axes_0 = const()[name = tensor("input_111_axes_0"), val = tensor([-1])]; + tensor norm9_1_weight_to_fp16 = const()[name = tensor("norm9_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226589760)))]; + tensor norm9_1_bias_to_fp16 = const()[name = tensor("norm9_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226591872)))]; + tensor var_2395_to_fp16 = const()[name = tensor("op_2395_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_111_cast_fp16 = layer_norm(axes = input_111_axes_0, beta = norm9_1_bias_to_fp16, epsilon = var_2395_to_fp16, gamma = norm9_1_weight_to_fp16, x = input_109_cast_fp16)[name = tensor("input_111_cast_fp16")]; + tensor attn9_in_proj_weight_to_fp16 = const()[name = tensor("attn9_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226593984)))]; + tensor linear_36_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn9_in_proj_weight_to_fp16, x = input_111_cast_fp16)[name = tensor("linear_36_cast_fp16")]; + tensor var_2409 = const()[name = tensor("op_2409"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_19_cast_fp16 = reshape(shape = var_2409, x = linear_36_cast_fp16)[name = tensor("qkv_19_cast_fp16")]; + tensor x_37_begin_0 = const()[name = tensor("x_37_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_37_end_0 = const()[name = tensor("x_37_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_37_end_mask_0 = const()[name = tensor("x_37_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_37_squeeze_mask_0 = const()[name = tensor("x_37_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_37_cast_fp16 = slice_by_index(begin = x_37_begin_0, end = x_37_end_0, end_mask = x_37_end_mask_0, squeeze_mask = x_37_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("x_37_cast_fp16")]; + tensor x_39_begin_0 = const()[name = tensor("x_39_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_39_end_0 = const()[name = tensor("x_39_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_39_end_mask_0 = const()[name = tensor("x_39_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_39_squeeze_mask_0 = const()[name = tensor("x_39_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_39_cast_fp16 = slice_by_index(begin = x_39_begin_0, end = x_39_end_0, end_mask = x_39_end_mask_0, squeeze_mask = x_39_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("x_39_cast_fp16")]; + tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_19_cast_fp16 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("v_19_cast_fp16")]; + tensor var_2454 = const()[name = tensor("op_2454"), val = tensor([256, 16, 32, 2])]; + tensor pairs_37_cast_fp16 = reshape(shape = var_2454, x = x_37_cast_fp16)[name = tensor("pairs_37_cast_fp16")]; + tensor xr_37_begin_0 = const()[name = tensor("xr_37_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_37_end_0 = const()[name = tensor("xr_37_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_37_end_mask_0 = const()[name = tensor("xr_37_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_37_squeeze_mask_0 = const()[name = tensor("xr_37_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_37_cast_fp16 = slice_by_index(begin = xr_37_begin_0, end = xr_37_end_0, end_mask = xr_37_end_mask_0, squeeze_mask = xr_37_squeeze_mask_0, x = pairs_37_cast_fp16)[name = tensor("xr_37_cast_fp16")]; + tensor xi_37_begin_0 = const()[name = tensor("xi_37_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_37_end_0 = const()[name = tensor("xi_37_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_37_end_mask_0 = const()[name = tensor("xi_37_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_37_squeeze_mask_0 = const()[name = tensor("xi_37_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_37_cast_fp16 = slice_by_index(begin = xi_37_begin_0, end = xi_37_end_0, end_mask = xi_37_end_mask_0, squeeze_mask = xi_37_squeeze_mask_0, x = pairs_37_cast_fp16)[name = tensor("xi_37_cast_fp16")]; + tensor var_2462_cast_fp16 = mul(x = xr_37_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2462_cast_fp16")]; + tensor var_2463_cast_fp16 = mul(x = xi_37_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2463_cast_fp16")]; + tensor xor__37_cast_fp16 = sub(x = var_2462_cast_fp16, y = var_2463_cast_fp16)[name = tensor("xor__37_cast_fp16")]; + tensor var_2466_cast_fp16 = mul(x = xr_37_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2466_cast_fp16")]; + tensor var_2467_cast_fp16 = mul(x = xi_37_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2467_cast_fp16")]; + tensor xoi_37_cast_fp16 = add(x = var_2466_cast_fp16, y = var_2467_cast_fp16)[name = tensor("xoi_37_cast_fp16")]; + tensor out_37_axis_0 = const()[name = tensor("out_37_axis_0"), val = tensor(-1)]; + tensor out_37_cast_fp16 = stack(axis = out_37_axis_0, values = (xor__37_cast_fp16, xoi_37_cast_fp16))[name = tensor("out_37_cast_fp16")]; + tensor var_2477 = const()[name = tensor("op_2477"), val = tensor([1, 256, 16, 64])]; + tensor q_19_cast_fp16 = reshape(shape = var_2477, x = out_37_cast_fp16)[name = tensor("q_19_cast_fp16")]; + tensor var_2483 = const()[name = tensor("op_2483"), val = tensor([256, 16, 32, 2])]; + tensor pairs_39_cast_fp16 = reshape(shape = var_2483, x = x_39_cast_fp16)[name = tensor("pairs_39_cast_fp16")]; + tensor xr_39_begin_0 = const()[name = tensor("xr_39_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_39_end_0 = const()[name = tensor("xr_39_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_39_end_mask_0 = const()[name = tensor("xr_39_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_39_squeeze_mask_0 = const()[name = tensor("xr_39_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_39_cast_fp16 = slice_by_index(begin = xr_39_begin_0, end = xr_39_end_0, end_mask = xr_39_end_mask_0, squeeze_mask = xr_39_squeeze_mask_0, x = pairs_39_cast_fp16)[name = tensor("xr_39_cast_fp16")]; + tensor xi_39_begin_0 = const()[name = tensor("xi_39_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_39_end_0 = const()[name = tensor("xi_39_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_39_end_mask_0 = const()[name = tensor("xi_39_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_39_squeeze_mask_0 = const()[name = tensor("xi_39_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_39_cast_fp16 = slice_by_index(begin = xi_39_begin_0, end = xi_39_end_0, end_mask = xi_39_end_mask_0, squeeze_mask = xi_39_squeeze_mask_0, x = pairs_39_cast_fp16)[name = tensor("xi_39_cast_fp16")]; + tensor var_2491_cast_fp16 = mul(x = xr_39_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2491_cast_fp16")]; + tensor var_2492_cast_fp16 = mul(x = xi_39_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2492_cast_fp16")]; + tensor xor__39_cast_fp16 = sub(x = var_2491_cast_fp16, y = var_2492_cast_fp16)[name = tensor("xor__39_cast_fp16")]; + tensor var_2495_cast_fp16 = mul(x = xr_39_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2495_cast_fp16")]; + tensor var_2496_cast_fp16 = mul(x = xi_39_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2496_cast_fp16")]; + tensor xoi_39_cast_fp16 = add(x = var_2495_cast_fp16, y = var_2496_cast_fp16)[name = tensor("xoi_39_cast_fp16")]; + tensor out_39_axis_0 = const()[name = tensor("out_39_axis_0"), val = tensor(-1)]; + tensor out_39_cast_fp16 = stack(axis = out_39_axis_0, values = (xor__39_cast_fp16, xoi_39_cast_fp16))[name = tensor("out_39_cast_fp16")]; + tensor var_2513 = const()[name = tensor("op_2513"), val = tensor([256, 1024])]; + tensor var_2514_cast_fp16 = reshape(shape = var_2513, x = out_39_cast_fp16)[name = tensor("op_2514_cast_fp16")]; + tensor written_k_19_transpose_x_1 = const()[name = tensor("written_k_19_transpose_x_1"), val = tensor(true)]; + tensor written_k_19_transpose_y_1 = const()[name = tensor("written_k_19_transpose_y_1"), val = tensor(false)]; + tensor written_k_19_cast_fp16 = matmul(transpose_x = written_k_19_transpose_x_1, transpose_y = written_k_19_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_2514_cast_fp16)[name = tensor("written_k_19_cast_fp16")]; + tensor var_2521 = const()[name = tensor("op_2521"), val = tensor([256, 1024])]; + tensor var_2522_cast_fp16 = reshape(shape = var_2521, x = v_19_cast_fp16)[name = tensor("op_2522_cast_fp16")]; + tensor written_v_19_transpose_x_1 = const()[name = tensor("written_v_19_transpose_x_1"), val = tensor(true)]; + tensor written_v_19_transpose_y_1 = const()[name = tensor("written_v_19_transpose_y_1"), val = tensor(false)]; + tensor written_v_19_cast_fp16 = matmul(transpose_x = written_v_19_transpose_x_1, transpose_y = written_v_19_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_2522_cast_fp16)[name = tensor("written_v_19_cast_fp16")]; + tensor k_cache9_to_fp16_dtype_0 = const()[name = tensor("k_cache9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache9_to_fp16 = cast(dtype = k_cache9_to_fp16_dtype_0, x = k_cache9)[name = tensor("cast_411")]; + tensor var_2533_cast_fp16 = mul(x = k_cache9_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_2533_cast_fp16")]; + tensor var_2538 = const()[name = tensor("op_2538"), val = tensor([1, 512, 16, 64])]; + tensor var_2539_cast_fp16 = reshape(shape = var_2538, x = written_k_19_cast_fp16)[name = tensor("op_2539_cast_fp16")]; + tensor new_k_cache_19_cast_fp16 = add(x = var_2533_cast_fp16, y = var_2539_cast_fp16)[name = tensor("new_k_cache_19_cast_fp16")]; + tensor new_k_cache_19_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_19_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache9_to_fp16_dtype_0 = const()[name = tensor("v_cache9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache9_to_fp16 = cast(dtype = v_cache9_to_fp16_dtype_0, x = v_cache9)[name = tensor("cast_409")]; + tensor var_2542_cast_fp16 = mul(x = v_cache9_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_2542_cast_fp16")]; + tensor var_2547 = const()[name = tensor("op_2547"), val = tensor([1, 512, 16, 64])]; + tensor var_2548_cast_fp16 = reshape(shape = var_2547, x = written_v_19_cast_fp16)[name = tensor("op_2548_cast_fp16")]; + tensor new_v_cache_19_cast_fp16 = add(x = var_2542_cast_fp16, y = var_2548_cast_fp16)[name = tensor("new_v_cache_19_cast_fp16")]; + tensor new_v_cache_19_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_19_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_2561 = const()[name = tensor("op_2561"), val = tensor([0, 2, 1, 3])]; + tensor var_2572_transpose_x_0 = const()[name = tensor("op_2572_transpose_x_0"), val = tensor(false)]; + tensor var_2572_transpose_y_0 = const()[name = tensor("op_2572_transpose_y_0"), val = tensor(false)]; + tensor transpose_87_perm_0 = const()[name = tensor("transpose_87_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_88_perm_0 = const()[name = tensor("transpose_88_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_88 = transpose(perm = transpose_88_perm_0, x = new_k_cache_19_cast_fp16)[name = tensor("transpose_168")]; + tensor transpose_87 = transpose(perm = transpose_87_perm_0, x = q_19_cast_fp16)[name = tensor("transpose_169")]; + tensor var_2572_cast_fp16 = matmul(transpose_x = var_2572_transpose_x_0, transpose_y = var_2572_transpose_y_0, x = transpose_87, y = transpose_88)[name = tensor("op_2572_cast_fp16")]; + tensor var_2573_to_fp16 = const()[name = tensor("op_2573_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_37_cast_fp16 = mul(x = var_2572_cast_fp16, y = var_2573_to_fp16)[name = tensor("scores_37_cast_fp16")]; + tensor scores_39_cast_fp16 = add(x = scores_37_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_39_cast_fp16")]; + tensor var_2583 = const()[name = tensor("op_2583"), val = tensor(-1)]; + tensor weights_19_cast_fp16 = softmax(axis = var_2583, x = scores_39_cast_fp16)[name = tensor("weights_19_cast_fp16")]; + tensor attn_19_transpose_x_0 = const()[name = tensor("attn_19_transpose_x_0"), val = tensor(false)]; + tensor attn_19_transpose_y_0 = const()[name = tensor("attn_19_transpose_y_0"), val = tensor(false)]; + tensor values_19_cast_fp16 = transpose(perm = var_2561, x = new_v_cache_19_cast_fp16)[name = tensor("transpose_170")]; + tensor attn_19_cast_fp16 = matmul(transpose_x = attn_19_transpose_x_0, transpose_y = attn_19_transpose_y_0, x = weights_19_cast_fp16, y = values_19_cast_fp16)[name = tensor("attn_19_cast_fp16")]; + tensor var_2591 = const()[name = tensor("op_2591"), val = tensor([0, 2, 1, 3])]; + tensor var_2596 = const()[name = tensor("op_2596"), val = tensor([1, 256, 1024])]; + tensor var_2592_cast_fp16 = transpose(perm = var_2591, x = attn_19_cast_fp16)[name = tensor("transpose_167")]; + tensor input_113_cast_fp16 = reshape(shape = var_2596, x = var_2592_cast_fp16)[name = tensor("input_113_cast_fp16")]; + tensor attn9_out_proj_weight_to_fp16 = const()[name = tensor("attn9_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(232885504)))]; + tensor linear_37_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn9_out_proj_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("linear_37_cast_fp16")]; + tensor position9_to_fp16_dtype_0 = const()[name = tensor("position9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position9_to_fp16 = cast(dtype = position9_to_fp16_dtype_0, x = position9)[name = tensor("cast_407")]; + tensor var_2605_cast_fp16 = add(x = position9_to_fp16, y = valid_len_to_fp16)[name = tensor("op_2605_cast_fp16")]; + tensor var_2605_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2605_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_115_cast_fp16 = add(x = input_109_cast_fp16, y = linear_37_cast_fp16)[name = tensor("input_115_cast_fp16")]; + tensor input_117_axes_0 = const()[name = tensor("input_117_axes_0"), val = tensor([-1])]; + tensor norm9_2_weight_to_fp16 = const()[name = tensor("norm9_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(234982720)))]; + tensor norm9_2_bias_to_fp16 = const()[name = tensor("norm9_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(234984832)))]; + tensor var_2609_to_fp16 = const()[name = tensor("op_2609_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_117_cast_fp16 = layer_norm(axes = input_117_axes_0, beta = norm9_2_bias_to_fp16, epsilon = var_2609_to_fp16, gamma = norm9_2_weight_to_fp16, x = input_115_cast_fp16)[name = tensor("input_117_cast_fp16")]; + tensor linear9_1_weight_to_fp16 = const()[name = tensor("linear9_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(234986944)))]; + tensor linear_38_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear9_1_weight_to_fp16, x = input_117_cast_fp16)[name = tensor("linear_38_cast_fp16")]; + tensor input_119_mode_0 = const()[name = tensor("input_119_mode_0"), val = tensor("EXACT")]; + tensor input_119_cast_fp16 = gelu(mode = input_119_mode_0, x = linear_38_cast_fp16)[name = tensor("input_119_cast_fp16")]; + tensor linear9_2_weight_to_fp16 = const()[name = tensor("linear9_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(243375616)))]; + tensor linear_39_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear9_2_weight_to_fp16, x = input_119_cast_fp16)[name = tensor("linear_39_cast_fp16")]; + tensor input_121_cast_fp16 = add(x = input_115_cast_fp16, y = linear_39_cast_fp16)[name = tensor("input_121_cast_fp16")]; + tensor input_123_axes_0 = const()[name = tensor("input_123_axes_0"), val = tensor([-1])]; + tensor norm10_1_weight_to_fp16 = const()[name = tensor("norm10_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(251764288)))]; + tensor norm10_1_bias_to_fp16 = const()[name = tensor("norm10_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(251766400)))]; + tensor var_2626_to_fp16 = const()[name = tensor("op_2626_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_123_cast_fp16 = layer_norm(axes = input_123_axes_0, beta = norm10_1_bias_to_fp16, epsilon = var_2626_to_fp16, gamma = norm10_1_weight_to_fp16, x = input_121_cast_fp16)[name = tensor("input_123_cast_fp16")]; + tensor attn10_in_proj_weight_to_fp16 = const()[name = tensor("attn10_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(251768512)))]; + tensor linear_40_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn10_in_proj_weight_to_fp16, x = input_123_cast_fp16)[name = tensor("linear_40_cast_fp16")]; + tensor var_2640 = const()[name = tensor("op_2640"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_21_cast_fp16 = reshape(shape = var_2640, x = linear_40_cast_fp16)[name = tensor("qkv_21_cast_fp16")]; + tensor x_41_begin_0 = const()[name = tensor("x_41_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_41_end_0 = const()[name = tensor("x_41_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_41_end_mask_0 = const()[name = tensor("x_41_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_41_squeeze_mask_0 = const()[name = tensor("x_41_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_41_cast_fp16 = slice_by_index(begin = x_41_begin_0, end = x_41_end_0, end_mask = x_41_end_mask_0, squeeze_mask = x_41_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("x_41_cast_fp16")]; + tensor x_43_begin_0 = const()[name = tensor("x_43_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_43_end_0 = const()[name = tensor("x_43_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_43_end_mask_0 = const()[name = tensor("x_43_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_43_squeeze_mask_0 = const()[name = tensor("x_43_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_43_cast_fp16 = slice_by_index(begin = x_43_begin_0, end = x_43_end_0, end_mask = x_43_end_mask_0, squeeze_mask = x_43_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("x_43_cast_fp16")]; + tensor v_21_begin_0 = const()[name = tensor("v_21_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_21_end_0 = const()[name = tensor("v_21_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_21_end_mask_0 = const()[name = tensor("v_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_21_squeeze_mask_0 = const()[name = tensor("v_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_21_cast_fp16 = slice_by_index(begin = v_21_begin_0, end = v_21_end_0, end_mask = v_21_end_mask_0, squeeze_mask = v_21_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("v_21_cast_fp16")]; + tensor var_2685 = const()[name = tensor("op_2685"), val = tensor([256, 16, 32, 2])]; + tensor pairs_41_cast_fp16 = reshape(shape = var_2685, x = x_41_cast_fp16)[name = tensor("pairs_41_cast_fp16")]; + tensor xr_41_begin_0 = const()[name = tensor("xr_41_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_41_end_0 = const()[name = tensor("xr_41_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_41_end_mask_0 = const()[name = tensor("xr_41_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_41_squeeze_mask_0 = const()[name = tensor("xr_41_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_41_cast_fp16 = slice_by_index(begin = xr_41_begin_0, end = xr_41_end_0, end_mask = xr_41_end_mask_0, squeeze_mask = xr_41_squeeze_mask_0, x = pairs_41_cast_fp16)[name = tensor("xr_41_cast_fp16")]; + tensor xi_41_begin_0 = const()[name = tensor("xi_41_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_41_end_0 = const()[name = tensor("xi_41_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_41_end_mask_0 = const()[name = tensor("xi_41_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_41_squeeze_mask_0 = const()[name = tensor("xi_41_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_41_cast_fp16 = slice_by_index(begin = xi_41_begin_0, end = xi_41_end_0, end_mask = xi_41_end_mask_0, squeeze_mask = xi_41_squeeze_mask_0, x = pairs_41_cast_fp16)[name = tensor("xi_41_cast_fp16")]; + tensor var_2693_cast_fp16 = mul(x = xr_41_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2693_cast_fp16")]; + tensor var_2694_cast_fp16 = mul(x = xi_41_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2694_cast_fp16")]; + tensor xor__41_cast_fp16 = sub(x = var_2693_cast_fp16, y = var_2694_cast_fp16)[name = tensor("xor__41_cast_fp16")]; + tensor var_2697_cast_fp16 = mul(x = xr_41_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2697_cast_fp16")]; + tensor var_2698_cast_fp16 = mul(x = xi_41_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2698_cast_fp16")]; + tensor xoi_41_cast_fp16 = add(x = var_2697_cast_fp16, y = var_2698_cast_fp16)[name = tensor("xoi_41_cast_fp16")]; + tensor out_41_axis_0 = const()[name = tensor("out_41_axis_0"), val = tensor(-1)]; + tensor out_41_cast_fp16 = stack(axis = out_41_axis_0, values = (xor__41_cast_fp16, xoi_41_cast_fp16))[name = tensor("out_41_cast_fp16")]; + tensor var_2708 = const()[name = tensor("op_2708"), val = tensor([1, 256, 16, 64])]; + tensor q_21_cast_fp16 = reshape(shape = var_2708, x = out_41_cast_fp16)[name = tensor("q_21_cast_fp16")]; + tensor var_2714 = const()[name = tensor("op_2714"), val = tensor([256, 16, 32, 2])]; + tensor pairs_43_cast_fp16 = reshape(shape = var_2714, x = x_43_cast_fp16)[name = tensor("pairs_43_cast_fp16")]; + tensor xr_43_begin_0 = const()[name = tensor("xr_43_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_43_end_0 = const()[name = tensor("xr_43_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_43_end_mask_0 = const()[name = tensor("xr_43_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_43_squeeze_mask_0 = const()[name = tensor("xr_43_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_43_cast_fp16 = slice_by_index(begin = xr_43_begin_0, end = xr_43_end_0, end_mask = xr_43_end_mask_0, squeeze_mask = xr_43_squeeze_mask_0, x = pairs_43_cast_fp16)[name = tensor("xr_43_cast_fp16")]; + tensor xi_43_begin_0 = const()[name = tensor("xi_43_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_43_end_0 = const()[name = tensor("xi_43_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_43_end_mask_0 = const()[name = tensor("xi_43_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_43_squeeze_mask_0 = const()[name = tensor("xi_43_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_43_cast_fp16 = slice_by_index(begin = xi_43_begin_0, end = xi_43_end_0, end_mask = xi_43_end_mask_0, squeeze_mask = xi_43_squeeze_mask_0, x = pairs_43_cast_fp16)[name = tensor("xi_43_cast_fp16")]; + tensor var_2722_cast_fp16 = mul(x = xr_43_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2722_cast_fp16")]; + tensor var_2723_cast_fp16 = mul(x = xi_43_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2723_cast_fp16")]; + tensor xor__43_cast_fp16 = sub(x = var_2722_cast_fp16, y = var_2723_cast_fp16)[name = tensor("xor__43_cast_fp16")]; + tensor var_2726_cast_fp16 = mul(x = xr_43_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2726_cast_fp16")]; + tensor var_2727_cast_fp16 = mul(x = xi_43_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2727_cast_fp16")]; + tensor xoi_43_cast_fp16 = add(x = var_2726_cast_fp16, y = var_2727_cast_fp16)[name = tensor("xoi_43_cast_fp16")]; + tensor out_43_axis_0 = const()[name = tensor("out_43_axis_0"), val = tensor(-1)]; + tensor out_43_cast_fp16 = stack(axis = out_43_axis_0, values = (xor__43_cast_fp16, xoi_43_cast_fp16))[name = tensor("out_43_cast_fp16")]; + tensor var_2744 = const()[name = tensor("op_2744"), val = tensor([256, 1024])]; + tensor var_2745_cast_fp16 = reshape(shape = var_2744, x = out_43_cast_fp16)[name = tensor("op_2745_cast_fp16")]; + tensor written_k_21_transpose_x_1 = const()[name = tensor("written_k_21_transpose_x_1"), val = tensor(true)]; + tensor written_k_21_transpose_y_1 = const()[name = tensor("written_k_21_transpose_y_1"), val = tensor(false)]; + tensor written_k_21_cast_fp16 = matmul(transpose_x = written_k_21_transpose_x_1, transpose_y = written_k_21_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_2745_cast_fp16)[name = tensor("written_k_21_cast_fp16")]; + tensor var_2752 = const()[name = tensor("op_2752"), val = tensor([256, 1024])]; + tensor var_2753_cast_fp16 = reshape(shape = var_2752, x = v_21_cast_fp16)[name = tensor("op_2753_cast_fp16")]; + tensor written_v_21_transpose_x_1 = const()[name = tensor("written_v_21_transpose_x_1"), val = tensor(true)]; + tensor written_v_21_transpose_y_1 = const()[name = tensor("written_v_21_transpose_y_1"), val = tensor(false)]; + tensor written_v_21_cast_fp16 = matmul(transpose_x = written_v_21_transpose_x_1, transpose_y = written_v_21_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_2753_cast_fp16)[name = tensor("written_v_21_cast_fp16")]; + tensor k_cache10_to_fp16_dtype_0 = const()[name = tensor("k_cache10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache10_to_fp16 = cast(dtype = k_cache10_to_fp16_dtype_0, x = k_cache10)[name = tensor("cast_405")]; + tensor var_2764_cast_fp16 = mul(x = k_cache10_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_2764_cast_fp16")]; + tensor var_2769 = const()[name = tensor("op_2769"), val = tensor([1, 512, 16, 64])]; + tensor var_2770_cast_fp16 = reshape(shape = var_2769, x = written_k_21_cast_fp16)[name = tensor("op_2770_cast_fp16")]; + tensor new_k_cache_21_cast_fp16 = add(x = var_2764_cast_fp16, y = var_2770_cast_fp16)[name = tensor("new_k_cache_21_cast_fp16")]; + tensor new_k_cache_21_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_21_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache10_to_fp16_dtype_0 = const()[name = tensor("v_cache10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache10_to_fp16 = cast(dtype = v_cache10_to_fp16_dtype_0, x = v_cache10)[name = tensor("cast_403")]; + tensor var_2773_cast_fp16 = mul(x = v_cache10_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_2773_cast_fp16")]; + tensor var_2778 = const()[name = tensor("op_2778"), val = tensor([1, 512, 16, 64])]; + tensor var_2779_cast_fp16 = reshape(shape = var_2778, x = written_v_21_cast_fp16)[name = tensor("op_2779_cast_fp16")]; + tensor new_v_cache_21_cast_fp16 = add(x = var_2773_cast_fp16, y = var_2779_cast_fp16)[name = tensor("new_v_cache_21_cast_fp16")]; + tensor new_v_cache_21_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_21_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_2792 = const()[name = tensor("op_2792"), val = tensor([0, 2, 1, 3])]; + tensor var_2803_transpose_x_0 = const()[name = tensor("op_2803_transpose_x_0"), val = tensor(false)]; + tensor var_2803_transpose_y_0 = const()[name = tensor("op_2803_transpose_y_0"), val = tensor(false)]; + tensor transpose_89_perm_0 = const()[name = tensor("transpose_89_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_90_perm_0 = const()[name = tensor("transpose_90_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_90 = transpose(perm = transpose_90_perm_0, x = new_k_cache_21_cast_fp16)[name = tensor("transpose_164")]; + tensor transpose_89 = transpose(perm = transpose_89_perm_0, x = q_21_cast_fp16)[name = tensor("transpose_165")]; + tensor var_2803_cast_fp16 = matmul(transpose_x = var_2803_transpose_x_0, transpose_y = var_2803_transpose_y_0, x = transpose_89, y = transpose_90)[name = tensor("op_2803_cast_fp16")]; + tensor var_2804_to_fp16 = const()[name = tensor("op_2804_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_41_cast_fp16 = mul(x = var_2803_cast_fp16, y = var_2804_to_fp16)[name = tensor("scores_41_cast_fp16")]; + tensor scores_43_cast_fp16 = add(x = scores_41_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_43_cast_fp16")]; + tensor var_2814 = const()[name = tensor("op_2814"), val = tensor(-1)]; + tensor weights_21_cast_fp16 = softmax(axis = var_2814, x = scores_43_cast_fp16)[name = tensor("weights_21_cast_fp16")]; + tensor attn_21_transpose_x_0 = const()[name = tensor("attn_21_transpose_x_0"), val = tensor(false)]; + tensor attn_21_transpose_y_0 = const()[name = tensor("attn_21_transpose_y_0"), val = tensor(false)]; + tensor values_21_cast_fp16 = transpose(perm = var_2792, x = new_v_cache_21_cast_fp16)[name = tensor("transpose_166")]; + tensor attn_21_cast_fp16 = matmul(transpose_x = attn_21_transpose_x_0, transpose_y = attn_21_transpose_y_0, x = weights_21_cast_fp16, y = values_21_cast_fp16)[name = tensor("attn_21_cast_fp16")]; + tensor var_2822 = const()[name = tensor("op_2822"), val = tensor([0, 2, 1, 3])]; + tensor var_2827 = const()[name = tensor("op_2827"), val = tensor([1, 256, 1024])]; + tensor var_2823_cast_fp16 = transpose(perm = var_2822, x = attn_21_cast_fp16)[name = tensor("transpose_163")]; + tensor input_125_cast_fp16 = reshape(shape = var_2827, x = var_2823_cast_fp16)[name = tensor("input_125_cast_fp16")]; + tensor attn10_out_proj_weight_to_fp16 = const()[name = tensor("attn10_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(258060032)))]; + tensor linear_41_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn10_out_proj_weight_to_fp16, x = input_125_cast_fp16)[name = tensor("linear_41_cast_fp16")]; + tensor position10_to_fp16_dtype_0 = const()[name = tensor("position10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position10_to_fp16 = cast(dtype = position10_to_fp16_dtype_0, x = position10)[name = tensor("cast_401")]; + tensor var_2836_cast_fp16 = add(x = position10_to_fp16, y = valid_len_to_fp16)[name = tensor("op_2836_cast_fp16")]; + tensor var_2836_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2836_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_127_cast_fp16 = add(x = input_121_cast_fp16, y = linear_41_cast_fp16)[name = tensor("input_127_cast_fp16")]; + tensor input_129_axes_0 = const()[name = tensor("input_129_axes_0"), val = tensor([-1])]; + tensor norm10_2_weight_to_fp16 = const()[name = tensor("norm10_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260157248)))]; + tensor norm10_2_bias_to_fp16 = const()[name = tensor("norm10_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260159360)))]; + tensor var_2840_to_fp16 = const()[name = tensor("op_2840_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_129_cast_fp16 = layer_norm(axes = input_129_axes_0, beta = norm10_2_bias_to_fp16, epsilon = var_2840_to_fp16, gamma = norm10_2_weight_to_fp16, x = input_127_cast_fp16)[name = tensor("input_129_cast_fp16")]; + tensor linear10_1_weight_to_fp16 = const()[name = tensor("linear10_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260161472)))]; + tensor linear_42_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear10_1_weight_to_fp16, x = input_129_cast_fp16)[name = tensor("linear_42_cast_fp16")]; + tensor input_131_mode_0 = const()[name = tensor("input_131_mode_0"), val = tensor("EXACT")]; + tensor input_131_cast_fp16 = gelu(mode = input_131_mode_0, x = linear_42_cast_fp16)[name = tensor("input_131_cast_fp16")]; + tensor linear10_2_weight_to_fp16 = const()[name = tensor("linear10_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(268550144)))]; + tensor linear_43_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear10_2_weight_to_fp16, x = input_131_cast_fp16)[name = tensor("linear_43_cast_fp16")]; + tensor input_133_cast_fp16 = add(x = input_127_cast_fp16, y = linear_43_cast_fp16)[name = tensor("input_133_cast_fp16")]; + tensor input_135_axes_0 = const()[name = tensor("input_135_axes_0"), val = tensor([-1])]; + tensor norm11_1_weight_to_fp16 = const()[name = tensor("norm11_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(276938816)))]; + tensor norm11_1_bias_to_fp16 = const()[name = tensor("norm11_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(276940928)))]; + tensor var_2857_to_fp16 = const()[name = tensor("op_2857_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_135_cast_fp16 = layer_norm(axes = input_135_axes_0, beta = norm11_1_bias_to_fp16, epsilon = var_2857_to_fp16, gamma = norm11_1_weight_to_fp16, x = input_133_cast_fp16)[name = tensor("input_135_cast_fp16")]; + tensor attn11_in_proj_weight_to_fp16 = const()[name = tensor("attn11_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(276943040)))]; + tensor linear_44_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn11_in_proj_weight_to_fp16, x = input_135_cast_fp16)[name = tensor("linear_44_cast_fp16")]; + tensor var_2871 = const()[name = tensor("op_2871"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_23_cast_fp16 = reshape(shape = var_2871, x = linear_44_cast_fp16)[name = tensor("qkv_23_cast_fp16")]; + tensor x_45_begin_0 = const()[name = tensor("x_45_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_45_end_0 = const()[name = tensor("x_45_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_45_end_mask_0 = const()[name = tensor("x_45_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_45_squeeze_mask_0 = const()[name = tensor("x_45_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_45_cast_fp16 = slice_by_index(begin = x_45_begin_0, end = x_45_end_0, end_mask = x_45_end_mask_0, squeeze_mask = x_45_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("x_45_cast_fp16")]; + tensor x_47_begin_0 = const()[name = tensor("x_47_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_47_end_0 = const()[name = tensor("x_47_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_47_end_mask_0 = const()[name = tensor("x_47_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_47_squeeze_mask_0 = const()[name = tensor("x_47_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_47_cast_fp16 = slice_by_index(begin = x_47_begin_0, end = x_47_end_0, end_mask = x_47_end_mask_0, squeeze_mask = x_47_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("x_47_cast_fp16")]; + tensor v_23_begin_0 = const()[name = tensor("v_23_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_23_end_0 = const()[name = tensor("v_23_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_23_end_mask_0 = const()[name = tensor("v_23_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_23_squeeze_mask_0 = const()[name = tensor("v_23_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_23_cast_fp16 = slice_by_index(begin = v_23_begin_0, end = v_23_end_0, end_mask = v_23_end_mask_0, squeeze_mask = v_23_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("v_23_cast_fp16")]; + tensor var_2916 = const()[name = tensor("op_2916"), val = tensor([256, 16, 32, 2])]; + tensor pairs_45_cast_fp16 = reshape(shape = var_2916, x = x_45_cast_fp16)[name = tensor("pairs_45_cast_fp16")]; + tensor xr_45_begin_0 = const()[name = tensor("xr_45_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_45_end_0 = const()[name = tensor("xr_45_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_45_end_mask_0 = const()[name = tensor("xr_45_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_45_squeeze_mask_0 = const()[name = tensor("xr_45_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_45_cast_fp16 = slice_by_index(begin = xr_45_begin_0, end = xr_45_end_0, end_mask = xr_45_end_mask_0, squeeze_mask = xr_45_squeeze_mask_0, x = pairs_45_cast_fp16)[name = tensor("xr_45_cast_fp16")]; + tensor xi_45_begin_0 = const()[name = tensor("xi_45_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_45_end_0 = const()[name = tensor("xi_45_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_45_end_mask_0 = const()[name = tensor("xi_45_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_45_squeeze_mask_0 = const()[name = tensor("xi_45_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_45_cast_fp16 = slice_by_index(begin = xi_45_begin_0, end = xi_45_end_0, end_mask = xi_45_end_mask_0, squeeze_mask = xi_45_squeeze_mask_0, x = pairs_45_cast_fp16)[name = tensor("xi_45_cast_fp16")]; + tensor var_2924_cast_fp16 = mul(x = xr_45_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2924_cast_fp16")]; + tensor var_2925_cast_fp16 = mul(x = xi_45_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2925_cast_fp16")]; + tensor xor__45_cast_fp16 = sub(x = var_2924_cast_fp16, y = var_2925_cast_fp16)[name = tensor("xor__45_cast_fp16")]; + tensor var_2928_cast_fp16 = mul(x = xr_45_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2928_cast_fp16")]; + tensor var_2929_cast_fp16 = mul(x = xi_45_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2929_cast_fp16")]; + tensor xoi_45_cast_fp16 = add(x = var_2928_cast_fp16, y = var_2929_cast_fp16)[name = tensor("xoi_45_cast_fp16")]; + tensor out_45_axis_0 = const()[name = tensor("out_45_axis_0"), val = tensor(-1)]; + tensor out_45_cast_fp16 = stack(axis = out_45_axis_0, values = (xor__45_cast_fp16, xoi_45_cast_fp16))[name = tensor("out_45_cast_fp16")]; + tensor var_2939 = const()[name = tensor("op_2939"), val = tensor([1, 256, 16, 64])]; + tensor q_23_cast_fp16 = reshape(shape = var_2939, x = out_45_cast_fp16)[name = tensor("q_23_cast_fp16")]; + tensor var_2945 = const()[name = tensor("op_2945"), val = tensor([256, 16, 32, 2])]; + tensor pairs_47_cast_fp16 = reshape(shape = var_2945, x = x_47_cast_fp16)[name = tensor("pairs_47_cast_fp16")]; + tensor xr_47_begin_0 = const()[name = tensor("xr_47_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_47_end_0 = const()[name = tensor("xr_47_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_47_end_mask_0 = const()[name = tensor("xr_47_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_47_squeeze_mask_0 = const()[name = tensor("xr_47_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_47_cast_fp16 = slice_by_index(begin = xr_47_begin_0, end = xr_47_end_0, end_mask = xr_47_end_mask_0, squeeze_mask = xr_47_squeeze_mask_0, x = pairs_47_cast_fp16)[name = tensor("xr_47_cast_fp16")]; + tensor xi_47_begin_0 = const()[name = tensor("xi_47_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_47_end_0 = const()[name = tensor("xi_47_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_47_end_mask_0 = const()[name = tensor("xi_47_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_47_squeeze_mask_0 = const()[name = tensor("xi_47_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_47_cast_fp16 = slice_by_index(begin = xi_47_begin_0, end = xi_47_end_0, end_mask = xi_47_end_mask_0, squeeze_mask = xi_47_squeeze_mask_0, x = pairs_47_cast_fp16)[name = tensor("xi_47_cast_fp16")]; + tensor var_2953_cast_fp16 = mul(x = xr_47_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2953_cast_fp16")]; + tensor var_2954_cast_fp16 = mul(x = xi_47_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2954_cast_fp16")]; + tensor xor__47_cast_fp16 = sub(x = var_2953_cast_fp16, y = var_2954_cast_fp16)[name = tensor("xor__47_cast_fp16")]; + tensor var_2957_cast_fp16 = mul(x = xr_47_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2957_cast_fp16")]; + tensor var_2958_cast_fp16 = mul(x = xi_47_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2958_cast_fp16")]; + tensor xoi_47_cast_fp16 = add(x = var_2957_cast_fp16, y = var_2958_cast_fp16)[name = tensor("xoi_47_cast_fp16")]; + tensor out_47_axis_0 = const()[name = tensor("out_47_axis_0"), val = tensor(-1)]; + tensor out_47_cast_fp16 = stack(axis = out_47_axis_0, values = (xor__47_cast_fp16, xoi_47_cast_fp16))[name = tensor("out_47_cast_fp16")]; + tensor var_2975 = const()[name = tensor("op_2975"), val = tensor([256, 1024])]; + tensor var_2976_cast_fp16 = reshape(shape = var_2975, x = out_47_cast_fp16)[name = tensor("op_2976_cast_fp16")]; + tensor written_k_23_transpose_x_1 = const()[name = tensor("written_k_23_transpose_x_1"), val = tensor(true)]; + tensor written_k_23_transpose_y_1 = const()[name = tensor("written_k_23_transpose_y_1"), val = tensor(false)]; + tensor written_k_23_cast_fp16 = matmul(transpose_x = written_k_23_transpose_x_1, transpose_y = written_k_23_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_2976_cast_fp16)[name = tensor("written_k_23_cast_fp16")]; + tensor var_2983 = const()[name = tensor("op_2983"), val = tensor([256, 1024])]; + tensor var_2984_cast_fp16 = reshape(shape = var_2983, x = v_23_cast_fp16)[name = tensor("op_2984_cast_fp16")]; + tensor written_v_23_transpose_x_1 = const()[name = tensor("written_v_23_transpose_x_1"), val = tensor(true)]; + tensor written_v_23_transpose_y_1 = const()[name = tensor("written_v_23_transpose_y_1"), val = tensor(false)]; + tensor written_v_23_cast_fp16 = matmul(transpose_x = written_v_23_transpose_x_1, transpose_y = written_v_23_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_2984_cast_fp16)[name = tensor("written_v_23_cast_fp16")]; + tensor k_cache11_to_fp16_dtype_0 = const()[name = tensor("k_cache11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache11_to_fp16 = cast(dtype = k_cache11_to_fp16_dtype_0, x = k_cache11)[name = tensor("cast_399")]; + tensor var_2995_cast_fp16 = mul(x = k_cache11_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_2995_cast_fp16")]; + tensor var_3000 = const()[name = tensor("op_3000"), val = tensor([1, 512, 16, 64])]; + tensor var_3001_cast_fp16 = reshape(shape = var_3000, x = written_k_23_cast_fp16)[name = tensor("op_3001_cast_fp16")]; + tensor new_k_cache_23_cast_fp16 = add(x = var_2995_cast_fp16, y = var_3001_cast_fp16)[name = tensor("new_k_cache_23_cast_fp16")]; + tensor new_k_cache_23_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_23_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache11_to_fp16_dtype_0 = const()[name = tensor("v_cache11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache11_to_fp16 = cast(dtype = v_cache11_to_fp16_dtype_0, x = v_cache11)[name = tensor("cast_397")]; + tensor var_3004_cast_fp16 = mul(x = v_cache11_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_3004_cast_fp16")]; + tensor var_3009 = const()[name = tensor("op_3009"), val = tensor([1, 512, 16, 64])]; + tensor var_3010_cast_fp16 = reshape(shape = var_3009, x = written_v_23_cast_fp16)[name = tensor("op_3010_cast_fp16")]; + tensor new_v_cache_23_cast_fp16 = add(x = var_3004_cast_fp16, y = var_3010_cast_fp16)[name = tensor("new_v_cache_23_cast_fp16")]; + tensor new_v_cache_23_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_23_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_3023 = const()[name = tensor("op_3023"), val = tensor([0, 2, 1, 3])]; + tensor var_3034_transpose_x_0 = const()[name = tensor("op_3034_transpose_x_0"), val = tensor(false)]; + tensor var_3034_transpose_y_0 = const()[name = tensor("op_3034_transpose_y_0"), val = tensor(false)]; + tensor transpose_91_perm_0 = const()[name = tensor("transpose_91_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_92_perm_0 = const()[name = tensor("transpose_92_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_92 = transpose(perm = transpose_92_perm_0, x = new_k_cache_23_cast_fp16)[name = tensor("transpose_160")]; + tensor transpose_91 = transpose(perm = transpose_91_perm_0, x = q_23_cast_fp16)[name = tensor("transpose_161")]; + tensor var_3034_cast_fp16 = matmul(transpose_x = var_3034_transpose_x_0, transpose_y = var_3034_transpose_y_0, x = transpose_91, y = transpose_92)[name = tensor("op_3034_cast_fp16")]; + tensor var_3035_to_fp16 = const()[name = tensor("op_3035_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_45_cast_fp16 = mul(x = var_3034_cast_fp16, y = var_3035_to_fp16)[name = tensor("scores_45_cast_fp16")]; + tensor scores_47_cast_fp16 = add(x = scores_45_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_47_cast_fp16")]; + tensor var_3045 = const()[name = tensor("op_3045"), val = tensor(-1)]; + tensor weights_23_cast_fp16 = softmax(axis = var_3045, x = scores_47_cast_fp16)[name = tensor("weights_23_cast_fp16")]; + tensor attn_23_transpose_x_0 = const()[name = tensor("attn_23_transpose_x_0"), val = tensor(false)]; + tensor attn_23_transpose_y_0 = const()[name = tensor("attn_23_transpose_y_0"), val = tensor(false)]; + tensor values_23_cast_fp16 = transpose(perm = var_3023, x = new_v_cache_23_cast_fp16)[name = tensor("transpose_162")]; + tensor attn_23_cast_fp16 = matmul(transpose_x = attn_23_transpose_x_0, transpose_y = attn_23_transpose_y_0, x = weights_23_cast_fp16, y = values_23_cast_fp16)[name = tensor("attn_23_cast_fp16")]; + tensor var_3053 = const()[name = tensor("op_3053"), val = tensor([0, 2, 1, 3])]; + tensor var_3058 = const()[name = tensor("op_3058"), val = tensor([1, 256, 1024])]; + tensor var_3054_cast_fp16 = transpose(perm = var_3053, x = attn_23_cast_fp16)[name = tensor("transpose_159")]; + tensor input_137_cast_fp16 = reshape(shape = var_3058, x = var_3054_cast_fp16)[name = tensor("input_137_cast_fp16")]; + tensor attn11_out_proj_weight_to_fp16 = const()[name = tensor("attn11_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(283234560)))]; + tensor linear_45_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn11_out_proj_weight_to_fp16, x = input_137_cast_fp16)[name = tensor("linear_45_cast_fp16")]; + tensor position11_to_fp16_dtype_0 = const()[name = tensor("position11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position11_to_fp16 = cast(dtype = position11_to_fp16_dtype_0, x = position11)[name = tensor("cast_395")]; + tensor var_3067_cast_fp16 = add(x = position11_to_fp16, y = valid_len_to_fp16)[name = tensor("op_3067_cast_fp16")]; + tensor var_3067_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3067_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_139_cast_fp16 = add(x = input_133_cast_fp16, y = linear_45_cast_fp16)[name = tensor("input_139_cast_fp16")]; + tensor input_141_axes_0 = const()[name = tensor("input_141_axes_0"), val = tensor([-1])]; + tensor norm11_2_weight_to_fp16 = const()[name = tensor("norm11_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285331776)))]; + tensor norm11_2_bias_to_fp16 = const()[name = tensor("norm11_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285333888)))]; + tensor var_3071_to_fp16 = const()[name = tensor("op_3071_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_141_cast_fp16 = layer_norm(axes = input_141_axes_0, beta = norm11_2_bias_to_fp16, epsilon = var_3071_to_fp16, gamma = norm11_2_weight_to_fp16, x = input_139_cast_fp16)[name = tensor("input_141_cast_fp16")]; + tensor linear11_1_weight_to_fp16 = const()[name = tensor("linear11_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285336000)))]; + tensor linear_46_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear11_1_weight_to_fp16, x = input_141_cast_fp16)[name = tensor("linear_46_cast_fp16")]; + tensor input_143_mode_0 = const()[name = tensor("input_143_mode_0"), val = tensor("EXACT")]; + tensor input_143_cast_fp16 = gelu(mode = input_143_mode_0, x = linear_46_cast_fp16)[name = tensor("input_143_cast_fp16")]; + tensor linear11_2_weight_to_fp16 = const()[name = tensor("linear11_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293724672)))]; + tensor linear_47_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear11_2_weight_to_fp16, x = input_143_cast_fp16)[name = tensor("linear_47_cast_fp16")]; + tensor input_145_cast_fp16 = add(x = input_139_cast_fp16, y = linear_47_cast_fp16)[name = tensor("input_145_cast_fp16")]; + tensor input_147_axes_0 = const()[name = tensor("input_147_axes_0"), val = tensor([-1])]; + tensor norm12_1_weight_to_fp16 = const()[name = tensor("norm12_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302113344)))]; + tensor norm12_1_bias_to_fp16 = const()[name = tensor("norm12_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302115456)))]; + tensor var_3088_to_fp16 = const()[name = tensor("op_3088_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_147_cast_fp16 = layer_norm(axes = input_147_axes_0, beta = norm12_1_bias_to_fp16, epsilon = var_3088_to_fp16, gamma = norm12_1_weight_to_fp16, x = input_145_cast_fp16)[name = tensor("input_147_cast_fp16")]; + tensor attn12_in_proj_weight_to_fp16 = const()[name = tensor("attn12_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302117568)))]; + tensor linear_48_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn12_in_proj_weight_to_fp16, x = input_147_cast_fp16)[name = tensor("linear_48_cast_fp16")]; + tensor var_3102 = const()[name = tensor("op_3102"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_25_cast_fp16 = reshape(shape = var_3102, x = linear_48_cast_fp16)[name = tensor("qkv_25_cast_fp16")]; + tensor x_49_begin_0 = const()[name = tensor("x_49_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_49_end_0 = const()[name = tensor("x_49_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_49_end_mask_0 = const()[name = tensor("x_49_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_49_squeeze_mask_0 = const()[name = tensor("x_49_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_49_cast_fp16 = slice_by_index(begin = x_49_begin_0, end = x_49_end_0, end_mask = x_49_end_mask_0, squeeze_mask = x_49_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("x_49_cast_fp16")]; + tensor x_51_begin_0 = const()[name = tensor("x_51_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_51_end_0 = const()[name = tensor("x_51_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_51_end_mask_0 = const()[name = tensor("x_51_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_51_squeeze_mask_0 = const()[name = tensor("x_51_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_51_cast_fp16 = slice_by_index(begin = x_51_begin_0, end = x_51_end_0, end_mask = x_51_end_mask_0, squeeze_mask = x_51_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("x_51_cast_fp16")]; + tensor v_25_begin_0 = const()[name = tensor("v_25_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_25_end_0 = const()[name = tensor("v_25_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_25_end_mask_0 = const()[name = tensor("v_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_25_squeeze_mask_0 = const()[name = tensor("v_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_25_cast_fp16 = slice_by_index(begin = v_25_begin_0, end = v_25_end_0, end_mask = v_25_end_mask_0, squeeze_mask = v_25_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("v_25_cast_fp16")]; + tensor var_3147 = const()[name = tensor("op_3147"), val = tensor([256, 16, 32, 2])]; + tensor pairs_49_cast_fp16 = reshape(shape = var_3147, x = x_49_cast_fp16)[name = tensor("pairs_49_cast_fp16")]; + tensor xr_49_begin_0 = const()[name = tensor("xr_49_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_49_end_0 = const()[name = tensor("xr_49_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_49_end_mask_0 = const()[name = tensor("xr_49_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_49_squeeze_mask_0 = const()[name = tensor("xr_49_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_49_cast_fp16 = slice_by_index(begin = xr_49_begin_0, end = xr_49_end_0, end_mask = xr_49_end_mask_0, squeeze_mask = xr_49_squeeze_mask_0, x = pairs_49_cast_fp16)[name = tensor("xr_49_cast_fp16")]; + tensor xi_49_begin_0 = const()[name = tensor("xi_49_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_49_end_0 = const()[name = tensor("xi_49_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_49_end_mask_0 = const()[name = tensor("xi_49_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_49_squeeze_mask_0 = const()[name = tensor("xi_49_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_49_cast_fp16 = slice_by_index(begin = xi_49_begin_0, end = xi_49_end_0, end_mask = xi_49_end_mask_0, squeeze_mask = xi_49_squeeze_mask_0, x = pairs_49_cast_fp16)[name = tensor("xi_49_cast_fp16")]; + tensor var_3155_cast_fp16 = mul(x = xr_49_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3155_cast_fp16")]; + tensor var_3156_cast_fp16 = mul(x = xi_49_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3156_cast_fp16")]; + tensor xor__49_cast_fp16 = sub(x = var_3155_cast_fp16, y = var_3156_cast_fp16)[name = tensor("xor__49_cast_fp16")]; + tensor var_3159_cast_fp16 = mul(x = xr_49_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3159_cast_fp16")]; + tensor var_3160_cast_fp16 = mul(x = xi_49_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3160_cast_fp16")]; + tensor xoi_49_cast_fp16 = add(x = var_3159_cast_fp16, y = var_3160_cast_fp16)[name = tensor("xoi_49_cast_fp16")]; + tensor out_49_axis_0 = const()[name = tensor("out_49_axis_0"), val = tensor(-1)]; + tensor out_49_cast_fp16 = stack(axis = out_49_axis_0, values = (xor__49_cast_fp16, xoi_49_cast_fp16))[name = tensor("out_49_cast_fp16")]; + tensor var_3170 = const()[name = tensor("op_3170"), val = tensor([1, 256, 16, 64])]; + tensor q_25_cast_fp16 = reshape(shape = var_3170, x = out_49_cast_fp16)[name = tensor("q_25_cast_fp16")]; + tensor var_3176 = const()[name = tensor("op_3176"), val = tensor([256, 16, 32, 2])]; + tensor pairs_51_cast_fp16 = reshape(shape = var_3176, x = x_51_cast_fp16)[name = tensor("pairs_51_cast_fp16")]; + tensor xr_51_begin_0 = const()[name = tensor("xr_51_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_51_end_0 = const()[name = tensor("xr_51_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_51_end_mask_0 = const()[name = tensor("xr_51_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_51_squeeze_mask_0 = const()[name = tensor("xr_51_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_51_cast_fp16 = slice_by_index(begin = xr_51_begin_0, end = xr_51_end_0, end_mask = xr_51_end_mask_0, squeeze_mask = xr_51_squeeze_mask_0, x = pairs_51_cast_fp16)[name = tensor("xr_51_cast_fp16")]; + tensor xi_51_begin_0 = const()[name = tensor("xi_51_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_51_end_0 = const()[name = tensor("xi_51_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_51_end_mask_0 = const()[name = tensor("xi_51_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_51_squeeze_mask_0 = const()[name = tensor("xi_51_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_51_cast_fp16 = slice_by_index(begin = xi_51_begin_0, end = xi_51_end_0, end_mask = xi_51_end_mask_0, squeeze_mask = xi_51_squeeze_mask_0, x = pairs_51_cast_fp16)[name = tensor("xi_51_cast_fp16")]; + tensor var_3184_cast_fp16 = mul(x = xr_51_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3184_cast_fp16")]; + tensor var_3185_cast_fp16 = mul(x = xi_51_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3185_cast_fp16")]; + tensor xor__51_cast_fp16 = sub(x = var_3184_cast_fp16, y = var_3185_cast_fp16)[name = tensor("xor__51_cast_fp16")]; + tensor var_3188_cast_fp16 = mul(x = xr_51_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3188_cast_fp16")]; + tensor var_3189_cast_fp16 = mul(x = xi_51_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3189_cast_fp16")]; + tensor xoi_51_cast_fp16 = add(x = var_3188_cast_fp16, y = var_3189_cast_fp16)[name = tensor("xoi_51_cast_fp16")]; + tensor out_51_axis_0 = const()[name = tensor("out_51_axis_0"), val = tensor(-1)]; + tensor out_51_cast_fp16 = stack(axis = out_51_axis_0, values = (xor__51_cast_fp16, xoi_51_cast_fp16))[name = tensor("out_51_cast_fp16")]; + tensor var_3206 = const()[name = tensor("op_3206"), val = tensor([256, 1024])]; + tensor var_3207_cast_fp16 = reshape(shape = var_3206, x = out_51_cast_fp16)[name = tensor("op_3207_cast_fp16")]; + tensor written_k_25_transpose_x_1 = const()[name = tensor("written_k_25_transpose_x_1"), val = tensor(true)]; + tensor written_k_25_transpose_y_1 = const()[name = tensor("written_k_25_transpose_y_1"), val = tensor(false)]; + tensor written_k_25_cast_fp16 = matmul(transpose_x = written_k_25_transpose_x_1, transpose_y = written_k_25_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_3207_cast_fp16)[name = tensor("written_k_25_cast_fp16")]; + tensor var_3214 = const()[name = tensor("op_3214"), val = tensor([256, 1024])]; + tensor var_3215_cast_fp16 = reshape(shape = var_3214, x = v_25_cast_fp16)[name = tensor("op_3215_cast_fp16")]; + tensor written_v_25_transpose_x_1 = const()[name = tensor("written_v_25_transpose_x_1"), val = tensor(true)]; + tensor written_v_25_transpose_y_1 = const()[name = tensor("written_v_25_transpose_y_1"), val = tensor(false)]; + tensor written_v_25_cast_fp16 = matmul(transpose_x = written_v_25_transpose_x_1, transpose_y = written_v_25_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_3215_cast_fp16)[name = tensor("written_v_25_cast_fp16")]; + tensor k_cache12_to_fp16_dtype_0 = const()[name = tensor("k_cache12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache12_to_fp16 = cast(dtype = k_cache12_to_fp16_dtype_0, x = k_cache12)[name = tensor("cast_393")]; + tensor var_3226_cast_fp16 = mul(x = k_cache12_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_3226_cast_fp16")]; + tensor var_3231 = const()[name = tensor("op_3231"), val = tensor([1, 512, 16, 64])]; + tensor var_3232_cast_fp16 = reshape(shape = var_3231, x = written_k_25_cast_fp16)[name = tensor("op_3232_cast_fp16")]; + tensor new_k_cache_25_cast_fp16 = add(x = var_3226_cast_fp16, y = var_3232_cast_fp16)[name = tensor("new_k_cache_25_cast_fp16")]; + tensor new_k_cache_25_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_25_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache12_to_fp16_dtype_0 = const()[name = tensor("v_cache12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache12_to_fp16 = cast(dtype = v_cache12_to_fp16_dtype_0, x = v_cache12)[name = tensor("cast_391")]; + tensor var_3235_cast_fp16 = mul(x = v_cache12_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_3235_cast_fp16")]; + tensor var_3240 = const()[name = tensor("op_3240"), val = tensor([1, 512, 16, 64])]; + tensor var_3241_cast_fp16 = reshape(shape = var_3240, x = written_v_25_cast_fp16)[name = tensor("op_3241_cast_fp16")]; + tensor new_v_cache_25_cast_fp16 = add(x = var_3235_cast_fp16, y = var_3241_cast_fp16)[name = tensor("new_v_cache_25_cast_fp16")]; + tensor new_v_cache_25_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_25_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_3254 = const()[name = tensor("op_3254"), val = tensor([0, 2, 1, 3])]; + tensor var_3265_transpose_x_0 = const()[name = tensor("op_3265_transpose_x_0"), val = tensor(false)]; + tensor var_3265_transpose_y_0 = const()[name = tensor("op_3265_transpose_y_0"), val = tensor(false)]; + tensor transpose_93_perm_0 = const()[name = tensor("transpose_93_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_94_perm_0 = const()[name = tensor("transpose_94_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_94 = transpose(perm = transpose_94_perm_0, x = new_k_cache_25_cast_fp16)[name = tensor("transpose_156")]; + tensor transpose_93 = transpose(perm = transpose_93_perm_0, x = q_25_cast_fp16)[name = tensor("transpose_157")]; + tensor var_3265_cast_fp16 = matmul(transpose_x = var_3265_transpose_x_0, transpose_y = var_3265_transpose_y_0, x = transpose_93, y = transpose_94)[name = tensor("op_3265_cast_fp16")]; + tensor var_3266_to_fp16 = const()[name = tensor("op_3266_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_49_cast_fp16 = mul(x = var_3265_cast_fp16, y = var_3266_to_fp16)[name = tensor("scores_49_cast_fp16")]; + tensor scores_51_cast_fp16 = add(x = scores_49_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_51_cast_fp16")]; + tensor var_3276 = const()[name = tensor("op_3276"), val = tensor(-1)]; + tensor weights_25_cast_fp16 = softmax(axis = var_3276, x = scores_51_cast_fp16)[name = tensor("weights_25_cast_fp16")]; + tensor attn_25_transpose_x_0 = const()[name = tensor("attn_25_transpose_x_0"), val = tensor(false)]; + tensor attn_25_transpose_y_0 = const()[name = tensor("attn_25_transpose_y_0"), val = tensor(false)]; + tensor values_25_cast_fp16 = transpose(perm = var_3254, x = new_v_cache_25_cast_fp16)[name = tensor("transpose_158")]; + tensor attn_25_cast_fp16 = matmul(transpose_x = attn_25_transpose_x_0, transpose_y = attn_25_transpose_y_0, x = weights_25_cast_fp16, y = values_25_cast_fp16)[name = tensor("attn_25_cast_fp16")]; + tensor var_3284 = const()[name = tensor("op_3284"), val = tensor([0, 2, 1, 3])]; + tensor var_3289 = const()[name = tensor("op_3289"), val = tensor([1, 256, 1024])]; + tensor var_3285_cast_fp16 = transpose(perm = var_3284, x = attn_25_cast_fp16)[name = tensor("transpose_155")]; + tensor input_149_cast_fp16 = reshape(shape = var_3289, x = var_3285_cast_fp16)[name = tensor("input_149_cast_fp16")]; + tensor attn12_out_proj_weight_to_fp16 = const()[name = tensor("attn12_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(308409088)))]; + tensor linear_49_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn12_out_proj_weight_to_fp16, x = input_149_cast_fp16)[name = tensor("linear_49_cast_fp16")]; + tensor position12_to_fp16_dtype_0 = const()[name = tensor("position12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position12_to_fp16 = cast(dtype = position12_to_fp16_dtype_0, x = position12)[name = tensor("cast_389")]; + tensor var_3298_cast_fp16 = add(x = position12_to_fp16, y = valid_len_to_fp16)[name = tensor("op_3298_cast_fp16")]; + tensor var_3298_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3298_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_151_cast_fp16 = add(x = input_145_cast_fp16, y = linear_49_cast_fp16)[name = tensor("input_151_cast_fp16")]; + tensor input_153_axes_0 = const()[name = tensor("input_153_axes_0"), val = tensor([-1])]; + tensor norm12_2_weight_to_fp16 = const()[name = tensor("norm12_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(310506304)))]; + tensor norm12_2_bias_to_fp16 = const()[name = tensor("norm12_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(310508416)))]; + tensor var_3302_to_fp16 = const()[name = tensor("op_3302_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_153_cast_fp16 = layer_norm(axes = input_153_axes_0, beta = norm12_2_bias_to_fp16, epsilon = var_3302_to_fp16, gamma = norm12_2_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("input_153_cast_fp16")]; + tensor linear12_1_weight_to_fp16 = const()[name = tensor("linear12_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(310510528)))]; + tensor linear_50_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear12_1_weight_to_fp16, x = input_153_cast_fp16)[name = tensor("linear_50_cast_fp16")]; + tensor input_155_mode_0 = const()[name = tensor("input_155_mode_0"), val = tensor("EXACT")]; + tensor input_155_cast_fp16 = gelu(mode = input_155_mode_0, x = linear_50_cast_fp16)[name = tensor("input_155_cast_fp16")]; + tensor linear12_2_weight_to_fp16 = const()[name = tensor("linear12_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(318899200)))]; + tensor linear_51_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear12_2_weight_to_fp16, x = input_155_cast_fp16)[name = tensor("linear_51_cast_fp16")]; + tensor input_157_cast_fp16 = add(x = input_151_cast_fp16, y = linear_51_cast_fp16)[name = tensor("input_157_cast_fp16")]; + tensor input_159_axes_0 = const()[name = tensor("input_159_axes_0"), val = tensor([-1])]; + tensor norm13_1_weight_to_fp16 = const()[name = tensor("norm13_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(327287872)))]; + tensor norm13_1_bias_to_fp16 = const()[name = tensor("norm13_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(327289984)))]; + tensor var_3319_to_fp16 = const()[name = tensor("op_3319_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_159_cast_fp16 = layer_norm(axes = input_159_axes_0, beta = norm13_1_bias_to_fp16, epsilon = var_3319_to_fp16, gamma = norm13_1_weight_to_fp16, x = input_157_cast_fp16)[name = tensor("input_159_cast_fp16")]; + tensor attn13_in_proj_weight_to_fp16 = const()[name = tensor("attn13_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(327292096)))]; + tensor linear_52_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn13_in_proj_weight_to_fp16, x = input_159_cast_fp16)[name = tensor("linear_52_cast_fp16")]; + tensor var_3333 = const()[name = tensor("op_3333"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_27_cast_fp16 = reshape(shape = var_3333, x = linear_52_cast_fp16)[name = tensor("qkv_27_cast_fp16")]; + tensor x_53_begin_0 = const()[name = tensor("x_53_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_53_end_0 = const()[name = tensor("x_53_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_53_end_mask_0 = const()[name = tensor("x_53_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_53_squeeze_mask_0 = const()[name = tensor("x_53_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_53_cast_fp16 = slice_by_index(begin = x_53_begin_0, end = x_53_end_0, end_mask = x_53_end_mask_0, squeeze_mask = x_53_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("x_53_cast_fp16")]; + tensor x_55_begin_0 = const()[name = tensor("x_55_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_55_end_0 = const()[name = tensor("x_55_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_55_end_mask_0 = const()[name = tensor("x_55_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_55_squeeze_mask_0 = const()[name = tensor("x_55_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_55_cast_fp16 = slice_by_index(begin = x_55_begin_0, end = x_55_end_0, end_mask = x_55_end_mask_0, squeeze_mask = x_55_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("x_55_cast_fp16")]; + tensor v_27_begin_0 = const()[name = tensor("v_27_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_27_end_0 = const()[name = tensor("v_27_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_27_end_mask_0 = const()[name = tensor("v_27_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_27_squeeze_mask_0 = const()[name = tensor("v_27_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_27_cast_fp16 = slice_by_index(begin = v_27_begin_0, end = v_27_end_0, end_mask = v_27_end_mask_0, squeeze_mask = v_27_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("v_27_cast_fp16")]; + tensor var_3378 = const()[name = tensor("op_3378"), val = tensor([256, 16, 32, 2])]; + tensor pairs_53_cast_fp16 = reshape(shape = var_3378, x = x_53_cast_fp16)[name = tensor("pairs_53_cast_fp16")]; + tensor xr_53_begin_0 = const()[name = tensor("xr_53_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_53_end_0 = const()[name = tensor("xr_53_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_53_end_mask_0 = const()[name = tensor("xr_53_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_53_squeeze_mask_0 = const()[name = tensor("xr_53_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_53_cast_fp16 = slice_by_index(begin = xr_53_begin_0, end = xr_53_end_0, end_mask = xr_53_end_mask_0, squeeze_mask = xr_53_squeeze_mask_0, x = pairs_53_cast_fp16)[name = tensor("xr_53_cast_fp16")]; + tensor xi_53_begin_0 = const()[name = tensor("xi_53_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_53_end_0 = const()[name = tensor("xi_53_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_53_end_mask_0 = const()[name = tensor("xi_53_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_53_squeeze_mask_0 = const()[name = tensor("xi_53_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_53_cast_fp16 = slice_by_index(begin = xi_53_begin_0, end = xi_53_end_0, end_mask = xi_53_end_mask_0, squeeze_mask = xi_53_squeeze_mask_0, x = pairs_53_cast_fp16)[name = tensor("xi_53_cast_fp16")]; + tensor var_3386_cast_fp16 = mul(x = xr_53_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3386_cast_fp16")]; + tensor var_3387_cast_fp16 = mul(x = xi_53_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3387_cast_fp16")]; + tensor xor__53_cast_fp16 = sub(x = var_3386_cast_fp16, y = var_3387_cast_fp16)[name = tensor("xor__53_cast_fp16")]; + tensor var_3390_cast_fp16 = mul(x = xr_53_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3390_cast_fp16")]; + tensor var_3391_cast_fp16 = mul(x = xi_53_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3391_cast_fp16")]; + tensor xoi_53_cast_fp16 = add(x = var_3390_cast_fp16, y = var_3391_cast_fp16)[name = tensor("xoi_53_cast_fp16")]; + tensor out_53_axis_0 = const()[name = tensor("out_53_axis_0"), val = tensor(-1)]; + tensor out_53_cast_fp16 = stack(axis = out_53_axis_0, values = (xor__53_cast_fp16, xoi_53_cast_fp16))[name = tensor("out_53_cast_fp16")]; + tensor var_3401 = const()[name = tensor("op_3401"), val = tensor([1, 256, 16, 64])]; + tensor q_27_cast_fp16 = reshape(shape = var_3401, x = out_53_cast_fp16)[name = tensor("q_27_cast_fp16")]; + tensor var_3407 = const()[name = tensor("op_3407"), val = tensor([256, 16, 32, 2])]; + tensor pairs_55_cast_fp16 = reshape(shape = var_3407, x = x_55_cast_fp16)[name = tensor("pairs_55_cast_fp16")]; + tensor xr_55_begin_0 = const()[name = tensor("xr_55_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_55_end_0 = const()[name = tensor("xr_55_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_55_end_mask_0 = const()[name = tensor("xr_55_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_55_squeeze_mask_0 = const()[name = tensor("xr_55_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_55_cast_fp16 = slice_by_index(begin = xr_55_begin_0, end = xr_55_end_0, end_mask = xr_55_end_mask_0, squeeze_mask = xr_55_squeeze_mask_0, x = pairs_55_cast_fp16)[name = tensor("xr_55_cast_fp16")]; + tensor xi_55_begin_0 = const()[name = tensor("xi_55_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_55_end_0 = const()[name = tensor("xi_55_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_55_end_mask_0 = const()[name = tensor("xi_55_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_55_squeeze_mask_0 = const()[name = tensor("xi_55_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_55_cast_fp16 = slice_by_index(begin = xi_55_begin_0, end = xi_55_end_0, end_mask = xi_55_end_mask_0, squeeze_mask = xi_55_squeeze_mask_0, x = pairs_55_cast_fp16)[name = tensor("xi_55_cast_fp16")]; + tensor var_3415_cast_fp16 = mul(x = xr_55_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3415_cast_fp16")]; + tensor var_3416_cast_fp16 = mul(x = xi_55_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3416_cast_fp16")]; + tensor xor__55_cast_fp16 = sub(x = var_3415_cast_fp16, y = var_3416_cast_fp16)[name = tensor("xor__55_cast_fp16")]; + tensor var_3419_cast_fp16 = mul(x = xr_55_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3419_cast_fp16")]; + tensor var_3420_cast_fp16 = mul(x = xi_55_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3420_cast_fp16")]; + tensor xoi_55_cast_fp16 = add(x = var_3419_cast_fp16, y = var_3420_cast_fp16)[name = tensor("xoi_55_cast_fp16")]; + tensor out_55_axis_0 = const()[name = tensor("out_55_axis_0"), val = tensor(-1)]; + tensor out_55_cast_fp16 = stack(axis = out_55_axis_0, values = (xor__55_cast_fp16, xoi_55_cast_fp16))[name = tensor("out_55_cast_fp16")]; + tensor var_3437 = const()[name = tensor("op_3437"), val = tensor([256, 1024])]; + tensor var_3438_cast_fp16 = reshape(shape = var_3437, x = out_55_cast_fp16)[name = tensor("op_3438_cast_fp16")]; + tensor written_k_27_transpose_x_1 = const()[name = tensor("written_k_27_transpose_x_1"), val = tensor(true)]; + tensor written_k_27_transpose_y_1 = const()[name = tensor("written_k_27_transpose_y_1"), val = tensor(false)]; + tensor written_k_27_cast_fp16 = matmul(transpose_x = written_k_27_transpose_x_1, transpose_y = written_k_27_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_3438_cast_fp16)[name = tensor("written_k_27_cast_fp16")]; + tensor var_3445 = const()[name = tensor("op_3445"), val = tensor([256, 1024])]; + tensor var_3446_cast_fp16 = reshape(shape = var_3445, x = v_27_cast_fp16)[name = tensor("op_3446_cast_fp16")]; + tensor written_v_27_transpose_x_1 = const()[name = tensor("written_v_27_transpose_x_1"), val = tensor(true)]; + tensor written_v_27_transpose_y_1 = const()[name = tensor("written_v_27_transpose_y_1"), val = tensor(false)]; + tensor written_v_27_cast_fp16 = matmul(transpose_x = written_v_27_transpose_x_1, transpose_y = written_v_27_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_3446_cast_fp16)[name = tensor("written_v_27_cast_fp16")]; + tensor k_cache13_to_fp16_dtype_0 = const()[name = tensor("k_cache13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache13_to_fp16 = cast(dtype = k_cache13_to_fp16_dtype_0, x = k_cache13)[name = tensor("cast_387")]; + tensor var_3457_cast_fp16 = mul(x = k_cache13_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_3457_cast_fp16")]; + tensor var_3462 = const()[name = tensor("op_3462"), val = tensor([1, 512, 16, 64])]; + tensor var_3463_cast_fp16 = reshape(shape = var_3462, x = written_k_27_cast_fp16)[name = tensor("op_3463_cast_fp16")]; + tensor new_k_cache_27_cast_fp16 = add(x = var_3457_cast_fp16, y = var_3463_cast_fp16)[name = tensor("new_k_cache_27_cast_fp16")]; + tensor new_k_cache_27_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_27_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache13_to_fp16_dtype_0 = const()[name = tensor("v_cache13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache13_to_fp16 = cast(dtype = v_cache13_to_fp16_dtype_0, x = v_cache13)[name = tensor("cast_385")]; + tensor var_3466_cast_fp16 = mul(x = v_cache13_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_3466_cast_fp16")]; + tensor var_3471 = const()[name = tensor("op_3471"), val = tensor([1, 512, 16, 64])]; + tensor var_3472_cast_fp16 = reshape(shape = var_3471, x = written_v_27_cast_fp16)[name = tensor("op_3472_cast_fp16")]; + tensor new_v_cache_27_cast_fp16 = add(x = var_3466_cast_fp16, y = var_3472_cast_fp16)[name = tensor("new_v_cache_27_cast_fp16")]; + tensor new_v_cache_27_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_27_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_3485 = const()[name = tensor("op_3485"), val = tensor([0, 2, 1, 3])]; + tensor var_3496_transpose_x_0 = const()[name = tensor("op_3496_transpose_x_0"), val = tensor(false)]; + tensor var_3496_transpose_y_0 = const()[name = tensor("op_3496_transpose_y_0"), val = tensor(false)]; + tensor transpose_95_perm_0 = const()[name = tensor("transpose_95_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_96_perm_0 = const()[name = tensor("transpose_96_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_96 = transpose(perm = transpose_96_perm_0, x = new_k_cache_27_cast_fp16)[name = tensor("transpose_152")]; + tensor transpose_95 = transpose(perm = transpose_95_perm_0, x = q_27_cast_fp16)[name = tensor("transpose_153")]; + tensor var_3496_cast_fp16 = matmul(transpose_x = var_3496_transpose_x_0, transpose_y = var_3496_transpose_y_0, x = transpose_95, y = transpose_96)[name = tensor("op_3496_cast_fp16")]; + tensor var_3497_to_fp16 = const()[name = tensor("op_3497_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_53_cast_fp16 = mul(x = var_3496_cast_fp16, y = var_3497_to_fp16)[name = tensor("scores_53_cast_fp16")]; + tensor scores_55_cast_fp16 = add(x = scores_53_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_55_cast_fp16")]; + tensor var_3507 = const()[name = tensor("op_3507"), val = tensor(-1)]; + tensor weights_27_cast_fp16 = softmax(axis = var_3507, x = scores_55_cast_fp16)[name = tensor("weights_27_cast_fp16")]; + tensor attn_27_transpose_x_0 = const()[name = tensor("attn_27_transpose_x_0"), val = tensor(false)]; + tensor attn_27_transpose_y_0 = const()[name = tensor("attn_27_transpose_y_0"), val = tensor(false)]; + tensor values_27_cast_fp16 = transpose(perm = var_3485, x = new_v_cache_27_cast_fp16)[name = tensor("transpose_154")]; + tensor attn_27_cast_fp16 = matmul(transpose_x = attn_27_transpose_x_0, transpose_y = attn_27_transpose_y_0, x = weights_27_cast_fp16, y = values_27_cast_fp16)[name = tensor("attn_27_cast_fp16")]; + tensor var_3515 = const()[name = tensor("op_3515"), val = tensor([0, 2, 1, 3])]; + tensor var_3520 = const()[name = tensor("op_3520"), val = tensor([1, 256, 1024])]; + tensor var_3516_cast_fp16 = transpose(perm = var_3515, x = attn_27_cast_fp16)[name = tensor("transpose_151")]; + tensor input_161_cast_fp16 = reshape(shape = var_3520, x = var_3516_cast_fp16)[name = tensor("input_161_cast_fp16")]; + tensor attn13_out_proj_weight_to_fp16 = const()[name = tensor("attn13_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(333583616)))]; + tensor linear_53_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn13_out_proj_weight_to_fp16, x = input_161_cast_fp16)[name = tensor("linear_53_cast_fp16")]; + tensor position13_to_fp16_dtype_0 = const()[name = tensor("position13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position13_to_fp16 = cast(dtype = position13_to_fp16_dtype_0, x = position13)[name = tensor("cast_383")]; + tensor var_3529_cast_fp16 = add(x = position13_to_fp16, y = valid_len_to_fp16)[name = tensor("op_3529_cast_fp16")]; + tensor var_3529_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3529_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_163_cast_fp16 = add(x = input_157_cast_fp16, y = linear_53_cast_fp16)[name = tensor("input_163_cast_fp16")]; + tensor input_165_axes_0 = const()[name = tensor("input_165_axes_0"), val = tensor([-1])]; + tensor norm13_2_weight_to_fp16 = const()[name = tensor("norm13_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335680832)))]; + tensor norm13_2_bias_to_fp16 = const()[name = tensor("norm13_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335682944)))]; + tensor var_3533_to_fp16 = const()[name = tensor("op_3533_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_165_cast_fp16 = layer_norm(axes = input_165_axes_0, beta = norm13_2_bias_to_fp16, epsilon = var_3533_to_fp16, gamma = norm13_2_weight_to_fp16, x = input_163_cast_fp16)[name = tensor("input_165_cast_fp16")]; + tensor linear13_1_weight_to_fp16 = const()[name = tensor("linear13_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335685056)))]; + tensor linear_54_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear13_1_weight_to_fp16, x = input_165_cast_fp16)[name = tensor("linear_54_cast_fp16")]; + tensor input_167_mode_0 = const()[name = tensor("input_167_mode_0"), val = tensor("EXACT")]; + tensor input_167_cast_fp16 = gelu(mode = input_167_mode_0, x = linear_54_cast_fp16)[name = tensor("input_167_cast_fp16")]; + tensor linear13_2_weight_to_fp16 = const()[name = tensor("linear13_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(344073728)))]; + tensor linear_55_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear13_2_weight_to_fp16, x = input_167_cast_fp16)[name = tensor("linear_55_cast_fp16")]; + tensor input_169_cast_fp16 = add(x = input_163_cast_fp16, y = linear_55_cast_fp16)[name = tensor("input_169_cast_fp16")]; + tensor input_171_axes_0 = const()[name = tensor("input_171_axes_0"), val = tensor([-1])]; + tensor norm14_1_weight_to_fp16 = const()[name = tensor("norm14_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(352462400)))]; + tensor norm14_1_bias_to_fp16 = const()[name = tensor("norm14_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(352464512)))]; + tensor var_3550_to_fp16 = const()[name = tensor("op_3550_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_171_cast_fp16 = layer_norm(axes = input_171_axes_0, beta = norm14_1_bias_to_fp16, epsilon = var_3550_to_fp16, gamma = norm14_1_weight_to_fp16, x = input_169_cast_fp16)[name = tensor("input_171_cast_fp16")]; + tensor attn14_in_proj_weight_to_fp16 = const()[name = tensor("attn14_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(352466624)))]; + tensor linear_56_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn14_in_proj_weight_to_fp16, x = input_171_cast_fp16)[name = tensor("linear_56_cast_fp16")]; + tensor var_3564 = const()[name = tensor("op_3564"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_29_cast_fp16 = reshape(shape = var_3564, x = linear_56_cast_fp16)[name = tensor("qkv_29_cast_fp16")]; + tensor x_57_begin_0 = const()[name = tensor("x_57_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_57_end_0 = const()[name = tensor("x_57_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_57_end_mask_0 = const()[name = tensor("x_57_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_57_squeeze_mask_0 = const()[name = tensor("x_57_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_57_cast_fp16 = slice_by_index(begin = x_57_begin_0, end = x_57_end_0, end_mask = x_57_end_mask_0, squeeze_mask = x_57_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("x_57_cast_fp16")]; + tensor x_59_begin_0 = const()[name = tensor("x_59_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_59_end_0 = const()[name = tensor("x_59_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_59_end_mask_0 = const()[name = tensor("x_59_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_59_squeeze_mask_0 = const()[name = tensor("x_59_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_59_cast_fp16 = slice_by_index(begin = x_59_begin_0, end = x_59_end_0, end_mask = x_59_end_mask_0, squeeze_mask = x_59_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("x_59_cast_fp16")]; + tensor v_29_begin_0 = const()[name = tensor("v_29_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_29_end_0 = const()[name = tensor("v_29_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_29_end_mask_0 = const()[name = tensor("v_29_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_29_squeeze_mask_0 = const()[name = tensor("v_29_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_29_cast_fp16 = slice_by_index(begin = v_29_begin_0, end = v_29_end_0, end_mask = v_29_end_mask_0, squeeze_mask = v_29_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("v_29_cast_fp16")]; + tensor var_3609 = const()[name = tensor("op_3609"), val = tensor([256, 16, 32, 2])]; + tensor pairs_57_cast_fp16 = reshape(shape = var_3609, x = x_57_cast_fp16)[name = tensor("pairs_57_cast_fp16")]; + tensor xr_57_begin_0 = const()[name = tensor("xr_57_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_57_end_0 = const()[name = tensor("xr_57_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_57_end_mask_0 = const()[name = tensor("xr_57_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_57_squeeze_mask_0 = const()[name = tensor("xr_57_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_57_cast_fp16 = slice_by_index(begin = xr_57_begin_0, end = xr_57_end_0, end_mask = xr_57_end_mask_0, squeeze_mask = xr_57_squeeze_mask_0, x = pairs_57_cast_fp16)[name = tensor("xr_57_cast_fp16")]; + tensor xi_57_begin_0 = const()[name = tensor("xi_57_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_57_end_0 = const()[name = tensor("xi_57_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_57_end_mask_0 = const()[name = tensor("xi_57_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_57_squeeze_mask_0 = const()[name = tensor("xi_57_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_57_cast_fp16 = slice_by_index(begin = xi_57_begin_0, end = xi_57_end_0, end_mask = xi_57_end_mask_0, squeeze_mask = xi_57_squeeze_mask_0, x = pairs_57_cast_fp16)[name = tensor("xi_57_cast_fp16")]; + tensor var_3617_cast_fp16 = mul(x = xr_57_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3617_cast_fp16")]; + tensor var_3618_cast_fp16 = mul(x = xi_57_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3618_cast_fp16")]; + tensor xor__57_cast_fp16 = sub(x = var_3617_cast_fp16, y = var_3618_cast_fp16)[name = tensor("xor__57_cast_fp16")]; + tensor var_3621_cast_fp16 = mul(x = xr_57_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3621_cast_fp16")]; + tensor var_3622_cast_fp16 = mul(x = xi_57_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3622_cast_fp16")]; + tensor xoi_57_cast_fp16 = add(x = var_3621_cast_fp16, y = var_3622_cast_fp16)[name = tensor("xoi_57_cast_fp16")]; + tensor out_57_axis_0 = const()[name = tensor("out_57_axis_0"), val = tensor(-1)]; + tensor out_57_cast_fp16 = stack(axis = out_57_axis_0, values = (xor__57_cast_fp16, xoi_57_cast_fp16))[name = tensor("out_57_cast_fp16")]; + tensor var_3632 = const()[name = tensor("op_3632"), val = tensor([1, 256, 16, 64])]; + tensor q_29_cast_fp16 = reshape(shape = var_3632, x = out_57_cast_fp16)[name = tensor("q_29_cast_fp16")]; + tensor var_3638 = const()[name = tensor("op_3638"), val = tensor([256, 16, 32, 2])]; + tensor pairs_59_cast_fp16 = reshape(shape = var_3638, x = x_59_cast_fp16)[name = tensor("pairs_59_cast_fp16")]; + tensor xr_59_begin_0 = const()[name = tensor("xr_59_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_59_end_0 = const()[name = tensor("xr_59_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_59_end_mask_0 = const()[name = tensor("xr_59_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_59_squeeze_mask_0 = const()[name = tensor("xr_59_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_59_cast_fp16 = slice_by_index(begin = xr_59_begin_0, end = xr_59_end_0, end_mask = xr_59_end_mask_0, squeeze_mask = xr_59_squeeze_mask_0, x = pairs_59_cast_fp16)[name = tensor("xr_59_cast_fp16")]; + tensor xi_59_begin_0 = const()[name = tensor("xi_59_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_59_end_0 = const()[name = tensor("xi_59_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_59_end_mask_0 = const()[name = tensor("xi_59_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_59_squeeze_mask_0 = const()[name = tensor("xi_59_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_59_cast_fp16 = slice_by_index(begin = xi_59_begin_0, end = xi_59_end_0, end_mask = xi_59_end_mask_0, squeeze_mask = xi_59_squeeze_mask_0, x = pairs_59_cast_fp16)[name = tensor("xi_59_cast_fp16")]; + tensor var_3646_cast_fp16 = mul(x = xr_59_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3646_cast_fp16")]; + tensor var_3647_cast_fp16 = mul(x = xi_59_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3647_cast_fp16")]; + tensor xor__59_cast_fp16 = sub(x = var_3646_cast_fp16, y = var_3647_cast_fp16)[name = tensor("xor__59_cast_fp16")]; + tensor var_3650_cast_fp16 = mul(x = xr_59_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3650_cast_fp16")]; + tensor var_3651_cast_fp16 = mul(x = xi_59_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3651_cast_fp16")]; + tensor xoi_59_cast_fp16 = add(x = var_3650_cast_fp16, y = var_3651_cast_fp16)[name = tensor("xoi_59_cast_fp16")]; + tensor out_59_axis_0 = const()[name = tensor("out_59_axis_0"), val = tensor(-1)]; + tensor out_59_cast_fp16 = stack(axis = out_59_axis_0, values = (xor__59_cast_fp16, xoi_59_cast_fp16))[name = tensor("out_59_cast_fp16")]; + tensor var_3668 = const()[name = tensor("op_3668"), val = tensor([256, 1024])]; + tensor var_3669_cast_fp16 = reshape(shape = var_3668, x = out_59_cast_fp16)[name = tensor("op_3669_cast_fp16")]; + tensor written_k_29_transpose_x_1 = const()[name = tensor("written_k_29_transpose_x_1"), val = tensor(true)]; + tensor written_k_29_transpose_y_1 = const()[name = tensor("written_k_29_transpose_y_1"), val = tensor(false)]; + tensor written_k_29_cast_fp16 = matmul(transpose_x = written_k_29_transpose_x_1, transpose_y = written_k_29_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_3669_cast_fp16)[name = tensor("written_k_29_cast_fp16")]; + tensor var_3676 = const()[name = tensor("op_3676"), val = tensor([256, 1024])]; + tensor var_3677_cast_fp16 = reshape(shape = var_3676, x = v_29_cast_fp16)[name = tensor("op_3677_cast_fp16")]; + tensor written_v_29_transpose_x_1 = const()[name = tensor("written_v_29_transpose_x_1"), val = tensor(true)]; + tensor written_v_29_transpose_y_1 = const()[name = tensor("written_v_29_transpose_y_1"), val = tensor(false)]; + tensor written_v_29_cast_fp16 = matmul(transpose_x = written_v_29_transpose_x_1, transpose_y = written_v_29_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_3677_cast_fp16)[name = tensor("written_v_29_cast_fp16")]; + tensor k_cache14_to_fp16_dtype_0 = const()[name = tensor("k_cache14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache14_to_fp16 = cast(dtype = k_cache14_to_fp16_dtype_0, x = k_cache14)[name = tensor("cast_381")]; + tensor var_3688_cast_fp16 = mul(x = k_cache14_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_3688_cast_fp16")]; + tensor var_3693 = const()[name = tensor("op_3693"), val = tensor([1, 512, 16, 64])]; + tensor var_3694_cast_fp16 = reshape(shape = var_3693, x = written_k_29_cast_fp16)[name = tensor("op_3694_cast_fp16")]; + tensor new_k_cache_29_cast_fp16 = add(x = var_3688_cast_fp16, y = var_3694_cast_fp16)[name = tensor("new_k_cache_29_cast_fp16")]; + tensor new_k_cache_29_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_29_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache14_to_fp16_dtype_0 = const()[name = tensor("v_cache14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache14_to_fp16 = cast(dtype = v_cache14_to_fp16_dtype_0, x = v_cache14)[name = tensor("cast_379")]; + tensor var_3697_cast_fp16 = mul(x = v_cache14_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_3697_cast_fp16")]; + tensor var_3702 = const()[name = tensor("op_3702"), val = tensor([1, 512, 16, 64])]; + tensor var_3703_cast_fp16 = reshape(shape = var_3702, x = written_v_29_cast_fp16)[name = tensor("op_3703_cast_fp16")]; + tensor new_v_cache_29_cast_fp16 = add(x = var_3697_cast_fp16, y = var_3703_cast_fp16)[name = tensor("new_v_cache_29_cast_fp16")]; + tensor new_v_cache_29_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_29_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_3716 = const()[name = tensor("op_3716"), val = tensor([0, 2, 1, 3])]; + tensor var_3727_transpose_x_0 = const()[name = tensor("op_3727_transpose_x_0"), val = tensor(false)]; + tensor var_3727_transpose_y_0 = const()[name = tensor("op_3727_transpose_y_0"), val = tensor(false)]; + tensor transpose_97_perm_0 = const()[name = tensor("transpose_97_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_98_perm_0 = const()[name = tensor("transpose_98_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_98 = transpose(perm = transpose_98_perm_0, x = new_k_cache_29_cast_fp16)[name = tensor("transpose_148")]; + tensor transpose_97 = transpose(perm = transpose_97_perm_0, x = q_29_cast_fp16)[name = tensor("transpose_149")]; + tensor var_3727_cast_fp16 = matmul(transpose_x = var_3727_transpose_x_0, transpose_y = var_3727_transpose_y_0, x = transpose_97, y = transpose_98)[name = tensor("op_3727_cast_fp16")]; + tensor var_3728_to_fp16 = const()[name = tensor("op_3728_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_57_cast_fp16 = mul(x = var_3727_cast_fp16, y = var_3728_to_fp16)[name = tensor("scores_57_cast_fp16")]; + tensor scores_59_cast_fp16 = add(x = scores_57_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_59_cast_fp16")]; + tensor var_3738 = const()[name = tensor("op_3738"), val = tensor(-1)]; + tensor weights_29_cast_fp16 = softmax(axis = var_3738, x = scores_59_cast_fp16)[name = tensor("weights_29_cast_fp16")]; + tensor attn_29_transpose_x_0 = const()[name = tensor("attn_29_transpose_x_0"), val = tensor(false)]; + tensor attn_29_transpose_y_0 = const()[name = tensor("attn_29_transpose_y_0"), val = tensor(false)]; + tensor values_29_cast_fp16 = transpose(perm = var_3716, x = new_v_cache_29_cast_fp16)[name = tensor("transpose_150")]; + tensor attn_29_cast_fp16 = matmul(transpose_x = attn_29_transpose_x_0, transpose_y = attn_29_transpose_y_0, x = weights_29_cast_fp16, y = values_29_cast_fp16)[name = tensor("attn_29_cast_fp16")]; + tensor var_3746 = const()[name = tensor("op_3746"), val = tensor([0, 2, 1, 3])]; + tensor var_3751 = const()[name = tensor("op_3751"), val = tensor([1, 256, 1024])]; + tensor var_3747_cast_fp16 = transpose(perm = var_3746, x = attn_29_cast_fp16)[name = tensor("transpose_147")]; + tensor input_173_cast_fp16 = reshape(shape = var_3751, x = var_3747_cast_fp16)[name = tensor("input_173_cast_fp16")]; + tensor attn14_out_proj_weight_to_fp16 = const()[name = tensor("attn14_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(358758144)))]; + tensor linear_57_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn14_out_proj_weight_to_fp16, x = input_173_cast_fp16)[name = tensor("linear_57_cast_fp16")]; + tensor position14_to_fp16_dtype_0 = const()[name = tensor("position14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position14_to_fp16 = cast(dtype = position14_to_fp16_dtype_0, x = position14)[name = tensor("cast_377")]; + tensor var_3760_cast_fp16 = add(x = position14_to_fp16, y = valid_len_to_fp16)[name = tensor("op_3760_cast_fp16")]; + tensor var_3760_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3760_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_175_cast_fp16 = add(x = input_169_cast_fp16, y = linear_57_cast_fp16)[name = tensor("input_175_cast_fp16")]; + tensor input_177_axes_0 = const()[name = tensor("input_177_axes_0"), val = tensor([-1])]; + tensor norm14_2_weight_to_fp16 = const()[name = tensor("norm14_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(360855360)))]; + tensor norm14_2_bias_to_fp16 = const()[name = tensor("norm14_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(360857472)))]; + tensor var_3764_to_fp16 = const()[name = tensor("op_3764_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_177_cast_fp16 = layer_norm(axes = input_177_axes_0, beta = norm14_2_bias_to_fp16, epsilon = var_3764_to_fp16, gamma = norm14_2_weight_to_fp16, x = input_175_cast_fp16)[name = tensor("input_177_cast_fp16")]; + tensor linear14_1_weight_to_fp16 = const()[name = tensor("linear14_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(360859584)))]; + tensor linear_58_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear14_1_weight_to_fp16, x = input_177_cast_fp16)[name = tensor("linear_58_cast_fp16")]; + tensor input_179_mode_0 = const()[name = tensor("input_179_mode_0"), val = tensor("EXACT")]; + tensor input_179_cast_fp16 = gelu(mode = input_179_mode_0, x = linear_58_cast_fp16)[name = tensor("input_179_cast_fp16")]; + tensor linear14_2_weight_to_fp16 = const()[name = tensor("linear14_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(369248256)))]; + tensor linear_59_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear14_2_weight_to_fp16, x = input_179_cast_fp16)[name = tensor("linear_59_cast_fp16")]; + tensor input_181_cast_fp16 = add(x = input_175_cast_fp16, y = linear_59_cast_fp16)[name = tensor("input_181_cast_fp16")]; + tensor input_183_axes_0 = const()[name = tensor("input_183_axes_0"), val = tensor([-1])]; + tensor norm15_1_weight_to_fp16 = const()[name = tensor("norm15_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(377636928)))]; + tensor norm15_1_bias_to_fp16 = const()[name = tensor("norm15_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(377639040)))]; + tensor var_3781_to_fp16 = const()[name = tensor("op_3781_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_183_cast_fp16 = layer_norm(axes = input_183_axes_0, beta = norm15_1_bias_to_fp16, epsilon = var_3781_to_fp16, gamma = norm15_1_weight_to_fp16, x = input_181_cast_fp16)[name = tensor("input_183_cast_fp16")]; + tensor attn15_in_proj_weight_to_fp16 = const()[name = tensor("attn15_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(377641152)))]; + tensor linear_60_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn15_in_proj_weight_to_fp16, x = input_183_cast_fp16)[name = tensor("linear_60_cast_fp16")]; + tensor var_3795 = const()[name = tensor("op_3795"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_31_cast_fp16 = reshape(shape = var_3795, x = linear_60_cast_fp16)[name = tensor("qkv_31_cast_fp16")]; + tensor x_61_begin_0 = const()[name = tensor("x_61_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_61_end_0 = const()[name = tensor("x_61_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_61_end_mask_0 = const()[name = tensor("x_61_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_61_squeeze_mask_0 = const()[name = tensor("x_61_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_61_cast_fp16 = slice_by_index(begin = x_61_begin_0, end = x_61_end_0, end_mask = x_61_end_mask_0, squeeze_mask = x_61_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("x_61_cast_fp16")]; + tensor x_63_begin_0 = const()[name = tensor("x_63_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_63_end_0 = const()[name = tensor("x_63_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_63_end_mask_0 = const()[name = tensor("x_63_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_63_squeeze_mask_0 = const()[name = tensor("x_63_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_63_cast_fp16 = slice_by_index(begin = x_63_begin_0, end = x_63_end_0, end_mask = x_63_end_mask_0, squeeze_mask = x_63_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("x_63_cast_fp16")]; + tensor v_31_begin_0 = const()[name = tensor("v_31_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_31_end_0 = const()[name = tensor("v_31_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_31_end_mask_0 = const()[name = tensor("v_31_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_31_squeeze_mask_0 = const()[name = tensor("v_31_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_31_cast_fp16 = slice_by_index(begin = v_31_begin_0, end = v_31_end_0, end_mask = v_31_end_mask_0, squeeze_mask = v_31_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("v_31_cast_fp16")]; + tensor var_3840 = const()[name = tensor("op_3840"), val = tensor([256, 16, 32, 2])]; + tensor pairs_61_cast_fp16 = reshape(shape = var_3840, x = x_61_cast_fp16)[name = tensor("pairs_61_cast_fp16")]; + tensor xr_61_begin_0 = const()[name = tensor("xr_61_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_61_end_0 = const()[name = tensor("xr_61_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_61_end_mask_0 = const()[name = tensor("xr_61_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_61_squeeze_mask_0 = const()[name = tensor("xr_61_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_61_cast_fp16 = slice_by_index(begin = xr_61_begin_0, end = xr_61_end_0, end_mask = xr_61_end_mask_0, squeeze_mask = xr_61_squeeze_mask_0, x = pairs_61_cast_fp16)[name = tensor("xr_61_cast_fp16")]; + tensor xi_61_begin_0 = const()[name = tensor("xi_61_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_61_end_0 = const()[name = tensor("xi_61_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_61_end_mask_0 = const()[name = tensor("xi_61_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_61_squeeze_mask_0 = const()[name = tensor("xi_61_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_61_cast_fp16 = slice_by_index(begin = xi_61_begin_0, end = xi_61_end_0, end_mask = xi_61_end_mask_0, squeeze_mask = xi_61_squeeze_mask_0, x = pairs_61_cast_fp16)[name = tensor("xi_61_cast_fp16")]; + tensor var_3848_cast_fp16 = mul(x = xr_61_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3848_cast_fp16")]; + tensor var_3849_cast_fp16 = mul(x = xi_61_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3849_cast_fp16")]; + tensor xor__61_cast_fp16 = sub(x = var_3848_cast_fp16, y = var_3849_cast_fp16)[name = tensor("xor__61_cast_fp16")]; + tensor var_3852_cast_fp16 = mul(x = xr_61_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3852_cast_fp16")]; + tensor var_3853_cast_fp16 = mul(x = xi_61_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3853_cast_fp16")]; + tensor xoi_61_cast_fp16 = add(x = var_3852_cast_fp16, y = var_3853_cast_fp16)[name = tensor("xoi_61_cast_fp16")]; + tensor out_61_axis_0 = const()[name = tensor("out_61_axis_0"), val = tensor(-1)]; + tensor out_61_cast_fp16 = stack(axis = out_61_axis_0, values = (xor__61_cast_fp16, xoi_61_cast_fp16))[name = tensor("out_61_cast_fp16")]; + tensor var_3863 = const()[name = tensor("op_3863"), val = tensor([1, 256, 16, 64])]; + tensor q_31_cast_fp16 = reshape(shape = var_3863, x = out_61_cast_fp16)[name = tensor("q_31_cast_fp16")]; + tensor var_3869 = const()[name = tensor("op_3869"), val = tensor([256, 16, 32, 2])]; + tensor pairs_63_cast_fp16 = reshape(shape = var_3869, x = x_63_cast_fp16)[name = tensor("pairs_63_cast_fp16")]; + tensor xr_63_begin_0 = const()[name = tensor("xr_63_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_63_end_0 = const()[name = tensor("xr_63_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_63_end_mask_0 = const()[name = tensor("xr_63_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_63_squeeze_mask_0 = const()[name = tensor("xr_63_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_63_cast_fp16 = slice_by_index(begin = xr_63_begin_0, end = xr_63_end_0, end_mask = xr_63_end_mask_0, squeeze_mask = xr_63_squeeze_mask_0, x = pairs_63_cast_fp16)[name = tensor("xr_63_cast_fp16")]; + tensor xi_63_begin_0 = const()[name = tensor("xi_63_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_63_end_0 = const()[name = tensor("xi_63_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_63_end_mask_0 = const()[name = tensor("xi_63_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_63_squeeze_mask_0 = const()[name = tensor("xi_63_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_63_cast_fp16 = slice_by_index(begin = xi_63_begin_0, end = xi_63_end_0, end_mask = xi_63_end_mask_0, squeeze_mask = xi_63_squeeze_mask_0, x = pairs_63_cast_fp16)[name = tensor("xi_63_cast_fp16")]; + tensor var_3877_cast_fp16 = mul(x = xr_63_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3877_cast_fp16")]; + tensor var_3878_cast_fp16 = mul(x = xi_63_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3878_cast_fp16")]; + tensor xor__63_cast_fp16 = sub(x = var_3877_cast_fp16, y = var_3878_cast_fp16)[name = tensor("xor__63_cast_fp16")]; + tensor var_3881_cast_fp16 = mul(x = xr_63_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3881_cast_fp16")]; + tensor var_3882_cast_fp16 = mul(x = xi_63_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3882_cast_fp16")]; + tensor xoi_63_cast_fp16 = add(x = var_3881_cast_fp16, y = var_3882_cast_fp16)[name = tensor("xoi_63_cast_fp16")]; + tensor out_63_axis_0 = const()[name = tensor("out_63_axis_0"), val = tensor(-1)]; + tensor out_63_cast_fp16 = stack(axis = out_63_axis_0, values = (xor__63_cast_fp16, xoi_63_cast_fp16))[name = tensor("out_63_cast_fp16")]; + tensor var_3899 = const()[name = tensor("op_3899"), val = tensor([256, 1024])]; + tensor var_3900_cast_fp16 = reshape(shape = var_3899, x = out_63_cast_fp16)[name = tensor("op_3900_cast_fp16")]; + tensor written_k_31_transpose_x_1 = const()[name = tensor("written_k_31_transpose_x_1"), val = tensor(true)]; + tensor written_k_31_transpose_y_1 = const()[name = tensor("written_k_31_transpose_y_1"), val = tensor(false)]; + tensor written_k_31_cast_fp16 = matmul(transpose_x = written_k_31_transpose_x_1, transpose_y = written_k_31_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_3900_cast_fp16)[name = tensor("written_k_31_cast_fp16")]; + tensor var_3907 = const()[name = tensor("op_3907"), val = tensor([256, 1024])]; + tensor var_3908_cast_fp16 = reshape(shape = var_3907, x = v_31_cast_fp16)[name = tensor("op_3908_cast_fp16")]; + tensor written_v_31_transpose_x_1 = const()[name = tensor("written_v_31_transpose_x_1"), val = tensor(true)]; + tensor written_v_31_transpose_y_1 = const()[name = tensor("written_v_31_transpose_y_1"), val = tensor(false)]; + tensor written_v_31_cast_fp16 = matmul(transpose_x = written_v_31_transpose_x_1, transpose_y = written_v_31_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_3908_cast_fp16)[name = tensor("written_v_31_cast_fp16")]; + tensor k_cache15_to_fp16_dtype_0 = const()[name = tensor("k_cache15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache15_to_fp16 = cast(dtype = k_cache15_to_fp16_dtype_0, x = k_cache15)[name = tensor("cast_375")]; + tensor var_3919_cast_fp16 = mul(x = k_cache15_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_3919_cast_fp16")]; + tensor var_3924 = const()[name = tensor("op_3924"), val = tensor([1, 512, 16, 64])]; + tensor var_3925_cast_fp16 = reshape(shape = var_3924, x = written_k_31_cast_fp16)[name = tensor("op_3925_cast_fp16")]; + tensor new_k_cache_31_cast_fp16 = add(x = var_3919_cast_fp16, y = var_3925_cast_fp16)[name = tensor("new_k_cache_31_cast_fp16")]; + tensor new_k_cache_31_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_31_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache15_to_fp16_dtype_0 = const()[name = tensor("v_cache15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache15_to_fp16 = cast(dtype = v_cache15_to_fp16_dtype_0, x = v_cache15)[name = tensor("cast_373")]; + tensor var_3928_cast_fp16 = mul(x = v_cache15_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_3928_cast_fp16")]; + tensor var_3933 = const()[name = tensor("op_3933"), val = tensor([1, 512, 16, 64])]; + tensor var_3934_cast_fp16 = reshape(shape = var_3933, x = written_v_31_cast_fp16)[name = tensor("op_3934_cast_fp16")]; + tensor new_v_cache_31_cast_fp16 = add(x = var_3928_cast_fp16, y = var_3934_cast_fp16)[name = tensor("new_v_cache_31_cast_fp16")]; + tensor new_v_cache_31_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_31_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_3947 = const()[name = tensor("op_3947"), val = tensor([0, 2, 1, 3])]; + tensor var_3958_transpose_x_0 = const()[name = tensor("op_3958_transpose_x_0"), val = tensor(false)]; + tensor var_3958_transpose_y_0 = const()[name = tensor("op_3958_transpose_y_0"), val = tensor(false)]; + tensor transpose_99_perm_0 = const()[name = tensor("transpose_99_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_100_perm_0 = const()[name = tensor("transpose_100_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_100 = transpose(perm = transpose_100_perm_0, x = new_k_cache_31_cast_fp16)[name = tensor("transpose_144")]; + tensor transpose_99 = transpose(perm = transpose_99_perm_0, x = q_31_cast_fp16)[name = tensor("transpose_145")]; + tensor var_3958_cast_fp16 = matmul(transpose_x = var_3958_transpose_x_0, transpose_y = var_3958_transpose_y_0, x = transpose_99, y = transpose_100)[name = tensor("op_3958_cast_fp16")]; + tensor var_3959_to_fp16 = const()[name = tensor("op_3959_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_61_cast_fp16 = mul(x = var_3958_cast_fp16, y = var_3959_to_fp16)[name = tensor("scores_61_cast_fp16")]; + tensor scores_63_cast_fp16 = add(x = scores_61_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_63_cast_fp16")]; + tensor var_3969 = const()[name = tensor("op_3969"), val = tensor(-1)]; + tensor weights_31_cast_fp16 = softmax(axis = var_3969, x = scores_63_cast_fp16)[name = tensor("weights_31_cast_fp16")]; + tensor attn_31_transpose_x_0 = const()[name = tensor("attn_31_transpose_x_0"), val = tensor(false)]; + tensor attn_31_transpose_y_0 = const()[name = tensor("attn_31_transpose_y_0"), val = tensor(false)]; + tensor values_31_cast_fp16 = transpose(perm = var_3947, x = new_v_cache_31_cast_fp16)[name = tensor("transpose_146")]; + tensor attn_31_cast_fp16 = matmul(transpose_x = attn_31_transpose_x_0, transpose_y = attn_31_transpose_y_0, x = weights_31_cast_fp16, y = values_31_cast_fp16)[name = tensor("attn_31_cast_fp16")]; + tensor var_3977 = const()[name = tensor("op_3977"), val = tensor([0, 2, 1, 3])]; + tensor var_3982 = const()[name = tensor("op_3982"), val = tensor([1, 256, 1024])]; + tensor var_3978_cast_fp16 = transpose(perm = var_3977, x = attn_31_cast_fp16)[name = tensor("transpose_143")]; + tensor input_185_cast_fp16 = reshape(shape = var_3982, x = var_3978_cast_fp16)[name = tensor("input_185_cast_fp16")]; + tensor attn15_out_proj_weight_to_fp16 = const()[name = tensor("attn15_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(383932672)))]; + tensor linear_61_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn15_out_proj_weight_to_fp16, x = input_185_cast_fp16)[name = tensor("linear_61_cast_fp16")]; + tensor position15_to_fp16_dtype_0 = const()[name = tensor("position15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position15_to_fp16 = cast(dtype = position15_to_fp16_dtype_0, x = position15)[name = tensor("cast_371")]; + tensor var_3991_cast_fp16 = add(x = position15_to_fp16, y = valid_len_to_fp16)[name = tensor("op_3991_cast_fp16")]; + tensor var_3991_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3991_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_187_cast_fp16 = add(x = input_181_cast_fp16, y = linear_61_cast_fp16)[name = tensor("input_187_cast_fp16")]; + tensor input_189_axes_0 = const()[name = tensor("input_189_axes_0"), val = tensor([-1])]; + tensor norm15_2_weight_to_fp16 = const()[name = tensor("norm15_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(386029888)))]; + tensor norm15_2_bias_to_fp16 = const()[name = tensor("norm15_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(386032000)))]; + tensor var_3995_to_fp16 = const()[name = tensor("op_3995_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_189_cast_fp16 = layer_norm(axes = input_189_axes_0, beta = norm15_2_bias_to_fp16, epsilon = var_3995_to_fp16, gamma = norm15_2_weight_to_fp16, x = input_187_cast_fp16)[name = tensor("input_189_cast_fp16")]; + tensor linear15_1_weight_to_fp16 = const()[name = tensor("linear15_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(386034112)))]; + tensor linear_62_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear15_1_weight_to_fp16, x = input_189_cast_fp16)[name = tensor("linear_62_cast_fp16")]; + tensor input_191_mode_0 = const()[name = tensor("input_191_mode_0"), val = tensor("EXACT")]; + tensor input_191_cast_fp16 = gelu(mode = input_191_mode_0, x = linear_62_cast_fp16)[name = tensor("input_191_cast_fp16")]; + tensor linear15_2_weight_to_fp16 = const()[name = tensor("linear15_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(394422784)))]; + tensor linear_63_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear15_2_weight_to_fp16, x = input_191_cast_fp16)[name = tensor("linear_63_cast_fp16")]; + tensor input_193_cast_fp16 = add(x = input_187_cast_fp16, y = linear_63_cast_fp16)[name = tensor("input_193_cast_fp16")]; + tensor input_195_axes_0 = const()[name = tensor("input_195_axes_0"), val = tensor([-1])]; + tensor norm16_1_weight_to_fp16 = const()[name = tensor("norm16_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(402811456)))]; + tensor norm16_1_bias_to_fp16 = const()[name = tensor("norm16_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(402813568)))]; + tensor var_4012_to_fp16 = const()[name = tensor("op_4012_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_195_cast_fp16 = layer_norm(axes = input_195_axes_0, beta = norm16_1_bias_to_fp16, epsilon = var_4012_to_fp16, gamma = norm16_1_weight_to_fp16, x = input_193_cast_fp16)[name = tensor("input_195_cast_fp16")]; + tensor attn16_in_proj_weight_to_fp16 = const()[name = tensor("attn16_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(402815680)))]; + tensor linear_64_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn16_in_proj_weight_to_fp16, x = input_195_cast_fp16)[name = tensor("linear_64_cast_fp16")]; + tensor var_4026 = const()[name = tensor("op_4026"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_33_cast_fp16 = reshape(shape = var_4026, x = linear_64_cast_fp16)[name = tensor("qkv_33_cast_fp16")]; + tensor x_65_begin_0 = const()[name = tensor("x_65_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_65_end_0 = const()[name = tensor("x_65_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_65_end_mask_0 = const()[name = tensor("x_65_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_65_squeeze_mask_0 = const()[name = tensor("x_65_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_65_cast_fp16 = slice_by_index(begin = x_65_begin_0, end = x_65_end_0, end_mask = x_65_end_mask_0, squeeze_mask = x_65_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("x_65_cast_fp16")]; + tensor x_67_begin_0 = const()[name = tensor("x_67_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_67_end_0 = const()[name = tensor("x_67_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_67_end_mask_0 = const()[name = tensor("x_67_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_67_squeeze_mask_0 = const()[name = tensor("x_67_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_67_cast_fp16 = slice_by_index(begin = x_67_begin_0, end = x_67_end_0, end_mask = x_67_end_mask_0, squeeze_mask = x_67_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("x_67_cast_fp16")]; + tensor v_33_begin_0 = const()[name = tensor("v_33_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_33_end_0 = const()[name = tensor("v_33_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_33_end_mask_0 = const()[name = tensor("v_33_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_33_squeeze_mask_0 = const()[name = tensor("v_33_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_33_cast_fp16 = slice_by_index(begin = v_33_begin_0, end = v_33_end_0, end_mask = v_33_end_mask_0, squeeze_mask = v_33_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("v_33_cast_fp16")]; + tensor var_4071 = const()[name = tensor("op_4071"), val = tensor([256, 16, 32, 2])]; + tensor pairs_65_cast_fp16 = reshape(shape = var_4071, x = x_65_cast_fp16)[name = tensor("pairs_65_cast_fp16")]; + tensor xr_65_begin_0 = const()[name = tensor("xr_65_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_65_end_0 = const()[name = tensor("xr_65_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_65_end_mask_0 = const()[name = tensor("xr_65_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_65_squeeze_mask_0 = const()[name = tensor("xr_65_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_65_cast_fp16 = slice_by_index(begin = xr_65_begin_0, end = xr_65_end_0, end_mask = xr_65_end_mask_0, squeeze_mask = xr_65_squeeze_mask_0, x = pairs_65_cast_fp16)[name = tensor("xr_65_cast_fp16")]; + tensor xi_65_begin_0 = const()[name = tensor("xi_65_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_65_end_0 = const()[name = tensor("xi_65_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_65_end_mask_0 = const()[name = tensor("xi_65_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_65_squeeze_mask_0 = const()[name = tensor("xi_65_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_65_cast_fp16 = slice_by_index(begin = xi_65_begin_0, end = xi_65_end_0, end_mask = xi_65_end_mask_0, squeeze_mask = xi_65_squeeze_mask_0, x = pairs_65_cast_fp16)[name = tensor("xi_65_cast_fp16")]; + tensor var_4079_cast_fp16 = mul(x = xr_65_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4079_cast_fp16")]; + tensor var_4080_cast_fp16 = mul(x = xi_65_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4080_cast_fp16")]; + tensor xor__65_cast_fp16 = sub(x = var_4079_cast_fp16, y = var_4080_cast_fp16)[name = tensor("xor__65_cast_fp16")]; + tensor var_4083_cast_fp16 = mul(x = xr_65_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4083_cast_fp16")]; + tensor var_4084_cast_fp16 = mul(x = xi_65_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4084_cast_fp16")]; + tensor xoi_65_cast_fp16 = add(x = var_4083_cast_fp16, y = var_4084_cast_fp16)[name = tensor("xoi_65_cast_fp16")]; + tensor out_65_axis_0 = const()[name = tensor("out_65_axis_0"), val = tensor(-1)]; + tensor out_65_cast_fp16 = stack(axis = out_65_axis_0, values = (xor__65_cast_fp16, xoi_65_cast_fp16))[name = tensor("out_65_cast_fp16")]; + tensor var_4094 = const()[name = tensor("op_4094"), val = tensor([1, 256, 16, 64])]; + tensor q_33_cast_fp16 = reshape(shape = var_4094, x = out_65_cast_fp16)[name = tensor("q_33_cast_fp16")]; + tensor var_4100 = const()[name = tensor("op_4100"), val = tensor([256, 16, 32, 2])]; + tensor pairs_67_cast_fp16 = reshape(shape = var_4100, x = x_67_cast_fp16)[name = tensor("pairs_67_cast_fp16")]; + tensor xr_67_begin_0 = const()[name = tensor("xr_67_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_67_end_0 = const()[name = tensor("xr_67_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_67_end_mask_0 = const()[name = tensor("xr_67_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_67_squeeze_mask_0 = const()[name = tensor("xr_67_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_67_cast_fp16 = slice_by_index(begin = xr_67_begin_0, end = xr_67_end_0, end_mask = xr_67_end_mask_0, squeeze_mask = xr_67_squeeze_mask_0, x = pairs_67_cast_fp16)[name = tensor("xr_67_cast_fp16")]; + tensor xi_67_begin_0 = const()[name = tensor("xi_67_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_67_end_0 = const()[name = tensor("xi_67_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_67_end_mask_0 = const()[name = tensor("xi_67_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_67_squeeze_mask_0 = const()[name = tensor("xi_67_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_67_cast_fp16 = slice_by_index(begin = xi_67_begin_0, end = xi_67_end_0, end_mask = xi_67_end_mask_0, squeeze_mask = xi_67_squeeze_mask_0, x = pairs_67_cast_fp16)[name = tensor("xi_67_cast_fp16")]; + tensor var_4108_cast_fp16 = mul(x = xr_67_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4108_cast_fp16")]; + tensor var_4109_cast_fp16 = mul(x = xi_67_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4109_cast_fp16")]; + tensor xor__67_cast_fp16 = sub(x = var_4108_cast_fp16, y = var_4109_cast_fp16)[name = tensor("xor__67_cast_fp16")]; + tensor var_4112_cast_fp16 = mul(x = xr_67_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4112_cast_fp16")]; + tensor var_4113_cast_fp16 = mul(x = xi_67_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4113_cast_fp16")]; + tensor xoi_67_cast_fp16 = add(x = var_4112_cast_fp16, y = var_4113_cast_fp16)[name = tensor("xoi_67_cast_fp16")]; + tensor out_67_axis_0 = const()[name = tensor("out_67_axis_0"), val = tensor(-1)]; + tensor out_67_cast_fp16 = stack(axis = out_67_axis_0, values = (xor__67_cast_fp16, xoi_67_cast_fp16))[name = tensor("out_67_cast_fp16")]; + tensor var_4130 = const()[name = tensor("op_4130"), val = tensor([256, 1024])]; + tensor var_4131_cast_fp16 = reshape(shape = var_4130, x = out_67_cast_fp16)[name = tensor("op_4131_cast_fp16")]; + tensor written_k_33_transpose_x_1 = const()[name = tensor("written_k_33_transpose_x_1"), val = tensor(true)]; + tensor written_k_33_transpose_y_1 = const()[name = tensor("written_k_33_transpose_y_1"), val = tensor(false)]; + tensor written_k_33_cast_fp16 = matmul(transpose_x = written_k_33_transpose_x_1, transpose_y = written_k_33_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_4131_cast_fp16)[name = tensor("written_k_33_cast_fp16")]; + tensor var_4138 = const()[name = tensor("op_4138"), val = tensor([256, 1024])]; + tensor var_4139_cast_fp16 = reshape(shape = var_4138, x = v_33_cast_fp16)[name = tensor("op_4139_cast_fp16")]; + tensor written_v_33_transpose_x_1 = const()[name = tensor("written_v_33_transpose_x_1"), val = tensor(true)]; + tensor written_v_33_transpose_y_1 = const()[name = tensor("written_v_33_transpose_y_1"), val = tensor(false)]; + tensor written_v_33_cast_fp16 = matmul(transpose_x = written_v_33_transpose_x_1, transpose_y = written_v_33_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_4139_cast_fp16)[name = tensor("written_v_33_cast_fp16")]; + tensor k_cache16_to_fp16_dtype_0 = const()[name = tensor("k_cache16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache16_to_fp16 = cast(dtype = k_cache16_to_fp16_dtype_0, x = k_cache16)[name = tensor("cast_369")]; + tensor var_4150_cast_fp16 = mul(x = k_cache16_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_4150_cast_fp16")]; + tensor var_4155 = const()[name = tensor("op_4155"), val = tensor([1, 512, 16, 64])]; + tensor var_4156_cast_fp16 = reshape(shape = var_4155, x = written_k_33_cast_fp16)[name = tensor("op_4156_cast_fp16")]; + tensor new_k_cache_33_cast_fp16 = add(x = var_4150_cast_fp16, y = var_4156_cast_fp16)[name = tensor("new_k_cache_33_cast_fp16")]; + tensor new_k_cache_33_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_33_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache16_to_fp16_dtype_0 = const()[name = tensor("v_cache16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache16_to_fp16 = cast(dtype = v_cache16_to_fp16_dtype_0, x = v_cache16)[name = tensor("cast_367")]; + tensor var_4159_cast_fp16 = mul(x = v_cache16_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_4159_cast_fp16")]; + tensor var_4164 = const()[name = tensor("op_4164"), val = tensor([1, 512, 16, 64])]; + tensor var_4165_cast_fp16 = reshape(shape = var_4164, x = written_v_33_cast_fp16)[name = tensor("op_4165_cast_fp16")]; + tensor new_v_cache_33_cast_fp16 = add(x = var_4159_cast_fp16, y = var_4165_cast_fp16)[name = tensor("new_v_cache_33_cast_fp16")]; + tensor new_v_cache_33_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_33_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_4178 = const()[name = tensor("op_4178"), val = tensor([0, 2, 1, 3])]; + tensor var_4189_transpose_x_0 = const()[name = tensor("op_4189_transpose_x_0"), val = tensor(false)]; + tensor var_4189_transpose_y_0 = const()[name = tensor("op_4189_transpose_y_0"), val = tensor(false)]; + tensor transpose_101_perm_0 = const()[name = tensor("transpose_101_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_102_perm_0 = const()[name = tensor("transpose_102_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_102 = transpose(perm = transpose_102_perm_0, x = new_k_cache_33_cast_fp16)[name = tensor("transpose_140")]; + tensor transpose_101 = transpose(perm = transpose_101_perm_0, x = q_33_cast_fp16)[name = tensor("transpose_141")]; + tensor var_4189_cast_fp16 = matmul(transpose_x = var_4189_transpose_x_0, transpose_y = var_4189_transpose_y_0, x = transpose_101, y = transpose_102)[name = tensor("op_4189_cast_fp16")]; + tensor var_4190_to_fp16 = const()[name = tensor("op_4190_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_65_cast_fp16 = mul(x = var_4189_cast_fp16, y = var_4190_to_fp16)[name = tensor("scores_65_cast_fp16")]; + tensor scores_67_cast_fp16 = add(x = scores_65_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_67_cast_fp16")]; + tensor var_4200 = const()[name = tensor("op_4200"), val = tensor(-1)]; + tensor weights_33_cast_fp16 = softmax(axis = var_4200, x = scores_67_cast_fp16)[name = tensor("weights_33_cast_fp16")]; + tensor attn_33_transpose_x_0 = const()[name = tensor("attn_33_transpose_x_0"), val = tensor(false)]; + tensor attn_33_transpose_y_0 = const()[name = tensor("attn_33_transpose_y_0"), val = tensor(false)]; + tensor values_33_cast_fp16 = transpose(perm = var_4178, x = new_v_cache_33_cast_fp16)[name = tensor("transpose_142")]; + tensor attn_33_cast_fp16 = matmul(transpose_x = attn_33_transpose_x_0, transpose_y = attn_33_transpose_y_0, x = weights_33_cast_fp16, y = values_33_cast_fp16)[name = tensor("attn_33_cast_fp16")]; + tensor var_4208 = const()[name = tensor("op_4208"), val = tensor([0, 2, 1, 3])]; + tensor var_4213 = const()[name = tensor("op_4213"), val = tensor([1, 256, 1024])]; + tensor var_4209_cast_fp16 = transpose(perm = var_4208, x = attn_33_cast_fp16)[name = tensor("transpose_139")]; + tensor input_197_cast_fp16 = reshape(shape = var_4213, x = var_4209_cast_fp16)[name = tensor("input_197_cast_fp16")]; + tensor attn16_out_proj_weight_to_fp16 = const()[name = tensor("attn16_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(409107200)))]; + tensor linear_65_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn16_out_proj_weight_to_fp16, x = input_197_cast_fp16)[name = tensor("linear_65_cast_fp16")]; + tensor position16_to_fp16_dtype_0 = const()[name = tensor("position16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position16_to_fp16 = cast(dtype = position16_to_fp16_dtype_0, x = position16)[name = tensor("cast_365")]; + tensor var_4222_cast_fp16 = add(x = position16_to_fp16, y = valid_len_to_fp16)[name = tensor("op_4222_cast_fp16")]; + tensor var_4222_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4222_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_199_cast_fp16 = add(x = input_193_cast_fp16, y = linear_65_cast_fp16)[name = tensor("input_199_cast_fp16")]; + tensor input_201_axes_0 = const()[name = tensor("input_201_axes_0"), val = tensor([-1])]; + tensor norm16_2_weight_to_fp16 = const()[name = tensor("norm16_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411204416)))]; + tensor norm16_2_bias_to_fp16 = const()[name = tensor("norm16_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411206528)))]; + tensor var_4226_to_fp16 = const()[name = tensor("op_4226_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_201_cast_fp16 = layer_norm(axes = input_201_axes_0, beta = norm16_2_bias_to_fp16, epsilon = var_4226_to_fp16, gamma = norm16_2_weight_to_fp16, x = input_199_cast_fp16)[name = tensor("input_201_cast_fp16")]; + tensor linear16_1_weight_to_fp16 = const()[name = tensor("linear16_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411208640)))]; + tensor linear_66_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear16_1_weight_to_fp16, x = input_201_cast_fp16)[name = tensor("linear_66_cast_fp16")]; + tensor input_203_mode_0 = const()[name = tensor("input_203_mode_0"), val = tensor("EXACT")]; + tensor input_203_cast_fp16 = gelu(mode = input_203_mode_0, x = linear_66_cast_fp16)[name = tensor("input_203_cast_fp16")]; + tensor linear16_2_weight_to_fp16 = const()[name = tensor("linear16_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(419597312)))]; + tensor linear_67_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear16_2_weight_to_fp16, x = input_203_cast_fp16)[name = tensor("linear_67_cast_fp16")]; + tensor input_205_cast_fp16 = add(x = input_199_cast_fp16, y = linear_67_cast_fp16)[name = tensor("input_205_cast_fp16")]; + tensor input_207_axes_0 = const()[name = tensor("input_207_axes_0"), val = tensor([-1])]; + tensor norm17_1_weight_to_fp16 = const()[name = tensor("norm17_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(427985984)))]; + tensor norm17_1_bias_to_fp16 = const()[name = tensor("norm17_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(427988096)))]; + tensor var_4243_to_fp16 = const()[name = tensor("op_4243_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_207_cast_fp16 = layer_norm(axes = input_207_axes_0, beta = norm17_1_bias_to_fp16, epsilon = var_4243_to_fp16, gamma = norm17_1_weight_to_fp16, x = input_205_cast_fp16)[name = tensor("input_207_cast_fp16")]; + tensor attn17_in_proj_weight_to_fp16 = const()[name = tensor("attn17_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(427990208)))]; + tensor linear_68_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn17_in_proj_weight_to_fp16, x = input_207_cast_fp16)[name = tensor("linear_68_cast_fp16")]; + tensor var_4257 = const()[name = tensor("op_4257"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_35_cast_fp16 = reshape(shape = var_4257, x = linear_68_cast_fp16)[name = tensor("qkv_35_cast_fp16")]; + tensor x_69_begin_0 = const()[name = tensor("x_69_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_69_end_0 = const()[name = tensor("x_69_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_69_end_mask_0 = const()[name = tensor("x_69_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_69_squeeze_mask_0 = const()[name = tensor("x_69_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_69_cast_fp16 = slice_by_index(begin = x_69_begin_0, end = x_69_end_0, end_mask = x_69_end_mask_0, squeeze_mask = x_69_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("x_69_cast_fp16")]; + tensor x_71_begin_0 = const()[name = tensor("x_71_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_71_end_0 = const()[name = tensor("x_71_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_71_end_mask_0 = const()[name = tensor("x_71_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_71_squeeze_mask_0 = const()[name = tensor("x_71_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_71_cast_fp16 = slice_by_index(begin = x_71_begin_0, end = x_71_end_0, end_mask = x_71_end_mask_0, squeeze_mask = x_71_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("x_71_cast_fp16")]; + tensor v_35_begin_0 = const()[name = tensor("v_35_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_35_end_0 = const()[name = tensor("v_35_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_35_end_mask_0 = const()[name = tensor("v_35_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_35_squeeze_mask_0 = const()[name = tensor("v_35_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_35_cast_fp16 = slice_by_index(begin = v_35_begin_0, end = v_35_end_0, end_mask = v_35_end_mask_0, squeeze_mask = v_35_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("v_35_cast_fp16")]; + tensor var_4302 = const()[name = tensor("op_4302"), val = tensor([256, 16, 32, 2])]; + tensor pairs_69_cast_fp16 = reshape(shape = var_4302, x = x_69_cast_fp16)[name = tensor("pairs_69_cast_fp16")]; + tensor xr_69_begin_0 = const()[name = tensor("xr_69_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_69_end_0 = const()[name = tensor("xr_69_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_69_end_mask_0 = const()[name = tensor("xr_69_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_69_squeeze_mask_0 = const()[name = tensor("xr_69_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_69_cast_fp16 = slice_by_index(begin = xr_69_begin_0, end = xr_69_end_0, end_mask = xr_69_end_mask_0, squeeze_mask = xr_69_squeeze_mask_0, x = pairs_69_cast_fp16)[name = tensor("xr_69_cast_fp16")]; + tensor xi_69_begin_0 = const()[name = tensor("xi_69_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_69_end_0 = const()[name = tensor("xi_69_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_69_end_mask_0 = const()[name = tensor("xi_69_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_69_squeeze_mask_0 = const()[name = tensor("xi_69_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_69_cast_fp16 = slice_by_index(begin = xi_69_begin_0, end = xi_69_end_0, end_mask = xi_69_end_mask_0, squeeze_mask = xi_69_squeeze_mask_0, x = pairs_69_cast_fp16)[name = tensor("xi_69_cast_fp16")]; + tensor var_4310_cast_fp16 = mul(x = xr_69_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4310_cast_fp16")]; + tensor var_4311_cast_fp16 = mul(x = xi_69_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4311_cast_fp16")]; + tensor xor__69_cast_fp16 = sub(x = var_4310_cast_fp16, y = var_4311_cast_fp16)[name = tensor("xor__69_cast_fp16")]; + tensor var_4314_cast_fp16 = mul(x = xr_69_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4314_cast_fp16")]; + tensor var_4315_cast_fp16 = mul(x = xi_69_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4315_cast_fp16")]; + tensor xoi_69_cast_fp16 = add(x = var_4314_cast_fp16, y = var_4315_cast_fp16)[name = tensor("xoi_69_cast_fp16")]; + tensor out_69_axis_0 = const()[name = tensor("out_69_axis_0"), val = tensor(-1)]; + tensor out_69_cast_fp16 = stack(axis = out_69_axis_0, values = (xor__69_cast_fp16, xoi_69_cast_fp16))[name = tensor("out_69_cast_fp16")]; + tensor var_4325 = const()[name = tensor("op_4325"), val = tensor([1, 256, 16, 64])]; + tensor q_35_cast_fp16 = reshape(shape = var_4325, x = out_69_cast_fp16)[name = tensor("q_35_cast_fp16")]; + tensor var_4331 = const()[name = tensor("op_4331"), val = tensor([256, 16, 32, 2])]; + tensor pairs_71_cast_fp16 = reshape(shape = var_4331, x = x_71_cast_fp16)[name = tensor("pairs_71_cast_fp16")]; + tensor xr_71_begin_0 = const()[name = tensor("xr_71_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_71_end_0 = const()[name = tensor("xr_71_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_71_end_mask_0 = const()[name = tensor("xr_71_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_71_squeeze_mask_0 = const()[name = tensor("xr_71_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_71_cast_fp16 = slice_by_index(begin = xr_71_begin_0, end = xr_71_end_0, end_mask = xr_71_end_mask_0, squeeze_mask = xr_71_squeeze_mask_0, x = pairs_71_cast_fp16)[name = tensor("xr_71_cast_fp16")]; + tensor xi_71_begin_0 = const()[name = tensor("xi_71_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_71_end_0 = const()[name = tensor("xi_71_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_71_end_mask_0 = const()[name = tensor("xi_71_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_71_squeeze_mask_0 = const()[name = tensor("xi_71_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_71_cast_fp16 = slice_by_index(begin = xi_71_begin_0, end = xi_71_end_0, end_mask = xi_71_end_mask_0, squeeze_mask = xi_71_squeeze_mask_0, x = pairs_71_cast_fp16)[name = tensor("xi_71_cast_fp16")]; + tensor var_4339_cast_fp16 = mul(x = xr_71_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4339_cast_fp16")]; + tensor var_4340_cast_fp16 = mul(x = xi_71_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4340_cast_fp16")]; + tensor xor__71_cast_fp16 = sub(x = var_4339_cast_fp16, y = var_4340_cast_fp16)[name = tensor("xor__71_cast_fp16")]; + tensor var_4343_cast_fp16 = mul(x = xr_71_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4343_cast_fp16")]; + tensor var_4344_cast_fp16 = mul(x = xi_71_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4344_cast_fp16")]; + tensor xoi_71_cast_fp16 = add(x = var_4343_cast_fp16, y = var_4344_cast_fp16)[name = tensor("xoi_71_cast_fp16")]; + tensor out_71_axis_0 = const()[name = tensor("out_71_axis_0"), val = tensor(-1)]; + tensor out_71_cast_fp16 = stack(axis = out_71_axis_0, values = (xor__71_cast_fp16, xoi_71_cast_fp16))[name = tensor("out_71_cast_fp16")]; + tensor var_4361 = const()[name = tensor("op_4361"), val = tensor([256, 1024])]; + tensor var_4362_cast_fp16 = reshape(shape = var_4361, x = out_71_cast_fp16)[name = tensor("op_4362_cast_fp16")]; + tensor written_k_35_transpose_x_1 = const()[name = tensor("written_k_35_transpose_x_1"), val = tensor(true)]; + tensor written_k_35_transpose_y_1 = const()[name = tensor("written_k_35_transpose_y_1"), val = tensor(false)]; + tensor written_k_35_cast_fp16 = matmul(transpose_x = written_k_35_transpose_x_1, transpose_y = written_k_35_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_4362_cast_fp16)[name = tensor("written_k_35_cast_fp16")]; + tensor var_4369 = const()[name = tensor("op_4369"), val = tensor([256, 1024])]; + tensor var_4370_cast_fp16 = reshape(shape = var_4369, x = v_35_cast_fp16)[name = tensor("op_4370_cast_fp16")]; + tensor written_v_35_transpose_x_1 = const()[name = tensor("written_v_35_transpose_x_1"), val = tensor(true)]; + tensor written_v_35_transpose_y_1 = const()[name = tensor("written_v_35_transpose_y_1"), val = tensor(false)]; + tensor written_v_35_cast_fp16 = matmul(transpose_x = written_v_35_transpose_x_1, transpose_y = written_v_35_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_4370_cast_fp16)[name = tensor("written_v_35_cast_fp16")]; + tensor k_cache17_to_fp16_dtype_0 = const()[name = tensor("k_cache17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache17_to_fp16 = cast(dtype = k_cache17_to_fp16_dtype_0, x = k_cache17)[name = tensor("cast_363")]; + tensor var_4381_cast_fp16 = mul(x = k_cache17_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_4381_cast_fp16")]; + tensor var_4386 = const()[name = tensor("op_4386"), val = tensor([1, 512, 16, 64])]; + tensor var_4387_cast_fp16 = reshape(shape = var_4386, x = written_k_35_cast_fp16)[name = tensor("op_4387_cast_fp16")]; + tensor new_k_cache_35_cast_fp16 = add(x = var_4381_cast_fp16, y = var_4387_cast_fp16)[name = tensor("new_k_cache_35_cast_fp16")]; + tensor new_k_cache_35_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_35_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache17_to_fp16_dtype_0 = const()[name = tensor("v_cache17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache17_to_fp16 = cast(dtype = v_cache17_to_fp16_dtype_0, x = v_cache17)[name = tensor("cast_361")]; + tensor var_4390_cast_fp16 = mul(x = v_cache17_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_4390_cast_fp16")]; + tensor var_4395 = const()[name = tensor("op_4395"), val = tensor([1, 512, 16, 64])]; + tensor var_4396_cast_fp16 = reshape(shape = var_4395, x = written_v_35_cast_fp16)[name = tensor("op_4396_cast_fp16")]; + tensor new_v_cache_35_cast_fp16 = add(x = var_4390_cast_fp16, y = var_4396_cast_fp16)[name = tensor("new_v_cache_35_cast_fp16")]; + tensor new_v_cache_35_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_35_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_4409 = const()[name = tensor("op_4409"), val = tensor([0, 2, 1, 3])]; + tensor var_4420_transpose_x_0 = const()[name = tensor("op_4420_transpose_x_0"), val = tensor(false)]; + tensor var_4420_transpose_y_0 = const()[name = tensor("op_4420_transpose_y_0"), val = tensor(false)]; + tensor transpose_103_perm_0 = const()[name = tensor("transpose_103_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_104_perm_0 = const()[name = tensor("transpose_104_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_104 = transpose(perm = transpose_104_perm_0, x = new_k_cache_35_cast_fp16)[name = tensor("transpose_136")]; + tensor transpose_103 = transpose(perm = transpose_103_perm_0, x = q_35_cast_fp16)[name = tensor("transpose_137")]; + tensor var_4420_cast_fp16 = matmul(transpose_x = var_4420_transpose_x_0, transpose_y = var_4420_transpose_y_0, x = transpose_103, y = transpose_104)[name = tensor("op_4420_cast_fp16")]; + tensor var_4421_to_fp16 = const()[name = tensor("op_4421_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_69_cast_fp16 = mul(x = var_4420_cast_fp16, y = var_4421_to_fp16)[name = tensor("scores_69_cast_fp16")]; + tensor scores_71_cast_fp16 = add(x = scores_69_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_71_cast_fp16")]; + tensor var_4431 = const()[name = tensor("op_4431"), val = tensor(-1)]; + tensor weights_35_cast_fp16 = softmax(axis = var_4431, x = scores_71_cast_fp16)[name = tensor("weights_35_cast_fp16")]; + tensor attn_35_transpose_x_0 = const()[name = tensor("attn_35_transpose_x_0"), val = tensor(false)]; + tensor attn_35_transpose_y_0 = const()[name = tensor("attn_35_transpose_y_0"), val = tensor(false)]; + tensor values_35_cast_fp16 = transpose(perm = var_4409, x = new_v_cache_35_cast_fp16)[name = tensor("transpose_138")]; + tensor attn_35_cast_fp16 = matmul(transpose_x = attn_35_transpose_x_0, transpose_y = attn_35_transpose_y_0, x = weights_35_cast_fp16, y = values_35_cast_fp16)[name = tensor("attn_35_cast_fp16")]; + tensor var_4439 = const()[name = tensor("op_4439"), val = tensor([0, 2, 1, 3])]; + tensor var_4444 = const()[name = tensor("op_4444"), val = tensor([1, 256, 1024])]; + tensor var_4440_cast_fp16 = transpose(perm = var_4439, x = attn_35_cast_fp16)[name = tensor("transpose_135")]; + tensor input_209_cast_fp16 = reshape(shape = var_4444, x = var_4440_cast_fp16)[name = tensor("input_209_cast_fp16")]; + tensor attn17_out_proj_weight_to_fp16 = const()[name = tensor("attn17_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(434281728)))]; + tensor linear_69_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn17_out_proj_weight_to_fp16, x = input_209_cast_fp16)[name = tensor("linear_69_cast_fp16")]; + tensor position17_to_fp16_dtype_0 = const()[name = tensor("position17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position17_to_fp16 = cast(dtype = position17_to_fp16_dtype_0, x = position17)[name = tensor("cast_359")]; + tensor var_4453_cast_fp16 = add(x = position17_to_fp16, y = valid_len_to_fp16)[name = tensor("op_4453_cast_fp16")]; + tensor var_4453_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4453_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_211_cast_fp16 = add(x = input_205_cast_fp16, y = linear_69_cast_fp16)[name = tensor("input_211_cast_fp16")]; + tensor input_213_axes_0 = const()[name = tensor("input_213_axes_0"), val = tensor([-1])]; + tensor norm17_2_weight_to_fp16 = const()[name = tensor("norm17_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(436378944)))]; + tensor norm17_2_bias_to_fp16 = const()[name = tensor("norm17_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(436381056)))]; + tensor var_4457_to_fp16 = const()[name = tensor("op_4457_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_213_cast_fp16 = layer_norm(axes = input_213_axes_0, beta = norm17_2_bias_to_fp16, epsilon = var_4457_to_fp16, gamma = norm17_2_weight_to_fp16, x = input_211_cast_fp16)[name = tensor("input_213_cast_fp16")]; + tensor linear17_1_weight_to_fp16 = const()[name = tensor("linear17_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(436383168)))]; + tensor linear_70_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear17_1_weight_to_fp16, x = input_213_cast_fp16)[name = tensor("linear_70_cast_fp16")]; + tensor input_215_mode_0 = const()[name = tensor("input_215_mode_0"), val = tensor("EXACT")]; + tensor input_215_cast_fp16 = gelu(mode = input_215_mode_0, x = linear_70_cast_fp16)[name = tensor("input_215_cast_fp16")]; + tensor linear17_2_weight_to_fp16 = const()[name = tensor("linear17_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(444771840)))]; + tensor linear_71_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear17_2_weight_to_fp16, x = input_215_cast_fp16)[name = tensor("linear_71_cast_fp16")]; + tensor input_217_cast_fp16 = add(x = input_211_cast_fp16, y = linear_71_cast_fp16)[name = tensor("input_217_cast_fp16")]; + tensor input_219_axes_0 = const()[name = tensor("input_219_axes_0"), val = tensor([-1])]; + tensor norm18_1_weight_to_fp16 = const()[name = tensor("norm18_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(453160512)))]; + tensor norm18_1_bias_to_fp16 = const()[name = tensor("norm18_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(453162624)))]; + tensor var_4474_to_fp16 = const()[name = tensor("op_4474_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_219_cast_fp16 = layer_norm(axes = input_219_axes_0, beta = norm18_1_bias_to_fp16, epsilon = var_4474_to_fp16, gamma = norm18_1_weight_to_fp16, x = input_217_cast_fp16)[name = tensor("input_219_cast_fp16")]; + tensor attn18_in_proj_weight_to_fp16 = const()[name = tensor("attn18_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(453164736)))]; + tensor linear_72_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn18_in_proj_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_72_cast_fp16")]; + tensor var_4488 = const()[name = tensor("op_4488"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_37_cast_fp16 = reshape(shape = var_4488, x = linear_72_cast_fp16)[name = tensor("qkv_37_cast_fp16")]; + tensor x_73_begin_0 = const()[name = tensor("x_73_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_73_end_0 = const()[name = tensor("x_73_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_73_end_mask_0 = const()[name = tensor("x_73_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_73_squeeze_mask_0 = const()[name = tensor("x_73_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_73_cast_fp16 = slice_by_index(begin = x_73_begin_0, end = x_73_end_0, end_mask = x_73_end_mask_0, squeeze_mask = x_73_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("x_73_cast_fp16")]; + tensor x_75_begin_0 = const()[name = tensor("x_75_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_75_end_0 = const()[name = tensor("x_75_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_75_end_mask_0 = const()[name = tensor("x_75_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_75_squeeze_mask_0 = const()[name = tensor("x_75_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_75_cast_fp16 = slice_by_index(begin = x_75_begin_0, end = x_75_end_0, end_mask = x_75_end_mask_0, squeeze_mask = x_75_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("x_75_cast_fp16")]; + tensor v_37_begin_0 = const()[name = tensor("v_37_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_37_end_0 = const()[name = tensor("v_37_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_37_end_mask_0 = const()[name = tensor("v_37_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_37_squeeze_mask_0 = const()[name = tensor("v_37_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_37_cast_fp16 = slice_by_index(begin = v_37_begin_0, end = v_37_end_0, end_mask = v_37_end_mask_0, squeeze_mask = v_37_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("v_37_cast_fp16")]; + tensor var_4533 = const()[name = tensor("op_4533"), val = tensor([256, 16, 32, 2])]; + tensor pairs_73_cast_fp16 = reshape(shape = var_4533, x = x_73_cast_fp16)[name = tensor("pairs_73_cast_fp16")]; + tensor xr_73_begin_0 = const()[name = tensor("xr_73_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_73_end_0 = const()[name = tensor("xr_73_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_73_end_mask_0 = const()[name = tensor("xr_73_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_73_squeeze_mask_0 = const()[name = tensor("xr_73_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_73_cast_fp16 = slice_by_index(begin = xr_73_begin_0, end = xr_73_end_0, end_mask = xr_73_end_mask_0, squeeze_mask = xr_73_squeeze_mask_0, x = pairs_73_cast_fp16)[name = tensor("xr_73_cast_fp16")]; + tensor xi_73_begin_0 = const()[name = tensor("xi_73_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_73_end_0 = const()[name = tensor("xi_73_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_73_end_mask_0 = const()[name = tensor("xi_73_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_73_squeeze_mask_0 = const()[name = tensor("xi_73_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_73_cast_fp16 = slice_by_index(begin = xi_73_begin_0, end = xi_73_end_0, end_mask = xi_73_end_mask_0, squeeze_mask = xi_73_squeeze_mask_0, x = pairs_73_cast_fp16)[name = tensor("xi_73_cast_fp16")]; + tensor var_4541_cast_fp16 = mul(x = xr_73_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4541_cast_fp16")]; + tensor var_4542_cast_fp16 = mul(x = xi_73_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4542_cast_fp16")]; + tensor xor__73_cast_fp16 = sub(x = var_4541_cast_fp16, y = var_4542_cast_fp16)[name = tensor("xor__73_cast_fp16")]; + tensor var_4545_cast_fp16 = mul(x = xr_73_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4545_cast_fp16")]; + tensor var_4546_cast_fp16 = mul(x = xi_73_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4546_cast_fp16")]; + tensor xoi_73_cast_fp16 = add(x = var_4545_cast_fp16, y = var_4546_cast_fp16)[name = tensor("xoi_73_cast_fp16")]; + tensor out_73_axis_0 = const()[name = tensor("out_73_axis_0"), val = tensor(-1)]; + tensor out_73_cast_fp16 = stack(axis = out_73_axis_0, values = (xor__73_cast_fp16, xoi_73_cast_fp16))[name = tensor("out_73_cast_fp16")]; + tensor var_4556 = const()[name = tensor("op_4556"), val = tensor([1, 256, 16, 64])]; + tensor q_37_cast_fp16 = reshape(shape = var_4556, x = out_73_cast_fp16)[name = tensor("q_37_cast_fp16")]; + tensor var_4562 = const()[name = tensor("op_4562"), val = tensor([256, 16, 32, 2])]; + tensor pairs_75_cast_fp16 = reshape(shape = var_4562, x = x_75_cast_fp16)[name = tensor("pairs_75_cast_fp16")]; + tensor xr_75_begin_0 = const()[name = tensor("xr_75_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_75_end_0 = const()[name = tensor("xr_75_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_75_end_mask_0 = const()[name = tensor("xr_75_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_75_squeeze_mask_0 = const()[name = tensor("xr_75_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_75_cast_fp16 = slice_by_index(begin = xr_75_begin_0, end = xr_75_end_0, end_mask = xr_75_end_mask_0, squeeze_mask = xr_75_squeeze_mask_0, x = pairs_75_cast_fp16)[name = tensor("xr_75_cast_fp16")]; + tensor xi_75_begin_0 = const()[name = tensor("xi_75_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_75_end_0 = const()[name = tensor("xi_75_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_75_end_mask_0 = const()[name = tensor("xi_75_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_75_squeeze_mask_0 = const()[name = tensor("xi_75_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_75_cast_fp16 = slice_by_index(begin = xi_75_begin_0, end = xi_75_end_0, end_mask = xi_75_end_mask_0, squeeze_mask = xi_75_squeeze_mask_0, x = pairs_75_cast_fp16)[name = tensor("xi_75_cast_fp16")]; + tensor var_4570_cast_fp16 = mul(x = xr_75_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4570_cast_fp16")]; + tensor var_4571_cast_fp16 = mul(x = xi_75_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4571_cast_fp16")]; + tensor xor__75_cast_fp16 = sub(x = var_4570_cast_fp16, y = var_4571_cast_fp16)[name = tensor("xor__75_cast_fp16")]; + tensor var_4574_cast_fp16 = mul(x = xr_75_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4574_cast_fp16")]; + tensor var_4575_cast_fp16 = mul(x = xi_75_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4575_cast_fp16")]; + tensor xoi_75_cast_fp16 = add(x = var_4574_cast_fp16, y = var_4575_cast_fp16)[name = tensor("xoi_75_cast_fp16")]; + tensor out_75_axis_0 = const()[name = tensor("out_75_axis_0"), val = tensor(-1)]; + tensor out_75_cast_fp16 = stack(axis = out_75_axis_0, values = (xor__75_cast_fp16, xoi_75_cast_fp16))[name = tensor("out_75_cast_fp16")]; + tensor var_4592 = const()[name = tensor("op_4592"), val = tensor([256, 1024])]; + tensor var_4593_cast_fp16 = reshape(shape = var_4592, x = out_75_cast_fp16)[name = tensor("op_4593_cast_fp16")]; + tensor written_k_37_transpose_x_1 = const()[name = tensor("written_k_37_transpose_x_1"), val = tensor(true)]; + tensor written_k_37_transpose_y_1 = const()[name = tensor("written_k_37_transpose_y_1"), val = tensor(false)]; + tensor written_k_37_cast_fp16 = matmul(transpose_x = written_k_37_transpose_x_1, transpose_y = written_k_37_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_4593_cast_fp16)[name = tensor("written_k_37_cast_fp16")]; + tensor var_4600 = const()[name = tensor("op_4600"), val = tensor([256, 1024])]; + tensor var_4601_cast_fp16 = reshape(shape = var_4600, x = v_37_cast_fp16)[name = tensor("op_4601_cast_fp16")]; + tensor written_v_37_transpose_x_1 = const()[name = tensor("written_v_37_transpose_x_1"), val = tensor(true)]; + tensor written_v_37_transpose_y_1 = const()[name = tensor("written_v_37_transpose_y_1"), val = tensor(false)]; + tensor written_v_37_cast_fp16 = matmul(transpose_x = written_v_37_transpose_x_1, transpose_y = written_v_37_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_4601_cast_fp16)[name = tensor("written_v_37_cast_fp16")]; + tensor k_cache18_to_fp16_dtype_0 = const()[name = tensor("k_cache18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache18_to_fp16 = cast(dtype = k_cache18_to_fp16_dtype_0, x = k_cache18)[name = tensor("cast_357")]; + tensor var_4612_cast_fp16 = mul(x = k_cache18_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_4612_cast_fp16")]; + tensor var_4617 = const()[name = tensor("op_4617"), val = tensor([1, 512, 16, 64])]; + tensor var_4618_cast_fp16 = reshape(shape = var_4617, x = written_k_37_cast_fp16)[name = tensor("op_4618_cast_fp16")]; + tensor new_k_cache_37_cast_fp16 = add(x = var_4612_cast_fp16, y = var_4618_cast_fp16)[name = tensor("new_k_cache_37_cast_fp16")]; + tensor new_k_cache_37_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_37_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache18_to_fp16_dtype_0 = const()[name = tensor("v_cache18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache18_to_fp16 = cast(dtype = v_cache18_to_fp16_dtype_0, x = v_cache18)[name = tensor("cast_355")]; + tensor var_4621_cast_fp16 = mul(x = v_cache18_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_4621_cast_fp16")]; + tensor var_4626 = const()[name = tensor("op_4626"), val = tensor([1, 512, 16, 64])]; + tensor var_4627_cast_fp16 = reshape(shape = var_4626, x = written_v_37_cast_fp16)[name = tensor("op_4627_cast_fp16")]; + tensor new_v_cache_37_cast_fp16 = add(x = var_4621_cast_fp16, y = var_4627_cast_fp16)[name = tensor("new_v_cache_37_cast_fp16")]; + tensor new_v_cache_37_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_37_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_4640 = const()[name = tensor("op_4640"), val = tensor([0, 2, 1, 3])]; + tensor var_4651_transpose_x_0 = const()[name = tensor("op_4651_transpose_x_0"), val = tensor(false)]; + tensor var_4651_transpose_y_0 = const()[name = tensor("op_4651_transpose_y_0"), val = tensor(false)]; + tensor transpose_105_perm_0 = const()[name = tensor("transpose_105_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_106_perm_0 = const()[name = tensor("transpose_106_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_106 = transpose(perm = transpose_106_perm_0, x = new_k_cache_37_cast_fp16)[name = tensor("transpose_132")]; + tensor transpose_105 = transpose(perm = transpose_105_perm_0, x = q_37_cast_fp16)[name = tensor("transpose_133")]; + tensor var_4651_cast_fp16 = matmul(transpose_x = var_4651_transpose_x_0, transpose_y = var_4651_transpose_y_0, x = transpose_105, y = transpose_106)[name = tensor("op_4651_cast_fp16")]; + tensor var_4652_to_fp16 = const()[name = tensor("op_4652_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_73_cast_fp16 = mul(x = var_4651_cast_fp16, y = var_4652_to_fp16)[name = tensor("scores_73_cast_fp16")]; + tensor scores_75_cast_fp16 = add(x = scores_73_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_75_cast_fp16")]; + tensor var_4662 = const()[name = tensor("op_4662"), val = tensor(-1)]; + tensor weights_37_cast_fp16 = softmax(axis = var_4662, x = scores_75_cast_fp16)[name = tensor("weights_37_cast_fp16")]; + tensor attn_37_transpose_x_0 = const()[name = tensor("attn_37_transpose_x_0"), val = tensor(false)]; + tensor attn_37_transpose_y_0 = const()[name = tensor("attn_37_transpose_y_0"), val = tensor(false)]; + tensor values_37_cast_fp16 = transpose(perm = var_4640, x = new_v_cache_37_cast_fp16)[name = tensor("transpose_134")]; + tensor attn_37_cast_fp16 = matmul(transpose_x = attn_37_transpose_x_0, transpose_y = attn_37_transpose_y_0, x = weights_37_cast_fp16, y = values_37_cast_fp16)[name = tensor("attn_37_cast_fp16")]; + tensor var_4670 = const()[name = tensor("op_4670"), val = tensor([0, 2, 1, 3])]; + tensor var_4675 = const()[name = tensor("op_4675"), val = tensor([1, 256, 1024])]; + tensor var_4671_cast_fp16 = transpose(perm = var_4670, x = attn_37_cast_fp16)[name = tensor("transpose_131")]; + tensor input_221_cast_fp16 = reshape(shape = var_4675, x = var_4671_cast_fp16)[name = tensor("input_221_cast_fp16")]; + tensor attn18_out_proj_weight_to_fp16 = const()[name = tensor("attn18_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(459456256)))]; + tensor linear_73_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn18_out_proj_weight_to_fp16, x = input_221_cast_fp16)[name = tensor("linear_73_cast_fp16")]; + tensor position18_to_fp16_dtype_0 = const()[name = tensor("position18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position18_to_fp16 = cast(dtype = position18_to_fp16_dtype_0, x = position18)[name = tensor("cast_353")]; + tensor var_4684_cast_fp16 = add(x = position18_to_fp16, y = valid_len_to_fp16)[name = tensor("op_4684_cast_fp16")]; + tensor var_4684_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4684_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_223_cast_fp16 = add(x = input_217_cast_fp16, y = linear_73_cast_fp16)[name = tensor("input_223_cast_fp16")]; + tensor input_225_axes_0 = const()[name = tensor("input_225_axes_0"), val = tensor([-1])]; + tensor norm18_2_weight_to_fp16 = const()[name = tensor("norm18_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(461553472)))]; + tensor norm18_2_bias_to_fp16 = const()[name = tensor("norm18_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(461555584)))]; + tensor var_4688_to_fp16 = const()[name = tensor("op_4688_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_225_cast_fp16 = layer_norm(axes = input_225_axes_0, beta = norm18_2_bias_to_fp16, epsilon = var_4688_to_fp16, gamma = norm18_2_weight_to_fp16, x = input_223_cast_fp16)[name = tensor("input_225_cast_fp16")]; + tensor linear18_1_weight_to_fp16 = const()[name = tensor("linear18_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(461557696)))]; + tensor linear_74_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear18_1_weight_to_fp16, x = input_225_cast_fp16)[name = tensor("linear_74_cast_fp16")]; + tensor input_227_mode_0 = const()[name = tensor("input_227_mode_0"), val = tensor("EXACT")]; + tensor input_227_cast_fp16 = gelu(mode = input_227_mode_0, x = linear_74_cast_fp16)[name = tensor("input_227_cast_fp16")]; + tensor linear18_2_weight_to_fp16 = const()[name = tensor("linear18_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(469946368)))]; + tensor linear_75_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear18_2_weight_to_fp16, x = input_227_cast_fp16)[name = tensor("linear_75_cast_fp16")]; + tensor input_229_cast_fp16 = add(x = input_223_cast_fp16, y = linear_75_cast_fp16)[name = tensor("input_229_cast_fp16")]; + tensor input_231_axes_0 = const()[name = tensor("input_231_axes_0"), val = tensor([-1])]; + tensor norm19_1_weight_to_fp16 = const()[name = tensor("norm19_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(478335040)))]; + tensor norm19_1_bias_to_fp16 = const()[name = tensor("norm19_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(478337152)))]; + tensor var_4705_to_fp16 = const()[name = tensor("op_4705_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_231_cast_fp16 = layer_norm(axes = input_231_axes_0, beta = norm19_1_bias_to_fp16, epsilon = var_4705_to_fp16, gamma = norm19_1_weight_to_fp16, x = input_229_cast_fp16)[name = tensor("input_231_cast_fp16")]; + tensor attn19_in_proj_weight_to_fp16 = const()[name = tensor("attn19_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(478339264)))]; + tensor linear_76_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn19_in_proj_weight_to_fp16, x = input_231_cast_fp16)[name = tensor("linear_76_cast_fp16")]; + tensor var_4719 = const()[name = tensor("op_4719"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_39_cast_fp16 = reshape(shape = var_4719, x = linear_76_cast_fp16)[name = tensor("qkv_39_cast_fp16")]; + tensor x_77_begin_0 = const()[name = tensor("x_77_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_77_end_0 = const()[name = tensor("x_77_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_77_end_mask_0 = const()[name = tensor("x_77_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_77_squeeze_mask_0 = const()[name = tensor("x_77_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_77_cast_fp16 = slice_by_index(begin = x_77_begin_0, end = x_77_end_0, end_mask = x_77_end_mask_0, squeeze_mask = x_77_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("x_77_cast_fp16")]; + tensor x_79_begin_0 = const()[name = tensor("x_79_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_79_end_0 = const()[name = tensor("x_79_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_79_end_mask_0 = const()[name = tensor("x_79_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_79_squeeze_mask_0 = const()[name = tensor("x_79_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_79_cast_fp16 = slice_by_index(begin = x_79_begin_0, end = x_79_end_0, end_mask = x_79_end_mask_0, squeeze_mask = x_79_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("x_79_cast_fp16")]; + tensor v_39_begin_0 = const()[name = tensor("v_39_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_39_end_0 = const()[name = tensor("v_39_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_39_end_mask_0 = const()[name = tensor("v_39_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_39_squeeze_mask_0 = const()[name = tensor("v_39_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_39_cast_fp16 = slice_by_index(begin = v_39_begin_0, end = v_39_end_0, end_mask = v_39_end_mask_0, squeeze_mask = v_39_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("v_39_cast_fp16")]; + tensor var_4764 = const()[name = tensor("op_4764"), val = tensor([256, 16, 32, 2])]; + tensor pairs_77_cast_fp16 = reshape(shape = var_4764, x = x_77_cast_fp16)[name = tensor("pairs_77_cast_fp16")]; + tensor xr_77_begin_0 = const()[name = tensor("xr_77_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_77_end_0 = const()[name = tensor("xr_77_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_77_end_mask_0 = const()[name = tensor("xr_77_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_77_squeeze_mask_0 = const()[name = tensor("xr_77_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_77_cast_fp16 = slice_by_index(begin = xr_77_begin_0, end = xr_77_end_0, end_mask = xr_77_end_mask_0, squeeze_mask = xr_77_squeeze_mask_0, x = pairs_77_cast_fp16)[name = tensor("xr_77_cast_fp16")]; + tensor xi_77_begin_0 = const()[name = tensor("xi_77_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_77_end_0 = const()[name = tensor("xi_77_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_77_end_mask_0 = const()[name = tensor("xi_77_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_77_squeeze_mask_0 = const()[name = tensor("xi_77_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_77_cast_fp16 = slice_by_index(begin = xi_77_begin_0, end = xi_77_end_0, end_mask = xi_77_end_mask_0, squeeze_mask = xi_77_squeeze_mask_0, x = pairs_77_cast_fp16)[name = tensor("xi_77_cast_fp16")]; + tensor var_4772_cast_fp16 = mul(x = xr_77_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4772_cast_fp16")]; + tensor var_4773_cast_fp16 = mul(x = xi_77_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4773_cast_fp16")]; + tensor xor__77_cast_fp16 = sub(x = var_4772_cast_fp16, y = var_4773_cast_fp16)[name = tensor("xor__77_cast_fp16")]; + tensor var_4776_cast_fp16 = mul(x = xr_77_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4776_cast_fp16")]; + tensor var_4777_cast_fp16 = mul(x = xi_77_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4777_cast_fp16")]; + tensor xoi_77_cast_fp16 = add(x = var_4776_cast_fp16, y = var_4777_cast_fp16)[name = tensor("xoi_77_cast_fp16")]; + tensor out_77_axis_0 = const()[name = tensor("out_77_axis_0"), val = tensor(-1)]; + tensor out_77_cast_fp16 = stack(axis = out_77_axis_0, values = (xor__77_cast_fp16, xoi_77_cast_fp16))[name = tensor("out_77_cast_fp16")]; + tensor var_4787 = const()[name = tensor("op_4787"), val = tensor([1, 256, 16, 64])]; + tensor q_39_cast_fp16 = reshape(shape = var_4787, x = out_77_cast_fp16)[name = tensor("q_39_cast_fp16")]; + tensor var_4793 = const()[name = tensor("op_4793"), val = tensor([256, 16, 32, 2])]; + tensor pairs_79_cast_fp16 = reshape(shape = var_4793, x = x_79_cast_fp16)[name = tensor("pairs_79_cast_fp16")]; + tensor xr_79_begin_0 = const()[name = tensor("xr_79_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_79_end_0 = const()[name = tensor("xr_79_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_79_end_mask_0 = const()[name = tensor("xr_79_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_79_squeeze_mask_0 = const()[name = tensor("xr_79_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_79_cast_fp16 = slice_by_index(begin = xr_79_begin_0, end = xr_79_end_0, end_mask = xr_79_end_mask_0, squeeze_mask = xr_79_squeeze_mask_0, x = pairs_79_cast_fp16)[name = tensor("xr_79_cast_fp16")]; + tensor xi_79_begin_0 = const()[name = tensor("xi_79_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_79_end_0 = const()[name = tensor("xi_79_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_79_end_mask_0 = const()[name = tensor("xi_79_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_79_squeeze_mask_0 = const()[name = tensor("xi_79_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_79_cast_fp16 = slice_by_index(begin = xi_79_begin_0, end = xi_79_end_0, end_mask = xi_79_end_mask_0, squeeze_mask = xi_79_squeeze_mask_0, x = pairs_79_cast_fp16)[name = tensor("xi_79_cast_fp16")]; + tensor var_4801_cast_fp16 = mul(x = xr_79_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4801_cast_fp16")]; + tensor var_4802_cast_fp16 = mul(x = xi_79_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4802_cast_fp16")]; + tensor xor__79_cast_fp16 = sub(x = var_4801_cast_fp16, y = var_4802_cast_fp16)[name = tensor("xor__79_cast_fp16")]; + tensor var_4805_cast_fp16 = mul(x = xr_79_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4805_cast_fp16")]; + tensor var_4806_cast_fp16 = mul(x = xi_79_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4806_cast_fp16")]; + tensor xoi_79_cast_fp16 = add(x = var_4805_cast_fp16, y = var_4806_cast_fp16)[name = tensor("xoi_79_cast_fp16")]; + tensor out_79_axis_0 = const()[name = tensor("out_79_axis_0"), val = tensor(-1)]; + tensor out_79_cast_fp16 = stack(axis = out_79_axis_0, values = (xor__79_cast_fp16, xoi_79_cast_fp16))[name = tensor("out_79_cast_fp16")]; + tensor var_4823 = const()[name = tensor("op_4823"), val = tensor([256, 1024])]; + tensor var_4824_cast_fp16 = reshape(shape = var_4823, x = out_79_cast_fp16)[name = tensor("op_4824_cast_fp16")]; + tensor written_k_39_transpose_x_1 = const()[name = tensor("written_k_39_transpose_x_1"), val = tensor(true)]; + tensor written_k_39_transpose_y_1 = const()[name = tensor("written_k_39_transpose_y_1"), val = tensor(false)]; + tensor written_k_39_cast_fp16 = matmul(transpose_x = written_k_39_transpose_x_1, transpose_y = written_k_39_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_4824_cast_fp16)[name = tensor("written_k_39_cast_fp16")]; + tensor var_4831 = const()[name = tensor("op_4831"), val = tensor([256, 1024])]; + tensor var_4832_cast_fp16 = reshape(shape = var_4831, x = v_39_cast_fp16)[name = tensor("op_4832_cast_fp16")]; + tensor written_v_39_transpose_x_1 = const()[name = tensor("written_v_39_transpose_x_1"), val = tensor(true)]; + tensor written_v_39_transpose_y_1 = const()[name = tensor("written_v_39_transpose_y_1"), val = tensor(false)]; + tensor written_v_39_cast_fp16 = matmul(transpose_x = written_v_39_transpose_x_1, transpose_y = written_v_39_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_4832_cast_fp16)[name = tensor("written_v_39_cast_fp16")]; + tensor k_cache19_to_fp16_dtype_0 = const()[name = tensor("k_cache19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache19_to_fp16 = cast(dtype = k_cache19_to_fp16_dtype_0, x = k_cache19)[name = tensor("cast_351")]; + tensor var_4843_cast_fp16 = mul(x = k_cache19_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_4843_cast_fp16")]; + tensor var_4848 = const()[name = tensor("op_4848"), val = tensor([1, 512, 16, 64])]; + tensor var_4849_cast_fp16 = reshape(shape = var_4848, x = written_k_39_cast_fp16)[name = tensor("op_4849_cast_fp16")]; + tensor new_k_cache_39_cast_fp16 = add(x = var_4843_cast_fp16, y = var_4849_cast_fp16)[name = tensor("new_k_cache_39_cast_fp16")]; + tensor new_k_cache_39_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_39_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache19_to_fp16_dtype_0 = const()[name = tensor("v_cache19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache19_to_fp16 = cast(dtype = v_cache19_to_fp16_dtype_0, x = v_cache19)[name = tensor("cast_349")]; + tensor var_4852_cast_fp16 = mul(x = v_cache19_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_4852_cast_fp16")]; + tensor var_4857 = const()[name = tensor("op_4857"), val = tensor([1, 512, 16, 64])]; + tensor var_4858_cast_fp16 = reshape(shape = var_4857, x = written_v_39_cast_fp16)[name = tensor("op_4858_cast_fp16")]; + tensor new_v_cache_39_cast_fp16 = add(x = var_4852_cast_fp16, y = var_4858_cast_fp16)[name = tensor("new_v_cache_39_cast_fp16")]; + tensor new_v_cache_39_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_39_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_4871 = const()[name = tensor("op_4871"), val = tensor([0, 2, 1, 3])]; + tensor var_4882_transpose_x_0 = const()[name = tensor("op_4882_transpose_x_0"), val = tensor(false)]; + tensor var_4882_transpose_y_0 = const()[name = tensor("op_4882_transpose_y_0"), val = tensor(false)]; + tensor transpose_107_perm_0 = const()[name = tensor("transpose_107_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_108_perm_0 = const()[name = tensor("transpose_108_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_108 = transpose(perm = transpose_108_perm_0, x = new_k_cache_39_cast_fp16)[name = tensor("transpose_128")]; + tensor transpose_107 = transpose(perm = transpose_107_perm_0, x = q_39_cast_fp16)[name = tensor("transpose_129")]; + tensor var_4882_cast_fp16 = matmul(transpose_x = var_4882_transpose_x_0, transpose_y = var_4882_transpose_y_0, x = transpose_107, y = transpose_108)[name = tensor("op_4882_cast_fp16")]; + tensor var_4883_to_fp16 = const()[name = tensor("op_4883_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_77_cast_fp16 = mul(x = var_4882_cast_fp16, y = var_4883_to_fp16)[name = tensor("scores_77_cast_fp16")]; + tensor scores_79_cast_fp16 = add(x = scores_77_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_79_cast_fp16")]; + tensor var_4893 = const()[name = tensor("op_4893"), val = tensor(-1)]; + tensor weights_39_cast_fp16 = softmax(axis = var_4893, x = scores_79_cast_fp16)[name = tensor("weights_39_cast_fp16")]; + tensor attn_39_transpose_x_0 = const()[name = tensor("attn_39_transpose_x_0"), val = tensor(false)]; + tensor attn_39_transpose_y_0 = const()[name = tensor("attn_39_transpose_y_0"), val = tensor(false)]; + tensor values_39_cast_fp16 = transpose(perm = var_4871, x = new_v_cache_39_cast_fp16)[name = tensor("transpose_130")]; + tensor attn_39_cast_fp16 = matmul(transpose_x = attn_39_transpose_x_0, transpose_y = attn_39_transpose_y_0, x = weights_39_cast_fp16, y = values_39_cast_fp16)[name = tensor("attn_39_cast_fp16")]; + tensor var_4901 = const()[name = tensor("op_4901"), val = tensor([0, 2, 1, 3])]; + tensor var_4906 = const()[name = tensor("op_4906"), val = tensor([1, 256, 1024])]; + tensor var_4902_cast_fp16 = transpose(perm = var_4901, x = attn_39_cast_fp16)[name = tensor("transpose_127")]; + tensor input_233_cast_fp16 = reshape(shape = var_4906, x = var_4902_cast_fp16)[name = tensor("input_233_cast_fp16")]; + tensor attn19_out_proj_weight_to_fp16 = const()[name = tensor("attn19_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(484630784)))]; + tensor linear_77_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn19_out_proj_weight_to_fp16, x = input_233_cast_fp16)[name = tensor("linear_77_cast_fp16")]; + tensor position19_to_fp16_dtype_0 = const()[name = tensor("position19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position19_to_fp16 = cast(dtype = position19_to_fp16_dtype_0, x = position19)[name = tensor("cast_347")]; + tensor var_4915_cast_fp16 = add(x = position19_to_fp16, y = valid_len_to_fp16)[name = tensor("op_4915_cast_fp16")]; + tensor var_4915_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4915_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_235_cast_fp16 = add(x = input_229_cast_fp16, y = linear_77_cast_fp16)[name = tensor("input_235_cast_fp16")]; + tensor input_237_axes_0 = const()[name = tensor("input_237_axes_0"), val = tensor([-1])]; + tensor norm19_2_weight_to_fp16 = const()[name = tensor("norm19_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486728000)))]; + tensor norm19_2_bias_to_fp16 = const()[name = tensor("norm19_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486730112)))]; + tensor var_4919_to_fp16 = const()[name = tensor("op_4919_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_237_cast_fp16 = layer_norm(axes = input_237_axes_0, beta = norm19_2_bias_to_fp16, epsilon = var_4919_to_fp16, gamma = norm19_2_weight_to_fp16, x = input_235_cast_fp16)[name = tensor("input_237_cast_fp16")]; + tensor linear19_1_weight_to_fp16 = const()[name = tensor("linear19_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486732224)))]; + tensor linear_78_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear19_1_weight_to_fp16, x = input_237_cast_fp16)[name = tensor("linear_78_cast_fp16")]; + tensor input_239_mode_0 = const()[name = tensor("input_239_mode_0"), val = tensor("EXACT")]; + tensor input_239_cast_fp16 = gelu(mode = input_239_mode_0, x = linear_78_cast_fp16)[name = tensor("input_239_cast_fp16")]; + tensor linear19_2_weight_to_fp16 = const()[name = tensor("linear19_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(495120896)))]; + tensor linear_79_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear19_2_weight_to_fp16, x = input_239_cast_fp16)[name = tensor("linear_79_cast_fp16")]; + tensor input_241_cast_fp16 = add(x = input_235_cast_fp16, y = linear_79_cast_fp16)[name = tensor("input_241_cast_fp16")]; + tensor input_243_axes_0 = const()[name = tensor("input_243_axes_0"), val = tensor([-1])]; + tensor norm20_1_weight_to_fp16 = const()[name = tensor("norm20_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(503509568)))]; + tensor norm20_1_bias_to_fp16 = const()[name = tensor("norm20_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(503511680)))]; + tensor var_4936_to_fp16 = const()[name = tensor("op_4936_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_243_cast_fp16 = layer_norm(axes = input_243_axes_0, beta = norm20_1_bias_to_fp16, epsilon = var_4936_to_fp16, gamma = norm20_1_weight_to_fp16, x = input_241_cast_fp16)[name = tensor("input_243_cast_fp16")]; + tensor attn20_in_proj_weight_to_fp16 = const()[name = tensor("attn20_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(503513792)))]; + tensor linear_80_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn20_in_proj_weight_to_fp16, x = input_243_cast_fp16)[name = tensor("linear_80_cast_fp16")]; + tensor var_4950 = const()[name = tensor("op_4950"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_41_cast_fp16 = reshape(shape = var_4950, x = linear_80_cast_fp16)[name = tensor("qkv_41_cast_fp16")]; + tensor x_81_begin_0 = const()[name = tensor("x_81_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_81_end_0 = const()[name = tensor("x_81_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_81_end_mask_0 = const()[name = tensor("x_81_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_81_squeeze_mask_0 = const()[name = tensor("x_81_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_81_cast_fp16 = slice_by_index(begin = x_81_begin_0, end = x_81_end_0, end_mask = x_81_end_mask_0, squeeze_mask = x_81_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("x_81_cast_fp16")]; + tensor x_83_begin_0 = const()[name = tensor("x_83_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_83_end_0 = const()[name = tensor("x_83_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_83_end_mask_0 = const()[name = tensor("x_83_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_83_squeeze_mask_0 = const()[name = tensor("x_83_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_83_cast_fp16 = slice_by_index(begin = x_83_begin_0, end = x_83_end_0, end_mask = x_83_end_mask_0, squeeze_mask = x_83_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("x_83_cast_fp16")]; + tensor v_41_begin_0 = const()[name = tensor("v_41_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_41_end_0 = const()[name = tensor("v_41_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_41_end_mask_0 = const()[name = tensor("v_41_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_41_squeeze_mask_0 = const()[name = tensor("v_41_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_41_cast_fp16 = slice_by_index(begin = v_41_begin_0, end = v_41_end_0, end_mask = v_41_end_mask_0, squeeze_mask = v_41_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("v_41_cast_fp16")]; + tensor var_4995 = const()[name = tensor("op_4995"), val = tensor([256, 16, 32, 2])]; + tensor pairs_81_cast_fp16 = reshape(shape = var_4995, x = x_81_cast_fp16)[name = tensor("pairs_81_cast_fp16")]; + tensor xr_81_begin_0 = const()[name = tensor("xr_81_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_81_end_0 = const()[name = tensor("xr_81_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_81_end_mask_0 = const()[name = tensor("xr_81_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_81_squeeze_mask_0 = const()[name = tensor("xr_81_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_81_cast_fp16 = slice_by_index(begin = xr_81_begin_0, end = xr_81_end_0, end_mask = xr_81_end_mask_0, squeeze_mask = xr_81_squeeze_mask_0, x = pairs_81_cast_fp16)[name = tensor("xr_81_cast_fp16")]; + tensor xi_81_begin_0 = const()[name = tensor("xi_81_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_81_end_0 = const()[name = tensor("xi_81_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_81_end_mask_0 = const()[name = tensor("xi_81_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_81_squeeze_mask_0 = const()[name = tensor("xi_81_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_81_cast_fp16 = slice_by_index(begin = xi_81_begin_0, end = xi_81_end_0, end_mask = xi_81_end_mask_0, squeeze_mask = xi_81_squeeze_mask_0, x = pairs_81_cast_fp16)[name = tensor("xi_81_cast_fp16")]; + tensor var_5003_cast_fp16 = mul(x = xr_81_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5003_cast_fp16")]; + tensor var_5004_cast_fp16 = mul(x = xi_81_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5004_cast_fp16")]; + tensor xor__81_cast_fp16 = sub(x = var_5003_cast_fp16, y = var_5004_cast_fp16)[name = tensor("xor__81_cast_fp16")]; + tensor var_5007_cast_fp16 = mul(x = xr_81_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5007_cast_fp16")]; + tensor var_5008_cast_fp16 = mul(x = xi_81_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5008_cast_fp16")]; + tensor xoi_81_cast_fp16 = add(x = var_5007_cast_fp16, y = var_5008_cast_fp16)[name = tensor("xoi_81_cast_fp16")]; + tensor out_81_axis_0 = const()[name = tensor("out_81_axis_0"), val = tensor(-1)]; + tensor out_81_cast_fp16 = stack(axis = out_81_axis_0, values = (xor__81_cast_fp16, xoi_81_cast_fp16))[name = tensor("out_81_cast_fp16")]; + tensor var_5018 = const()[name = tensor("op_5018"), val = tensor([1, 256, 16, 64])]; + tensor q_41_cast_fp16 = reshape(shape = var_5018, x = out_81_cast_fp16)[name = tensor("q_41_cast_fp16")]; + tensor var_5024 = const()[name = tensor("op_5024"), val = tensor([256, 16, 32, 2])]; + tensor pairs_83_cast_fp16 = reshape(shape = var_5024, x = x_83_cast_fp16)[name = tensor("pairs_83_cast_fp16")]; + tensor xr_83_begin_0 = const()[name = tensor("xr_83_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_83_end_0 = const()[name = tensor("xr_83_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_83_end_mask_0 = const()[name = tensor("xr_83_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_83_squeeze_mask_0 = const()[name = tensor("xr_83_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_83_cast_fp16 = slice_by_index(begin = xr_83_begin_0, end = xr_83_end_0, end_mask = xr_83_end_mask_0, squeeze_mask = xr_83_squeeze_mask_0, x = pairs_83_cast_fp16)[name = tensor("xr_83_cast_fp16")]; + tensor xi_83_begin_0 = const()[name = tensor("xi_83_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_83_end_0 = const()[name = tensor("xi_83_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_83_end_mask_0 = const()[name = tensor("xi_83_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_83_squeeze_mask_0 = const()[name = tensor("xi_83_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_83_cast_fp16 = slice_by_index(begin = xi_83_begin_0, end = xi_83_end_0, end_mask = xi_83_end_mask_0, squeeze_mask = xi_83_squeeze_mask_0, x = pairs_83_cast_fp16)[name = tensor("xi_83_cast_fp16")]; + tensor var_5032_cast_fp16 = mul(x = xr_83_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5032_cast_fp16")]; + tensor var_5033_cast_fp16 = mul(x = xi_83_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5033_cast_fp16")]; + tensor xor__83_cast_fp16 = sub(x = var_5032_cast_fp16, y = var_5033_cast_fp16)[name = tensor("xor__83_cast_fp16")]; + tensor var_5036_cast_fp16 = mul(x = xr_83_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5036_cast_fp16")]; + tensor var_5037_cast_fp16 = mul(x = xi_83_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5037_cast_fp16")]; + tensor xoi_83_cast_fp16 = add(x = var_5036_cast_fp16, y = var_5037_cast_fp16)[name = tensor("xoi_83_cast_fp16")]; + tensor out_83_axis_0 = const()[name = tensor("out_83_axis_0"), val = tensor(-1)]; + tensor out_83_cast_fp16 = stack(axis = out_83_axis_0, values = (xor__83_cast_fp16, xoi_83_cast_fp16))[name = tensor("out_83_cast_fp16")]; + tensor var_5054 = const()[name = tensor("op_5054"), val = tensor([256, 1024])]; + tensor var_5055_cast_fp16 = reshape(shape = var_5054, x = out_83_cast_fp16)[name = tensor("op_5055_cast_fp16")]; + tensor written_k_41_transpose_x_1 = const()[name = tensor("written_k_41_transpose_x_1"), val = tensor(true)]; + tensor written_k_41_transpose_y_1 = const()[name = tensor("written_k_41_transpose_y_1"), val = tensor(false)]; + tensor written_k_41_cast_fp16 = matmul(transpose_x = written_k_41_transpose_x_1, transpose_y = written_k_41_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_5055_cast_fp16)[name = tensor("written_k_41_cast_fp16")]; + tensor var_5062 = const()[name = tensor("op_5062"), val = tensor([256, 1024])]; + tensor var_5063_cast_fp16 = reshape(shape = var_5062, x = v_41_cast_fp16)[name = tensor("op_5063_cast_fp16")]; + tensor written_v_41_transpose_x_1 = const()[name = tensor("written_v_41_transpose_x_1"), val = tensor(true)]; + tensor written_v_41_transpose_y_1 = const()[name = tensor("written_v_41_transpose_y_1"), val = tensor(false)]; + tensor written_v_41_cast_fp16 = matmul(transpose_x = written_v_41_transpose_x_1, transpose_y = written_v_41_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_5063_cast_fp16)[name = tensor("written_v_41_cast_fp16")]; + tensor k_cache20_to_fp16_dtype_0 = const()[name = tensor("k_cache20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache20_to_fp16 = cast(dtype = k_cache20_to_fp16_dtype_0, x = k_cache20)[name = tensor("cast_345")]; + tensor var_5074_cast_fp16 = mul(x = k_cache20_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_5074_cast_fp16")]; + tensor var_5079 = const()[name = tensor("op_5079"), val = tensor([1, 512, 16, 64])]; + tensor var_5080_cast_fp16 = reshape(shape = var_5079, x = written_k_41_cast_fp16)[name = tensor("op_5080_cast_fp16")]; + tensor new_k_cache_41_cast_fp16 = add(x = var_5074_cast_fp16, y = var_5080_cast_fp16)[name = tensor("new_k_cache_41_cast_fp16")]; + tensor new_k_cache_41_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_41_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache20_to_fp16_dtype_0 = const()[name = tensor("v_cache20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache20_to_fp16 = cast(dtype = v_cache20_to_fp16_dtype_0, x = v_cache20)[name = tensor("cast_343")]; + tensor var_5083_cast_fp16 = mul(x = v_cache20_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_5083_cast_fp16")]; + tensor var_5088 = const()[name = tensor("op_5088"), val = tensor([1, 512, 16, 64])]; + tensor var_5089_cast_fp16 = reshape(shape = var_5088, x = written_v_41_cast_fp16)[name = tensor("op_5089_cast_fp16")]; + tensor new_v_cache_41_cast_fp16 = add(x = var_5083_cast_fp16, y = var_5089_cast_fp16)[name = tensor("new_v_cache_41_cast_fp16")]; + tensor new_v_cache_41_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_41_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_5102 = const()[name = tensor("op_5102"), val = tensor([0, 2, 1, 3])]; + tensor var_5113_transpose_x_0 = const()[name = tensor("op_5113_transpose_x_0"), val = tensor(false)]; + tensor var_5113_transpose_y_0 = const()[name = tensor("op_5113_transpose_y_0"), val = tensor(false)]; + tensor transpose_109_perm_0 = const()[name = tensor("transpose_109_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_110_perm_0 = const()[name = tensor("transpose_110_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_110 = transpose(perm = transpose_110_perm_0, x = new_k_cache_41_cast_fp16)[name = tensor("transpose_124")]; + tensor transpose_109 = transpose(perm = transpose_109_perm_0, x = q_41_cast_fp16)[name = tensor("transpose_125")]; + tensor var_5113_cast_fp16 = matmul(transpose_x = var_5113_transpose_x_0, transpose_y = var_5113_transpose_y_0, x = transpose_109, y = transpose_110)[name = tensor("op_5113_cast_fp16")]; + tensor var_5114_to_fp16 = const()[name = tensor("op_5114_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_81_cast_fp16 = mul(x = var_5113_cast_fp16, y = var_5114_to_fp16)[name = tensor("scores_81_cast_fp16")]; + tensor scores_83_cast_fp16 = add(x = scores_81_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_83_cast_fp16")]; + tensor var_5124 = const()[name = tensor("op_5124"), val = tensor(-1)]; + tensor weights_41_cast_fp16 = softmax(axis = var_5124, x = scores_83_cast_fp16)[name = tensor("weights_41_cast_fp16")]; + tensor attn_41_transpose_x_0 = const()[name = tensor("attn_41_transpose_x_0"), val = tensor(false)]; + tensor attn_41_transpose_y_0 = const()[name = tensor("attn_41_transpose_y_0"), val = tensor(false)]; + tensor values_41_cast_fp16 = transpose(perm = var_5102, x = new_v_cache_41_cast_fp16)[name = tensor("transpose_126")]; + tensor attn_41_cast_fp16 = matmul(transpose_x = attn_41_transpose_x_0, transpose_y = attn_41_transpose_y_0, x = weights_41_cast_fp16, y = values_41_cast_fp16)[name = tensor("attn_41_cast_fp16")]; + tensor var_5132 = const()[name = tensor("op_5132"), val = tensor([0, 2, 1, 3])]; + tensor var_5137 = const()[name = tensor("op_5137"), val = tensor([1, 256, 1024])]; + tensor var_5133_cast_fp16 = transpose(perm = var_5132, x = attn_41_cast_fp16)[name = tensor("transpose_123")]; + tensor input_245_cast_fp16 = reshape(shape = var_5137, x = var_5133_cast_fp16)[name = tensor("input_245_cast_fp16")]; + tensor attn20_out_proj_weight_to_fp16 = const()[name = tensor("attn20_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(509805312)))]; + tensor linear_81_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn20_out_proj_weight_to_fp16, x = input_245_cast_fp16)[name = tensor("linear_81_cast_fp16")]; + tensor position20_to_fp16_dtype_0 = const()[name = tensor("position20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position20_to_fp16 = cast(dtype = position20_to_fp16_dtype_0, x = position20)[name = tensor("cast_341")]; + tensor var_5146_cast_fp16 = add(x = position20_to_fp16, y = valid_len_to_fp16)[name = tensor("op_5146_cast_fp16")]; + tensor var_5146_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5146_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_247_cast_fp16 = add(x = input_241_cast_fp16, y = linear_81_cast_fp16)[name = tensor("input_247_cast_fp16")]; + tensor input_249_axes_0 = const()[name = tensor("input_249_axes_0"), val = tensor([-1])]; + tensor norm20_2_weight_to_fp16 = const()[name = tensor("norm20_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(511902528)))]; + tensor norm20_2_bias_to_fp16 = const()[name = tensor("norm20_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(511904640)))]; + tensor var_5150_to_fp16 = const()[name = tensor("op_5150_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_249_cast_fp16 = layer_norm(axes = input_249_axes_0, beta = norm20_2_bias_to_fp16, epsilon = var_5150_to_fp16, gamma = norm20_2_weight_to_fp16, x = input_247_cast_fp16)[name = tensor("input_249_cast_fp16")]; + tensor linear20_1_weight_to_fp16 = const()[name = tensor("linear20_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(511906752)))]; + tensor linear_82_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear20_1_weight_to_fp16, x = input_249_cast_fp16)[name = tensor("linear_82_cast_fp16")]; + tensor input_251_mode_0 = const()[name = tensor("input_251_mode_0"), val = tensor("EXACT")]; + tensor input_251_cast_fp16 = gelu(mode = input_251_mode_0, x = linear_82_cast_fp16)[name = tensor("input_251_cast_fp16")]; + tensor linear20_2_weight_to_fp16 = const()[name = tensor("linear20_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(520295424)))]; + tensor linear_83_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear20_2_weight_to_fp16, x = input_251_cast_fp16)[name = tensor("linear_83_cast_fp16")]; + tensor input_253_cast_fp16 = add(x = input_247_cast_fp16, y = linear_83_cast_fp16)[name = tensor("input_253_cast_fp16")]; + tensor input_255_axes_0 = const()[name = tensor("input_255_axes_0"), val = tensor([-1])]; + tensor norm21_1_weight_to_fp16 = const()[name = tensor("norm21_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(528684096)))]; + tensor norm21_1_bias_to_fp16 = const()[name = tensor("norm21_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(528686208)))]; + tensor var_5167_to_fp16 = const()[name = tensor("op_5167_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_255_cast_fp16 = layer_norm(axes = input_255_axes_0, beta = norm21_1_bias_to_fp16, epsilon = var_5167_to_fp16, gamma = norm21_1_weight_to_fp16, x = input_253_cast_fp16)[name = tensor("input_255_cast_fp16")]; + tensor attn21_in_proj_weight_to_fp16 = const()[name = tensor("attn21_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(528688320)))]; + tensor linear_84_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn21_in_proj_weight_to_fp16, x = input_255_cast_fp16)[name = tensor("linear_84_cast_fp16")]; + tensor var_5181 = const()[name = tensor("op_5181"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_43_cast_fp16 = reshape(shape = var_5181, x = linear_84_cast_fp16)[name = tensor("qkv_43_cast_fp16")]; + tensor x_85_begin_0 = const()[name = tensor("x_85_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_85_end_0 = const()[name = tensor("x_85_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_85_end_mask_0 = const()[name = tensor("x_85_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_85_squeeze_mask_0 = const()[name = tensor("x_85_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_85_cast_fp16 = slice_by_index(begin = x_85_begin_0, end = x_85_end_0, end_mask = x_85_end_mask_0, squeeze_mask = x_85_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("x_85_cast_fp16")]; + tensor x_87_begin_0 = const()[name = tensor("x_87_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_87_end_0 = const()[name = tensor("x_87_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_87_end_mask_0 = const()[name = tensor("x_87_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_87_squeeze_mask_0 = const()[name = tensor("x_87_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_87_cast_fp16 = slice_by_index(begin = x_87_begin_0, end = x_87_end_0, end_mask = x_87_end_mask_0, squeeze_mask = x_87_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("x_87_cast_fp16")]; + tensor v_43_begin_0 = const()[name = tensor("v_43_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_43_end_0 = const()[name = tensor("v_43_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_43_end_mask_0 = const()[name = tensor("v_43_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_43_squeeze_mask_0 = const()[name = tensor("v_43_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_43_cast_fp16 = slice_by_index(begin = v_43_begin_0, end = v_43_end_0, end_mask = v_43_end_mask_0, squeeze_mask = v_43_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("v_43_cast_fp16")]; + tensor var_5226 = const()[name = tensor("op_5226"), val = tensor([256, 16, 32, 2])]; + tensor pairs_85_cast_fp16 = reshape(shape = var_5226, x = x_85_cast_fp16)[name = tensor("pairs_85_cast_fp16")]; + tensor xr_85_begin_0 = const()[name = tensor("xr_85_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_85_end_0 = const()[name = tensor("xr_85_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_85_end_mask_0 = const()[name = tensor("xr_85_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_85_squeeze_mask_0 = const()[name = tensor("xr_85_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_85_cast_fp16 = slice_by_index(begin = xr_85_begin_0, end = xr_85_end_0, end_mask = xr_85_end_mask_0, squeeze_mask = xr_85_squeeze_mask_0, x = pairs_85_cast_fp16)[name = tensor("xr_85_cast_fp16")]; + tensor xi_85_begin_0 = const()[name = tensor("xi_85_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_85_end_0 = const()[name = tensor("xi_85_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_85_end_mask_0 = const()[name = tensor("xi_85_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_85_squeeze_mask_0 = const()[name = tensor("xi_85_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_85_cast_fp16 = slice_by_index(begin = xi_85_begin_0, end = xi_85_end_0, end_mask = xi_85_end_mask_0, squeeze_mask = xi_85_squeeze_mask_0, x = pairs_85_cast_fp16)[name = tensor("xi_85_cast_fp16")]; + tensor var_5234_cast_fp16 = mul(x = xr_85_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5234_cast_fp16")]; + tensor var_5235_cast_fp16 = mul(x = xi_85_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5235_cast_fp16")]; + tensor xor__85_cast_fp16 = sub(x = var_5234_cast_fp16, y = var_5235_cast_fp16)[name = tensor("xor__85_cast_fp16")]; + tensor var_5238_cast_fp16 = mul(x = xr_85_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5238_cast_fp16")]; + tensor var_5239_cast_fp16 = mul(x = xi_85_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5239_cast_fp16")]; + tensor xoi_85_cast_fp16 = add(x = var_5238_cast_fp16, y = var_5239_cast_fp16)[name = tensor("xoi_85_cast_fp16")]; + tensor out_85_axis_0 = const()[name = tensor("out_85_axis_0"), val = tensor(-1)]; + tensor out_85_cast_fp16 = stack(axis = out_85_axis_0, values = (xor__85_cast_fp16, xoi_85_cast_fp16))[name = tensor("out_85_cast_fp16")]; + tensor var_5249 = const()[name = tensor("op_5249"), val = tensor([1, 256, 16, 64])]; + tensor q_43_cast_fp16 = reshape(shape = var_5249, x = out_85_cast_fp16)[name = tensor("q_43_cast_fp16")]; + tensor var_5255 = const()[name = tensor("op_5255"), val = tensor([256, 16, 32, 2])]; + tensor pairs_87_cast_fp16 = reshape(shape = var_5255, x = x_87_cast_fp16)[name = tensor("pairs_87_cast_fp16")]; + tensor xr_87_begin_0 = const()[name = tensor("xr_87_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_87_end_0 = const()[name = tensor("xr_87_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_87_end_mask_0 = const()[name = tensor("xr_87_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_87_squeeze_mask_0 = const()[name = tensor("xr_87_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_87_cast_fp16 = slice_by_index(begin = xr_87_begin_0, end = xr_87_end_0, end_mask = xr_87_end_mask_0, squeeze_mask = xr_87_squeeze_mask_0, x = pairs_87_cast_fp16)[name = tensor("xr_87_cast_fp16")]; + tensor xi_87_begin_0 = const()[name = tensor("xi_87_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_87_end_0 = const()[name = tensor("xi_87_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_87_end_mask_0 = const()[name = tensor("xi_87_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_87_squeeze_mask_0 = const()[name = tensor("xi_87_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_87_cast_fp16 = slice_by_index(begin = xi_87_begin_0, end = xi_87_end_0, end_mask = xi_87_end_mask_0, squeeze_mask = xi_87_squeeze_mask_0, x = pairs_87_cast_fp16)[name = tensor("xi_87_cast_fp16")]; + tensor var_5263_cast_fp16 = mul(x = xr_87_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5263_cast_fp16")]; + tensor var_5264_cast_fp16 = mul(x = xi_87_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5264_cast_fp16")]; + tensor xor__87_cast_fp16 = sub(x = var_5263_cast_fp16, y = var_5264_cast_fp16)[name = tensor("xor__87_cast_fp16")]; + tensor var_5267_cast_fp16 = mul(x = xr_87_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5267_cast_fp16")]; + tensor var_5268_cast_fp16 = mul(x = xi_87_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5268_cast_fp16")]; + tensor xoi_87_cast_fp16 = add(x = var_5267_cast_fp16, y = var_5268_cast_fp16)[name = tensor("xoi_87_cast_fp16")]; + tensor out_87_axis_0 = const()[name = tensor("out_87_axis_0"), val = tensor(-1)]; + tensor out_87_cast_fp16 = stack(axis = out_87_axis_0, values = (xor__87_cast_fp16, xoi_87_cast_fp16))[name = tensor("out_87_cast_fp16")]; + tensor var_5285 = const()[name = tensor("op_5285"), val = tensor([256, 1024])]; + tensor var_5286_cast_fp16 = reshape(shape = var_5285, x = out_87_cast_fp16)[name = tensor("op_5286_cast_fp16")]; + tensor written_k_43_transpose_x_1 = const()[name = tensor("written_k_43_transpose_x_1"), val = tensor(true)]; + tensor written_k_43_transpose_y_1 = const()[name = tensor("written_k_43_transpose_y_1"), val = tensor(false)]; + tensor written_k_43_cast_fp16 = matmul(transpose_x = written_k_43_transpose_x_1, transpose_y = written_k_43_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_5286_cast_fp16)[name = tensor("written_k_43_cast_fp16")]; + tensor var_5293 = const()[name = tensor("op_5293"), val = tensor([256, 1024])]; + tensor var_5294_cast_fp16 = reshape(shape = var_5293, x = v_43_cast_fp16)[name = tensor("op_5294_cast_fp16")]; + tensor written_v_43_transpose_x_1 = const()[name = tensor("written_v_43_transpose_x_1"), val = tensor(true)]; + tensor written_v_43_transpose_y_1 = const()[name = tensor("written_v_43_transpose_y_1"), val = tensor(false)]; + tensor written_v_43_cast_fp16 = matmul(transpose_x = written_v_43_transpose_x_1, transpose_y = written_v_43_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_5294_cast_fp16)[name = tensor("written_v_43_cast_fp16")]; + tensor k_cache21_to_fp16_dtype_0 = const()[name = tensor("k_cache21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache21_to_fp16 = cast(dtype = k_cache21_to_fp16_dtype_0, x = k_cache21)[name = tensor("cast_339")]; + tensor var_5305_cast_fp16 = mul(x = k_cache21_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_5305_cast_fp16")]; + tensor var_5310 = const()[name = tensor("op_5310"), val = tensor([1, 512, 16, 64])]; + tensor var_5311_cast_fp16 = reshape(shape = var_5310, x = written_k_43_cast_fp16)[name = tensor("op_5311_cast_fp16")]; + tensor new_k_cache_43_cast_fp16 = add(x = var_5305_cast_fp16, y = var_5311_cast_fp16)[name = tensor("new_k_cache_43_cast_fp16")]; + tensor new_k_cache_43_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_43_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache21_to_fp16_dtype_0 = const()[name = tensor("v_cache21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache21_to_fp16 = cast(dtype = v_cache21_to_fp16_dtype_0, x = v_cache21)[name = tensor("cast_337")]; + tensor var_5314_cast_fp16 = mul(x = v_cache21_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_5314_cast_fp16")]; + tensor var_5319 = const()[name = tensor("op_5319"), val = tensor([1, 512, 16, 64])]; + tensor var_5320_cast_fp16 = reshape(shape = var_5319, x = written_v_43_cast_fp16)[name = tensor("op_5320_cast_fp16")]; + tensor new_v_cache_43_cast_fp16 = add(x = var_5314_cast_fp16, y = var_5320_cast_fp16)[name = tensor("new_v_cache_43_cast_fp16")]; + tensor new_v_cache_43_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_43_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_5333 = const()[name = tensor("op_5333"), val = tensor([0, 2, 1, 3])]; + tensor var_5344_transpose_x_0 = const()[name = tensor("op_5344_transpose_x_0"), val = tensor(false)]; + tensor var_5344_transpose_y_0 = const()[name = tensor("op_5344_transpose_y_0"), val = tensor(false)]; + tensor transpose_111_perm_0 = const()[name = tensor("transpose_111_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_112_perm_0 = const()[name = tensor("transpose_112_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_112 = transpose(perm = transpose_112_perm_0, x = new_k_cache_43_cast_fp16)[name = tensor("transpose_120")]; + tensor transpose_111 = transpose(perm = transpose_111_perm_0, x = q_43_cast_fp16)[name = tensor("transpose_121")]; + tensor var_5344_cast_fp16 = matmul(transpose_x = var_5344_transpose_x_0, transpose_y = var_5344_transpose_y_0, x = transpose_111, y = transpose_112)[name = tensor("op_5344_cast_fp16")]; + tensor var_5345_to_fp16 = const()[name = tensor("op_5345_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_85_cast_fp16 = mul(x = var_5344_cast_fp16, y = var_5345_to_fp16)[name = tensor("scores_85_cast_fp16")]; + tensor scores_87_cast_fp16 = add(x = scores_85_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_87_cast_fp16")]; + tensor var_5355 = const()[name = tensor("op_5355"), val = tensor(-1)]; + tensor weights_43_cast_fp16 = softmax(axis = var_5355, x = scores_87_cast_fp16)[name = tensor("weights_43_cast_fp16")]; + tensor attn_43_transpose_x_0 = const()[name = tensor("attn_43_transpose_x_0"), val = tensor(false)]; + tensor attn_43_transpose_y_0 = const()[name = tensor("attn_43_transpose_y_0"), val = tensor(false)]; + tensor values_43_cast_fp16 = transpose(perm = var_5333, x = new_v_cache_43_cast_fp16)[name = tensor("transpose_122")]; + tensor attn_43_cast_fp16 = matmul(transpose_x = attn_43_transpose_x_0, transpose_y = attn_43_transpose_y_0, x = weights_43_cast_fp16, y = values_43_cast_fp16)[name = tensor("attn_43_cast_fp16")]; + tensor var_5363 = const()[name = tensor("op_5363"), val = tensor([0, 2, 1, 3])]; + tensor var_5368 = const()[name = tensor("op_5368"), val = tensor([1, 256, 1024])]; + tensor var_5364_cast_fp16 = transpose(perm = var_5363, x = attn_43_cast_fp16)[name = tensor("transpose_119")]; + tensor input_257_cast_fp16 = reshape(shape = var_5368, x = var_5364_cast_fp16)[name = tensor("input_257_cast_fp16")]; + tensor attn21_out_proj_weight_to_fp16 = const()[name = tensor("attn21_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(534979840)))]; + tensor linear_85_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn21_out_proj_weight_to_fp16, x = input_257_cast_fp16)[name = tensor("linear_85_cast_fp16")]; + tensor position21_to_fp16_dtype_0 = const()[name = tensor("position21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position21_to_fp16 = cast(dtype = position21_to_fp16_dtype_0, x = position21)[name = tensor("cast_335")]; + tensor var_5377_cast_fp16 = add(x = position21_to_fp16, y = valid_len_to_fp16)[name = tensor("op_5377_cast_fp16")]; + tensor var_5377_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5377_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_259_cast_fp16 = add(x = input_253_cast_fp16, y = linear_85_cast_fp16)[name = tensor("input_259_cast_fp16")]; + tensor input_261_axes_0 = const()[name = tensor("input_261_axes_0"), val = tensor([-1])]; + tensor norm21_2_weight_to_fp16 = const()[name = tensor("norm21_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(537077056)))]; + tensor norm21_2_bias_to_fp16 = const()[name = tensor("norm21_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(537079168)))]; + tensor var_5381_to_fp16 = const()[name = tensor("op_5381_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_261_cast_fp16 = layer_norm(axes = input_261_axes_0, beta = norm21_2_bias_to_fp16, epsilon = var_5381_to_fp16, gamma = norm21_2_weight_to_fp16, x = input_259_cast_fp16)[name = tensor("input_261_cast_fp16")]; + tensor linear21_1_weight_to_fp16 = const()[name = tensor("linear21_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(537081280)))]; + tensor linear_86_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear21_1_weight_to_fp16, x = input_261_cast_fp16)[name = tensor("linear_86_cast_fp16")]; + tensor input_263_mode_0 = const()[name = tensor("input_263_mode_0"), val = tensor("EXACT")]; + tensor input_263_cast_fp16 = gelu(mode = input_263_mode_0, x = linear_86_cast_fp16)[name = tensor("input_263_cast_fp16")]; + tensor linear21_2_weight_to_fp16 = const()[name = tensor("linear21_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(545469952)))]; + tensor linear_87_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear21_2_weight_to_fp16, x = input_263_cast_fp16)[name = tensor("linear_87_cast_fp16")]; + tensor input_265_cast_fp16 = add(x = input_259_cast_fp16, y = linear_87_cast_fp16)[name = tensor("input_265_cast_fp16")]; + tensor input_267_axes_0 = const()[name = tensor("input_267_axes_0"), val = tensor([-1])]; + tensor norm22_1_weight_to_fp16 = const()[name = tensor("norm22_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(553858624)))]; + tensor norm22_1_bias_to_fp16 = const()[name = tensor("norm22_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(553860736)))]; + tensor var_5398_to_fp16 = const()[name = tensor("op_5398_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_267_cast_fp16 = layer_norm(axes = input_267_axes_0, beta = norm22_1_bias_to_fp16, epsilon = var_5398_to_fp16, gamma = norm22_1_weight_to_fp16, x = input_265_cast_fp16)[name = tensor("input_267_cast_fp16")]; + tensor attn22_in_proj_weight_to_fp16 = const()[name = tensor("attn22_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(553862848)))]; + tensor linear_88_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn22_in_proj_weight_to_fp16, x = input_267_cast_fp16)[name = tensor("linear_88_cast_fp16")]; + tensor var_5412 = const()[name = tensor("op_5412"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_45_cast_fp16 = reshape(shape = var_5412, x = linear_88_cast_fp16)[name = tensor("qkv_45_cast_fp16")]; + tensor x_89_begin_0 = const()[name = tensor("x_89_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_89_end_0 = const()[name = tensor("x_89_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_89_end_mask_0 = const()[name = tensor("x_89_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_89_squeeze_mask_0 = const()[name = tensor("x_89_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_89_cast_fp16 = slice_by_index(begin = x_89_begin_0, end = x_89_end_0, end_mask = x_89_end_mask_0, squeeze_mask = x_89_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("x_89_cast_fp16")]; + tensor x_91_begin_0 = const()[name = tensor("x_91_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_91_end_0 = const()[name = tensor("x_91_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_91_end_mask_0 = const()[name = tensor("x_91_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_91_squeeze_mask_0 = const()[name = tensor("x_91_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_91_cast_fp16 = slice_by_index(begin = x_91_begin_0, end = x_91_end_0, end_mask = x_91_end_mask_0, squeeze_mask = x_91_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("x_91_cast_fp16")]; + tensor v_45_begin_0 = const()[name = tensor("v_45_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_45_end_0 = const()[name = tensor("v_45_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_45_end_mask_0 = const()[name = tensor("v_45_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_45_squeeze_mask_0 = const()[name = tensor("v_45_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_45_cast_fp16 = slice_by_index(begin = v_45_begin_0, end = v_45_end_0, end_mask = v_45_end_mask_0, squeeze_mask = v_45_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("v_45_cast_fp16")]; + tensor var_5457 = const()[name = tensor("op_5457"), val = tensor([256, 16, 32, 2])]; + tensor pairs_89_cast_fp16 = reshape(shape = var_5457, x = x_89_cast_fp16)[name = tensor("pairs_89_cast_fp16")]; + tensor xr_89_begin_0 = const()[name = tensor("xr_89_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_89_end_0 = const()[name = tensor("xr_89_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_89_end_mask_0 = const()[name = tensor("xr_89_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_89_squeeze_mask_0 = const()[name = tensor("xr_89_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_89_cast_fp16 = slice_by_index(begin = xr_89_begin_0, end = xr_89_end_0, end_mask = xr_89_end_mask_0, squeeze_mask = xr_89_squeeze_mask_0, x = pairs_89_cast_fp16)[name = tensor("xr_89_cast_fp16")]; + tensor xi_89_begin_0 = const()[name = tensor("xi_89_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_89_end_0 = const()[name = tensor("xi_89_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_89_end_mask_0 = const()[name = tensor("xi_89_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_89_squeeze_mask_0 = const()[name = tensor("xi_89_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_89_cast_fp16 = slice_by_index(begin = xi_89_begin_0, end = xi_89_end_0, end_mask = xi_89_end_mask_0, squeeze_mask = xi_89_squeeze_mask_0, x = pairs_89_cast_fp16)[name = tensor("xi_89_cast_fp16")]; + tensor var_5465_cast_fp16 = mul(x = xr_89_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5465_cast_fp16")]; + tensor var_5466_cast_fp16 = mul(x = xi_89_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5466_cast_fp16")]; + tensor xor__89_cast_fp16 = sub(x = var_5465_cast_fp16, y = var_5466_cast_fp16)[name = tensor("xor__89_cast_fp16")]; + tensor var_5469_cast_fp16 = mul(x = xr_89_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5469_cast_fp16")]; + tensor var_5470_cast_fp16 = mul(x = xi_89_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5470_cast_fp16")]; + tensor xoi_89_cast_fp16 = add(x = var_5469_cast_fp16, y = var_5470_cast_fp16)[name = tensor("xoi_89_cast_fp16")]; + tensor out_89_axis_0 = const()[name = tensor("out_89_axis_0"), val = tensor(-1)]; + tensor out_89_cast_fp16 = stack(axis = out_89_axis_0, values = (xor__89_cast_fp16, xoi_89_cast_fp16))[name = tensor("out_89_cast_fp16")]; + tensor var_5480 = const()[name = tensor("op_5480"), val = tensor([1, 256, 16, 64])]; + tensor q_45_cast_fp16 = reshape(shape = var_5480, x = out_89_cast_fp16)[name = tensor("q_45_cast_fp16")]; + tensor var_5486 = const()[name = tensor("op_5486"), val = tensor([256, 16, 32, 2])]; + tensor pairs_91_cast_fp16 = reshape(shape = var_5486, x = x_91_cast_fp16)[name = tensor("pairs_91_cast_fp16")]; + tensor xr_91_begin_0 = const()[name = tensor("xr_91_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_91_end_0 = const()[name = tensor("xr_91_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_91_end_mask_0 = const()[name = tensor("xr_91_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_91_squeeze_mask_0 = const()[name = tensor("xr_91_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_91_cast_fp16 = slice_by_index(begin = xr_91_begin_0, end = xr_91_end_0, end_mask = xr_91_end_mask_0, squeeze_mask = xr_91_squeeze_mask_0, x = pairs_91_cast_fp16)[name = tensor("xr_91_cast_fp16")]; + tensor xi_91_begin_0 = const()[name = tensor("xi_91_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_91_end_0 = const()[name = tensor("xi_91_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_91_end_mask_0 = const()[name = tensor("xi_91_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_91_squeeze_mask_0 = const()[name = tensor("xi_91_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_91_cast_fp16 = slice_by_index(begin = xi_91_begin_0, end = xi_91_end_0, end_mask = xi_91_end_mask_0, squeeze_mask = xi_91_squeeze_mask_0, x = pairs_91_cast_fp16)[name = tensor("xi_91_cast_fp16")]; + tensor var_5494_cast_fp16 = mul(x = xr_91_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5494_cast_fp16")]; + tensor var_5495_cast_fp16 = mul(x = xi_91_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5495_cast_fp16")]; + tensor xor__91_cast_fp16 = sub(x = var_5494_cast_fp16, y = var_5495_cast_fp16)[name = tensor("xor__91_cast_fp16")]; + tensor var_5498_cast_fp16 = mul(x = xr_91_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5498_cast_fp16")]; + tensor var_5499_cast_fp16 = mul(x = xi_91_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5499_cast_fp16")]; + tensor xoi_91_cast_fp16 = add(x = var_5498_cast_fp16, y = var_5499_cast_fp16)[name = tensor("xoi_91_cast_fp16")]; + tensor out_91_axis_0 = const()[name = tensor("out_91_axis_0"), val = tensor(-1)]; + tensor out_91_cast_fp16 = stack(axis = out_91_axis_0, values = (xor__91_cast_fp16, xoi_91_cast_fp16))[name = tensor("out_91_cast_fp16")]; + tensor var_5516 = const()[name = tensor("op_5516"), val = tensor([256, 1024])]; + tensor var_5517_cast_fp16 = reshape(shape = var_5516, x = out_91_cast_fp16)[name = tensor("op_5517_cast_fp16")]; + tensor written_k_45_transpose_x_1 = const()[name = tensor("written_k_45_transpose_x_1"), val = tensor(true)]; + tensor written_k_45_transpose_y_1 = const()[name = tensor("written_k_45_transpose_y_1"), val = tensor(false)]; + tensor written_k_45_cast_fp16 = matmul(transpose_x = written_k_45_transpose_x_1, transpose_y = written_k_45_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_5517_cast_fp16)[name = tensor("written_k_45_cast_fp16")]; + tensor var_5524 = const()[name = tensor("op_5524"), val = tensor([256, 1024])]; + tensor var_5525_cast_fp16 = reshape(shape = var_5524, x = v_45_cast_fp16)[name = tensor("op_5525_cast_fp16")]; + tensor written_v_45_transpose_x_1 = const()[name = tensor("written_v_45_transpose_x_1"), val = tensor(true)]; + tensor written_v_45_transpose_y_1 = const()[name = tensor("written_v_45_transpose_y_1"), val = tensor(false)]; + tensor written_v_45_cast_fp16 = matmul(transpose_x = written_v_45_transpose_x_1, transpose_y = written_v_45_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_5525_cast_fp16)[name = tensor("written_v_45_cast_fp16")]; + tensor k_cache22_to_fp16_dtype_0 = const()[name = tensor("k_cache22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache22_to_fp16 = cast(dtype = k_cache22_to_fp16_dtype_0, x = k_cache22)[name = tensor("cast_333")]; + tensor var_5536_cast_fp16 = mul(x = k_cache22_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_5536_cast_fp16")]; + tensor var_5541 = const()[name = tensor("op_5541"), val = tensor([1, 512, 16, 64])]; + tensor var_5542_cast_fp16 = reshape(shape = var_5541, x = written_k_45_cast_fp16)[name = tensor("op_5542_cast_fp16")]; + tensor new_k_cache_45_cast_fp16 = add(x = var_5536_cast_fp16, y = var_5542_cast_fp16)[name = tensor("new_k_cache_45_cast_fp16")]; + tensor new_k_cache_45_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_45_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache22_to_fp16_dtype_0 = const()[name = tensor("v_cache22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache22_to_fp16 = cast(dtype = v_cache22_to_fp16_dtype_0, x = v_cache22)[name = tensor("cast_331")]; + tensor var_5545_cast_fp16 = mul(x = v_cache22_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_5545_cast_fp16")]; + tensor var_5550 = const()[name = tensor("op_5550"), val = tensor([1, 512, 16, 64])]; + tensor var_5551_cast_fp16 = reshape(shape = var_5550, x = written_v_45_cast_fp16)[name = tensor("op_5551_cast_fp16")]; + tensor new_v_cache_45_cast_fp16 = add(x = var_5545_cast_fp16, y = var_5551_cast_fp16)[name = tensor("new_v_cache_45_cast_fp16")]; + tensor new_v_cache_45_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_45_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_5564 = const()[name = tensor("op_5564"), val = tensor([0, 2, 1, 3])]; + tensor var_5575_transpose_x_0 = const()[name = tensor("op_5575_transpose_x_0"), val = tensor(false)]; + tensor var_5575_transpose_y_0 = const()[name = tensor("op_5575_transpose_y_0"), val = tensor(false)]; + tensor transpose_113_perm_0 = const()[name = tensor("transpose_113_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_114_perm_0 = const()[name = tensor("transpose_114_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_114 = transpose(perm = transpose_114_perm_0, x = new_k_cache_45_cast_fp16)[name = tensor("transpose_116")]; + tensor transpose_113 = transpose(perm = transpose_113_perm_0, x = q_45_cast_fp16)[name = tensor("transpose_117")]; + tensor var_5575_cast_fp16 = matmul(transpose_x = var_5575_transpose_x_0, transpose_y = var_5575_transpose_y_0, x = transpose_113, y = transpose_114)[name = tensor("op_5575_cast_fp16")]; + tensor var_5576_to_fp16 = const()[name = tensor("op_5576_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_89_cast_fp16 = mul(x = var_5575_cast_fp16, y = var_5576_to_fp16)[name = tensor("scores_89_cast_fp16")]; + tensor scores_91_cast_fp16 = add(x = scores_89_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_91_cast_fp16")]; + tensor var_5586 = const()[name = tensor("op_5586"), val = tensor(-1)]; + tensor weights_45_cast_fp16 = softmax(axis = var_5586, x = scores_91_cast_fp16)[name = tensor("weights_45_cast_fp16")]; + tensor attn_45_transpose_x_0 = const()[name = tensor("attn_45_transpose_x_0"), val = tensor(false)]; + tensor attn_45_transpose_y_0 = const()[name = tensor("attn_45_transpose_y_0"), val = tensor(false)]; + tensor values_45_cast_fp16 = transpose(perm = var_5564, x = new_v_cache_45_cast_fp16)[name = tensor("transpose_118")]; + tensor attn_45_cast_fp16 = matmul(transpose_x = attn_45_transpose_x_0, transpose_y = attn_45_transpose_y_0, x = weights_45_cast_fp16, y = values_45_cast_fp16)[name = tensor("attn_45_cast_fp16")]; + tensor var_5594 = const()[name = tensor("op_5594"), val = tensor([0, 2, 1, 3])]; + tensor var_5599 = const()[name = tensor("op_5599"), val = tensor([1, 256, 1024])]; + tensor var_5595_cast_fp16 = transpose(perm = var_5594, x = attn_45_cast_fp16)[name = tensor("transpose_115")]; + tensor input_269_cast_fp16 = reshape(shape = var_5599, x = var_5595_cast_fp16)[name = tensor("input_269_cast_fp16")]; + tensor attn22_out_proj_weight_to_fp16 = const()[name = tensor("attn22_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(560154368)))]; + tensor linear_89_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn22_out_proj_weight_to_fp16, x = input_269_cast_fp16)[name = tensor("linear_89_cast_fp16")]; + tensor position22_to_fp16_dtype_0 = const()[name = tensor("position22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position22_to_fp16 = cast(dtype = position22_to_fp16_dtype_0, x = position22)[name = tensor("cast_329")]; + tensor var_5608_cast_fp16 = add(x = position22_to_fp16, y = valid_len_to_fp16)[name = tensor("op_5608_cast_fp16")]; + tensor var_5608_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5608_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_271_cast_fp16 = add(x = input_265_cast_fp16, y = linear_89_cast_fp16)[name = tensor("input_271_cast_fp16")]; + tensor input_273_axes_0 = const()[name = tensor("input_273_axes_0"), val = tensor([-1])]; + tensor norm22_2_weight_to_fp16 = const()[name = tensor("norm22_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(562251584)))]; + tensor norm22_2_bias_to_fp16 = const()[name = tensor("norm22_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(562253696)))]; + tensor var_5612_to_fp16 = const()[name = tensor("op_5612_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_273_cast_fp16 = layer_norm(axes = input_273_axes_0, beta = norm22_2_bias_to_fp16, epsilon = var_5612_to_fp16, gamma = norm22_2_weight_to_fp16, x = input_271_cast_fp16)[name = tensor("input_273_cast_fp16")]; + tensor linear22_1_weight_to_fp16 = const()[name = tensor("linear22_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(562255808)))]; + tensor linear_90_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear22_1_weight_to_fp16, x = input_273_cast_fp16)[name = tensor("linear_90_cast_fp16")]; + tensor input_275_mode_0 = const()[name = tensor("input_275_mode_0"), val = tensor("EXACT")]; + tensor input_275_cast_fp16 = gelu(mode = input_275_mode_0, x = linear_90_cast_fp16)[name = tensor("input_275_cast_fp16")]; + tensor linear22_2_weight_to_fp16 = const()[name = tensor("linear22_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(570644480)))]; + tensor linear_91_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear22_2_weight_to_fp16, x = input_275_cast_fp16)[name = tensor("linear_91_cast_fp16")]; + tensor input_277_cast_fp16 = add(x = input_271_cast_fp16, y = linear_91_cast_fp16)[name = tensor("input_277_cast_fp16")]; + tensor input_279_axes_0 = const()[name = tensor("input_279_axes_0"), val = tensor([-1])]; + tensor norm23_1_weight_to_fp16 = const()[name = tensor("norm23_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(579033152)))]; + tensor norm23_1_bias_to_fp16 = const()[name = tensor("norm23_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(579035264)))]; + tensor var_5629_to_fp16 = const()[name = tensor("op_5629_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_279_cast_fp16 = layer_norm(axes = input_279_axes_0, beta = norm23_1_bias_to_fp16, epsilon = var_5629_to_fp16, gamma = norm23_1_weight_to_fp16, x = input_277_cast_fp16)[name = tensor("input_279_cast_fp16")]; + tensor attn23_in_proj_weight_to_fp16 = const()[name = tensor("attn23_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(579037376)))]; + tensor linear_92_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn23_in_proj_weight_to_fp16, x = input_279_cast_fp16)[name = tensor("linear_92_cast_fp16")]; + tensor var_5643 = const()[name = tensor("op_5643"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_cast_fp16 = reshape(shape = var_5643, x = linear_92_cast_fp16)[name = tensor("qkv_cast_fp16")]; + tensor x_begin_0 = const()[name = tensor("x_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_end_0 = const()[name = tensor("x_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_end_mask_0 = const()[name = tensor("x_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_squeeze_mask_0 = const()[name = tensor("x_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_cast_fp16 = slice_by_index(begin = x_begin_0, end = x_end_0, end_mask = x_end_mask_0, squeeze_mask = x_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("x_cast_fp16")]; + tensor v_begin_0 = const()[name = tensor("v_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_end_0 = const()[name = tensor("v_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_end_mask_0 = const()[name = tensor("v_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_squeeze_mask_0 = const()[name = tensor("v_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_cast_fp16 = slice_by_index(begin = v_begin_0, end = v_end_0, end_mask = v_end_mask_0, squeeze_mask = v_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("v_cast_fp16")]; + tensor var_5675 = const()[name = tensor("op_5675"), val = tensor([256, 16, 32, 2])]; + tensor pairs_cast_fp16 = reshape(shape = var_5675, x = x_cast_fp16)[name = tensor("pairs_cast_fp16")]; + tensor xr_begin_0 = const()[name = tensor("xr_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_end_0 = const()[name = tensor("xr_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_end_mask_0 = const()[name = tensor("xr_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_squeeze_mask_0 = const()[name = tensor("xr_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_cast_fp16 = slice_by_index(begin = xr_begin_0, end = xr_end_0, end_mask = xr_end_mask_0, squeeze_mask = xr_squeeze_mask_0, x = pairs_cast_fp16)[name = tensor("xr_cast_fp16")]; + tensor xi_begin_0 = const()[name = tensor("xi_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_end_0 = const()[name = tensor("xi_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_end_mask_0 = const()[name = tensor("xi_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_squeeze_mask_0 = const()[name = tensor("xi_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_cast_fp16 = slice_by_index(begin = xi_begin_0, end = xi_end_0, end_mask = xi_end_mask_0, squeeze_mask = xi_squeeze_mask_0, x = pairs_cast_fp16)[name = tensor("xi_cast_fp16")]; + tensor var_5683_cast_fp16 = mul(x = xr_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5683_cast_fp16")]; + tensor var_5684_cast_fp16 = mul(x = xi_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5684_cast_fp16")]; + tensor xor__cast_fp16 = sub(x = var_5683_cast_fp16, y = var_5684_cast_fp16)[name = tensor("xor__cast_fp16")]; + tensor var_5687_cast_fp16 = mul(x = xr_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5687_cast_fp16")]; + tensor var_5688_cast_fp16 = mul(x = xi_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5688_cast_fp16")]; + tensor xoi_cast_fp16 = add(x = var_5687_cast_fp16, y = var_5688_cast_fp16)[name = tensor("xoi_cast_fp16")]; + tensor out_axis_0 = const()[name = tensor("out_axis_0"), val = tensor(-1)]; + tensor out_cast_fp16 = stack(axis = out_axis_0, values = (xor__cast_fp16, xoi_cast_fp16))[name = tensor("out_cast_fp16")]; + tensor var_5705 = const()[name = tensor("op_5705"), val = tensor([256, 1024])]; + tensor var_5706_cast_fp16 = reshape(shape = var_5705, x = out_cast_fp16)[name = tensor("op_5706_cast_fp16")]; + tensor written_k_transpose_x_1 = const()[name = tensor("written_k_transpose_x_1"), val = tensor(true)]; + tensor written_k_transpose_y_1 = const()[name = tensor("written_k_transpose_y_1"), val = tensor(false)]; + tensor written_k_cast_fp16 = matmul(transpose_x = written_k_transpose_x_1, transpose_y = written_k_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_5706_cast_fp16)[name = tensor("written_k_cast_fp16")]; + tensor var_5713 = const()[name = tensor("op_5713"), val = tensor([256, 1024])]; + tensor var_5714_cast_fp16 = reshape(shape = var_5713, x = v_cast_fp16)[name = tensor("op_5714_cast_fp16")]; + tensor written_v_transpose_x_1 = const()[name = tensor("written_v_transpose_x_1"), val = tensor(true)]; + tensor written_v_transpose_y_1 = const()[name = tensor("written_v_transpose_y_1"), val = tensor(false)]; + tensor written_v_cast_fp16 = matmul(transpose_x = written_v_transpose_x_1, transpose_y = written_v_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_5714_cast_fp16)[name = tensor("written_v_cast_fp16")]; + tensor k_cache23_to_fp16_dtype_0 = const()[name = tensor("k_cache23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache23_to_fp16 = cast(dtype = k_cache23_to_fp16_dtype_0, x = k_cache23)[name = tensor("cast_327")]; + tensor var_5725_cast_fp16 = mul(x = k_cache23_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_5725_cast_fp16")]; + tensor var_5730 = const()[name = tensor("op_5730"), val = tensor([1, 512, 16, 64])]; + tensor var_5731_cast_fp16 = reshape(shape = var_5730, x = written_k_cast_fp16)[name = tensor("op_5731_cast_fp16")]; + tensor new_k_cache_cast_fp16 = add(x = var_5725_cast_fp16, y = var_5731_cast_fp16)[name = tensor("new_k_cache_cast_fp16")]; + tensor new_k_cache_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache23_to_fp16_dtype_0 = const()[name = tensor("v_cache23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache23_to_fp16 = cast(dtype = v_cache23_to_fp16_dtype_0, x = v_cache23)[name = tensor("cast_325")]; + tensor var_5734_cast_fp16 = mul(x = v_cache23_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_5734_cast_fp16")]; + tensor var_5739 = const()[name = tensor("op_5739"), val = tensor([1, 512, 16, 64])]; + tensor var_5740_cast_fp16 = reshape(shape = var_5739, x = written_v_cast_fp16)[name = tensor("op_5740_cast_fp16")]; + tensor new_v_cache_cast_fp16 = add(x = var_5734_cast_fp16, y = var_5740_cast_fp16)[name = tensor("new_v_cache_cast_fp16")]; + tensor new_v_cache_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor position23_to_fp16_dtype_0 = const()[name = tensor("position23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position23_to_fp16 = cast(dtype = position23_to_fp16_dtype_0, x = position23)[name = tensor("cast_323")]; + tensor var_5747_cast_fp16 = add(x = position23_to_fp16, y = valid_len_to_fp16)[name = tensor("op_5747_cast_fp16")]; + tensor var_5747_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5747_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor new_position23 = cast(dtype = var_5747_cast_fp16_to_fp32_dtype_0, x = var_5747_cast_fp16)[name = tensor("cast_322")]; + tensor new_v_cache23 = cast(dtype = new_v_cache_cast_fp16_to_fp32_dtype_0, x = new_v_cache_cast_fp16)[name = tensor("cast_324")]; + tensor new_k_cache23 = cast(dtype = new_k_cache_cast_fp16_to_fp32_dtype_0, x = new_k_cache_cast_fp16)[name = tensor("cast_326")]; + tensor new_position22 = cast(dtype = var_5608_cast_fp16_to_fp32_dtype_0, x = var_5608_cast_fp16)[name = tensor("cast_328")]; + tensor new_v_cache22 = cast(dtype = new_v_cache_45_cast_fp16_to_fp32_dtype_0, x = new_v_cache_45_cast_fp16)[name = tensor("cast_330")]; + tensor new_k_cache22 = cast(dtype = new_k_cache_45_cast_fp16_to_fp32_dtype_0, x = new_k_cache_45_cast_fp16)[name = tensor("cast_332")]; + tensor new_position21 = cast(dtype = var_5377_cast_fp16_to_fp32_dtype_0, x = var_5377_cast_fp16)[name = tensor("cast_334")]; + tensor new_v_cache21 = cast(dtype = new_v_cache_43_cast_fp16_to_fp32_dtype_0, x = new_v_cache_43_cast_fp16)[name = tensor("cast_336")]; + tensor new_k_cache21 = cast(dtype = new_k_cache_43_cast_fp16_to_fp32_dtype_0, x = new_k_cache_43_cast_fp16)[name = tensor("cast_338")]; + tensor new_position20 = cast(dtype = var_5146_cast_fp16_to_fp32_dtype_0, x = var_5146_cast_fp16)[name = tensor("cast_340")]; + tensor new_v_cache20 = cast(dtype = new_v_cache_41_cast_fp16_to_fp32_dtype_0, x = new_v_cache_41_cast_fp16)[name = tensor("cast_342")]; + tensor new_k_cache20 = cast(dtype = new_k_cache_41_cast_fp16_to_fp32_dtype_0, x = new_k_cache_41_cast_fp16)[name = tensor("cast_344")]; + tensor new_position19 = cast(dtype = var_4915_cast_fp16_to_fp32_dtype_0, x = var_4915_cast_fp16)[name = tensor("cast_346")]; + tensor new_v_cache19 = cast(dtype = new_v_cache_39_cast_fp16_to_fp32_dtype_0, x = new_v_cache_39_cast_fp16)[name = tensor("cast_348")]; + tensor new_k_cache19 = cast(dtype = new_k_cache_39_cast_fp16_to_fp32_dtype_0, x = new_k_cache_39_cast_fp16)[name = tensor("cast_350")]; + tensor new_position18 = cast(dtype = var_4684_cast_fp16_to_fp32_dtype_0, x = var_4684_cast_fp16)[name = tensor("cast_352")]; + tensor new_v_cache18 = cast(dtype = new_v_cache_37_cast_fp16_to_fp32_dtype_0, x = new_v_cache_37_cast_fp16)[name = tensor("cast_354")]; + tensor new_k_cache18 = cast(dtype = new_k_cache_37_cast_fp16_to_fp32_dtype_0, x = new_k_cache_37_cast_fp16)[name = tensor("cast_356")]; + tensor new_position17 = cast(dtype = var_4453_cast_fp16_to_fp32_dtype_0, x = var_4453_cast_fp16)[name = tensor("cast_358")]; + tensor new_v_cache17 = cast(dtype = new_v_cache_35_cast_fp16_to_fp32_dtype_0, x = new_v_cache_35_cast_fp16)[name = tensor("cast_360")]; + tensor new_k_cache17 = cast(dtype = new_k_cache_35_cast_fp16_to_fp32_dtype_0, x = new_k_cache_35_cast_fp16)[name = tensor("cast_362")]; + tensor new_position16 = cast(dtype = var_4222_cast_fp16_to_fp32_dtype_0, x = var_4222_cast_fp16)[name = tensor("cast_364")]; + tensor new_v_cache16 = cast(dtype = new_v_cache_33_cast_fp16_to_fp32_dtype_0, x = new_v_cache_33_cast_fp16)[name = tensor("cast_366")]; + tensor new_k_cache16 = cast(dtype = new_k_cache_33_cast_fp16_to_fp32_dtype_0, x = new_k_cache_33_cast_fp16)[name = tensor("cast_368")]; + tensor new_position15 = cast(dtype = var_3991_cast_fp16_to_fp32_dtype_0, x = var_3991_cast_fp16)[name = tensor("cast_370")]; + tensor new_v_cache15 = cast(dtype = new_v_cache_31_cast_fp16_to_fp32_dtype_0, x = new_v_cache_31_cast_fp16)[name = tensor("cast_372")]; + tensor new_k_cache15 = cast(dtype = new_k_cache_31_cast_fp16_to_fp32_dtype_0, x = new_k_cache_31_cast_fp16)[name = tensor("cast_374")]; + tensor new_position14 = cast(dtype = var_3760_cast_fp16_to_fp32_dtype_0, x = var_3760_cast_fp16)[name = tensor("cast_376")]; + tensor new_v_cache14 = cast(dtype = new_v_cache_29_cast_fp16_to_fp32_dtype_0, x = new_v_cache_29_cast_fp16)[name = tensor("cast_378")]; + tensor new_k_cache14 = cast(dtype = new_k_cache_29_cast_fp16_to_fp32_dtype_0, x = new_k_cache_29_cast_fp16)[name = tensor("cast_380")]; + tensor new_position13 = cast(dtype = var_3529_cast_fp16_to_fp32_dtype_0, x = var_3529_cast_fp16)[name = tensor("cast_382")]; + tensor new_v_cache13 = cast(dtype = new_v_cache_27_cast_fp16_to_fp32_dtype_0, x = new_v_cache_27_cast_fp16)[name = tensor("cast_384")]; + tensor new_k_cache13 = cast(dtype = new_k_cache_27_cast_fp16_to_fp32_dtype_0, x = new_k_cache_27_cast_fp16)[name = tensor("cast_386")]; + tensor new_position12 = cast(dtype = var_3298_cast_fp16_to_fp32_dtype_0, x = var_3298_cast_fp16)[name = tensor("cast_388")]; + tensor new_v_cache12 = cast(dtype = new_v_cache_25_cast_fp16_to_fp32_dtype_0, x = new_v_cache_25_cast_fp16)[name = tensor("cast_390")]; + tensor new_k_cache12 = cast(dtype = new_k_cache_25_cast_fp16_to_fp32_dtype_0, x = new_k_cache_25_cast_fp16)[name = tensor("cast_392")]; + tensor new_position11 = cast(dtype = var_3067_cast_fp16_to_fp32_dtype_0, x = var_3067_cast_fp16)[name = tensor("cast_394")]; + tensor new_v_cache11 = cast(dtype = new_v_cache_23_cast_fp16_to_fp32_dtype_0, x = new_v_cache_23_cast_fp16)[name = tensor("cast_396")]; + tensor new_k_cache11 = cast(dtype = new_k_cache_23_cast_fp16_to_fp32_dtype_0, x = new_k_cache_23_cast_fp16)[name = tensor("cast_398")]; + tensor new_position10 = cast(dtype = var_2836_cast_fp16_to_fp32_dtype_0, x = var_2836_cast_fp16)[name = tensor("cast_400")]; + tensor new_v_cache10 = cast(dtype = new_v_cache_21_cast_fp16_to_fp32_dtype_0, x = new_v_cache_21_cast_fp16)[name = tensor("cast_402")]; + tensor new_k_cache10 = cast(dtype = new_k_cache_21_cast_fp16_to_fp32_dtype_0, x = new_k_cache_21_cast_fp16)[name = tensor("cast_404")]; + tensor new_position9 = cast(dtype = var_2605_cast_fp16_to_fp32_dtype_0, x = var_2605_cast_fp16)[name = tensor("cast_406")]; + tensor new_v_cache9 = cast(dtype = new_v_cache_19_cast_fp16_to_fp32_dtype_0, x = new_v_cache_19_cast_fp16)[name = tensor("cast_408")]; + tensor new_k_cache9 = cast(dtype = new_k_cache_19_cast_fp16_to_fp32_dtype_0, x = new_k_cache_19_cast_fp16)[name = tensor("cast_410")]; + tensor new_position8 = cast(dtype = var_2374_cast_fp16_to_fp32_dtype_0, x = var_2374_cast_fp16)[name = tensor("cast_412")]; + tensor new_v_cache8 = cast(dtype = new_v_cache_17_cast_fp16_to_fp32_dtype_0, x = new_v_cache_17_cast_fp16)[name = tensor("cast_414")]; + tensor new_k_cache8 = cast(dtype = new_k_cache_17_cast_fp16_to_fp32_dtype_0, x = new_k_cache_17_cast_fp16)[name = tensor("cast_416")]; + tensor new_position7 = cast(dtype = var_2143_cast_fp16_to_fp32_dtype_0, x = var_2143_cast_fp16)[name = tensor("cast_418")]; + tensor new_v_cache7 = cast(dtype = new_v_cache_15_cast_fp16_to_fp32_dtype_0, x = new_v_cache_15_cast_fp16)[name = tensor("cast_420")]; + tensor new_k_cache7 = cast(dtype = new_k_cache_15_cast_fp16_to_fp32_dtype_0, x = new_k_cache_15_cast_fp16)[name = tensor("cast_422")]; + tensor new_position6 = cast(dtype = var_1912_cast_fp16_to_fp32_dtype_0, x = var_1912_cast_fp16)[name = tensor("cast_424")]; + tensor new_v_cache6 = cast(dtype = new_v_cache_13_cast_fp16_to_fp32_dtype_0, x = new_v_cache_13_cast_fp16)[name = tensor("cast_426")]; + tensor new_k_cache6 = cast(dtype = new_k_cache_13_cast_fp16_to_fp32_dtype_0, x = new_k_cache_13_cast_fp16)[name = tensor("cast_428")]; + tensor new_position5 = cast(dtype = var_1681_cast_fp16_to_fp32_dtype_0, x = var_1681_cast_fp16)[name = tensor("cast_430")]; + tensor new_v_cache5 = cast(dtype = new_v_cache_11_cast_fp16_to_fp32_dtype_0, x = new_v_cache_11_cast_fp16)[name = tensor("cast_432")]; + tensor new_k_cache5 = cast(dtype = new_k_cache_11_cast_fp16_to_fp32_dtype_0, x = new_k_cache_11_cast_fp16)[name = tensor("cast_434")]; + tensor new_position4 = cast(dtype = var_1450_cast_fp16_to_fp32_dtype_0, x = var_1450_cast_fp16)[name = tensor("cast_436")]; + tensor new_v_cache4 = cast(dtype = new_v_cache_9_cast_fp16_to_fp32_dtype_0, x = new_v_cache_9_cast_fp16)[name = tensor("cast_438")]; + tensor new_k_cache4 = cast(dtype = new_k_cache_9_cast_fp16_to_fp32_dtype_0, x = new_k_cache_9_cast_fp16)[name = tensor("cast_440")]; + tensor new_position3 = cast(dtype = var_1219_cast_fp16_to_fp32_dtype_0, x = var_1219_cast_fp16)[name = tensor("cast_442")]; + tensor new_v_cache3 = cast(dtype = new_v_cache_7_cast_fp16_to_fp32_dtype_0, x = new_v_cache_7_cast_fp16)[name = tensor("cast_444")]; + tensor new_k_cache3 = cast(dtype = new_k_cache_7_cast_fp16_to_fp32_dtype_0, x = new_k_cache_7_cast_fp16)[name = tensor("cast_446")]; + tensor new_position2 = cast(dtype = var_988_cast_fp16_to_fp32_dtype_0, x = var_988_cast_fp16)[name = tensor("cast_448")]; + tensor new_v_cache2 = cast(dtype = new_v_cache_5_cast_fp16_to_fp32_dtype_0, x = new_v_cache_5_cast_fp16)[name = tensor("cast_450")]; + tensor new_k_cache2 = cast(dtype = new_k_cache_5_cast_fp16_to_fp32_dtype_0, x = new_k_cache_5_cast_fp16)[name = tensor("cast_452")]; + tensor new_position1 = cast(dtype = var_757_cast_fp16_to_fp32_dtype_0, x = var_757_cast_fp16)[name = tensor("cast_454")]; + tensor new_v_cache1 = cast(dtype = new_v_cache_3_cast_fp16_to_fp32_dtype_0, x = new_v_cache_3_cast_fp16)[name = tensor("cast_456")]; + tensor new_k_cache1 = cast(dtype = new_k_cache_3_cast_fp16_to_fp32_dtype_0, x = new_k_cache_3_cast_fp16)[name = tensor("cast_458")]; + tensor new_position0 = cast(dtype = var_526_cast_fp16_to_fp32_dtype_0, x = var_297_cast_fp16)[name = tensor("cast_460")]; + tensor new_v_cache0 = cast(dtype = new_v_cache_1_cast_fp16_to_fp32_dtype_0, x = new_v_cache_1_cast_fp16)[name = tensor("cast_461")]; + tensor new_k_cache0 = cast(dtype = new_k_cache_1_cast_fp16_to_fp32_dtype_0, x = new_k_cache_1_cast_fp16)[name = tensor("cast_463")]; + } -> (new_k_cache0, new_v_cache0, new_position0, new_k_cache1, new_v_cache1, new_position1, new_k_cache2, new_v_cache2, new_position2, new_k_cache3, new_v_cache3, new_position3, new_k_cache4, new_v_cache4, new_position4, new_k_cache5, new_v_cache5, new_position5, new_k_cache6, new_v_cache6, new_position6, new_k_cache7, new_v_cache7, new_position7, new_k_cache8, new_v_cache8, new_position8, new_k_cache9, new_v_cache9, new_position9, new_k_cache10, new_v_cache10, new_position10, new_k_cache11, new_v_cache11, new_position11, new_k_cache12, new_v_cache12, new_position12, new_k_cache13, new_v_cache13, new_position13, new_k_cache14, new_v_cache14, new_position14, new_k_cache15, new_v_cache15, new_position15, new_k_cache16, new_v_cache16, new_position16, new_k_cache17, new_v_cache17, new_position17, new_k_cache18, new_v_cache18, new_position18, new_k_cache19, new_v_cache19, new_position19, new_k_cache20, new_v_cache20, new_position20, new_k_cache21, new_v_cache21, new_position21, new_k_cache22, new_v_cache22, new_position22, new_k_cache23, new_v_cache23, new_position23); +} \ No newline at end of file diff --git a/v2.1/italian_24l/cond_prefill_ane.mlmodelc/weights/weight.bin b/v2.1/italian_24l/cond_prefill_ane.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..81ccb31e83d1d2181b8d9630cd90e2f77c87c433 --- /dev/null +++ b/v2.1/italian_24l/cond_prefill_ane.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2246bbe2b996f11142924a8bd867377a2e0c0b82b35714328f2b0d787554e5fe +size 585328896 diff --git a/v2.1/italian_24l/cond_prefill_ane.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/italian_24l/cond_prefill_ane.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..0bb3f7dc02832114675f0bc3519a3fba867aed6c --- /dev/null +++ b/v2.1/italian_24l/cond_prefill_ane.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3a8f6ea5c4d029ce7e93ad60f410e71f0fd746e02f670edc22c122af1113d99c +size 483717 diff --git a/v2.1/italian_24l/cond_prefill_ane.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/italian_24l/cond_prefill_ane.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..81ccb31e83d1d2181b8d9630cd90e2f77c87c433 --- /dev/null +++ b/v2.1/italian_24l/cond_prefill_ane.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2246bbe2b996f11142924a8bd867377a2e0c0b82b35714328f2b0d787554e5fe +size 585328896 diff --git a/v2.1/italian_24l/cond_prefill_ane.mlpackage/Manifest.json b/v2.1/italian_24l/cond_prefill_ane.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..197ceff0d3d304584babe8189e9dc6ce2954d25a --- /dev/null +++ b/v2.1/italian_24l/cond_prefill_ane.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "0839920D-2B5E-4210-B2B4-75C2065BD225": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "FAE07E1F-F700-49A1-8EE5-03A49AE0EE7A": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "0839920D-2B5E-4210-B2B4-75C2065BD225" +} diff --git a/v2.1/italian_24l/flowlm_step_ane.mlmodelc/analytics/coremldata.bin b/v2.1/italian_24l/flowlm_step_ane.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..f809e768931ecac84e69c2594e69c1b6277caa91 --- /dev/null +++ b/v2.1/italian_24l/flowlm_step_ane.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6451a7741a03d3db9017c4a5a7f2941f6b39b64106f6477d6f603c9e5843e925 +size 243 diff --git a/v2.1/italian_24l/flowlm_step_ane.mlmodelc/coremldata.bin b/v2.1/italian_24l/flowlm_step_ane.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..4c7785416a3f9866573e90c999268c2735f495fd --- /dev/null +++ b/v2.1/italian_24l/flowlm_step_ane.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:31403bb12161234a915aca846ef4467c2cffd20cf2b4051b0e2f527d5b1cedb2 +size 4522 diff --git a/v2.1/italian_24l/flowlm_step_ane.mlmodelc/model.mil b/v2.1/italian_24l/flowlm_step_ane.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..aba003fece24c433c711eda0a30588b0c20f7b54 --- /dev/null +++ b/v2.1/italian_24l/flowlm_step_ane.mlmodelc/model.mil @@ -0,0 +1,3699 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.9.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor k_cache0, tensor k_cache1, tensor k_cache10, tensor k_cache11, tensor k_cache12, tensor k_cache13, tensor k_cache14, tensor k_cache15, tensor k_cache16, tensor k_cache17, tensor k_cache18, tensor k_cache19, tensor k_cache2, tensor k_cache20, tensor k_cache21, tensor k_cache22, tensor k_cache23, tensor k_cache3, tensor k_cache4, tensor k_cache5, tensor k_cache6, tensor k_cache7, tensor k_cache8, tensor k_cache9, tensor position0, tensor position1, tensor position10, tensor position11, tensor position12, tensor position13, tensor position14, tensor position15, tensor position16, tensor position17, tensor position18, tensor position19, tensor position2, tensor position20, tensor position21, tensor position22, tensor position23, tensor position3, tensor position4, tensor position5, tensor position6, tensor position7, tensor position8, tensor position9, tensor sequence, tensor v_cache0, tensor v_cache1, tensor v_cache10, tensor v_cache11, tensor v_cache12, tensor v_cache13, tensor v_cache14, tensor v_cache15, tensor v_cache16, tensor v_cache17, tensor v_cache18, tensor v_cache19, tensor v_cache2, tensor v_cache20, tensor v_cache21, tensor v_cache22, tensor v_cache23, tensor v_cache3, tensor v_cache4, tensor v_cache5, tensor v_cache6, tensor v_cache7, tensor v_cache8, tensor v_cache9) { + tensor sequence_to_fp16_dtype_0 = const()[name = tensor("sequence_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor input_linear_weight_to_fp16 = const()[name = tensor("input_linear_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(65664)))]; + tensor sequence_to_fp16 = cast(dtype = sequence_to_fp16_dtype_0, x = sequence)[name = tensor("cast_578")]; + tensor linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = input_linear_weight_to_fp16, x = sequence_to_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; + tensor norm0_1_weight_to_fp16 = const()[name = tensor("norm0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67776)))]; + tensor norm0_1_bias_to_fp16 = const()[name = tensor("norm0_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(69888)))]; + tensor var_227_to_fp16 = const()[name = tensor("op_227_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_5_cast_fp16 = layer_norm(axes = input_5_axes_0, beta = norm0_1_bias_to_fp16, epsilon = var_227_to_fp16, gamma = norm0_1_weight_to_fp16, x = linear_0_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor attn0_in_proj_weight_to_fp16 = const()[name = tensor("attn0_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(72000)))]; + tensor linear_1_bias_0_to_fp16 = const()[name = tensor("linear_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6363520)))]; + tensor linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn0_in_proj_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("linear_1_cast_fp16")]; + tensor var_240 = const()[name = tensor("op_240"), val = tensor([1, 3, 16, 64])]; + tensor qkv_1_cast_fp16 = reshape(shape = var_240, x = linear_1_cast_fp16)[name = tensor("qkv_1_cast_fp16")]; + tensor x_1_begin_0 = const()[name = tensor("x_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_1_end_0 = const()[name = tensor("x_1_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_1_end_mask_0 = const()[name = tensor("x_1_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_1_squeeze_mask_0 = const()[name = tensor("x_1_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_1_cast_fp16 = slice_by_index(begin = x_1_begin_0, end = x_1_end_0, end_mask = x_1_end_mask_0, squeeze_mask = x_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("x_1_cast_fp16")]; + tensor x_3_begin_0 = const()[name = tensor("x_3_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_3_end_0 = const()[name = tensor("x_3_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_3_end_mask_0 = const()[name = tensor("x_3_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_3_squeeze_mask_0 = const()[name = tensor("x_3_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_3_cast_fp16 = slice_by_index(begin = x_3_begin_0, end = x_3_end_0, end_mask = x_3_end_mask_0, squeeze_mask = x_3_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("x_3_cast_fp16")]; + tensor v_1_begin_0 = const()[name = tensor("v_1_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_1_end_0 = const()[name = tensor("v_1_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_1_end_mask_0 = const()[name = tensor("v_1_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_1_squeeze_mask_0 = const()[name = tensor("v_1_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_1_cast_fp16 = slice_by_index(begin = v_1_begin_0, end = v_1_end_0, end_mask = v_1_end_mask_0, squeeze_mask = v_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("v_1_cast_fp16")]; + tensor var_269 = const()[name = tensor("op_269"), val = tensor([1, 1, 1])]; + tensor position0_to_fp16_dtype_0 = const()[name = tensor("position0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position0_to_fp16 = cast(dtype = position0_to_fp16_dtype_0, x = position0)[name = tensor("cast_577")]; + tensor var_270_cast_fp16 = reshape(shape = var_269, x = position0_to_fp16)[name = tensor("op_270_cast_fp16")]; + tensor rope_freqs_to_fp16 = const()[name = tensor("rope_freqs_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6369728)))]; + tensor angles_1_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_270_cast_fp16)[name = tensor("angles_1_cast_fp16")]; + tensor rotr_1_cast_fp16 = cos(x = angles_1_cast_fp16)[name = tensor("rotr_1_cast_fp16")]; + tensor roti_1_cast_fp16 = sin(x = angles_1_cast_fp16)[name = tensor("roti_1_cast_fp16")]; + tensor var_278 = const()[name = tensor("op_278"), val = tensor([1, 16, 32, 2])]; + tensor pairs_1_cast_fp16 = reshape(shape = var_278, x = x_1_cast_fp16)[name = tensor("pairs_1_cast_fp16")]; + tensor xr_1_begin_0 = const()[name = tensor("xr_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_1_end_0 = const()[name = tensor("xr_1_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_1_end_mask_0 = const()[name = tensor("xr_1_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_1_squeeze_mask_0 = const()[name = tensor("xr_1_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_1_cast_fp16 = slice_by_index(begin = xr_1_begin_0, end = xr_1_end_0, end_mask = xr_1_end_mask_0, squeeze_mask = xr_1_squeeze_mask_0, x = pairs_1_cast_fp16)[name = tensor("xr_1_cast_fp16")]; + tensor xi_1_begin_0 = const()[name = tensor("xi_1_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_1_end_0 = const()[name = tensor("xi_1_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_1_end_mask_0 = const()[name = tensor("xi_1_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_1_squeeze_mask_0 = const()[name = tensor("xi_1_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_1_cast_fp16 = slice_by_index(begin = xi_1_begin_0, end = xi_1_end_0, end_mask = xi_1_end_mask_0, squeeze_mask = xi_1_squeeze_mask_0, x = pairs_1_cast_fp16)[name = tensor("xi_1_cast_fp16")]; + tensor var_286_cast_fp16 = mul(x = xr_1_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_286_cast_fp16")]; + tensor var_287_cast_fp16 = mul(x = xi_1_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_287_cast_fp16")]; + tensor xor__1_cast_fp16 = sub(x = var_286_cast_fp16, y = var_287_cast_fp16)[name = tensor("xor__1_cast_fp16")]; + tensor var_290_cast_fp16 = mul(x = xr_1_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_290_cast_fp16")]; + tensor var_291_cast_fp16 = mul(x = xi_1_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_291_cast_fp16")]; + tensor xoi_1_cast_fp16 = add(x = var_290_cast_fp16, y = var_291_cast_fp16)[name = tensor("xoi_1_cast_fp16")]; + tensor out_1_axis_0 = const()[name = tensor("out_1_axis_0"), val = tensor(-1)]; + tensor out_1_cast_fp16 = stack(axis = out_1_axis_0, values = (xor__1_cast_fp16, xoi_1_cast_fp16))[name = tensor("out_1_cast_fp16")]; + tensor var_314 = const()[name = tensor("op_314"), val = tensor([1, 16, 32, 2])]; + tensor pairs_3_cast_fp16 = reshape(shape = var_314, x = x_3_cast_fp16)[name = tensor("pairs_3_cast_fp16")]; + tensor xr_3_begin_0 = const()[name = tensor("xr_3_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_3_end_0 = const()[name = tensor("xr_3_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_3_end_mask_0 = const()[name = tensor("xr_3_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_3_squeeze_mask_0 = const()[name = tensor("xr_3_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_3_cast_fp16 = slice_by_index(begin = xr_3_begin_0, end = xr_3_end_0, end_mask = xr_3_end_mask_0, squeeze_mask = xr_3_squeeze_mask_0, x = pairs_3_cast_fp16)[name = tensor("xr_3_cast_fp16")]; + tensor xi_3_begin_0 = const()[name = tensor("xi_3_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_3_end_0 = const()[name = tensor("xi_3_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_3_end_mask_0 = const()[name = tensor("xi_3_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_3_squeeze_mask_0 = const()[name = tensor("xi_3_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_3_cast_fp16 = slice_by_index(begin = xi_3_begin_0, end = xi_3_end_0, end_mask = xi_3_end_mask_0, squeeze_mask = xi_3_squeeze_mask_0, x = pairs_3_cast_fp16)[name = tensor("xi_3_cast_fp16")]; + tensor var_322_cast_fp16 = mul(x = xr_3_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_322_cast_fp16")]; + tensor var_323_cast_fp16 = mul(x = xi_3_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_323_cast_fp16")]; + tensor xor__3_cast_fp16 = sub(x = var_322_cast_fp16, y = var_323_cast_fp16)[name = tensor("xor__3_cast_fp16")]; + tensor var_326_cast_fp16 = mul(x = xr_3_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_326_cast_fp16")]; + tensor var_327_cast_fp16 = mul(x = xi_3_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_327_cast_fp16")]; + tensor xoi_3_cast_fp16 = add(x = var_326_cast_fp16, y = var_327_cast_fp16)[name = tensor("xoi_3_cast_fp16")]; + tensor out_3_axis_0 = const()[name = tensor("out_3_axis_0"), val = tensor(-1)]; + tensor out_3_cast_fp16 = stack(axis = out_3_axis_0, values = (xor__3_cast_fp16, xoi_3_cast_fp16))[name = tensor("out_3_cast_fp16")]; + tensor _inversed_339_y_0_to_fp16 = const()[name = tensor("_inversed_339_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_339_cast_fp16 = mul(x = position0_to_fp16, y = _inversed_339_y_0_to_fp16)[name = tensor("_inversed_339_cast_fp16")]; + tensor var_340_cast_fp16 = floor(x = _inversed_339_cast_fp16)[name = tensor("op_340_cast_fp16")]; + tensor var_341_to_fp16 = const()[name = tensor("op_341_to_fp16"), val = tensor(0x1p+9)]; + tensor var_342_cast_fp16 = mul(x = var_340_cast_fp16, y = var_341_to_fp16)[name = tensor("op_342_cast_fp16")]; + tensor write_idx_1_cast_fp16 = sub(x = position0_to_fp16, y = var_342_cast_fp16)[name = tensor("write_idx_1_cast_fp16")]; + tensor slot_indices_to_fp16 = const()[name = tensor("slot_indices_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6369856)))]; + tensor var_345_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_1_cast_fp16)[name = tensor("op_345_cast_fp16")]; + tensor var_355 = const()[name = tensor("op_355"), val = tensor([1, 512, 1, 1])]; + tensor onehot_1_to_fp16_dtype_0 = const()[name = tensor("onehot_1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_345_cast_fp16_to_fp16 = cast(dtype = onehot_1_to_fp16_dtype_0, x = var_345_cast_fp16)[name = tensor("cast_576")]; + tensor onehot_3_cast_fp16 = reshape(shape = var_355, x = var_345_cast_fp16_to_fp16)[name = tensor("onehot_3_cast_fp16")]; + tensor var_357_to_fp16 = const()[name = tensor("op_357_to_fp16"), val = tensor(0x1p+0)]; + tensor var_359_cast_fp16 = sub(x = var_357_to_fp16, y = onehot_3_cast_fp16)[name = tensor("op_359_cast_fp16")]; + tensor k_cache0_to_fp16_dtype_0 = const()[name = tensor("k_cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache0_to_fp16 = cast(dtype = k_cache0_to_fp16_dtype_0, x = k_cache0)[name = tensor("cast_575")]; + tensor var_360_cast_fp16 = mul(x = k_cache0_to_fp16, y = var_359_cast_fp16)[name = tensor("op_360_cast_fp16")]; + tensor var_365 = const()[name = tensor("op_365"), val = tensor([1, 1, 16, 64])]; + tensor var_366_cast_fp16 = reshape(shape = var_365, x = out_3_cast_fp16)[name = tensor("op_366_cast_fp16")]; + tensor var_367_cast_fp16 = mul(x = var_366_cast_fp16, y = onehot_3_cast_fp16)[name = tensor("op_367_cast_fp16")]; + tensor new_k_cache_1_cast_fp16 = add(x = var_360_cast_fp16, y = var_367_cast_fp16)[name = tensor("new_k_cache_1_cast_fp16")]; + tensor new_k_cache_1_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_1_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache0_to_fp16_dtype_0 = const()[name = tensor("v_cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache0_to_fp16 = cast(dtype = v_cache0_to_fp16_dtype_0, x = v_cache0)[name = tensor("cast_573")]; + tensor var_373_cast_fp16 = mul(x = v_cache0_to_fp16, y = var_359_cast_fp16)[name = tensor("op_373_cast_fp16")]; + tensor var_378 = const()[name = tensor("op_378"), val = tensor([1, 1, 16, 64])]; + tensor var_379_cast_fp16 = reshape(shape = var_378, x = v_1_cast_fp16)[name = tensor("op_379_cast_fp16")]; + tensor var_380_cast_fp16 = mul(x = var_379_cast_fp16, y = onehot_3_cast_fp16)[name = tensor("op_380_cast_fp16")]; + tensor new_v_cache_1_cast_fp16 = add(x = var_373_cast_fp16, y = var_380_cast_fp16)[name = tensor("new_v_cache_1_cast_fp16")]; + tensor new_v_cache_1_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_1_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_0_perm_0 = const()[name = tensor("transpose_0_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_393 = const()[name = tensor("op_393"), val = tensor([0, 2, 1, 3])]; + tensor var_399 = const()[name = tensor("op_399"), val = tensor([1, 16, 1, 64])]; + tensor q4_1_cast_fp16 = reshape(shape = var_399, x = out_1_cast_fp16)[name = tensor("q4_1_cast_fp16")]; + tensor var_404_transpose_x_0 = const()[name = tensor("op_404_transpose_x_0"), val = tensor(false)]; + tensor var_404_transpose_y_0 = const()[name = tensor("op_404_transpose_y_0"), val = tensor(false)]; + tensor transpose_0_cast_fp16 = transpose(perm = transpose_0_perm_0, x = new_k_cache_1_cast_fp16)[name = tensor("transpose_95")]; + tensor var_404_cast_fp16 = matmul(transpose_x = var_404_transpose_x_0, transpose_y = var_404_transpose_y_0, x = q4_1_cast_fp16, y = transpose_0_cast_fp16)[name = tensor("op_404_cast_fp16")]; + tensor var_405_to_fp16 = const()[name = tensor("op_405_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_1_cast_fp16 = mul(x = var_404_cast_fp16, y = var_405_to_fp16)[name = tensor("scores_1_cast_fp16")]; + tensor var_409 = const()[name = tensor("op_409"), val = tensor([1, 1])]; + tensor var_410_cast_fp16 = reshape(shape = var_409, x = position0_to_fp16)[name = tensor("op_410_cast_fp16")]; + tensor var_411_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_410_cast_fp16)[name = tensor("op_411_cast_fp16")]; + tensor mask_1_to_fp16_dtype_0 = const()[name = tensor("mask_1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_418_to_fp16 = const()[name = tensor("op_418_to_fp16"), val = tensor(0x1p+0)]; + tensor var_411_cast_fp16_to_fp16 = cast(dtype = mask_1_to_fp16_dtype_0, x = var_411_cast_fp16)[name = tensor("cast_571")]; + tensor var_419_cast_fp16 = sub(x = var_411_cast_fp16_to_fp16, y = var_418_to_fp16)[name = tensor("op_419_cast_fp16")]; + tensor var_424 = const()[name = tensor("op_424"), val = tensor([1, 1, 1, 512])]; + tensor var_425_cast_fp16 = reshape(shape = var_424, x = var_419_cast_fp16)[name = tensor("op_425_cast_fp16")]; + tensor var_426_to_fp16 = const()[name = tensor("op_426_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_427_cast_fp16 = mul(x = var_425_cast_fp16, y = var_426_to_fp16)[name = tensor("op_427_cast_fp16")]; + tensor scores_3_cast_fp16 = add(x = scores_1_cast_fp16, y = var_427_cast_fp16)[name = tensor("scores_3_cast_fp16")]; + tensor var_430 = const()[name = tensor("op_430"), val = tensor(-1)]; + tensor weights_1_cast_fp16 = softmax(axis = var_430, x = scores_3_cast_fp16)[name = tensor("weights_1_cast_fp16")]; + tensor attn_1_transpose_x_0 = const()[name = tensor("attn_1_transpose_x_0"), val = tensor(false)]; + tensor attn_1_transpose_y_0 = const()[name = tensor("attn_1_transpose_y_0"), val = tensor(false)]; + tensor values_1_cast_fp16 = transpose(perm = var_393, x = new_v_cache_1_cast_fp16)[name = tensor("transpose_94")]; + tensor attn_1_cast_fp16 = matmul(transpose_x = attn_1_transpose_x_0, transpose_y = attn_1_transpose_y_0, x = weights_1_cast_fp16, y = values_1_cast_fp16)[name = tensor("attn_1_cast_fp16")]; + tensor var_438 = const()[name = tensor("op_438"), val = tensor([0, 2, 1, 3])]; + tensor var_443 = const()[name = tensor("op_443"), val = tensor([1, 1, 1024])]; + tensor var_439_cast_fp16 = transpose(perm = var_438, x = attn_1_cast_fp16)[name = tensor("transpose_93")]; + tensor input_7_cast_fp16 = reshape(shape = var_443, x = var_439_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor attn0_out_proj_weight_to_fp16 = const()[name = tensor("attn0_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6370944)))]; + tensor linear_2_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn0_out_proj_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor var_449_to_fp16 = const()[name = tensor("op_449_to_fp16"), val = tensor(0x1p+0)]; + tensor var_450_cast_fp16 = add(x = position0_to_fp16, y = var_449_to_fp16)[name = tensor("op_450_cast_fp16")]; + tensor var_450_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_450_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_9_cast_fp16 = add(x = linear_0_cast_fp16, y = linear_2_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor input_11_axes_0 = const()[name = tensor("input_11_axes_0"), val = tensor([-1])]; + tensor norm0_2_weight_to_fp16 = const()[name = tensor("norm0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8468160)))]; + tensor norm0_2_bias_to_fp16 = const()[name = tensor("norm0_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8470272)))]; + tensor var_454_to_fp16 = const()[name = tensor("op_454_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_11_cast_fp16 = layer_norm(axes = input_11_axes_0, beta = norm0_2_bias_to_fp16, epsilon = var_454_to_fp16, gamma = norm0_2_weight_to_fp16, x = input_9_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor linear0_1_weight_to_fp16 = const()[name = tensor("linear0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8472384)))]; + tensor linear_3_bias_0_to_fp16 = const()[name = tensor("linear_3_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16861056)))]; + tensor linear_3_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear0_1_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("linear_3_cast_fp16")]; + tensor input_13_mode_0 = const()[name = tensor("input_13_mode_0"), val = tensor("EXACT")]; + tensor input_13_cast_fp16 = gelu(mode = input_13_mode_0, x = linear_3_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor linear0_2_weight_to_fp16 = const()[name = tensor("linear0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16869312)))]; + tensor linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear0_2_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor input_15_cast_fp16 = add(x = input_9_cast_fp16, y = linear_4_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor input_17_axes_0 = const()[name = tensor("input_17_axes_0"), val = tensor([-1])]; + tensor norm1_1_weight_to_fp16 = const()[name = tensor("norm1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25257984)))]; + tensor norm1_1_bias_to_fp16 = const()[name = tensor("norm1_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25260096)))]; + tensor var_471_to_fp16 = const()[name = tensor("op_471_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_17_cast_fp16 = layer_norm(axes = input_17_axes_0, beta = norm1_1_bias_to_fp16, epsilon = var_471_to_fp16, gamma = norm1_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor attn1_in_proj_weight_to_fp16 = const()[name = tensor("attn1_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25262208)))]; + tensor linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn1_in_proj_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor var_484 = const()[name = tensor("op_484"), val = tensor([1, 3, 16, 64])]; + tensor qkv_3_cast_fp16 = reshape(shape = var_484, x = linear_5_cast_fp16)[name = tensor("qkv_3_cast_fp16")]; + tensor x_5_begin_0 = const()[name = tensor("x_5_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_5_end_0 = const()[name = tensor("x_5_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_5_end_mask_0 = const()[name = tensor("x_5_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_5_squeeze_mask_0 = const()[name = tensor("x_5_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_5_cast_fp16 = slice_by_index(begin = x_5_begin_0, end = x_5_end_0, end_mask = x_5_end_mask_0, squeeze_mask = x_5_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("x_5_cast_fp16")]; + tensor x_7_begin_0 = const()[name = tensor("x_7_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_7_end_0 = const()[name = tensor("x_7_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_7_end_mask_0 = const()[name = tensor("x_7_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_7_squeeze_mask_0 = const()[name = tensor("x_7_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_7_cast_fp16 = slice_by_index(begin = x_7_begin_0, end = x_7_end_0, end_mask = x_7_end_mask_0, squeeze_mask = x_7_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("x_7_cast_fp16")]; + tensor v_3_begin_0 = const()[name = tensor("v_3_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_3_end_0 = const()[name = tensor("v_3_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_3_end_mask_0 = const()[name = tensor("v_3_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_3_squeeze_mask_0 = const()[name = tensor("v_3_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_3_cast_fp16 = slice_by_index(begin = v_3_begin_0, end = v_3_end_0, end_mask = v_3_end_mask_0, squeeze_mask = v_3_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("v_3_cast_fp16")]; + tensor var_513 = const()[name = tensor("op_513"), val = tensor([1, 1, 1])]; + tensor position1_to_fp16_dtype_0 = const()[name = tensor("position1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position1_to_fp16 = cast(dtype = position1_to_fp16_dtype_0, x = position1)[name = tensor("cast_569")]; + tensor var_514_cast_fp16 = reshape(shape = var_513, x = position1_to_fp16)[name = tensor("op_514_cast_fp16")]; + tensor angles_5_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_514_cast_fp16)[name = tensor("angles_5_cast_fp16")]; + tensor rotr_5_cast_fp16 = cos(x = angles_5_cast_fp16)[name = tensor("rotr_5_cast_fp16")]; + tensor roti_5_cast_fp16 = sin(x = angles_5_cast_fp16)[name = tensor("roti_5_cast_fp16")]; + tensor var_522 = const()[name = tensor("op_522"), val = tensor([1, 16, 32, 2])]; + tensor pairs_5_cast_fp16 = reshape(shape = var_522, x = x_5_cast_fp16)[name = tensor("pairs_5_cast_fp16")]; + tensor xr_5_begin_0 = const()[name = tensor("xr_5_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_5_end_0 = const()[name = tensor("xr_5_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_5_end_mask_0 = const()[name = tensor("xr_5_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_5_squeeze_mask_0 = const()[name = tensor("xr_5_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_5_cast_fp16 = slice_by_index(begin = xr_5_begin_0, end = xr_5_end_0, end_mask = xr_5_end_mask_0, squeeze_mask = xr_5_squeeze_mask_0, x = pairs_5_cast_fp16)[name = tensor("xr_5_cast_fp16")]; + tensor xi_5_begin_0 = const()[name = tensor("xi_5_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_5_end_0 = const()[name = tensor("xi_5_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_5_end_mask_0 = const()[name = tensor("xi_5_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_5_squeeze_mask_0 = const()[name = tensor("xi_5_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_5_cast_fp16 = slice_by_index(begin = xi_5_begin_0, end = xi_5_end_0, end_mask = xi_5_end_mask_0, squeeze_mask = xi_5_squeeze_mask_0, x = pairs_5_cast_fp16)[name = tensor("xi_5_cast_fp16")]; + tensor var_530_cast_fp16 = mul(x = xr_5_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_530_cast_fp16")]; + tensor var_531_cast_fp16 = mul(x = xi_5_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_531_cast_fp16")]; + tensor xor__5_cast_fp16 = sub(x = var_530_cast_fp16, y = var_531_cast_fp16)[name = tensor("xor__5_cast_fp16")]; + tensor var_534_cast_fp16 = mul(x = xr_5_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_534_cast_fp16")]; + tensor var_535_cast_fp16 = mul(x = xi_5_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_535_cast_fp16")]; + tensor xoi_5_cast_fp16 = add(x = var_534_cast_fp16, y = var_535_cast_fp16)[name = tensor("xoi_5_cast_fp16")]; + tensor out_5_axis_0 = const()[name = tensor("out_5_axis_0"), val = tensor(-1)]; + tensor out_5_cast_fp16 = stack(axis = out_5_axis_0, values = (xor__5_cast_fp16, xoi_5_cast_fp16))[name = tensor("out_5_cast_fp16")]; + tensor var_558 = const()[name = tensor("op_558"), val = tensor([1, 16, 32, 2])]; + tensor pairs_7_cast_fp16 = reshape(shape = var_558, x = x_7_cast_fp16)[name = tensor("pairs_7_cast_fp16")]; + tensor xr_7_begin_0 = const()[name = tensor("xr_7_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_7_end_0 = const()[name = tensor("xr_7_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_7_end_mask_0 = const()[name = tensor("xr_7_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_7_squeeze_mask_0 = const()[name = tensor("xr_7_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_7_cast_fp16 = slice_by_index(begin = xr_7_begin_0, end = xr_7_end_0, end_mask = xr_7_end_mask_0, squeeze_mask = xr_7_squeeze_mask_0, x = pairs_7_cast_fp16)[name = tensor("xr_7_cast_fp16")]; + tensor xi_7_begin_0 = const()[name = tensor("xi_7_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_7_end_0 = const()[name = tensor("xi_7_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_7_end_mask_0 = const()[name = tensor("xi_7_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_7_squeeze_mask_0 = const()[name = tensor("xi_7_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_7_cast_fp16 = slice_by_index(begin = xi_7_begin_0, end = xi_7_end_0, end_mask = xi_7_end_mask_0, squeeze_mask = xi_7_squeeze_mask_0, x = pairs_7_cast_fp16)[name = tensor("xi_7_cast_fp16")]; + tensor var_566_cast_fp16 = mul(x = xr_7_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_566_cast_fp16")]; + tensor var_567_cast_fp16 = mul(x = xi_7_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_567_cast_fp16")]; + tensor xor__7_cast_fp16 = sub(x = var_566_cast_fp16, y = var_567_cast_fp16)[name = tensor("xor__7_cast_fp16")]; + tensor var_570_cast_fp16 = mul(x = xr_7_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_570_cast_fp16")]; + tensor var_571_cast_fp16 = mul(x = xi_7_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_571_cast_fp16")]; + tensor xoi_7_cast_fp16 = add(x = var_570_cast_fp16, y = var_571_cast_fp16)[name = tensor("xoi_7_cast_fp16")]; + tensor out_7_axis_0 = const()[name = tensor("out_7_axis_0"), val = tensor(-1)]; + tensor out_7_cast_fp16 = stack(axis = out_7_axis_0, values = (xor__7_cast_fp16, xoi_7_cast_fp16))[name = tensor("out_7_cast_fp16")]; + tensor _inversed_583_y_0_to_fp16 = const()[name = tensor("_inversed_583_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_583_cast_fp16 = mul(x = position1_to_fp16, y = _inversed_583_y_0_to_fp16)[name = tensor("_inversed_583_cast_fp16")]; + tensor var_584_cast_fp16 = floor(x = _inversed_583_cast_fp16)[name = tensor("op_584_cast_fp16")]; + tensor var_585_to_fp16 = const()[name = tensor("op_585_to_fp16"), val = tensor(0x1p+9)]; + tensor var_586_cast_fp16 = mul(x = var_584_cast_fp16, y = var_585_to_fp16)[name = tensor("op_586_cast_fp16")]; + tensor write_idx_3_cast_fp16 = sub(x = position1_to_fp16, y = var_586_cast_fp16)[name = tensor("write_idx_3_cast_fp16")]; + tensor var_589_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_3_cast_fp16)[name = tensor("op_589_cast_fp16")]; + tensor var_599 = const()[name = tensor("op_599"), val = tensor([1, 512, 1, 1])]; + tensor onehot_5_to_fp16_dtype_0 = const()[name = tensor("onehot_5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_589_cast_fp16_to_fp16 = cast(dtype = onehot_5_to_fp16_dtype_0, x = var_589_cast_fp16)[name = tensor("cast_568")]; + tensor onehot_7_cast_fp16 = reshape(shape = var_599, x = var_589_cast_fp16_to_fp16)[name = tensor("onehot_7_cast_fp16")]; + tensor var_601_to_fp16 = const()[name = tensor("op_601_to_fp16"), val = tensor(0x1p+0)]; + tensor var_603_cast_fp16 = sub(x = var_601_to_fp16, y = onehot_7_cast_fp16)[name = tensor("op_603_cast_fp16")]; + tensor k_cache1_to_fp16_dtype_0 = const()[name = tensor("k_cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache1_to_fp16 = cast(dtype = k_cache1_to_fp16_dtype_0, x = k_cache1)[name = tensor("cast_567")]; + tensor var_604_cast_fp16 = mul(x = k_cache1_to_fp16, y = var_603_cast_fp16)[name = tensor("op_604_cast_fp16")]; + tensor var_609 = const()[name = tensor("op_609"), val = tensor([1, 1, 16, 64])]; + tensor var_610_cast_fp16 = reshape(shape = var_609, x = out_7_cast_fp16)[name = tensor("op_610_cast_fp16")]; + tensor var_611_cast_fp16 = mul(x = var_610_cast_fp16, y = onehot_7_cast_fp16)[name = tensor("op_611_cast_fp16")]; + tensor new_k_cache_3_cast_fp16 = add(x = var_604_cast_fp16, y = var_611_cast_fp16)[name = tensor("new_k_cache_3_cast_fp16")]; + tensor new_k_cache_3_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_3_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache1_to_fp16_dtype_0 = const()[name = tensor("v_cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache1_to_fp16 = cast(dtype = v_cache1_to_fp16_dtype_0, x = v_cache1)[name = tensor("cast_565")]; + tensor var_617_cast_fp16 = mul(x = v_cache1_to_fp16, y = var_603_cast_fp16)[name = tensor("op_617_cast_fp16")]; + tensor var_622 = const()[name = tensor("op_622"), val = tensor([1, 1, 16, 64])]; + tensor var_623_cast_fp16 = reshape(shape = var_622, x = v_3_cast_fp16)[name = tensor("op_623_cast_fp16")]; + tensor var_624_cast_fp16 = mul(x = var_623_cast_fp16, y = onehot_7_cast_fp16)[name = tensor("op_624_cast_fp16")]; + tensor new_v_cache_3_cast_fp16 = add(x = var_617_cast_fp16, y = var_624_cast_fp16)[name = tensor("new_v_cache_3_cast_fp16")]; + tensor new_v_cache_3_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_3_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_1_perm_0 = const()[name = tensor("transpose_1_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_637 = const()[name = tensor("op_637"), val = tensor([0, 2, 1, 3])]; + tensor var_643 = const()[name = tensor("op_643"), val = tensor([1, 16, 1, 64])]; + tensor q4_3_cast_fp16 = reshape(shape = var_643, x = out_5_cast_fp16)[name = tensor("q4_3_cast_fp16")]; + tensor var_648_transpose_x_0 = const()[name = tensor("op_648_transpose_x_0"), val = tensor(false)]; + tensor var_648_transpose_y_0 = const()[name = tensor("op_648_transpose_y_0"), val = tensor(false)]; + tensor transpose_1_cast_fp16 = transpose(perm = transpose_1_perm_0, x = new_k_cache_3_cast_fp16)[name = tensor("transpose_92")]; + tensor var_648_cast_fp16 = matmul(transpose_x = var_648_transpose_x_0, transpose_y = var_648_transpose_y_0, x = q4_3_cast_fp16, y = transpose_1_cast_fp16)[name = tensor("op_648_cast_fp16")]; + tensor var_649_to_fp16 = const()[name = tensor("op_649_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_5_cast_fp16 = mul(x = var_648_cast_fp16, y = var_649_to_fp16)[name = tensor("scores_5_cast_fp16")]; + tensor var_653 = const()[name = tensor("op_653"), val = tensor([1, 1])]; + tensor var_654_cast_fp16 = reshape(shape = var_653, x = position1_to_fp16)[name = tensor("op_654_cast_fp16")]; + tensor var_655_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_654_cast_fp16)[name = tensor("op_655_cast_fp16")]; + tensor mask_3_to_fp16_dtype_0 = const()[name = tensor("mask_3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_662_to_fp16 = const()[name = tensor("op_662_to_fp16"), val = tensor(0x1p+0)]; + tensor var_655_cast_fp16_to_fp16 = cast(dtype = mask_3_to_fp16_dtype_0, x = var_655_cast_fp16)[name = tensor("cast_563")]; + tensor var_663_cast_fp16 = sub(x = var_655_cast_fp16_to_fp16, y = var_662_to_fp16)[name = tensor("op_663_cast_fp16")]; + tensor var_668 = const()[name = tensor("op_668"), val = tensor([1, 1, 1, 512])]; + tensor var_669_cast_fp16 = reshape(shape = var_668, x = var_663_cast_fp16)[name = tensor("op_669_cast_fp16")]; + tensor var_670_to_fp16 = const()[name = tensor("op_670_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_671_cast_fp16 = mul(x = var_669_cast_fp16, y = var_670_to_fp16)[name = tensor("op_671_cast_fp16")]; + tensor scores_7_cast_fp16 = add(x = scores_5_cast_fp16, y = var_671_cast_fp16)[name = tensor("scores_7_cast_fp16")]; + tensor var_674 = const()[name = tensor("op_674"), val = tensor(-1)]; + tensor weights_3_cast_fp16 = softmax(axis = var_674, x = scores_7_cast_fp16)[name = tensor("weights_3_cast_fp16")]; + tensor attn_3_transpose_x_0 = const()[name = tensor("attn_3_transpose_x_0"), val = tensor(false)]; + tensor attn_3_transpose_y_0 = const()[name = tensor("attn_3_transpose_y_0"), val = tensor(false)]; + tensor values_3_cast_fp16 = transpose(perm = var_637, x = new_v_cache_3_cast_fp16)[name = tensor("transpose_91")]; + tensor attn_3_cast_fp16 = matmul(transpose_x = attn_3_transpose_x_0, transpose_y = attn_3_transpose_y_0, x = weights_3_cast_fp16, y = values_3_cast_fp16)[name = tensor("attn_3_cast_fp16")]; + tensor var_682 = const()[name = tensor("op_682"), val = tensor([0, 2, 1, 3])]; + tensor var_687 = const()[name = tensor("op_687"), val = tensor([1, 1, 1024])]; + tensor var_683_cast_fp16 = transpose(perm = var_682, x = attn_3_cast_fp16)[name = tensor("transpose_90")]; + tensor input_19_cast_fp16 = reshape(shape = var_687, x = var_683_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor attn1_out_proj_weight_to_fp16 = const()[name = tensor("attn1_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31553728)))]; + tensor linear_6_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn1_out_proj_weight_to_fp16, x = input_19_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor var_693_to_fp16 = const()[name = tensor("op_693_to_fp16"), val = tensor(0x1p+0)]; + tensor var_694_cast_fp16 = add(x = position1_to_fp16, y = var_693_to_fp16)[name = tensor("op_694_cast_fp16")]; + tensor var_694_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_694_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_21_cast_fp16 = add(x = input_15_cast_fp16, y = linear_6_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor input_23_axes_0 = const()[name = tensor("input_23_axes_0"), val = tensor([-1])]; + tensor norm1_2_weight_to_fp16 = const()[name = tensor("norm1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33650944)))]; + tensor norm1_2_bias_to_fp16 = const()[name = tensor("norm1_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33653056)))]; + tensor var_698_to_fp16 = const()[name = tensor("op_698_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_23_cast_fp16 = layer_norm(axes = input_23_axes_0, beta = norm1_2_bias_to_fp16, epsilon = var_698_to_fp16, gamma = norm1_2_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor linear1_1_weight_to_fp16 = const()[name = tensor("linear1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33655168)))]; + tensor linear_7_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear1_1_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_25_mode_0 = const()[name = tensor("input_25_mode_0"), val = tensor("EXACT")]; + tensor input_25_cast_fp16 = gelu(mode = input_25_mode_0, x = linear_7_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor linear1_2_weight_to_fp16 = const()[name = tensor("linear1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42043840)))]; + tensor linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear1_2_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor input_27_cast_fp16 = add(x = input_21_cast_fp16, y = linear_8_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor input_29_axes_0 = const()[name = tensor("input_29_axes_0"), val = tensor([-1])]; + tensor norm2_1_weight_to_fp16 = const()[name = tensor("norm2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50432512)))]; + tensor norm2_1_bias_to_fp16 = const()[name = tensor("norm2_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50434624)))]; + tensor var_715_to_fp16 = const()[name = tensor("op_715_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_29_cast_fp16 = layer_norm(axes = input_29_axes_0, beta = norm2_1_bias_to_fp16, epsilon = var_715_to_fp16, gamma = norm2_1_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor attn2_in_proj_weight_to_fp16 = const()[name = tensor("attn2_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50436736)))]; + tensor linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn2_in_proj_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor var_728 = const()[name = tensor("op_728"), val = tensor([1, 3, 16, 64])]; + tensor qkv_5_cast_fp16 = reshape(shape = var_728, x = linear_9_cast_fp16)[name = tensor("qkv_5_cast_fp16")]; + tensor x_9_begin_0 = const()[name = tensor("x_9_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_9_end_0 = const()[name = tensor("x_9_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_9_end_mask_0 = const()[name = tensor("x_9_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_9_squeeze_mask_0 = const()[name = tensor("x_9_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_9_cast_fp16 = slice_by_index(begin = x_9_begin_0, end = x_9_end_0, end_mask = x_9_end_mask_0, squeeze_mask = x_9_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("x_9_cast_fp16")]; + tensor x_11_begin_0 = const()[name = tensor("x_11_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_11_end_0 = const()[name = tensor("x_11_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_11_end_mask_0 = const()[name = tensor("x_11_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_11_squeeze_mask_0 = const()[name = tensor("x_11_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_11_cast_fp16 = slice_by_index(begin = x_11_begin_0, end = x_11_end_0, end_mask = x_11_end_mask_0, squeeze_mask = x_11_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("x_11_cast_fp16")]; + tensor v_5_begin_0 = const()[name = tensor("v_5_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_5_end_0 = const()[name = tensor("v_5_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_5_end_mask_0 = const()[name = tensor("v_5_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_5_squeeze_mask_0 = const()[name = tensor("v_5_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_5_cast_fp16 = slice_by_index(begin = v_5_begin_0, end = v_5_end_0, end_mask = v_5_end_mask_0, squeeze_mask = v_5_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("v_5_cast_fp16")]; + tensor var_757 = const()[name = tensor("op_757"), val = tensor([1, 1, 1])]; + tensor position2_to_fp16_dtype_0 = const()[name = tensor("position2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position2_to_fp16 = cast(dtype = position2_to_fp16_dtype_0, x = position2)[name = tensor("cast_561")]; + tensor var_758_cast_fp16 = reshape(shape = var_757, x = position2_to_fp16)[name = tensor("op_758_cast_fp16")]; + tensor angles_9_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_758_cast_fp16)[name = tensor("angles_9_cast_fp16")]; + tensor rotr_9_cast_fp16 = cos(x = angles_9_cast_fp16)[name = tensor("rotr_9_cast_fp16")]; + tensor roti_9_cast_fp16 = sin(x = angles_9_cast_fp16)[name = tensor("roti_9_cast_fp16")]; + tensor var_766 = const()[name = tensor("op_766"), val = tensor([1, 16, 32, 2])]; + tensor pairs_9_cast_fp16 = reshape(shape = var_766, x = x_9_cast_fp16)[name = tensor("pairs_9_cast_fp16")]; + tensor xr_9_begin_0 = const()[name = tensor("xr_9_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_9_end_0 = const()[name = tensor("xr_9_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_9_end_mask_0 = const()[name = tensor("xr_9_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_9_squeeze_mask_0 = const()[name = tensor("xr_9_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_9_cast_fp16 = slice_by_index(begin = xr_9_begin_0, end = xr_9_end_0, end_mask = xr_9_end_mask_0, squeeze_mask = xr_9_squeeze_mask_0, x = pairs_9_cast_fp16)[name = tensor("xr_9_cast_fp16")]; + tensor xi_9_begin_0 = const()[name = tensor("xi_9_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_9_end_0 = const()[name = tensor("xi_9_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_9_end_mask_0 = const()[name = tensor("xi_9_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_9_squeeze_mask_0 = const()[name = tensor("xi_9_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_9_cast_fp16 = slice_by_index(begin = xi_9_begin_0, end = xi_9_end_0, end_mask = xi_9_end_mask_0, squeeze_mask = xi_9_squeeze_mask_0, x = pairs_9_cast_fp16)[name = tensor("xi_9_cast_fp16")]; + tensor var_774_cast_fp16 = mul(x = xr_9_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_774_cast_fp16")]; + tensor var_775_cast_fp16 = mul(x = xi_9_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_775_cast_fp16")]; + tensor xor__9_cast_fp16 = sub(x = var_774_cast_fp16, y = var_775_cast_fp16)[name = tensor("xor__9_cast_fp16")]; + tensor var_778_cast_fp16 = mul(x = xr_9_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_778_cast_fp16")]; + tensor var_779_cast_fp16 = mul(x = xi_9_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_779_cast_fp16")]; + tensor xoi_9_cast_fp16 = add(x = var_778_cast_fp16, y = var_779_cast_fp16)[name = tensor("xoi_9_cast_fp16")]; + tensor out_9_axis_0 = const()[name = tensor("out_9_axis_0"), val = tensor(-1)]; + tensor out_9_cast_fp16 = stack(axis = out_9_axis_0, values = (xor__9_cast_fp16, xoi_9_cast_fp16))[name = tensor("out_9_cast_fp16")]; + tensor var_802 = const()[name = tensor("op_802"), val = tensor([1, 16, 32, 2])]; + tensor pairs_11_cast_fp16 = reshape(shape = var_802, x = x_11_cast_fp16)[name = tensor("pairs_11_cast_fp16")]; + tensor xr_11_begin_0 = const()[name = tensor("xr_11_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_11_end_0 = const()[name = tensor("xr_11_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_11_end_mask_0 = const()[name = tensor("xr_11_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_11_squeeze_mask_0 = const()[name = tensor("xr_11_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_11_cast_fp16 = slice_by_index(begin = xr_11_begin_0, end = xr_11_end_0, end_mask = xr_11_end_mask_0, squeeze_mask = xr_11_squeeze_mask_0, x = pairs_11_cast_fp16)[name = tensor("xr_11_cast_fp16")]; + tensor xi_11_begin_0 = const()[name = tensor("xi_11_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_11_end_0 = const()[name = tensor("xi_11_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_11_end_mask_0 = const()[name = tensor("xi_11_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_11_squeeze_mask_0 = const()[name = tensor("xi_11_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_11_cast_fp16 = slice_by_index(begin = xi_11_begin_0, end = xi_11_end_0, end_mask = xi_11_end_mask_0, squeeze_mask = xi_11_squeeze_mask_0, x = pairs_11_cast_fp16)[name = tensor("xi_11_cast_fp16")]; + tensor var_810_cast_fp16 = mul(x = xr_11_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_810_cast_fp16")]; + tensor var_811_cast_fp16 = mul(x = xi_11_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_811_cast_fp16")]; + tensor xor__11_cast_fp16 = sub(x = var_810_cast_fp16, y = var_811_cast_fp16)[name = tensor("xor__11_cast_fp16")]; + tensor var_814_cast_fp16 = mul(x = xr_11_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_814_cast_fp16")]; + tensor var_815_cast_fp16 = mul(x = xi_11_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_815_cast_fp16")]; + tensor xoi_11_cast_fp16 = add(x = var_814_cast_fp16, y = var_815_cast_fp16)[name = tensor("xoi_11_cast_fp16")]; + tensor out_11_axis_0 = const()[name = tensor("out_11_axis_0"), val = tensor(-1)]; + tensor out_11_cast_fp16 = stack(axis = out_11_axis_0, values = (xor__11_cast_fp16, xoi_11_cast_fp16))[name = tensor("out_11_cast_fp16")]; + tensor _inversed_827_y_0_to_fp16 = const()[name = tensor("_inversed_827_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_827_cast_fp16 = mul(x = position2_to_fp16, y = _inversed_827_y_0_to_fp16)[name = tensor("_inversed_827_cast_fp16")]; + tensor var_828_cast_fp16 = floor(x = _inversed_827_cast_fp16)[name = tensor("op_828_cast_fp16")]; + tensor var_829_to_fp16 = const()[name = tensor("op_829_to_fp16"), val = tensor(0x1p+9)]; + tensor var_830_cast_fp16 = mul(x = var_828_cast_fp16, y = var_829_to_fp16)[name = tensor("op_830_cast_fp16")]; + tensor write_idx_5_cast_fp16 = sub(x = position2_to_fp16, y = var_830_cast_fp16)[name = tensor("write_idx_5_cast_fp16")]; + tensor var_833_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_5_cast_fp16)[name = tensor("op_833_cast_fp16")]; + tensor var_843 = const()[name = tensor("op_843"), val = tensor([1, 512, 1, 1])]; + tensor onehot_9_to_fp16_dtype_0 = const()[name = tensor("onehot_9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_833_cast_fp16_to_fp16 = cast(dtype = onehot_9_to_fp16_dtype_0, x = var_833_cast_fp16)[name = tensor("cast_560")]; + tensor onehot_11_cast_fp16 = reshape(shape = var_843, x = var_833_cast_fp16_to_fp16)[name = tensor("onehot_11_cast_fp16")]; + tensor var_845_to_fp16 = const()[name = tensor("op_845_to_fp16"), val = tensor(0x1p+0)]; + tensor var_847_cast_fp16 = sub(x = var_845_to_fp16, y = onehot_11_cast_fp16)[name = tensor("op_847_cast_fp16")]; + tensor k_cache2_to_fp16_dtype_0 = const()[name = tensor("k_cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache2_to_fp16 = cast(dtype = k_cache2_to_fp16_dtype_0, x = k_cache2)[name = tensor("cast_559")]; + tensor var_848_cast_fp16 = mul(x = k_cache2_to_fp16, y = var_847_cast_fp16)[name = tensor("op_848_cast_fp16")]; + tensor var_853 = const()[name = tensor("op_853"), val = tensor([1, 1, 16, 64])]; + tensor var_854_cast_fp16 = reshape(shape = var_853, x = out_11_cast_fp16)[name = tensor("op_854_cast_fp16")]; + tensor var_855_cast_fp16 = mul(x = var_854_cast_fp16, y = onehot_11_cast_fp16)[name = tensor("op_855_cast_fp16")]; + tensor new_k_cache_5_cast_fp16 = add(x = var_848_cast_fp16, y = var_855_cast_fp16)[name = tensor("new_k_cache_5_cast_fp16")]; + tensor new_k_cache_5_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_5_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache2_to_fp16_dtype_0 = const()[name = tensor("v_cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache2_to_fp16 = cast(dtype = v_cache2_to_fp16_dtype_0, x = v_cache2)[name = tensor("cast_557")]; + tensor var_861_cast_fp16 = mul(x = v_cache2_to_fp16, y = var_847_cast_fp16)[name = tensor("op_861_cast_fp16")]; + tensor var_866 = const()[name = tensor("op_866"), val = tensor([1, 1, 16, 64])]; + tensor var_867_cast_fp16 = reshape(shape = var_866, x = v_5_cast_fp16)[name = tensor("op_867_cast_fp16")]; + tensor var_868_cast_fp16 = mul(x = var_867_cast_fp16, y = onehot_11_cast_fp16)[name = tensor("op_868_cast_fp16")]; + tensor new_v_cache_5_cast_fp16 = add(x = var_861_cast_fp16, y = var_868_cast_fp16)[name = tensor("new_v_cache_5_cast_fp16")]; + tensor new_v_cache_5_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_5_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_2_perm_0 = const()[name = tensor("transpose_2_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_881 = const()[name = tensor("op_881"), val = tensor([0, 2, 1, 3])]; + tensor var_887 = const()[name = tensor("op_887"), val = tensor([1, 16, 1, 64])]; + tensor q4_5_cast_fp16 = reshape(shape = var_887, x = out_9_cast_fp16)[name = tensor("q4_5_cast_fp16")]; + tensor var_892_transpose_x_0 = const()[name = tensor("op_892_transpose_x_0"), val = tensor(false)]; + tensor var_892_transpose_y_0 = const()[name = tensor("op_892_transpose_y_0"), val = tensor(false)]; + tensor transpose_2_cast_fp16 = transpose(perm = transpose_2_perm_0, x = new_k_cache_5_cast_fp16)[name = tensor("transpose_89")]; + tensor var_892_cast_fp16 = matmul(transpose_x = var_892_transpose_x_0, transpose_y = var_892_transpose_y_0, x = q4_5_cast_fp16, y = transpose_2_cast_fp16)[name = tensor("op_892_cast_fp16")]; + tensor var_893_to_fp16 = const()[name = tensor("op_893_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_9_cast_fp16 = mul(x = var_892_cast_fp16, y = var_893_to_fp16)[name = tensor("scores_9_cast_fp16")]; + tensor var_897 = const()[name = tensor("op_897"), val = tensor([1, 1])]; + tensor var_898_cast_fp16 = reshape(shape = var_897, x = position2_to_fp16)[name = tensor("op_898_cast_fp16")]; + tensor var_899_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_898_cast_fp16)[name = tensor("op_899_cast_fp16")]; + tensor mask_5_to_fp16_dtype_0 = const()[name = tensor("mask_5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_906_to_fp16 = const()[name = tensor("op_906_to_fp16"), val = tensor(0x1p+0)]; + tensor var_899_cast_fp16_to_fp16 = cast(dtype = mask_5_to_fp16_dtype_0, x = var_899_cast_fp16)[name = tensor("cast_555")]; + tensor var_907_cast_fp16 = sub(x = var_899_cast_fp16_to_fp16, y = var_906_to_fp16)[name = tensor("op_907_cast_fp16")]; + tensor var_912 = const()[name = tensor("op_912"), val = tensor([1, 1, 1, 512])]; + tensor var_913_cast_fp16 = reshape(shape = var_912, x = var_907_cast_fp16)[name = tensor("op_913_cast_fp16")]; + tensor var_914_to_fp16 = const()[name = tensor("op_914_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_915_cast_fp16 = mul(x = var_913_cast_fp16, y = var_914_to_fp16)[name = tensor("op_915_cast_fp16")]; + tensor scores_11_cast_fp16 = add(x = scores_9_cast_fp16, y = var_915_cast_fp16)[name = tensor("scores_11_cast_fp16")]; + tensor var_918 = const()[name = tensor("op_918"), val = tensor(-1)]; + tensor weights_5_cast_fp16 = softmax(axis = var_918, x = scores_11_cast_fp16)[name = tensor("weights_5_cast_fp16")]; + tensor attn_5_transpose_x_0 = const()[name = tensor("attn_5_transpose_x_0"), val = tensor(false)]; + tensor attn_5_transpose_y_0 = const()[name = tensor("attn_5_transpose_y_0"), val = tensor(false)]; + tensor values_5_cast_fp16 = transpose(perm = var_881, x = new_v_cache_5_cast_fp16)[name = tensor("transpose_88")]; + tensor attn_5_cast_fp16 = matmul(transpose_x = attn_5_transpose_x_0, transpose_y = attn_5_transpose_y_0, x = weights_5_cast_fp16, y = values_5_cast_fp16)[name = tensor("attn_5_cast_fp16")]; + tensor var_926 = const()[name = tensor("op_926"), val = tensor([0, 2, 1, 3])]; + tensor var_931 = const()[name = tensor("op_931"), val = tensor([1, 1, 1024])]; + tensor var_927_cast_fp16 = transpose(perm = var_926, x = attn_5_cast_fp16)[name = tensor("transpose_87")]; + tensor input_31_cast_fp16 = reshape(shape = var_931, x = var_927_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor attn2_out_proj_weight_to_fp16 = const()[name = tensor("attn2_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(56728256)))]; + tensor linear_10_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn2_out_proj_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor var_937_to_fp16 = const()[name = tensor("op_937_to_fp16"), val = tensor(0x1p+0)]; + tensor var_938_cast_fp16 = add(x = position2_to_fp16, y = var_937_to_fp16)[name = tensor("op_938_cast_fp16")]; + tensor var_938_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_938_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_33_cast_fp16 = add(x = input_27_cast_fp16, y = linear_10_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor input_35_axes_0 = const()[name = tensor("input_35_axes_0"), val = tensor([-1])]; + tensor norm2_2_weight_to_fp16 = const()[name = tensor("norm2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58825472)))]; + tensor norm2_2_bias_to_fp16 = const()[name = tensor("norm2_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58827584)))]; + tensor var_942_to_fp16 = const()[name = tensor("op_942_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_35_cast_fp16 = layer_norm(axes = input_35_axes_0, beta = norm2_2_bias_to_fp16, epsilon = var_942_to_fp16, gamma = norm2_2_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor linear2_1_weight_to_fp16 = const()[name = tensor("linear2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58829696)))]; + tensor linear_11_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear2_1_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor input_37_mode_0 = const()[name = tensor("input_37_mode_0"), val = tensor("EXACT")]; + tensor input_37_cast_fp16 = gelu(mode = input_37_mode_0, x = linear_11_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor linear2_2_weight_to_fp16 = const()[name = tensor("linear2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67218368)))]; + tensor linear_12_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear2_2_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor input_39_cast_fp16 = add(x = input_33_cast_fp16, y = linear_12_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor input_41_axes_0 = const()[name = tensor("input_41_axes_0"), val = tensor([-1])]; + tensor norm3_1_weight_to_fp16 = const()[name = tensor("norm3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75607040)))]; + tensor norm3_1_bias_to_fp16 = const()[name = tensor("norm3_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75609152)))]; + tensor var_959_to_fp16 = const()[name = tensor("op_959_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_41_cast_fp16 = layer_norm(axes = input_41_axes_0, beta = norm3_1_bias_to_fp16, epsilon = var_959_to_fp16, gamma = norm3_1_weight_to_fp16, x = input_39_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor attn3_in_proj_weight_to_fp16 = const()[name = tensor("attn3_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75611264)))]; + tensor linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn3_in_proj_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor var_972 = const()[name = tensor("op_972"), val = tensor([1, 3, 16, 64])]; + tensor qkv_7_cast_fp16 = reshape(shape = var_972, x = linear_13_cast_fp16)[name = tensor("qkv_7_cast_fp16")]; + tensor x_13_begin_0 = const()[name = tensor("x_13_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_13_end_0 = const()[name = tensor("x_13_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_13_end_mask_0 = const()[name = tensor("x_13_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_13_squeeze_mask_0 = const()[name = tensor("x_13_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_13_cast_fp16 = slice_by_index(begin = x_13_begin_0, end = x_13_end_0, end_mask = x_13_end_mask_0, squeeze_mask = x_13_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("x_13_cast_fp16")]; + tensor x_15_begin_0 = const()[name = tensor("x_15_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_15_end_0 = const()[name = tensor("x_15_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_15_end_mask_0 = const()[name = tensor("x_15_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_15_squeeze_mask_0 = const()[name = tensor("x_15_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_15_cast_fp16 = slice_by_index(begin = x_15_begin_0, end = x_15_end_0, end_mask = x_15_end_mask_0, squeeze_mask = x_15_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("x_15_cast_fp16")]; + tensor v_7_begin_0 = const()[name = tensor("v_7_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_7_end_0 = const()[name = tensor("v_7_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_7_end_mask_0 = const()[name = tensor("v_7_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_7_squeeze_mask_0 = const()[name = tensor("v_7_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_7_cast_fp16 = slice_by_index(begin = v_7_begin_0, end = v_7_end_0, end_mask = v_7_end_mask_0, squeeze_mask = v_7_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("v_7_cast_fp16")]; + tensor var_1001 = const()[name = tensor("op_1001"), val = tensor([1, 1, 1])]; + tensor position3_to_fp16_dtype_0 = const()[name = tensor("position3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position3_to_fp16 = cast(dtype = position3_to_fp16_dtype_0, x = position3)[name = tensor("cast_553")]; + tensor var_1002_cast_fp16 = reshape(shape = var_1001, x = position3_to_fp16)[name = tensor("op_1002_cast_fp16")]; + tensor angles_13_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_1002_cast_fp16)[name = tensor("angles_13_cast_fp16")]; + tensor rotr_13_cast_fp16 = cos(x = angles_13_cast_fp16)[name = tensor("rotr_13_cast_fp16")]; + tensor roti_13_cast_fp16 = sin(x = angles_13_cast_fp16)[name = tensor("roti_13_cast_fp16")]; + tensor var_1010 = const()[name = tensor("op_1010"), val = tensor([1, 16, 32, 2])]; + tensor pairs_13_cast_fp16 = reshape(shape = var_1010, x = x_13_cast_fp16)[name = tensor("pairs_13_cast_fp16")]; + tensor xr_13_begin_0 = const()[name = tensor("xr_13_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_13_end_0 = const()[name = tensor("xr_13_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_13_end_mask_0 = const()[name = tensor("xr_13_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_13_squeeze_mask_0 = const()[name = tensor("xr_13_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_13_cast_fp16 = slice_by_index(begin = xr_13_begin_0, end = xr_13_end_0, end_mask = xr_13_end_mask_0, squeeze_mask = xr_13_squeeze_mask_0, x = pairs_13_cast_fp16)[name = tensor("xr_13_cast_fp16")]; + tensor xi_13_begin_0 = const()[name = tensor("xi_13_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_13_end_0 = const()[name = tensor("xi_13_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_13_end_mask_0 = const()[name = tensor("xi_13_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_13_squeeze_mask_0 = const()[name = tensor("xi_13_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_13_cast_fp16 = slice_by_index(begin = xi_13_begin_0, end = xi_13_end_0, end_mask = xi_13_end_mask_0, squeeze_mask = xi_13_squeeze_mask_0, x = pairs_13_cast_fp16)[name = tensor("xi_13_cast_fp16")]; + tensor var_1018_cast_fp16 = mul(x = xr_13_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_1018_cast_fp16")]; + tensor var_1019_cast_fp16 = mul(x = xi_13_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_1019_cast_fp16")]; + tensor xor__13_cast_fp16 = sub(x = var_1018_cast_fp16, y = var_1019_cast_fp16)[name = tensor("xor__13_cast_fp16")]; + tensor var_1022_cast_fp16 = mul(x = xr_13_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_1022_cast_fp16")]; + tensor var_1023_cast_fp16 = mul(x = xi_13_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_1023_cast_fp16")]; + tensor xoi_13_cast_fp16 = add(x = var_1022_cast_fp16, y = var_1023_cast_fp16)[name = tensor("xoi_13_cast_fp16")]; + tensor out_13_axis_0 = const()[name = tensor("out_13_axis_0"), val = tensor(-1)]; + tensor out_13_cast_fp16 = stack(axis = out_13_axis_0, values = (xor__13_cast_fp16, xoi_13_cast_fp16))[name = tensor("out_13_cast_fp16")]; + tensor var_1046 = const()[name = tensor("op_1046"), val = tensor([1, 16, 32, 2])]; + tensor pairs_15_cast_fp16 = reshape(shape = var_1046, x = x_15_cast_fp16)[name = tensor("pairs_15_cast_fp16")]; + tensor xr_15_begin_0 = const()[name = tensor("xr_15_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_15_end_0 = const()[name = tensor("xr_15_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_15_end_mask_0 = const()[name = tensor("xr_15_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_15_squeeze_mask_0 = const()[name = tensor("xr_15_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_15_cast_fp16 = slice_by_index(begin = xr_15_begin_0, end = xr_15_end_0, end_mask = xr_15_end_mask_0, squeeze_mask = xr_15_squeeze_mask_0, x = pairs_15_cast_fp16)[name = tensor("xr_15_cast_fp16")]; + tensor xi_15_begin_0 = const()[name = tensor("xi_15_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_15_end_0 = const()[name = tensor("xi_15_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_15_end_mask_0 = const()[name = tensor("xi_15_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_15_squeeze_mask_0 = const()[name = tensor("xi_15_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_15_cast_fp16 = slice_by_index(begin = xi_15_begin_0, end = xi_15_end_0, end_mask = xi_15_end_mask_0, squeeze_mask = xi_15_squeeze_mask_0, x = pairs_15_cast_fp16)[name = tensor("xi_15_cast_fp16")]; + tensor var_1054_cast_fp16 = mul(x = xr_15_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_1054_cast_fp16")]; + tensor var_1055_cast_fp16 = mul(x = xi_15_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_1055_cast_fp16")]; + tensor xor__15_cast_fp16 = sub(x = var_1054_cast_fp16, y = var_1055_cast_fp16)[name = tensor("xor__15_cast_fp16")]; + tensor var_1058_cast_fp16 = mul(x = xr_15_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_1058_cast_fp16")]; + tensor var_1059_cast_fp16 = mul(x = xi_15_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_1059_cast_fp16")]; + tensor xoi_15_cast_fp16 = add(x = var_1058_cast_fp16, y = var_1059_cast_fp16)[name = tensor("xoi_15_cast_fp16")]; + tensor out_15_axis_0 = const()[name = tensor("out_15_axis_0"), val = tensor(-1)]; + tensor out_15_cast_fp16 = stack(axis = out_15_axis_0, values = (xor__15_cast_fp16, xoi_15_cast_fp16))[name = tensor("out_15_cast_fp16")]; + tensor _inversed_1071_y_0_to_fp16 = const()[name = tensor("_inversed_1071_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1071_cast_fp16 = mul(x = position3_to_fp16, y = _inversed_1071_y_0_to_fp16)[name = tensor("_inversed_1071_cast_fp16")]; + tensor var_1072_cast_fp16 = floor(x = _inversed_1071_cast_fp16)[name = tensor("op_1072_cast_fp16")]; + tensor var_1073_to_fp16 = const()[name = tensor("op_1073_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1074_cast_fp16 = mul(x = var_1072_cast_fp16, y = var_1073_to_fp16)[name = tensor("op_1074_cast_fp16")]; + tensor write_idx_7_cast_fp16 = sub(x = position3_to_fp16, y = var_1074_cast_fp16)[name = tensor("write_idx_7_cast_fp16")]; + tensor var_1077_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_7_cast_fp16)[name = tensor("op_1077_cast_fp16")]; + tensor var_1087 = const()[name = tensor("op_1087"), val = tensor([1, 512, 1, 1])]; + tensor onehot_13_to_fp16_dtype_0 = const()[name = tensor("onehot_13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_1077_cast_fp16_to_fp16 = cast(dtype = onehot_13_to_fp16_dtype_0, x = var_1077_cast_fp16)[name = tensor("cast_552")]; + tensor onehot_15_cast_fp16 = reshape(shape = var_1087, x = var_1077_cast_fp16_to_fp16)[name = tensor("onehot_15_cast_fp16")]; + tensor var_1089_to_fp16 = const()[name = tensor("op_1089_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1091_cast_fp16 = sub(x = var_1089_to_fp16, y = onehot_15_cast_fp16)[name = tensor("op_1091_cast_fp16")]; + tensor k_cache3_to_fp16_dtype_0 = const()[name = tensor("k_cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache3_to_fp16 = cast(dtype = k_cache3_to_fp16_dtype_0, x = k_cache3)[name = tensor("cast_551")]; + tensor var_1092_cast_fp16 = mul(x = k_cache3_to_fp16, y = var_1091_cast_fp16)[name = tensor("op_1092_cast_fp16")]; + tensor var_1097 = const()[name = tensor("op_1097"), val = tensor([1, 1, 16, 64])]; + tensor var_1098_cast_fp16 = reshape(shape = var_1097, x = out_15_cast_fp16)[name = tensor("op_1098_cast_fp16")]; + tensor var_1099_cast_fp16 = mul(x = var_1098_cast_fp16, y = onehot_15_cast_fp16)[name = tensor("op_1099_cast_fp16")]; + tensor new_k_cache_7_cast_fp16 = add(x = var_1092_cast_fp16, y = var_1099_cast_fp16)[name = tensor("new_k_cache_7_cast_fp16")]; + tensor new_k_cache_7_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_7_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache3_to_fp16_dtype_0 = const()[name = tensor("v_cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache3_to_fp16 = cast(dtype = v_cache3_to_fp16_dtype_0, x = v_cache3)[name = tensor("cast_549")]; + tensor var_1105_cast_fp16 = mul(x = v_cache3_to_fp16, y = var_1091_cast_fp16)[name = tensor("op_1105_cast_fp16")]; + tensor var_1110 = const()[name = tensor("op_1110"), val = tensor([1, 1, 16, 64])]; + tensor var_1111_cast_fp16 = reshape(shape = var_1110, x = v_7_cast_fp16)[name = tensor("op_1111_cast_fp16")]; + tensor var_1112_cast_fp16 = mul(x = var_1111_cast_fp16, y = onehot_15_cast_fp16)[name = tensor("op_1112_cast_fp16")]; + tensor new_v_cache_7_cast_fp16 = add(x = var_1105_cast_fp16, y = var_1112_cast_fp16)[name = tensor("new_v_cache_7_cast_fp16")]; + tensor new_v_cache_7_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_7_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_3_perm_0 = const()[name = tensor("transpose_3_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_1125 = const()[name = tensor("op_1125"), val = tensor([0, 2, 1, 3])]; + tensor var_1131 = const()[name = tensor("op_1131"), val = tensor([1, 16, 1, 64])]; + tensor q4_7_cast_fp16 = reshape(shape = var_1131, x = out_13_cast_fp16)[name = tensor("q4_7_cast_fp16")]; + tensor var_1136_transpose_x_0 = const()[name = tensor("op_1136_transpose_x_0"), val = tensor(false)]; + tensor var_1136_transpose_y_0 = const()[name = tensor("op_1136_transpose_y_0"), val = tensor(false)]; + tensor transpose_3_cast_fp16 = transpose(perm = transpose_3_perm_0, x = new_k_cache_7_cast_fp16)[name = tensor("transpose_86")]; + tensor var_1136_cast_fp16 = matmul(transpose_x = var_1136_transpose_x_0, transpose_y = var_1136_transpose_y_0, x = q4_7_cast_fp16, y = transpose_3_cast_fp16)[name = tensor("op_1136_cast_fp16")]; + tensor var_1137_to_fp16 = const()[name = tensor("op_1137_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_13_cast_fp16 = mul(x = var_1136_cast_fp16, y = var_1137_to_fp16)[name = tensor("scores_13_cast_fp16")]; + tensor var_1141 = const()[name = tensor("op_1141"), val = tensor([1, 1])]; + tensor var_1142_cast_fp16 = reshape(shape = var_1141, x = position3_to_fp16)[name = tensor("op_1142_cast_fp16")]; + tensor var_1143_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_1142_cast_fp16)[name = tensor("op_1143_cast_fp16")]; + tensor mask_7_to_fp16_dtype_0 = const()[name = tensor("mask_7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_1150_to_fp16 = const()[name = tensor("op_1150_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1143_cast_fp16_to_fp16 = cast(dtype = mask_7_to_fp16_dtype_0, x = var_1143_cast_fp16)[name = tensor("cast_547")]; + tensor var_1151_cast_fp16 = sub(x = var_1143_cast_fp16_to_fp16, y = var_1150_to_fp16)[name = tensor("op_1151_cast_fp16")]; + tensor var_1156 = const()[name = tensor("op_1156"), val = tensor([1, 1, 1, 512])]; + tensor var_1157_cast_fp16 = reshape(shape = var_1156, x = var_1151_cast_fp16)[name = tensor("op_1157_cast_fp16")]; + tensor var_1158_to_fp16 = const()[name = tensor("op_1158_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_1159_cast_fp16 = mul(x = var_1157_cast_fp16, y = var_1158_to_fp16)[name = tensor("op_1159_cast_fp16")]; + tensor scores_15_cast_fp16 = add(x = scores_13_cast_fp16, y = var_1159_cast_fp16)[name = tensor("scores_15_cast_fp16")]; + tensor var_1162 = const()[name = tensor("op_1162"), val = tensor(-1)]; + tensor weights_7_cast_fp16 = softmax(axis = var_1162, x = scores_15_cast_fp16)[name = tensor("weights_7_cast_fp16")]; + tensor attn_7_transpose_x_0 = const()[name = tensor("attn_7_transpose_x_0"), val = tensor(false)]; + tensor attn_7_transpose_y_0 = const()[name = tensor("attn_7_transpose_y_0"), val = tensor(false)]; + tensor values_7_cast_fp16 = transpose(perm = var_1125, x = new_v_cache_7_cast_fp16)[name = tensor("transpose_85")]; + tensor attn_7_cast_fp16 = matmul(transpose_x = attn_7_transpose_x_0, transpose_y = attn_7_transpose_y_0, x = weights_7_cast_fp16, y = values_7_cast_fp16)[name = tensor("attn_7_cast_fp16")]; + tensor var_1170 = const()[name = tensor("op_1170"), val = tensor([0, 2, 1, 3])]; + tensor var_1175 = const()[name = tensor("op_1175"), val = tensor([1, 1, 1024])]; + tensor var_1171_cast_fp16 = transpose(perm = var_1170, x = attn_7_cast_fp16)[name = tensor("transpose_84")]; + tensor input_43_cast_fp16 = reshape(shape = var_1175, x = var_1171_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor attn3_out_proj_weight_to_fp16 = const()[name = tensor("attn3_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(81902784)))]; + tensor linear_14_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn3_out_proj_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor var_1181_to_fp16 = const()[name = tensor("op_1181_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1182_cast_fp16 = add(x = position3_to_fp16, y = var_1181_to_fp16)[name = tensor("op_1182_cast_fp16")]; + tensor var_1182_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1182_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_45_cast_fp16 = add(x = input_39_cast_fp16, y = linear_14_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor input_47_axes_0 = const()[name = tensor("input_47_axes_0"), val = tensor([-1])]; + tensor norm3_2_weight_to_fp16 = const()[name = tensor("norm3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84000000)))]; + tensor norm3_2_bias_to_fp16 = const()[name = tensor("norm3_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84002112)))]; + tensor var_1186_to_fp16 = const()[name = tensor("op_1186_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_47_cast_fp16 = layer_norm(axes = input_47_axes_0, beta = norm3_2_bias_to_fp16, epsilon = var_1186_to_fp16, gamma = norm3_2_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor linear3_1_weight_to_fp16 = const()[name = tensor("linear3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84004224)))]; + tensor linear_15_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear3_1_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor input_49_mode_0 = const()[name = tensor("input_49_mode_0"), val = tensor("EXACT")]; + tensor input_49_cast_fp16 = gelu(mode = input_49_mode_0, x = linear_15_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor linear3_2_weight_to_fp16 = const()[name = tensor("linear3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(92392896)))]; + tensor linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear3_2_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor input_51_cast_fp16 = add(x = input_45_cast_fp16, y = linear_16_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor input_53_axes_0 = const()[name = tensor("input_53_axes_0"), val = tensor([-1])]; + tensor norm4_1_weight_to_fp16 = const()[name = tensor("norm4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100781568)))]; + tensor norm4_1_bias_to_fp16 = const()[name = tensor("norm4_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100783680)))]; + tensor var_1203_to_fp16 = const()[name = tensor("op_1203_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_53_cast_fp16 = layer_norm(axes = input_53_axes_0, beta = norm4_1_bias_to_fp16, epsilon = var_1203_to_fp16, gamma = norm4_1_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor attn4_in_proj_weight_to_fp16 = const()[name = tensor("attn4_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100785792)))]; + tensor linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn4_in_proj_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor var_1216 = const()[name = tensor("op_1216"), val = tensor([1, 3, 16, 64])]; + tensor qkv_9_cast_fp16 = reshape(shape = var_1216, x = linear_17_cast_fp16)[name = tensor("qkv_9_cast_fp16")]; + tensor x_17_begin_0 = const()[name = tensor("x_17_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_17_end_0 = const()[name = tensor("x_17_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_17_end_mask_0 = const()[name = tensor("x_17_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_17_squeeze_mask_0 = const()[name = tensor("x_17_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_17_cast_fp16 = slice_by_index(begin = x_17_begin_0, end = x_17_end_0, end_mask = x_17_end_mask_0, squeeze_mask = x_17_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("x_17_cast_fp16")]; + tensor x_19_begin_0 = const()[name = tensor("x_19_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_19_end_0 = const()[name = tensor("x_19_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_19_end_mask_0 = const()[name = tensor("x_19_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_19_squeeze_mask_0 = const()[name = tensor("x_19_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_19_cast_fp16 = slice_by_index(begin = x_19_begin_0, end = x_19_end_0, end_mask = x_19_end_mask_0, squeeze_mask = x_19_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("x_19_cast_fp16")]; + tensor v_9_begin_0 = const()[name = tensor("v_9_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_9_end_0 = const()[name = tensor("v_9_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_9_end_mask_0 = const()[name = tensor("v_9_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_9_squeeze_mask_0 = const()[name = tensor("v_9_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_9_cast_fp16 = slice_by_index(begin = v_9_begin_0, end = v_9_end_0, end_mask = v_9_end_mask_0, squeeze_mask = v_9_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("v_9_cast_fp16")]; + tensor var_1245 = const()[name = tensor("op_1245"), val = tensor([1, 1, 1])]; + tensor position4_to_fp16_dtype_0 = const()[name = tensor("position4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position4_to_fp16 = cast(dtype = position4_to_fp16_dtype_0, x = position4)[name = tensor("cast_545")]; + tensor var_1246_cast_fp16 = reshape(shape = var_1245, x = position4_to_fp16)[name = tensor("op_1246_cast_fp16")]; + tensor angles_17_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_1246_cast_fp16)[name = tensor("angles_17_cast_fp16")]; + tensor rotr_17_cast_fp16 = cos(x = angles_17_cast_fp16)[name = tensor("rotr_17_cast_fp16")]; + tensor roti_17_cast_fp16 = sin(x = angles_17_cast_fp16)[name = tensor("roti_17_cast_fp16")]; + tensor var_1254 = const()[name = tensor("op_1254"), val = tensor([1, 16, 32, 2])]; + tensor pairs_17_cast_fp16 = reshape(shape = var_1254, x = x_17_cast_fp16)[name = tensor("pairs_17_cast_fp16")]; + tensor xr_17_begin_0 = const()[name = tensor("xr_17_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_17_end_0 = const()[name = tensor("xr_17_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_17_end_mask_0 = const()[name = tensor("xr_17_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_17_squeeze_mask_0 = const()[name = tensor("xr_17_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_17_cast_fp16 = slice_by_index(begin = xr_17_begin_0, end = xr_17_end_0, end_mask = xr_17_end_mask_0, squeeze_mask = xr_17_squeeze_mask_0, x = pairs_17_cast_fp16)[name = tensor("xr_17_cast_fp16")]; + tensor xi_17_begin_0 = const()[name = tensor("xi_17_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_17_end_0 = const()[name = tensor("xi_17_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_17_end_mask_0 = const()[name = tensor("xi_17_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_17_squeeze_mask_0 = const()[name = tensor("xi_17_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_17_cast_fp16 = slice_by_index(begin = xi_17_begin_0, end = xi_17_end_0, end_mask = xi_17_end_mask_0, squeeze_mask = xi_17_squeeze_mask_0, x = pairs_17_cast_fp16)[name = tensor("xi_17_cast_fp16")]; + tensor var_1262_cast_fp16 = mul(x = xr_17_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_1262_cast_fp16")]; + tensor var_1263_cast_fp16 = mul(x = xi_17_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_1263_cast_fp16")]; + tensor xor__17_cast_fp16 = sub(x = var_1262_cast_fp16, y = var_1263_cast_fp16)[name = tensor("xor__17_cast_fp16")]; + tensor var_1266_cast_fp16 = mul(x = xr_17_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_1266_cast_fp16")]; + tensor var_1267_cast_fp16 = mul(x = xi_17_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_1267_cast_fp16")]; + tensor xoi_17_cast_fp16 = add(x = var_1266_cast_fp16, y = var_1267_cast_fp16)[name = tensor("xoi_17_cast_fp16")]; + tensor out_17_axis_0 = const()[name = tensor("out_17_axis_0"), val = tensor(-1)]; + tensor out_17_cast_fp16 = stack(axis = out_17_axis_0, values = (xor__17_cast_fp16, xoi_17_cast_fp16))[name = tensor("out_17_cast_fp16")]; + tensor var_1290 = const()[name = tensor("op_1290"), val = tensor([1, 16, 32, 2])]; + tensor pairs_19_cast_fp16 = reshape(shape = var_1290, x = x_19_cast_fp16)[name = tensor("pairs_19_cast_fp16")]; + tensor xr_19_begin_0 = const()[name = tensor("xr_19_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_19_end_0 = const()[name = tensor("xr_19_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_19_end_mask_0 = const()[name = tensor("xr_19_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_19_squeeze_mask_0 = const()[name = tensor("xr_19_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_19_cast_fp16 = slice_by_index(begin = xr_19_begin_0, end = xr_19_end_0, end_mask = xr_19_end_mask_0, squeeze_mask = xr_19_squeeze_mask_0, x = pairs_19_cast_fp16)[name = tensor("xr_19_cast_fp16")]; + tensor xi_19_begin_0 = const()[name = tensor("xi_19_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_19_end_0 = const()[name = tensor("xi_19_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_19_end_mask_0 = const()[name = tensor("xi_19_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_19_squeeze_mask_0 = const()[name = tensor("xi_19_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_19_cast_fp16 = slice_by_index(begin = xi_19_begin_0, end = xi_19_end_0, end_mask = xi_19_end_mask_0, squeeze_mask = xi_19_squeeze_mask_0, x = pairs_19_cast_fp16)[name = tensor("xi_19_cast_fp16")]; + tensor var_1298_cast_fp16 = mul(x = xr_19_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_1298_cast_fp16")]; + tensor var_1299_cast_fp16 = mul(x = xi_19_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_1299_cast_fp16")]; + tensor xor__19_cast_fp16 = sub(x = var_1298_cast_fp16, y = var_1299_cast_fp16)[name = tensor("xor__19_cast_fp16")]; + tensor var_1302_cast_fp16 = mul(x = xr_19_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_1302_cast_fp16")]; + tensor var_1303_cast_fp16 = mul(x = xi_19_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_1303_cast_fp16")]; + tensor xoi_19_cast_fp16 = add(x = var_1302_cast_fp16, y = var_1303_cast_fp16)[name = tensor("xoi_19_cast_fp16")]; + tensor out_19_axis_0 = const()[name = tensor("out_19_axis_0"), val = tensor(-1)]; + tensor out_19_cast_fp16 = stack(axis = out_19_axis_0, values = (xor__19_cast_fp16, xoi_19_cast_fp16))[name = tensor("out_19_cast_fp16")]; + tensor _inversed_1315_y_0_to_fp16 = const()[name = tensor("_inversed_1315_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1315_cast_fp16 = mul(x = position4_to_fp16, y = _inversed_1315_y_0_to_fp16)[name = tensor("_inversed_1315_cast_fp16")]; + tensor var_1316_cast_fp16 = floor(x = _inversed_1315_cast_fp16)[name = tensor("op_1316_cast_fp16")]; + tensor var_1317_to_fp16 = const()[name = tensor("op_1317_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1318_cast_fp16 = mul(x = var_1316_cast_fp16, y = var_1317_to_fp16)[name = tensor("op_1318_cast_fp16")]; + tensor write_idx_9_cast_fp16 = sub(x = position4_to_fp16, y = var_1318_cast_fp16)[name = tensor("write_idx_9_cast_fp16")]; + tensor var_1321_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_9_cast_fp16)[name = tensor("op_1321_cast_fp16")]; + tensor var_1331 = const()[name = tensor("op_1331"), val = tensor([1, 512, 1, 1])]; + tensor onehot_17_to_fp16_dtype_0 = const()[name = tensor("onehot_17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_1321_cast_fp16_to_fp16 = cast(dtype = onehot_17_to_fp16_dtype_0, x = var_1321_cast_fp16)[name = tensor("cast_544")]; + tensor onehot_19_cast_fp16 = reshape(shape = var_1331, x = var_1321_cast_fp16_to_fp16)[name = tensor("onehot_19_cast_fp16")]; + tensor var_1333_to_fp16 = const()[name = tensor("op_1333_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1335_cast_fp16 = sub(x = var_1333_to_fp16, y = onehot_19_cast_fp16)[name = tensor("op_1335_cast_fp16")]; + tensor k_cache4_to_fp16_dtype_0 = const()[name = tensor("k_cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache4_to_fp16 = cast(dtype = k_cache4_to_fp16_dtype_0, x = k_cache4)[name = tensor("cast_543")]; + tensor var_1336_cast_fp16 = mul(x = k_cache4_to_fp16, y = var_1335_cast_fp16)[name = tensor("op_1336_cast_fp16")]; + tensor var_1341 = const()[name = tensor("op_1341"), val = tensor([1, 1, 16, 64])]; + tensor var_1342_cast_fp16 = reshape(shape = var_1341, x = out_19_cast_fp16)[name = tensor("op_1342_cast_fp16")]; + tensor var_1343_cast_fp16 = mul(x = var_1342_cast_fp16, y = onehot_19_cast_fp16)[name = tensor("op_1343_cast_fp16")]; + tensor new_k_cache_9_cast_fp16 = add(x = var_1336_cast_fp16, y = var_1343_cast_fp16)[name = tensor("new_k_cache_9_cast_fp16")]; + tensor new_k_cache_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache4_to_fp16_dtype_0 = const()[name = tensor("v_cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache4_to_fp16 = cast(dtype = v_cache4_to_fp16_dtype_0, x = v_cache4)[name = tensor("cast_541")]; + tensor var_1349_cast_fp16 = mul(x = v_cache4_to_fp16, y = var_1335_cast_fp16)[name = tensor("op_1349_cast_fp16")]; + tensor var_1354 = const()[name = tensor("op_1354"), val = tensor([1, 1, 16, 64])]; + tensor var_1355_cast_fp16 = reshape(shape = var_1354, x = v_9_cast_fp16)[name = tensor("op_1355_cast_fp16")]; + tensor var_1356_cast_fp16 = mul(x = var_1355_cast_fp16, y = onehot_19_cast_fp16)[name = tensor("op_1356_cast_fp16")]; + tensor new_v_cache_9_cast_fp16 = add(x = var_1349_cast_fp16, y = var_1356_cast_fp16)[name = tensor("new_v_cache_9_cast_fp16")]; + tensor new_v_cache_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_4_perm_0 = const()[name = tensor("transpose_4_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_1369 = const()[name = tensor("op_1369"), val = tensor([0, 2, 1, 3])]; + tensor var_1375 = const()[name = tensor("op_1375"), val = tensor([1, 16, 1, 64])]; + tensor q4_9_cast_fp16 = reshape(shape = var_1375, x = out_17_cast_fp16)[name = tensor("q4_9_cast_fp16")]; + tensor var_1380_transpose_x_0 = const()[name = tensor("op_1380_transpose_x_0"), val = tensor(false)]; + tensor var_1380_transpose_y_0 = const()[name = tensor("op_1380_transpose_y_0"), val = tensor(false)]; + tensor transpose_4_cast_fp16 = transpose(perm = transpose_4_perm_0, x = new_k_cache_9_cast_fp16)[name = tensor("transpose_83")]; + tensor var_1380_cast_fp16 = matmul(transpose_x = var_1380_transpose_x_0, transpose_y = var_1380_transpose_y_0, x = q4_9_cast_fp16, y = transpose_4_cast_fp16)[name = tensor("op_1380_cast_fp16")]; + tensor var_1381_to_fp16 = const()[name = tensor("op_1381_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_17_cast_fp16 = mul(x = var_1380_cast_fp16, y = var_1381_to_fp16)[name = tensor("scores_17_cast_fp16")]; + tensor var_1385 = const()[name = tensor("op_1385"), val = tensor([1, 1])]; + tensor var_1386_cast_fp16 = reshape(shape = var_1385, x = position4_to_fp16)[name = tensor("op_1386_cast_fp16")]; + tensor var_1387_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_1386_cast_fp16)[name = tensor("op_1387_cast_fp16")]; + tensor mask_9_to_fp16_dtype_0 = const()[name = tensor("mask_9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_1394_to_fp16 = const()[name = tensor("op_1394_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1387_cast_fp16_to_fp16 = cast(dtype = mask_9_to_fp16_dtype_0, x = var_1387_cast_fp16)[name = tensor("cast_539")]; + tensor var_1395_cast_fp16 = sub(x = var_1387_cast_fp16_to_fp16, y = var_1394_to_fp16)[name = tensor("op_1395_cast_fp16")]; + tensor var_1400 = const()[name = tensor("op_1400"), val = tensor([1, 1, 1, 512])]; + tensor var_1401_cast_fp16 = reshape(shape = var_1400, x = var_1395_cast_fp16)[name = tensor("op_1401_cast_fp16")]; + tensor var_1402_to_fp16 = const()[name = tensor("op_1402_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_1403_cast_fp16 = mul(x = var_1401_cast_fp16, y = var_1402_to_fp16)[name = tensor("op_1403_cast_fp16")]; + tensor scores_19_cast_fp16 = add(x = scores_17_cast_fp16, y = var_1403_cast_fp16)[name = tensor("scores_19_cast_fp16")]; + tensor var_1406 = const()[name = tensor("op_1406"), val = tensor(-1)]; + tensor weights_9_cast_fp16 = softmax(axis = var_1406, x = scores_19_cast_fp16)[name = tensor("weights_9_cast_fp16")]; + tensor attn_9_transpose_x_0 = const()[name = tensor("attn_9_transpose_x_0"), val = tensor(false)]; + tensor attn_9_transpose_y_0 = const()[name = tensor("attn_9_transpose_y_0"), val = tensor(false)]; + tensor values_9_cast_fp16 = transpose(perm = var_1369, x = new_v_cache_9_cast_fp16)[name = tensor("transpose_82")]; + tensor attn_9_cast_fp16 = matmul(transpose_x = attn_9_transpose_x_0, transpose_y = attn_9_transpose_y_0, x = weights_9_cast_fp16, y = values_9_cast_fp16)[name = tensor("attn_9_cast_fp16")]; + tensor var_1414 = const()[name = tensor("op_1414"), val = tensor([0, 2, 1, 3])]; + tensor var_1419 = const()[name = tensor("op_1419"), val = tensor([1, 1, 1024])]; + tensor var_1415_cast_fp16 = transpose(perm = var_1414, x = attn_9_cast_fp16)[name = tensor("transpose_81")]; + tensor input_55_cast_fp16 = reshape(shape = var_1419, x = var_1415_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor attn4_out_proj_weight_to_fp16 = const()[name = tensor("attn4_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(107077312)))]; + tensor linear_18_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn4_out_proj_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor var_1425_to_fp16 = const()[name = tensor("op_1425_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1426_cast_fp16 = add(x = position4_to_fp16, y = var_1425_to_fp16)[name = tensor("op_1426_cast_fp16")]; + tensor var_1426_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1426_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_57_cast_fp16 = add(x = input_51_cast_fp16, y = linear_18_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor input_59_axes_0 = const()[name = tensor("input_59_axes_0"), val = tensor([-1])]; + tensor norm4_2_weight_to_fp16 = const()[name = tensor("norm4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109174528)))]; + tensor norm4_2_bias_to_fp16 = const()[name = tensor("norm4_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109176640)))]; + tensor var_1430_to_fp16 = const()[name = tensor("op_1430_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_59_cast_fp16 = layer_norm(axes = input_59_axes_0, beta = norm4_2_bias_to_fp16, epsilon = var_1430_to_fp16, gamma = norm4_2_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor linear4_1_weight_to_fp16 = const()[name = tensor("linear4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109178752)))]; + tensor linear_19_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear4_1_weight_to_fp16, x = input_59_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_61_mode_0 = const()[name = tensor("input_61_mode_0"), val = tensor("EXACT")]; + tensor input_61_cast_fp16 = gelu(mode = input_61_mode_0, x = linear_19_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor linear4_2_weight_to_fp16 = const()[name = tensor("linear4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(117567424)))]; + tensor linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear4_2_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor input_63_cast_fp16 = add(x = input_57_cast_fp16, y = linear_20_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor input_65_axes_0 = const()[name = tensor("input_65_axes_0"), val = tensor([-1])]; + tensor norm5_1_weight_to_fp16 = const()[name = tensor("norm5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125956096)))]; + tensor norm5_1_bias_to_fp16 = const()[name = tensor("norm5_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125958208)))]; + tensor var_1447_to_fp16 = const()[name = tensor("op_1447_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_65_cast_fp16 = layer_norm(axes = input_65_axes_0, beta = norm5_1_bias_to_fp16, epsilon = var_1447_to_fp16, gamma = norm5_1_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("input_65_cast_fp16")]; + tensor attn5_in_proj_weight_to_fp16 = const()[name = tensor("attn5_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125960320)))]; + tensor linear_21_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn5_in_proj_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("linear_21_cast_fp16")]; + tensor var_1460 = const()[name = tensor("op_1460"), val = tensor([1, 3, 16, 64])]; + tensor qkv_11_cast_fp16 = reshape(shape = var_1460, x = linear_21_cast_fp16)[name = tensor("qkv_11_cast_fp16")]; + tensor x_21_begin_0 = const()[name = tensor("x_21_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_21_end_0 = const()[name = tensor("x_21_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_21_end_mask_0 = const()[name = tensor("x_21_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_21_squeeze_mask_0 = const()[name = tensor("x_21_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_21_cast_fp16 = slice_by_index(begin = x_21_begin_0, end = x_21_end_0, end_mask = x_21_end_mask_0, squeeze_mask = x_21_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("x_21_cast_fp16")]; + tensor x_23_begin_0 = const()[name = tensor("x_23_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_23_end_0 = const()[name = tensor("x_23_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_23_end_mask_0 = const()[name = tensor("x_23_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_23_squeeze_mask_0 = const()[name = tensor("x_23_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_23_cast_fp16 = slice_by_index(begin = x_23_begin_0, end = x_23_end_0, end_mask = x_23_end_mask_0, squeeze_mask = x_23_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("x_23_cast_fp16")]; + tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_11_cast_fp16 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("v_11_cast_fp16")]; + tensor var_1489 = const()[name = tensor("op_1489"), val = tensor([1, 1, 1])]; + tensor position5_to_fp16_dtype_0 = const()[name = tensor("position5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position5_to_fp16 = cast(dtype = position5_to_fp16_dtype_0, x = position5)[name = tensor("cast_537")]; + tensor var_1490_cast_fp16 = reshape(shape = var_1489, x = position5_to_fp16)[name = tensor("op_1490_cast_fp16")]; + tensor angles_21_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_1490_cast_fp16)[name = tensor("angles_21_cast_fp16")]; + tensor rotr_21_cast_fp16 = cos(x = angles_21_cast_fp16)[name = tensor("rotr_21_cast_fp16")]; + tensor roti_21_cast_fp16 = sin(x = angles_21_cast_fp16)[name = tensor("roti_21_cast_fp16")]; + tensor var_1498 = const()[name = tensor("op_1498"), val = tensor([1, 16, 32, 2])]; + tensor pairs_21_cast_fp16 = reshape(shape = var_1498, x = x_21_cast_fp16)[name = tensor("pairs_21_cast_fp16")]; + tensor xr_21_begin_0 = const()[name = tensor("xr_21_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_21_end_0 = const()[name = tensor("xr_21_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_21_end_mask_0 = const()[name = tensor("xr_21_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_21_squeeze_mask_0 = const()[name = tensor("xr_21_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_21_cast_fp16 = slice_by_index(begin = xr_21_begin_0, end = xr_21_end_0, end_mask = xr_21_end_mask_0, squeeze_mask = xr_21_squeeze_mask_0, x = pairs_21_cast_fp16)[name = tensor("xr_21_cast_fp16")]; + tensor xi_21_begin_0 = const()[name = tensor("xi_21_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_21_end_0 = const()[name = tensor("xi_21_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_21_end_mask_0 = const()[name = tensor("xi_21_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_21_squeeze_mask_0 = const()[name = tensor("xi_21_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_21_cast_fp16 = slice_by_index(begin = xi_21_begin_0, end = xi_21_end_0, end_mask = xi_21_end_mask_0, squeeze_mask = xi_21_squeeze_mask_0, x = pairs_21_cast_fp16)[name = tensor("xi_21_cast_fp16")]; + tensor var_1506_cast_fp16 = mul(x = xr_21_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_1506_cast_fp16")]; + tensor var_1507_cast_fp16 = mul(x = xi_21_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_1507_cast_fp16")]; + tensor xor__21_cast_fp16 = sub(x = var_1506_cast_fp16, y = var_1507_cast_fp16)[name = tensor("xor__21_cast_fp16")]; + tensor var_1510_cast_fp16 = mul(x = xr_21_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_1510_cast_fp16")]; + tensor var_1511_cast_fp16 = mul(x = xi_21_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_1511_cast_fp16")]; + tensor xoi_21_cast_fp16 = add(x = var_1510_cast_fp16, y = var_1511_cast_fp16)[name = tensor("xoi_21_cast_fp16")]; + tensor out_21_axis_0 = const()[name = tensor("out_21_axis_0"), val = tensor(-1)]; + tensor out_21_cast_fp16 = stack(axis = out_21_axis_0, values = (xor__21_cast_fp16, xoi_21_cast_fp16))[name = tensor("out_21_cast_fp16")]; + tensor var_1534 = const()[name = tensor("op_1534"), val = tensor([1, 16, 32, 2])]; + tensor pairs_23_cast_fp16 = reshape(shape = var_1534, x = x_23_cast_fp16)[name = tensor("pairs_23_cast_fp16")]; + tensor xr_23_begin_0 = const()[name = tensor("xr_23_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_23_end_0 = const()[name = tensor("xr_23_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_23_end_mask_0 = const()[name = tensor("xr_23_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_23_squeeze_mask_0 = const()[name = tensor("xr_23_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_23_cast_fp16 = slice_by_index(begin = xr_23_begin_0, end = xr_23_end_0, end_mask = xr_23_end_mask_0, squeeze_mask = xr_23_squeeze_mask_0, x = pairs_23_cast_fp16)[name = tensor("xr_23_cast_fp16")]; + tensor xi_23_begin_0 = const()[name = tensor("xi_23_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_23_end_0 = const()[name = tensor("xi_23_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_23_end_mask_0 = const()[name = tensor("xi_23_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_23_squeeze_mask_0 = const()[name = tensor("xi_23_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_23_cast_fp16 = slice_by_index(begin = xi_23_begin_0, end = xi_23_end_0, end_mask = xi_23_end_mask_0, squeeze_mask = xi_23_squeeze_mask_0, x = pairs_23_cast_fp16)[name = tensor("xi_23_cast_fp16")]; + tensor var_1542_cast_fp16 = mul(x = xr_23_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_1542_cast_fp16")]; + tensor var_1543_cast_fp16 = mul(x = xi_23_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_1543_cast_fp16")]; + tensor xor__23_cast_fp16 = sub(x = var_1542_cast_fp16, y = var_1543_cast_fp16)[name = tensor("xor__23_cast_fp16")]; + tensor var_1546_cast_fp16 = mul(x = xr_23_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_1546_cast_fp16")]; + tensor var_1547_cast_fp16 = mul(x = xi_23_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_1547_cast_fp16")]; + tensor xoi_23_cast_fp16 = add(x = var_1546_cast_fp16, y = var_1547_cast_fp16)[name = tensor("xoi_23_cast_fp16")]; + tensor out_23_axis_0 = const()[name = tensor("out_23_axis_0"), val = tensor(-1)]; + tensor out_23_cast_fp16 = stack(axis = out_23_axis_0, values = (xor__23_cast_fp16, xoi_23_cast_fp16))[name = tensor("out_23_cast_fp16")]; + tensor _inversed_1559_y_0_to_fp16 = const()[name = tensor("_inversed_1559_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1559_cast_fp16 = mul(x = position5_to_fp16, y = _inversed_1559_y_0_to_fp16)[name = tensor("_inversed_1559_cast_fp16")]; + tensor var_1560_cast_fp16 = floor(x = _inversed_1559_cast_fp16)[name = tensor("op_1560_cast_fp16")]; + tensor var_1561_to_fp16 = const()[name = tensor("op_1561_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1562_cast_fp16 = mul(x = var_1560_cast_fp16, y = var_1561_to_fp16)[name = tensor("op_1562_cast_fp16")]; + tensor write_idx_11_cast_fp16 = sub(x = position5_to_fp16, y = var_1562_cast_fp16)[name = tensor("write_idx_11_cast_fp16")]; + tensor var_1565_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_11_cast_fp16)[name = tensor("op_1565_cast_fp16")]; + tensor var_1575 = const()[name = tensor("op_1575"), val = tensor([1, 512, 1, 1])]; + tensor onehot_21_to_fp16_dtype_0 = const()[name = tensor("onehot_21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_1565_cast_fp16_to_fp16 = cast(dtype = onehot_21_to_fp16_dtype_0, x = var_1565_cast_fp16)[name = tensor("cast_536")]; + tensor onehot_23_cast_fp16 = reshape(shape = var_1575, x = var_1565_cast_fp16_to_fp16)[name = tensor("onehot_23_cast_fp16")]; + tensor var_1577_to_fp16 = const()[name = tensor("op_1577_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1579_cast_fp16 = sub(x = var_1577_to_fp16, y = onehot_23_cast_fp16)[name = tensor("op_1579_cast_fp16")]; + tensor k_cache5_to_fp16_dtype_0 = const()[name = tensor("k_cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache5_to_fp16 = cast(dtype = k_cache5_to_fp16_dtype_0, x = k_cache5)[name = tensor("cast_535")]; + tensor var_1580_cast_fp16 = mul(x = k_cache5_to_fp16, y = var_1579_cast_fp16)[name = tensor("op_1580_cast_fp16")]; + tensor var_1585 = const()[name = tensor("op_1585"), val = tensor([1, 1, 16, 64])]; + tensor var_1586_cast_fp16 = reshape(shape = var_1585, x = out_23_cast_fp16)[name = tensor("op_1586_cast_fp16")]; + tensor var_1587_cast_fp16 = mul(x = var_1586_cast_fp16, y = onehot_23_cast_fp16)[name = tensor("op_1587_cast_fp16")]; + tensor new_k_cache_11_cast_fp16 = add(x = var_1580_cast_fp16, y = var_1587_cast_fp16)[name = tensor("new_k_cache_11_cast_fp16")]; + tensor new_k_cache_11_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_11_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache5_to_fp16_dtype_0 = const()[name = tensor("v_cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache5_to_fp16 = cast(dtype = v_cache5_to_fp16_dtype_0, x = v_cache5)[name = tensor("cast_533")]; + tensor var_1593_cast_fp16 = mul(x = v_cache5_to_fp16, y = var_1579_cast_fp16)[name = tensor("op_1593_cast_fp16")]; + tensor var_1598 = const()[name = tensor("op_1598"), val = tensor([1, 1, 16, 64])]; + tensor var_1599_cast_fp16 = reshape(shape = var_1598, x = v_11_cast_fp16)[name = tensor("op_1599_cast_fp16")]; + tensor var_1600_cast_fp16 = mul(x = var_1599_cast_fp16, y = onehot_23_cast_fp16)[name = tensor("op_1600_cast_fp16")]; + tensor new_v_cache_11_cast_fp16 = add(x = var_1593_cast_fp16, y = var_1600_cast_fp16)[name = tensor("new_v_cache_11_cast_fp16")]; + tensor new_v_cache_11_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_11_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_5_perm_0 = const()[name = tensor("transpose_5_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_1613 = const()[name = tensor("op_1613"), val = tensor([0, 2, 1, 3])]; + tensor var_1619 = const()[name = tensor("op_1619"), val = tensor([1, 16, 1, 64])]; + tensor q4_11_cast_fp16 = reshape(shape = var_1619, x = out_21_cast_fp16)[name = tensor("q4_11_cast_fp16")]; + tensor var_1624_transpose_x_0 = const()[name = tensor("op_1624_transpose_x_0"), val = tensor(false)]; + tensor var_1624_transpose_y_0 = const()[name = tensor("op_1624_transpose_y_0"), val = tensor(false)]; + tensor transpose_5_cast_fp16 = transpose(perm = transpose_5_perm_0, x = new_k_cache_11_cast_fp16)[name = tensor("transpose_80")]; + tensor var_1624_cast_fp16 = matmul(transpose_x = var_1624_transpose_x_0, transpose_y = var_1624_transpose_y_0, x = q4_11_cast_fp16, y = transpose_5_cast_fp16)[name = tensor("op_1624_cast_fp16")]; + tensor var_1625_to_fp16 = const()[name = tensor("op_1625_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_21_cast_fp16 = mul(x = var_1624_cast_fp16, y = var_1625_to_fp16)[name = tensor("scores_21_cast_fp16")]; + tensor var_1629 = const()[name = tensor("op_1629"), val = tensor([1, 1])]; + tensor var_1630_cast_fp16 = reshape(shape = var_1629, x = position5_to_fp16)[name = tensor("op_1630_cast_fp16")]; + tensor var_1631_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_1630_cast_fp16)[name = tensor("op_1631_cast_fp16")]; + tensor mask_11_to_fp16_dtype_0 = const()[name = tensor("mask_11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_1638_to_fp16 = const()[name = tensor("op_1638_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1631_cast_fp16_to_fp16 = cast(dtype = mask_11_to_fp16_dtype_0, x = var_1631_cast_fp16)[name = tensor("cast_531")]; + tensor var_1639_cast_fp16 = sub(x = var_1631_cast_fp16_to_fp16, y = var_1638_to_fp16)[name = tensor("op_1639_cast_fp16")]; + tensor var_1644 = const()[name = tensor("op_1644"), val = tensor([1, 1, 1, 512])]; + tensor var_1645_cast_fp16 = reshape(shape = var_1644, x = var_1639_cast_fp16)[name = tensor("op_1645_cast_fp16")]; + tensor var_1646_to_fp16 = const()[name = tensor("op_1646_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_1647_cast_fp16 = mul(x = var_1645_cast_fp16, y = var_1646_to_fp16)[name = tensor("op_1647_cast_fp16")]; + tensor scores_23_cast_fp16 = add(x = scores_21_cast_fp16, y = var_1647_cast_fp16)[name = tensor("scores_23_cast_fp16")]; + tensor var_1650 = const()[name = tensor("op_1650"), val = tensor(-1)]; + tensor weights_11_cast_fp16 = softmax(axis = var_1650, x = scores_23_cast_fp16)[name = tensor("weights_11_cast_fp16")]; + tensor attn_11_transpose_x_0 = const()[name = tensor("attn_11_transpose_x_0"), val = tensor(false)]; + tensor attn_11_transpose_y_0 = const()[name = tensor("attn_11_transpose_y_0"), val = tensor(false)]; + tensor values_11_cast_fp16 = transpose(perm = var_1613, x = new_v_cache_11_cast_fp16)[name = tensor("transpose_79")]; + tensor attn_11_cast_fp16 = matmul(transpose_x = attn_11_transpose_x_0, transpose_y = attn_11_transpose_y_0, x = weights_11_cast_fp16, y = values_11_cast_fp16)[name = tensor("attn_11_cast_fp16")]; + tensor var_1658 = const()[name = tensor("op_1658"), val = tensor([0, 2, 1, 3])]; + tensor var_1663 = const()[name = tensor("op_1663"), val = tensor([1, 1, 1024])]; + tensor var_1659_cast_fp16 = transpose(perm = var_1658, x = attn_11_cast_fp16)[name = tensor("transpose_78")]; + tensor input_67_cast_fp16 = reshape(shape = var_1663, x = var_1659_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor attn5_out_proj_weight_to_fp16 = const()[name = tensor("attn5_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132251840)))]; + tensor linear_22_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn5_out_proj_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("linear_22_cast_fp16")]; + tensor var_1669_to_fp16 = const()[name = tensor("op_1669_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1670_cast_fp16 = add(x = position5_to_fp16, y = var_1669_to_fp16)[name = tensor("op_1670_cast_fp16")]; + tensor var_1670_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1670_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_69_cast_fp16 = add(x = input_63_cast_fp16, y = linear_22_cast_fp16)[name = tensor("input_69_cast_fp16")]; + tensor input_71_axes_0 = const()[name = tensor("input_71_axes_0"), val = tensor([-1])]; + tensor norm5_2_weight_to_fp16 = const()[name = tensor("norm5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134349056)))]; + tensor norm5_2_bias_to_fp16 = const()[name = tensor("norm5_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134351168)))]; + tensor var_1674_to_fp16 = const()[name = tensor("op_1674_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_71_cast_fp16 = layer_norm(axes = input_71_axes_0, beta = norm5_2_bias_to_fp16, epsilon = var_1674_to_fp16, gamma = norm5_2_weight_to_fp16, x = input_69_cast_fp16)[name = tensor("input_71_cast_fp16")]; + tensor linear5_1_weight_to_fp16 = const()[name = tensor("linear5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134353280)))]; + tensor linear_23_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear5_1_weight_to_fp16, x = input_71_cast_fp16)[name = tensor("linear_23_cast_fp16")]; + tensor input_73_mode_0 = const()[name = tensor("input_73_mode_0"), val = tensor("EXACT")]; + tensor input_73_cast_fp16 = gelu(mode = input_73_mode_0, x = linear_23_cast_fp16)[name = tensor("input_73_cast_fp16")]; + tensor linear5_2_weight_to_fp16 = const()[name = tensor("linear5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142741952)))]; + tensor linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear5_2_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("linear_24_cast_fp16")]; + tensor input_75_cast_fp16 = add(x = input_69_cast_fp16, y = linear_24_cast_fp16)[name = tensor("input_75_cast_fp16")]; + tensor input_77_axes_0 = const()[name = tensor("input_77_axes_0"), val = tensor([-1])]; + tensor norm6_1_weight_to_fp16 = const()[name = tensor("norm6_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151130624)))]; + tensor norm6_1_bias_to_fp16 = const()[name = tensor("norm6_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151132736)))]; + tensor var_1691_to_fp16 = const()[name = tensor("op_1691_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_77_cast_fp16 = layer_norm(axes = input_77_axes_0, beta = norm6_1_bias_to_fp16, epsilon = var_1691_to_fp16, gamma = norm6_1_weight_to_fp16, x = input_75_cast_fp16)[name = tensor("input_77_cast_fp16")]; + tensor attn6_in_proj_weight_to_fp16 = const()[name = tensor("attn6_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151134848)))]; + tensor linear_25_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn6_in_proj_weight_to_fp16, x = input_77_cast_fp16)[name = tensor("linear_25_cast_fp16")]; + tensor var_1704 = const()[name = tensor("op_1704"), val = tensor([1, 3, 16, 64])]; + tensor qkv_13_cast_fp16 = reshape(shape = var_1704, x = linear_25_cast_fp16)[name = tensor("qkv_13_cast_fp16")]; + tensor x_25_begin_0 = const()[name = tensor("x_25_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_25_end_0 = const()[name = tensor("x_25_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_25_end_mask_0 = const()[name = tensor("x_25_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_25_squeeze_mask_0 = const()[name = tensor("x_25_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_25_cast_fp16 = slice_by_index(begin = x_25_begin_0, end = x_25_end_0, end_mask = x_25_end_mask_0, squeeze_mask = x_25_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("x_25_cast_fp16")]; + tensor x_27_begin_0 = const()[name = tensor("x_27_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_27_end_0 = const()[name = tensor("x_27_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_27_end_mask_0 = const()[name = tensor("x_27_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_27_squeeze_mask_0 = const()[name = tensor("x_27_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_27_cast_fp16 = slice_by_index(begin = x_27_begin_0, end = x_27_end_0, end_mask = x_27_end_mask_0, squeeze_mask = x_27_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("x_27_cast_fp16")]; + tensor v_13_begin_0 = const()[name = tensor("v_13_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_13_end_0 = const()[name = tensor("v_13_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_13_end_mask_0 = const()[name = tensor("v_13_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_13_squeeze_mask_0 = const()[name = tensor("v_13_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_13_cast_fp16 = slice_by_index(begin = v_13_begin_0, end = v_13_end_0, end_mask = v_13_end_mask_0, squeeze_mask = v_13_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("v_13_cast_fp16")]; + tensor var_1733 = const()[name = tensor("op_1733"), val = tensor([1, 1, 1])]; + tensor position6_to_fp16_dtype_0 = const()[name = tensor("position6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position6_to_fp16 = cast(dtype = position6_to_fp16_dtype_0, x = position6)[name = tensor("cast_529")]; + tensor var_1734_cast_fp16 = reshape(shape = var_1733, x = position6_to_fp16)[name = tensor("op_1734_cast_fp16")]; + tensor angles_25_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_1734_cast_fp16)[name = tensor("angles_25_cast_fp16")]; + tensor rotr_25_cast_fp16 = cos(x = angles_25_cast_fp16)[name = tensor("rotr_25_cast_fp16")]; + tensor roti_25_cast_fp16 = sin(x = angles_25_cast_fp16)[name = tensor("roti_25_cast_fp16")]; + tensor var_1742 = const()[name = tensor("op_1742"), val = tensor([1, 16, 32, 2])]; + tensor pairs_25_cast_fp16 = reshape(shape = var_1742, x = x_25_cast_fp16)[name = tensor("pairs_25_cast_fp16")]; + tensor xr_25_begin_0 = const()[name = tensor("xr_25_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_25_end_0 = const()[name = tensor("xr_25_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_25_end_mask_0 = const()[name = tensor("xr_25_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_25_squeeze_mask_0 = const()[name = tensor("xr_25_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_25_cast_fp16 = slice_by_index(begin = xr_25_begin_0, end = xr_25_end_0, end_mask = xr_25_end_mask_0, squeeze_mask = xr_25_squeeze_mask_0, x = pairs_25_cast_fp16)[name = tensor("xr_25_cast_fp16")]; + tensor xi_25_begin_0 = const()[name = tensor("xi_25_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_25_end_0 = const()[name = tensor("xi_25_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_25_end_mask_0 = const()[name = tensor("xi_25_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_25_squeeze_mask_0 = const()[name = tensor("xi_25_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_25_cast_fp16 = slice_by_index(begin = xi_25_begin_0, end = xi_25_end_0, end_mask = xi_25_end_mask_0, squeeze_mask = xi_25_squeeze_mask_0, x = pairs_25_cast_fp16)[name = tensor("xi_25_cast_fp16")]; + tensor var_1750_cast_fp16 = mul(x = xr_25_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_1750_cast_fp16")]; + tensor var_1751_cast_fp16 = mul(x = xi_25_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_1751_cast_fp16")]; + tensor xor__25_cast_fp16 = sub(x = var_1750_cast_fp16, y = var_1751_cast_fp16)[name = tensor("xor__25_cast_fp16")]; + tensor var_1754_cast_fp16 = mul(x = xr_25_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_1754_cast_fp16")]; + tensor var_1755_cast_fp16 = mul(x = xi_25_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_1755_cast_fp16")]; + tensor xoi_25_cast_fp16 = add(x = var_1754_cast_fp16, y = var_1755_cast_fp16)[name = tensor("xoi_25_cast_fp16")]; + tensor out_25_axis_0 = const()[name = tensor("out_25_axis_0"), val = tensor(-1)]; + tensor out_25_cast_fp16 = stack(axis = out_25_axis_0, values = (xor__25_cast_fp16, xoi_25_cast_fp16))[name = tensor("out_25_cast_fp16")]; + tensor var_1778 = const()[name = tensor("op_1778"), val = tensor([1, 16, 32, 2])]; + tensor pairs_27_cast_fp16 = reshape(shape = var_1778, x = x_27_cast_fp16)[name = tensor("pairs_27_cast_fp16")]; + tensor xr_27_begin_0 = const()[name = tensor("xr_27_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_27_end_0 = const()[name = tensor("xr_27_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_27_end_mask_0 = const()[name = tensor("xr_27_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_27_squeeze_mask_0 = const()[name = tensor("xr_27_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_27_cast_fp16 = slice_by_index(begin = xr_27_begin_0, end = xr_27_end_0, end_mask = xr_27_end_mask_0, squeeze_mask = xr_27_squeeze_mask_0, x = pairs_27_cast_fp16)[name = tensor("xr_27_cast_fp16")]; + tensor xi_27_begin_0 = const()[name = tensor("xi_27_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_27_end_0 = const()[name = tensor("xi_27_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_27_end_mask_0 = const()[name = tensor("xi_27_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_27_squeeze_mask_0 = const()[name = tensor("xi_27_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_27_cast_fp16 = slice_by_index(begin = xi_27_begin_0, end = xi_27_end_0, end_mask = xi_27_end_mask_0, squeeze_mask = xi_27_squeeze_mask_0, x = pairs_27_cast_fp16)[name = tensor("xi_27_cast_fp16")]; + tensor var_1786_cast_fp16 = mul(x = xr_27_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_1786_cast_fp16")]; + tensor var_1787_cast_fp16 = mul(x = xi_27_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_1787_cast_fp16")]; + tensor xor__27_cast_fp16 = sub(x = var_1786_cast_fp16, y = var_1787_cast_fp16)[name = tensor("xor__27_cast_fp16")]; + tensor var_1790_cast_fp16 = mul(x = xr_27_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_1790_cast_fp16")]; + tensor var_1791_cast_fp16 = mul(x = xi_27_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_1791_cast_fp16")]; + tensor xoi_27_cast_fp16 = add(x = var_1790_cast_fp16, y = var_1791_cast_fp16)[name = tensor("xoi_27_cast_fp16")]; + tensor out_27_axis_0 = const()[name = tensor("out_27_axis_0"), val = tensor(-1)]; + tensor out_27_cast_fp16 = stack(axis = out_27_axis_0, values = (xor__27_cast_fp16, xoi_27_cast_fp16))[name = tensor("out_27_cast_fp16")]; + tensor _inversed_1803_y_0_to_fp16 = const()[name = tensor("_inversed_1803_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1803_cast_fp16 = mul(x = position6_to_fp16, y = _inversed_1803_y_0_to_fp16)[name = tensor("_inversed_1803_cast_fp16")]; + tensor var_1804_cast_fp16 = floor(x = _inversed_1803_cast_fp16)[name = tensor("op_1804_cast_fp16")]; + tensor var_1805_to_fp16 = const()[name = tensor("op_1805_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1806_cast_fp16 = mul(x = var_1804_cast_fp16, y = var_1805_to_fp16)[name = tensor("op_1806_cast_fp16")]; + tensor write_idx_13_cast_fp16 = sub(x = position6_to_fp16, y = var_1806_cast_fp16)[name = tensor("write_idx_13_cast_fp16")]; + tensor var_1809_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_13_cast_fp16)[name = tensor("op_1809_cast_fp16")]; + tensor var_1819 = const()[name = tensor("op_1819"), val = tensor([1, 512, 1, 1])]; + tensor onehot_25_to_fp16_dtype_0 = const()[name = tensor("onehot_25_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_1809_cast_fp16_to_fp16 = cast(dtype = onehot_25_to_fp16_dtype_0, x = var_1809_cast_fp16)[name = tensor("cast_528")]; + tensor onehot_27_cast_fp16 = reshape(shape = var_1819, x = var_1809_cast_fp16_to_fp16)[name = tensor("onehot_27_cast_fp16")]; + tensor var_1821_to_fp16 = const()[name = tensor("op_1821_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1823_cast_fp16 = sub(x = var_1821_to_fp16, y = onehot_27_cast_fp16)[name = tensor("op_1823_cast_fp16")]; + tensor k_cache6_to_fp16_dtype_0 = const()[name = tensor("k_cache6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache6_to_fp16 = cast(dtype = k_cache6_to_fp16_dtype_0, x = k_cache6)[name = tensor("cast_527")]; + tensor var_1824_cast_fp16 = mul(x = k_cache6_to_fp16, y = var_1823_cast_fp16)[name = tensor("op_1824_cast_fp16")]; + tensor var_1829 = const()[name = tensor("op_1829"), val = tensor([1, 1, 16, 64])]; + tensor var_1830_cast_fp16 = reshape(shape = var_1829, x = out_27_cast_fp16)[name = tensor("op_1830_cast_fp16")]; + tensor var_1831_cast_fp16 = mul(x = var_1830_cast_fp16, y = onehot_27_cast_fp16)[name = tensor("op_1831_cast_fp16")]; + tensor new_k_cache_13_cast_fp16 = add(x = var_1824_cast_fp16, y = var_1831_cast_fp16)[name = tensor("new_k_cache_13_cast_fp16")]; + tensor new_k_cache_13_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_13_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache6_to_fp16_dtype_0 = const()[name = tensor("v_cache6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache6_to_fp16 = cast(dtype = v_cache6_to_fp16_dtype_0, x = v_cache6)[name = tensor("cast_525")]; + tensor var_1837_cast_fp16 = mul(x = v_cache6_to_fp16, y = var_1823_cast_fp16)[name = tensor("op_1837_cast_fp16")]; + tensor var_1842 = const()[name = tensor("op_1842"), val = tensor([1, 1, 16, 64])]; + tensor var_1843_cast_fp16 = reshape(shape = var_1842, x = v_13_cast_fp16)[name = tensor("op_1843_cast_fp16")]; + tensor var_1844_cast_fp16 = mul(x = var_1843_cast_fp16, y = onehot_27_cast_fp16)[name = tensor("op_1844_cast_fp16")]; + tensor new_v_cache_13_cast_fp16 = add(x = var_1837_cast_fp16, y = var_1844_cast_fp16)[name = tensor("new_v_cache_13_cast_fp16")]; + tensor new_v_cache_13_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_13_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_6_perm_0 = const()[name = tensor("transpose_6_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_1857 = const()[name = tensor("op_1857"), val = tensor([0, 2, 1, 3])]; + tensor var_1863 = const()[name = tensor("op_1863"), val = tensor([1, 16, 1, 64])]; + tensor q4_13_cast_fp16 = reshape(shape = var_1863, x = out_25_cast_fp16)[name = tensor("q4_13_cast_fp16")]; + tensor var_1868_transpose_x_0 = const()[name = tensor("op_1868_transpose_x_0"), val = tensor(false)]; + tensor var_1868_transpose_y_0 = const()[name = tensor("op_1868_transpose_y_0"), val = tensor(false)]; + tensor transpose_6_cast_fp16 = transpose(perm = transpose_6_perm_0, x = new_k_cache_13_cast_fp16)[name = tensor("transpose_77")]; + tensor var_1868_cast_fp16 = matmul(transpose_x = var_1868_transpose_x_0, transpose_y = var_1868_transpose_y_0, x = q4_13_cast_fp16, y = transpose_6_cast_fp16)[name = tensor("op_1868_cast_fp16")]; + tensor var_1869_to_fp16 = const()[name = tensor("op_1869_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_25_cast_fp16 = mul(x = var_1868_cast_fp16, y = var_1869_to_fp16)[name = tensor("scores_25_cast_fp16")]; + tensor var_1873 = const()[name = tensor("op_1873"), val = tensor([1, 1])]; + tensor var_1874_cast_fp16 = reshape(shape = var_1873, x = position6_to_fp16)[name = tensor("op_1874_cast_fp16")]; + tensor var_1875_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_1874_cast_fp16)[name = tensor("op_1875_cast_fp16")]; + tensor mask_13_to_fp16_dtype_0 = const()[name = tensor("mask_13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_1882_to_fp16 = const()[name = tensor("op_1882_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1875_cast_fp16_to_fp16 = cast(dtype = mask_13_to_fp16_dtype_0, x = var_1875_cast_fp16)[name = tensor("cast_523")]; + tensor var_1883_cast_fp16 = sub(x = var_1875_cast_fp16_to_fp16, y = var_1882_to_fp16)[name = tensor("op_1883_cast_fp16")]; + tensor var_1888 = const()[name = tensor("op_1888"), val = tensor([1, 1, 1, 512])]; + tensor var_1889_cast_fp16 = reshape(shape = var_1888, x = var_1883_cast_fp16)[name = tensor("op_1889_cast_fp16")]; + tensor var_1890_to_fp16 = const()[name = tensor("op_1890_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_1891_cast_fp16 = mul(x = var_1889_cast_fp16, y = var_1890_to_fp16)[name = tensor("op_1891_cast_fp16")]; + tensor scores_27_cast_fp16 = add(x = scores_25_cast_fp16, y = var_1891_cast_fp16)[name = tensor("scores_27_cast_fp16")]; + tensor var_1894 = const()[name = tensor("op_1894"), val = tensor(-1)]; + tensor weights_13_cast_fp16 = softmax(axis = var_1894, x = scores_27_cast_fp16)[name = tensor("weights_13_cast_fp16")]; + tensor attn_13_transpose_x_0 = const()[name = tensor("attn_13_transpose_x_0"), val = tensor(false)]; + tensor attn_13_transpose_y_0 = const()[name = tensor("attn_13_transpose_y_0"), val = tensor(false)]; + tensor values_13_cast_fp16 = transpose(perm = var_1857, x = new_v_cache_13_cast_fp16)[name = tensor("transpose_76")]; + tensor attn_13_cast_fp16 = matmul(transpose_x = attn_13_transpose_x_0, transpose_y = attn_13_transpose_y_0, x = weights_13_cast_fp16, y = values_13_cast_fp16)[name = tensor("attn_13_cast_fp16")]; + tensor var_1902 = const()[name = tensor("op_1902"), val = tensor([0, 2, 1, 3])]; + tensor var_1907 = const()[name = tensor("op_1907"), val = tensor([1, 1, 1024])]; + tensor var_1903_cast_fp16 = transpose(perm = var_1902, x = attn_13_cast_fp16)[name = tensor("transpose_75")]; + tensor input_79_cast_fp16 = reshape(shape = var_1907, x = var_1903_cast_fp16)[name = tensor("input_79_cast_fp16")]; + tensor attn6_out_proj_weight_to_fp16 = const()[name = tensor("attn6_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(157426368)))]; + tensor linear_26_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn6_out_proj_weight_to_fp16, x = input_79_cast_fp16)[name = tensor("linear_26_cast_fp16")]; + tensor var_1913_to_fp16 = const()[name = tensor("op_1913_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1914_cast_fp16 = add(x = position6_to_fp16, y = var_1913_to_fp16)[name = tensor("op_1914_cast_fp16")]; + tensor var_1914_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1914_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_81_cast_fp16 = add(x = input_75_cast_fp16, y = linear_26_cast_fp16)[name = tensor("input_81_cast_fp16")]; + tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([-1])]; + tensor norm6_2_weight_to_fp16 = const()[name = tensor("norm6_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159523584)))]; + tensor norm6_2_bias_to_fp16 = const()[name = tensor("norm6_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159525696)))]; + tensor var_1918_to_fp16 = const()[name = tensor("op_1918_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_83_cast_fp16 = layer_norm(axes = input_83_axes_0, beta = norm6_2_bias_to_fp16, epsilon = var_1918_to_fp16, gamma = norm6_2_weight_to_fp16, x = input_81_cast_fp16)[name = tensor("input_83_cast_fp16")]; + tensor linear6_1_weight_to_fp16 = const()[name = tensor("linear6_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159527808)))]; + tensor linear_27_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear6_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_27_cast_fp16")]; + tensor input_85_mode_0 = const()[name = tensor("input_85_mode_0"), val = tensor("EXACT")]; + tensor input_85_cast_fp16 = gelu(mode = input_85_mode_0, x = linear_27_cast_fp16)[name = tensor("input_85_cast_fp16")]; + tensor linear6_2_weight_to_fp16 = const()[name = tensor("linear6_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(167916480)))]; + tensor linear_28_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear6_2_weight_to_fp16, x = input_85_cast_fp16)[name = tensor("linear_28_cast_fp16")]; + tensor input_87_cast_fp16 = add(x = input_81_cast_fp16, y = linear_28_cast_fp16)[name = tensor("input_87_cast_fp16")]; + tensor input_89_axes_0 = const()[name = tensor("input_89_axes_0"), val = tensor([-1])]; + tensor norm7_1_weight_to_fp16 = const()[name = tensor("norm7_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176305152)))]; + tensor norm7_1_bias_to_fp16 = const()[name = tensor("norm7_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176307264)))]; + tensor var_1935_to_fp16 = const()[name = tensor("op_1935_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_89_cast_fp16 = layer_norm(axes = input_89_axes_0, beta = norm7_1_bias_to_fp16, epsilon = var_1935_to_fp16, gamma = norm7_1_weight_to_fp16, x = input_87_cast_fp16)[name = tensor("input_89_cast_fp16")]; + tensor attn7_in_proj_weight_to_fp16 = const()[name = tensor("attn7_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176309376)))]; + tensor linear_29_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn7_in_proj_weight_to_fp16, x = input_89_cast_fp16)[name = tensor("linear_29_cast_fp16")]; + tensor var_1948 = const()[name = tensor("op_1948"), val = tensor([1, 3, 16, 64])]; + tensor qkv_15_cast_fp16 = reshape(shape = var_1948, x = linear_29_cast_fp16)[name = tensor("qkv_15_cast_fp16")]; + tensor x_29_begin_0 = const()[name = tensor("x_29_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_29_end_0 = const()[name = tensor("x_29_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_29_end_mask_0 = const()[name = tensor("x_29_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_29_squeeze_mask_0 = const()[name = tensor("x_29_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_29_cast_fp16 = slice_by_index(begin = x_29_begin_0, end = x_29_end_0, end_mask = x_29_end_mask_0, squeeze_mask = x_29_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("x_29_cast_fp16")]; + tensor x_31_begin_0 = const()[name = tensor("x_31_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_31_end_0 = const()[name = tensor("x_31_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_31_end_mask_0 = const()[name = tensor("x_31_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_31_squeeze_mask_0 = const()[name = tensor("x_31_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_31_cast_fp16 = slice_by_index(begin = x_31_begin_0, end = x_31_end_0, end_mask = x_31_end_mask_0, squeeze_mask = x_31_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("x_31_cast_fp16")]; + tensor v_15_begin_0 = const()[name = tensor("v_15_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_15_end_0 = const()[name = tensor("v_15_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_15_end_mask_0 = const()[name = tensor("v_15_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_15_squeeze_mask_0 = const()[name = tensor("v_15_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_15_cast_fp16 = slice_by_index(begin = v_15_begin_0, end = v_15_end_0, end_mask = v_15_end_mask_0, squeeze_mask = v_15_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("v_15_cast_fp16")]; + tensor var_1977 = const()[name = tensor("op_1977"), val = tensor([1, 1, 1])]; + tensor position7_to_fp16_dtype_0 = const()[name = tensor("position7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position7_to_fp16 = cast(dtype = position7_to_fp16_dtype_0, x = position7)[name = tensor("cast_521")]; + tensor var_1978_cast_fp16 = reshape(shape = var_1977, x = position7_to_fp16)[name = tensor("op_1978_cast_fp16")]; + tensor angles_29_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_1978_cast_fp16)[name = tensor("angles_29_cast_fp16")]; + tensor rotr_29_cast_fp16 = cos(x = angles_29_cast_fp16)[name = tensor("rotr_29_cast_fp16")]; + tensor roti_29_cast_fp16 = sin(x = angles_29_cast_fp16)[name = tensor("roti_29_cast_fp16")]; + tensor var_1986 = const()[name = tensor("op_1986"), val = tensor([1, 16, 32, 2])]; + tensor pairs_29_cast_fp16 = reshape(shape = var_1986, x = x_29_cast_fp16)[name = tensor("pairs_29_cast_fp16")]; + tensor xr_29_begin_0 = const()[name = tensor("xr_29_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_29_end_0 = const()[name = tensor("xr_29_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_29_end_mask_0 = const()[name = tensor("xr_29_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_29_squeeze_mask_0 = const()[name = tensor("xr_29_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_29_cast_fp16 = slice_by_index(begin = xr_29_begin_0, end = xr_29_end_0, end_mask = xr_29_end_mask_0, squeeze_mask = xr_29_squeeze_mask_0, x = pairs_29_cast_fp16)[name = tensor("xr_29_cast_fp16")]; + tensor xi_29_begin_0 = const()[name = tensor("xi_29_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_29_end_0 = const()[name = tensor("xi_29_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_29_end_mask_0 = const()[name = tensor("xi_29_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_29_squeeze_mask_0 = const()[name = tensor("xi_29_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_29_cast_fp16 = slice_by_index(begin = xi_29_begin_0, end = xi_29_end_0, end_mask = xi_29_end_mask_0, squeeze_mask = xi_29_squeeze_mask_0, x = pairs_29_cast_fp16)[name = tensor("xi_29_cast_fp16")]; + tensor var_1994_cast_fp16 = mul(x = xr_29_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_1994_cast_fp16")]; + tensor var_1995_cast_fp16 = mul(x = xi_29_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_1995_cast_fp16")]; + tensor xor__29_cast_fp16 = sub(x = var_1994_cast_fp16, y = var_1995_cast_fp16)[name = tensor("xor__29_cast_fp16")]; + tensor var_1998_cast_fp16 = mul(x = xr_29_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_1998_cast_fp16")]; + tensor var_1999_cast_fp16 = mul(x = xi_29_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_1999_cast_fp16")]; + tensor xoi_29_cast_fp16 = add(x = var_1998_cast_fp16, y = var_1999_cast_fp16)[name = tensor("xoi_29_cast_fp16")]; + tensor out_29_axis_0 = const()[name = tensor("out_29_axis_0"), val = tensor(-1)]; + tensor out_29_cast_fp16 = stack(axis = out_29_axis_0, values = (xor__29_cast_fp16, xoi_29_cast_fp16))[name = tensor("out_29_cast_fp16")]; + tensor var_2022 = const()[name = tensor("op_2022"), val = tensor([1, 16, 32, 2])]; + tensor pairs_31_cast_fp16 = reshape(shape = var_2022, x = x_31_cast_fp16)[name = tensor("pairs_31_cast_fp16")]; + tensor xr_31_begin_0 = const()[name = tensor("xr_31_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_31_end_0 = const()[name = tensor("xr_31_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_31_end_mask_0 = const()[name = tensor("xr_31_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_31_squeeze_mask_0 = const()[name = tensor("xr_31_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_31_cast_fp16 = slice_by_index(begin = xr_31_begin_0, end = xr_31_end_0, end_mask = xr_31_end_mask_0, squeeze_mask = xr_31_squeeze_mask_0, x = pairs_31_cast_fp16)[name = tensor("xr_31_cast_fp16")]; + tensor xi_31_begin_0 = const()[name = tensor("xi_31_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_31_end_0 = const()[name = tensor("xi_31_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_31_end_mask_0 = const()[name = tensor("xi_31_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_31_squeeze_mask_0 = const()[name = tensor("xi_31_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_31_cast_fp16 = slice_by_index(begin = xi_31_begin_0, end = xi_31_end_0, end_mask = xi_31_end_mask_0, squeeze_mask = xi_31_squeeze_mask_0, x = pairs_31_cast_fp16)[name = tensor("xi_31_cast_fp16")]; + tensor var_2030_cast_fp16 = mul(x = xr_31_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_2030_cast_fp16")]; + tensor var_2031_cast_fp16 = mul(x = xi_31_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_2031_cast_fp16")]; + tensor xor__31_cast_fp16 = sub(x = var_2030_cast_fp16, y = var_2031_cast_fp16)[name = tensor("xor__31_cast_fp16")]; + tensor var_2034_cast_fp16 = mul(x = xr_31_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_2034_cast_fp16")]; + tensor var_2035_cast_fp16 = mul(x = xi_31_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_2035_cast_fp16")]; + tensor xoi_31_cast_fp16 = add(x = var_2034_cast_fp16, y = var_2035_cast_fp16)[name = tensor("xoi_31_cast_fp16")]; + tensor out_31_axis_0 = const()[name = tensor("out_31_axis_0"), val = tensor(-1)]; + tensor out_31_cast_fp16 = stack(axis = out_31_axis_0, values = (xor__31_cast_fp16, xoi_31_cast_fp16))[name = tensor("out_31_cast_fp16")]; + tensor _inversed_2047_y_0_to_fp16 = const()[name = tensor("_inversed_2047_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2047_cast_fp16 = mul(x = position7_to_fp16, y = _inversed_2047_y_0_to_fp16)[name = tensor("_inversed_2047_cast_fp16")]; + tensor var_2048_cast_fp16 = floor(x = _inversed_2047_cast_fp16)[name = tensor("op_2048_cast_fp16")]; + tensor var_2049_to_fp16 = const()[name = tensor("op_2049_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2050_cast_fp16 = mul(x = var_2048_cast_fp16, y = var_2049_to_fp16)[name = tensor("op_2050_cast_fp16")]; + tensor write_idx_15_cast_fp16 = sub(x = position7_to_fp16, y = var_2050_cast_fp16)[name = tensor("write_idx_15_cast_fp16")]; + tensor var_2053_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_15_cast_fp16)[name = tensor("op_2053_cast_fp16")]; + tensor var_2063 = const()[name = tensor("op_2063"), val = tensor([1, 512, 1, 1])]; + tensor onehot_29_to_fp16_dtype_0 = const()[name = tensor("onehot_29_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_2053_cast_fp16_to_fp16 = cast(dtype = onehot_29_to_fp16_dtype_0, x = var_2053_cast_fp16)[name = tensor("cast_520")]; + tensor onehot_31_cast_fp16 = reshape(shape = var_2063, x = var_2053_cast_fp16_to_fp16)[name = tensor("onehot_31_cast_fp16")]; + tensor var_2065_to_fp16 = const()[name = tensor("op_2065_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2067_cast_fp16 = sub(x = var_2065_to_fp16, y = onehot_31_cast_fp16)[name = tensor("op_2067_cast_fp16")]; + tensor k_cache7_to_fp16_dtype_0 = const()[name = tensor("k_cache7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache7_to_fp16 = cast(dtype = k_cache7_to_fp16_dtype_0, x = k_cache7)[name = tensor("cast_519")]; + tensor var_2068_cast_fp16 = mul(x = k_cache7_to_fp16, y = var_2067_cast_fp16)[name = tensor("op_2068_cast_fp16")]; + tensor var_2073 = const()[name = tensor("op_2073"), val = tensor([1, 1, 16, 64])]; + tensor var_2074_cast_fp16 = reshape(shape = var_2073, x = out_31_cast_fp16)[name = tensor("op_2074_cast_fp16")]; + tensor var_2075_cast_fp16 = mul(x = var_2074_cast_fp16, y = onehot_31_cast_fp16)[name = tensor("op_2075_cast_fp16")]; + tensor new_k_cache_15_cast_fp16 = add(x = var_2068_cast_fp16, y = var_2075_cast_fp16)[name = tensor("new_k_cache_15_cast_fp16")]; + tensor new_k_cache_15_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_15_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache7_to_fp16_dtype_0 = const()[name = tensor("v_cache7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache7_to_fp16 = cast(dtype = v_cache7_to_fp16_dtype_0, x = v_cache7)[name = tensor("cast_517")]; + tensor var_2081_cast_fp16 = mul(x = v_cache7_to_fp16, y = var_2067_cast_fp16)[name = tensor("op_2081_cast_fp16")]; + tensor var_2086 = const()[name = tensor("op_2086"), val = tensor([1, 1, 16, 64])]; + tensor var_2087_cast_fp16 = reshape(shape = var_2086, x = v_15_cast_fp16)[name = tensor("op_2087_cast_fp16")]; + tensor var_2088_cast_fp16 = mul(x = var_2087_cast_fp16, y = onehot_31_cast_fp16)[name = tensor("op_2088_cast_fp16")]; + tensor new_v_cache_15_cast_fp16 = add(x = var_2081_cast_fp16, y = var_2088_cast_fp16)[name = tensor("new_v_cache_15_cast_fp16")]; + tensor new_v_cache_15_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_15_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_7_perm_0 = const()[name = tensor("transpose_7_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_2101 = const()[name = tensor("op_2101"), val = tensor([0, 2, 1, 3])]; + tensor var_2107 = const()[name = tensor("op_2107"), val = tensor([1, 16, 1, 64])]; + tensor q4_15_cast_fp16 = reshape(shape = var_2107, x = out_29_cast_fp16)[name = tensor("q4_15_cast_fp16")]; + tensor var_2112_transpose_x_0 = const()[name = tensor("op_2112_transpose_x_0"), val = tensor(false)]; + tensor var_2112_transpose_y_0 = const()[name = tensor("op_2112_transpose_y_0"), val = tensor(false)]; + tensor transpose_7_cast_fp16 = transpose(perm = transpose_7_perm_0, x = new_k_cache_15_cast_fp16)[name = tensor("transpose_74")]; + tensor var_2112_cast_fp16 = matmul(transpose_x = var_2112_transpose_x_0, transpose_y = var_2112_transpose_y_0, x = q4_15_cast_fp16, y = transpose_7_cast_fp16)[name = tensor("op_2112_cast_fp16")]; + tensor var_2113_to_fp16 = const()[name = tensor("op_2113_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_29_cast_fp16 = mul(x = var_2112_cast_fp16, y = var_2113_to_fp16)[name = tensor("scores_29_cast_fp16")]; + tensor var_2117 = const()[name = tensor("op_2117"), val = tensor([1, 1])]; + tensor var_2118_cast_fp16 = reshape(shape = var_2117, x = position7_to_fp16)[name = tensor("op_2118_cast_fp16")]; + tensor var_2119_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_2118_cast_fp16)[name = tensor("op_2119_cast_fp16")]; + tensor mask_15_to_fp16_dtype_0 = const()[name = tensor("mask_15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_2126_to_fp16 = const()[name = tensor("op_2126_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2119_cast_fp16_to_fp16 = cast(dtype = mask_15_to_fp16_dtype_0, x = var_2119_cast_fp16)[name = tensor("cast_515")]; + tensor var_2127_cast_fp16 = sub(x = var_2119_cast_fp16_to_fp16, y = var_2126_to_fp16)[name = tensor("op_2127_cast_fp16")]; + tensor var_2132 = const()[name = tensor("op_2132"), val = tensor([1, 1, 1, 512])]; + tensor var_2133_cast_fp16 = reshape(shape = var_2132, x = var_2127_cast_fp16)[name = tensor("op_2133_cast_fp16")]; + tensor var_2134_to_fp16 = const()[name = tensor("op_2134_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_2135_cast_fp16 = mul(x = var_2133_cast_fp16, y = var_2134_to_fp16)[name = tensor("op_2135_cast_fp16")]; + tensor scores_31_cast_fp16 = add(x = scores_29_cast_fp16, y = var_2135_cast_fp16)[name = tensor("scores_31_cast_fp16")]; + tensor var_2138 = const()[name = tensor("op_2138"), val = tensor(-1)]; + tensor weights_15_cast_fp16 = softmax(axis = var_2138, x = scores_31_cast_fp16)[name = tensor("weights_15_cast_fp16")]; + tensor attn_15_transpose_x_0 = const()[name = tensor("attn_15_transpose_x_0"), val = tensor(false)]; + tensor attn_15_transpose_y_0 = const()[name = tensor("attn_15_transpose_y_0"), val = tensor(false)]; + tensor values_15_cast_fp16 = transpose(perm = var_2101, x = new_v_cache_15_cast_fp16)[name = tensor("transpose_73")]; + tensor attn_15_cast_fp16 = matmul(transpose_x = attn_15_transpose_x_0, transpose_y = attn_15_transpose_y_0, x = weights_15_cast_fp16, y = values_15_cast_fp16)[name = tensor("attn_15_cast_fp16")]; + tensor var_2146 = const()[name = tensor("op_2146"), val = tensor([0, 2, 1, 3])]; + tensor var_2151 = const()[name = tensor("op_2151"), val = tensor([1, 1, 1024])]; + tensor var_2147_cast_fp16 = transpose(perm = var_2146, x = attn_15_cast_fp16)[name = tensor("transpose_72")]; + tensor input_91_cast_fp16 = reshape(shape = var_2151, x = var_2147_cast_fp16)[name = tensor("input_91_cast_fp16")]; + tensor attn7_out_proj_weight_to_fp16 = const()[name = tensor("attn7_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(182600896)))]; + tensor linear_30_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn7_out_proj_weight_to_fp16, x = input_91_cast_fp16)[name = tensor("linear_30_cast_fp16")]; + tensor var_2157_to_fp16 = const()[name = tensor("op_2157_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2158_cast_fp16 = add(x = position7_to_fp16, y = var_2157_to_fp16)[name = tensor("op_2158_cast_fp16")]; + tensor var_2158_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2158_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_93_cast_fp16 = add(x = input_87_cast_fp16, y = linear_30_cast_fp16)[name = tensor("input_93_cast_fp16")]; + tensor input_95_axes_0 = const()[name = tensor("input_95_axes_0"), val = tensor([-1])]; + tensor norm7_2_weight_to_fp16 = const()[name = tensor("norm7_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184698112)))]; + tensor norm7_2_bias_to_fp16 = const()[name = tensor("norm7_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184700224)))]; + tensor var_2162_to_fp16 = const()[name = tensor("op_2162_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_95_cast_fp16 = layer_norm(axes = input_95_axes_0, beta = norm7_2_bias_to_fp16, epsilon = var_2162_to_fp16, gamma = norm7_2_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("input_95_cast_fp16")]; + tensor linear7_1_weight_to_fp16 = const()[name = tensor("linear7_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184702336)))]; + tensor linear_31_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear7_1_weight_to_fp16, x = input_95_cast_fp16)[name = tensor("linear_31_cast_fp16")]; + tensor input_97_mode_0 = const()[name = tensor("input_97_mode_0"), val = tensor("EXACT")]; + tensor input_97_cast_fp16 = gelu(mode = input_97_mode_0, x = linear_31_cast_fp16)[name = tensor("input_97_cast_fp16")]; + tensor linear7_2_weight_to_fp16 = const()[name = tensor("linear7_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(193091008)))]; + tensor linear_32_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear7_2_weight_to_fp16, x = input_97_cast_fp16)[name = tensor("linear_32_cast_fp16")]; + tensor input_99_cast_fp16 = add(x = input_93_cast_fp16, y = linear_32_cast_fp16)[name = tensor("input_99_cast_fp16")]; + tensor input_101_axes_0 = const()[name = tensor("input_101_axes_0"), val = tensor([-1])]; + tensor norm8_1_weight_to_fp16 = const()[name = tensor("norm8_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201479680)))]; + tensor norm8_1_bias_to_fp16 = const()[name = tensor("norm8_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201481792)))]; + tensor var_2179_to_fp16 = const()[name = tensor("op_2179_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_101_cast_fp16 = layer_norm(axes = input_101_axes_0, beta = norm8_1_bias_to_fp16, epsilon = var_2179_to_fp16, gamma = norm8_1_weight_to_fp16, x = input_99_cast_fp16)[name = tensor("input_101_cast_fp16")]; + tensor attn8_in_proj_weight_to_fp16 = const()[name = tensor("attn8_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201483904)))]; + tensor linear_33_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn8_in_proj_weight_to_fp16, x = input_101_cast_fp16)[name = tensor("linear_33_cast_fp16")]; + tensor var_2192 = const()[name = tensor("op_2192"), val = tensor([1, 3, 16, 64])]; + tensor qkv_17_cast_fp16 = reshape(shape = var_2192, x = linear_33_cast_fp16)[name = tensor("qkv_17_cast_fp16")]; + tensor x_33_begin_0 = const()[name = tensor("x_33_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_33_end_0 = const()[name = tensor("x_33_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_33_end_mask_0 = const()[name = tensor("x_33_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_33_squeeze_mask_0 = const()[name = tensor("x_33_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_33_cast_fp16 = slice_by_index(begin = x_33_begin_0, end = x_33_end_0, end_mask = x_33_end_mask_0, squeeze_mask = x_33_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("x_33_cast_fp16")]; + tensor x_35_begin_0 = const()[name = tensor("x_35_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_35_end_0 = const()[name = tensor("x_35_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_35_end_mask_0 = const()[name = tensor("x_35_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_35_squeeze_mask_0 = const()[name = tensor("x_35_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_35_cast_fp16 = slice_by_index(begin = x_35_begin_0, end = x_35_end_0, end_mask = x_35_end_mask_0, squeeze_mask = x_35_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("x_35_cast_fp16")]; + tensor v_17_begin_0 = const()[name = tensor("v_17_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_17_end_0 = const()[name = tensor("v_17_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_17_end_mask_0 = const()[name = tensor("v_17_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_17_squeeze_mask_0 = const()[name = tensor("v_17_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_17_cast_fp16 = slice_by_index(begin = v_17_begin_0, end = v_17_end_0, end_mask = v_17_end_mask_0, squeeze_mask = v_17_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("v_17_cast_fp16")]; + tensor var_2221 = const()[name = tensor("op_2221"), val = tensor([1, 1, 1])]; + tensor position8_to_fp16_dtype_0 = const()[name = tensor("position8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position8_to_fp16 = cast(dtype = position8_to_fp16_dtype_0, x = position8)[name = tensor("cast_513")]; + tensor var_2222_cast_fp16 = reshape(shape = var_2221, x = position8_to_fp16)[name = tensor("op_2222_cast_fp16")]; + tensor angles_33_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_2222_cast_fp16)[name = tensor("angles_33_cast_fp16")]; + tensor rotr_33_cast_fp16 = cos(x = angles_33_cast_fp16)[name = tensor("rotr_33_cast_fp16")]; + tensor roti_33_cast_fp16 = sin(x = angles_33_cast_fp16)[name = tensor("roti_33_cast_fp16")]; + tensor var_2230 = const()[name = tensor("op_2230"), val = tensor([1, 16, 32, 2])]; + tensor pairs_33_cast_fp16 = reshape(shape = var_2230, x = x_33_cast_fp16)[name = tensor("pairs_33_cast_fp16")]; + tensor xr_33_begin_0 = const()[name = tensor("xr_33_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_33_end_0 = const()[name = tensor("xr_33_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_33_end_mask_0 = const()[name = tensor("xr_33_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_33_squeeze_mask_0 = const()[name = tensor("xr_33_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_33_cast_fp16 = slice_by_index(begin = xr_33_begin_0, end = xr_33_end_0, end_mask = xr_33_end_mask_0, squeeze_mask = xr_33_squeeze_mask_0, x = pairs_33_cast_fp16)[name = tensor("xr_33_cast_fp16")]; + tensor xi_33_begin_0 = const()[name = tensor("xi_33_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_33_end_0 = const()[name = tensor("xi_33_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_33_end_mask_0 = const()[name = tensor("xi_33_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_33_squeeze_mask_0 = const()[name = tensor("xi_33_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_33_cast_fp16 = slice_by_index(begin = xi_33_begin_0, end = xi_33_end_0, end_mask = xi_33_end_mask_0, squeeze_mask = xi_33_squeeze_mask_0, x = pairs_33_cast_fp16)[name = tensor("xi_33_cast_fp16")]; + tensor var_2238_cast_fp16 = mul(x = xr_33_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_2238_cast_fp16")]; + tensor var_2239_cast_fp16 = mul(x = xi_33_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_2239_cast_fp16")]; + tensor xor__33_cast_fp16 = sub(x = var_2238_cast_fp16, y = var_2239_cast_fp16)[name = tensor("xor__33_cast_fp16")]; + tensor var_2242_cast_fp16 = mul(x = xr_33_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_2242_cast_fp16")]; + tensor var_2243_cast_fp16 = mul(x = xi_33_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_2243_cast_fp16")]; + tensor xoi_33_cast_fp16 = add(x = var_2242_cast_fp16, y = var_2243_cast_fp16)[name = tensor("xoi_33_cast_fp16")]; + tensor out_33_axis_0 = const()[name = tensor("out_33_axis_0"), val = tensor(-1)]; + tensor out_33_cast_fp16 = stack(axis = out_33_axis_0, values = (xor__33_cast_fp16, xoi_33_cast_fp16))[name = tensor("out_33_cast_fp16")]; + tensor var_2266 = const()[name = tensor("op_2266"), val = tensor([1, 16, 32, 2])]; + tensor pairs_35_cast_fp16 = reshape(shape = var_2266, x = x_35_cast_fp16)[name = tensor("pairs_35_cast_fp16")]; + tensor xr_35_begin_0 = const()[name = tensor("xr_35_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_35_end_0 = const()[name = tensor("xr_35_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_35_end_mask_0 = const()[name = tensor("xr_35_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_35_squeeze_mask_0 = const()[name = tensor("xr_35_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_35_cast_fp16 = slice_by_index(begin = xr_35_begin_0, end = xr_35_end_0, end_mask = xr_35_end_mask_0, squeeze_mask = xr_35_squeeze_mask_0, x = pairs_35_cast_fp16)[name = tensor("xr_35_cast_fp16")]; + tensor xi_35_begin_0 = const()[name = tensor("xi_35_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_35_end_0 = const()[name = tensor("xi_35_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_35_end_mask_0 = const()[name = tensor("xi_35_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_35_squeeze_mask_0 = const()[name = tensor("xi_35_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_35_cast_fp16 = slice_by_index(begin = xi_35_begin_0, end = xi_35_end_0, end_mask = xi_35_end_mask_0, squeeze_mask = xi_35_squeeze_mask_0, x = pairs_35_cast_fp16)[name = tensor("xi_35_cast_fp16")]; + tensor var_2274_cast_fp16 = mul(x = xr_35_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_2274_cast_fp16")]; + tensor var_2275_cast_fp16 = mul(x = xi_35_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_2275_cast_fp16")]; + tensor xor__35_cast_fp16 = sub(x = var_2274_cast_fp16, y = var_2275_cast_fp16)[name = tensor("xor__35_cast_fp16")]; + tensor var_2278_cast_fp16 = mul(x = xr_35_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_2278_cast_fp16")]; + tensor var_2279_cast_fp16 = mul(x = xi_35_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_2279_cast_fp16")]; + tensor xoi_35_cast_fp16 = add(x = var_2278_cast_fp16, y = var_2279_cast_fp16)[name = tensor("xoi_35_cast_fp16")]; + tensor out_35_axis_0 = const()[name = tensor("out_35_axis_0"), val = tensor(-1)]; + tensor out_35_cast_fp16 = stack(axis = out_35_axis_0, values = (xor__35_cast_fp16, xoi_35_cast_fp16))[name = tensor("out_35_cast_fp16")]; + tensor _inversed_2291_y_0_to_fp16 = const()[name = tensor("_inversed_2291_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2291_cast_fp16 = mul(x = position8_to_fp16, y = _inversed_2291_y_0_to_fp16)[name = tensor("_inversed_2291_cast_fp16")]; + tensor var_2292_cast_fp16 = floor(x = _inversed_2291_cast_fp16)[name = tensor("op_2292_cast_fp16")]; + tensor var_2293_to_fp16 = const()[name = tensor("op_2293_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2294_cast_fp16 = mul(x = var_2292_cast_fp16, y = var_2293_to_fp16)[name = tensor("op_2294_cast_fp16")]; + tensor write_idx_17_cast_fp16 = sub(x = position8_to_fp16, y = var_2294_cast_fp16)[name = tensor("write_idx_17_cast_fp16")]; + tensor var_2297_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_17_cast_fp16)[name = tensor("op_2297_cast_fp16")]; + tensor var_2307 = const()[name = tensor("op_2307"), val = tensor([1, 512, 1, 1])]; + tensor onehot_33_to_fp16_dtype_0 = const()[name = tensor("onehot_33_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_2297_cast_fp16_to_fp16 = cast(dtype = onehot_33_to_fp16_dtype_0, x = var_2297_cast_fp16)[name = tensor("cast_512")]; + tensor onehot_35_cast_fp16 = reshape(shape = var_2307, x = var_2297_cast_fp16_to_fp16)[name = tensor("onehot_35_cast_fp16")]; + tensor var_2309_to_fp16 = const()[name = tensor("op_2309_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2311_cast_fp16 = sub(x = var_2309_to_fp16, y = onehot_35_cast_fp16)[name = tensor("op_2311_cast_fp16")]; + tensor k_cache8_to_fp16_dtype_0 = const()[name = tensor("k_cache8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache8_to_fp16 = cast(dtype = k_cache8_to_fp16_dtype_0, x = k_cache8)[name = tensor("cast_511")]; + tensor var_2312_cast_fp16 = mul(x = k_cache8_to_fp16, y = var_2311_cast_fp16)[name = tensor("op_2312_cast_fp16")]; + tensor var_2317 = const()[name = tensor("op_2317"), val = tensor([1, 1, 16, 64])]; + tensor var_2318_cast_fp16 = reshape(shape = var_2317, x = out_35_cast_fp16)[name = tensor("op_2318_cast_fp16")]; + tensor var_2319_cast_fp16 = mul(x = var_2318_cast_fp16, y = onehot_35_cast_fp16)[name = tensor("op_2319_cast_fp16")]; + tensor new_k_cache_17_cast_fp16 = add(x = var_2312_cast_fp16, y = var_2319_cast_fp16)[name = tensor("new_k_cache_17_cast_fp16")]; + tensor new_k_cache_17_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_17_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache8_to_fp16_dtype_0 = const()[name = tensor("v_cache8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache8_to_fp16 = cast(dtype = v_cache8_to_fp16_dtype_0, x = v_cache8)[name = tensor("cast_509")]; + tensor var_2325_cast_fp16 = mul(x = v_cache8_to_fp16, y = var_2311_cast_fp16)[name = tensor("op_2325_cast_fp16")]; + tensor var_2330 = const()[name = tensor("op_2330"), val = tensor([1, 1, 16, 64])]; + tensor var_2331_cast_fp16 = reshape(shape = var_2330, x = v_17_cast_fp16)[name = tensor("op_2331_cast_fp16")]; + tensor var_2332_cast_fp16 = mul(x = var_2331_cast_fp16, y = onehot_35_cast_fp16)[name = tensor("op_2332_cast_fp16")]; + tensor new_v_cache_17_cast_fp16 = add(x = var_2325_cast_fp16, y = var_2332_cast_fp16)[name = tensor("new_v_cache_17_cast_fp16")]; + tensor new_v_cache_17_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_17_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_8_perm_0 = const()[name = tensor("transpose_8_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_2345 = const()[name = tensor("op_2345"), val = tensor([0, 2, 1, 3])]; + tensor var_2351 = const()[name = tensor("op_2351"), val = tensor([1, 16, 1, 64])]; + tensor q4_17_cast_fp16 = reshape(shape = var_2351, x = out_33_cast_fp16)[name = tensor("q4_17_cast_fp16")]; + tensor var_2356_transpose_x_0 = const()[name = tensor("op_2356_transpose_x_0"), val = tensor(false)]; + tensor var_2356_transpose_y_0 = const()[name = tensor("op_2356_transpose_y_0"), val = tensor(false)]; + tensor transpose_8_cast_fp16 = transpose(perm = transpose_8_perm_0, x = new_k_cache_17_cast_fp16)[name = tensor("transpose_71")]; + tensor var_2356_cast_fp16 = matmul(transpose_x = var_2356_transpose_x_0, transpose_y = var_2356_transpose_y_0, x = q4_17_cast_fp16, y = transpose_8_cast_fp16)[name = tensor("op_2356_cast_fp16")]; + tensor var_2357_to_fp16 = const()[name = tensor("op_2357_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_33_cast_fp16 = mul(x = var_2356_cast_fp16, y = var_2357_to_fp16)[name = tensor("scores_33_cast_fp16")]; + tensor var_2361 = const()[name = tensor("op_2361"), val = tensor([1, 1])]; + tensor var_2362_cast_fp16 = reshape(shape = var_2361, x = position8_to_fp16)[name = tensor("op_2362_cast_fp16")]; + tensor var_2363_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_2362_cast_fp16)[name = tensor("op_2363_cast_fp16")]; + tensor mask_17_to_fp16_dtype_0 = const()[name = tensor("mask_17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_2370_to_fp16 = const()[name = tensor("op_2370_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2363_cast_fp16_to_fp16 = cast(dtype = mask_17_to_fp16_dtype_0, x = var_2363_cast_fp16)[name = tensor("cast_507")]; + tensor var_2371_cast_fp16 = sub(x = var_2363_cast_fp16_to_fp16, y = var_2370_to_fp16)[name = tensor("op_2371_cast_fp16")]; + tensor var_2376 = const()[name = tensor("op_2376"), val = tensor([1, 1, 1, 512])]; + tensor var_2377_cast_fp16 = reshape(shape = var_2376, x = var_2371_cast_fp16)[name = tensor("op_2377_cast_fp16")]; + tensor var_2378_to_fp16 = const()[name = tensor("op_2378_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_2379_cast_fp16 = mul(x = var_2377_cast_fp16, y = var_2378_to_fp16)[name = tensor("op_2379_cast_fp16")]; + tensor scores_35_cast_fp16 = add(x = scores_33_cast_fp16, y = var_2379_cast_fp16)[name = tensor("scores_35_cast_fp16")]; + tensor var_2382 = const()[name = tensor("op_2382"), val = tensor(-1)]; + tensor weights_17_cast_fp16 = softmax(axis = var_2382, x = scores_35_cast_fp16)[name = tensor("weights_17_cast_fp16")]; + tensor attn_17_transpose_x_0 = const()[name = tensor("attn_17_transpose_x_0"), val = tensor(false)]; + tensor attn_17_transpose_y_0 = const()[name = tensor("attn_17_transpose_y_0"), val = tensor(false)]; + tensor values_17_cast_fp16 = transpose(perm = var_2345, x = new_v_cache_17_cast_fp16)[name = tensor("transpose_70")]; + tensor attn_17_cast_fp16 = matmul(transpose_x = attn_17_transpose_x_0, transpose_y = attn_17_transpose_y_0, x = weights_17_cast_fp16, y = values_17_cast_fp16)[name = tensor("attn_17_cast_fp16")]; + tensor var_2390 = const()[name = tensor("op_2390"), val = tensor([0, 2, 1, 3])]; + tensor var_2395 = const()[name = tensor("op_2395"), val = tensor([1, 1, 1024])]; + tensor var_2391_cast_fp16 = transpose(perm = var_2390, x = attn_17_cast_fp16)[name = tensor("transpose_69")]; + tensor input_103_cast_fp16 = reshape(shape = var_2395, x = var_2391_cast_fp16)[name = tensor("input_103_cast_fp16")]; + tensor attn8_out_proj_weight_to_fp16 = const()[name = tensor("attn8_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(207775424)))]; + tensor linear_34_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn8_out_proj_weight_to_fp16, x = input_103_cast_fp16)[name = tensor("linear_34_cast_fp16")]; + tensor var_2401_to_fp16 = const()[name = tensor("op_2401_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2402_cast_fp16 = add(x = position8_to_fp16, y = var_2401_to_fp16)[name = tensor("op_2402_cast_fp16")]; + tensor var_2402_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2402_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_105_cast_fp16 = add(x = input_99_cast_fp16, y = linear_34_cast_fp16)[name = tensor("input_105_cast_fp16")]; + tensor input_107_axes_0 = const()[name = tensor("input_107_axes_0"), val = tensor([-1])]; + tensor norm8_2_weight_to_fp16 = const()[name = tensor("norm8_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209872640)))]; + tensor norm8_2_bias_to_fp16 = const()[name = tensor("norm8_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209874752)))]; + tensor var_2406_to_fp16 = const()[name = tensor("op_2406_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_107_cast_fp16 = layer_norm(axes = input_107_axes_0, beta = norm8_2_bias_to_fp16, epsilon = var_2406_to_fp16, gamma = norm8_2_weight_to_fp16, x = input_105_cast_fp16)[name = tensor("input_107_cast_fp16")]; + tensor linear8_1_weight_to_fp16 = const()[name = tensor("linear8_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209876864)))]; + tensor linear_35_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear8_1_weight_to_fp16, x = input_107_cast_fp16)[name = tensor("linear_35_cast_fp16")]; + tensor input_109_mode_0 = const()[name = tensor("input_109_mode_0"), val = tensor("EXACT")]; + tensor input_109_cast_fp16 = gelu(mode = input_109_mode_0, x = linear_35_cast_fp16)[name = tensor("input_109_cast_fp16")]; + tensor linear8_2_weight_to_fp16 = const()[name = tensor("linear8_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(218265536)))]; + tensor linear_36_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear8_2_weight_to_fp16, x = input_109_cast_fp16)[name = tensor("linear_36_cast_fp16")]; + tensor input_111_cast_fp16 = add(x = input_105_cast_fp16, y = linear_36_cast_fp16)[name = tensor("input_111_cast_fp16")]; + tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([-1])]; + tensor norm9_1_weight_to_fp16 = const()[name = tensor("norm9_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226654208)))]; + tensor norm9_1_bias_to_fp16 = const()[name = tensor("norm9_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226656320)))]; + tensor var_2423_to_fp16 = const()[name = tensor("op_2423_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_113_cast_fp16 = layer_norm(axes = input_113_axes_0, beta = norm9_1_bias_to_fp16, epsilon = var_2423_to_fp16, gamma = norm9_1_weight_to_fp16, x = input_111_cast_fp16)[name = tensor("input_113_cast_fp16")]; + tensor attn9_in_proj_weight_to_fp16 = const()[name = tensor("attn9_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226658432)))]; + tensor linear_37_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn9_in_proj_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("linear_37_cast_fp16")]; + tensor var_2436 = const()[name = tensor("op_2436"), val = tensor([1, 3, 16, 64])]; + tensor qkv_19_cast_fp16 = reshape(shape = var_2436, x = linear_37_cast_fp16)[name = tensor("qkv_19_cast_fp16")]; + tensor x_37_begin_0 = const()[name = tensor("x_37_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_37_end_0 = const()[name = tensor("x_37_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_37_end_mask_0 = const()[name = tensor("x_37_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_37_squeeze_mask_0 = const()[name = tensor("x_37_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_37_cast_fp16 = slice_by_index(begin = x_37_begin_0, end = x_37_end_0, end_mask = x_37_end_mask_0, squeeze_mask = x_37_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("x_37_cast_fp16")]; + tensor x_39_begin_0 = const()[name = tensor("x_39_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_39_end_0 = const()[name = tensor("x_39_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_39_end_mask_0 = const()[name = tensor("x_39_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_39_squeeze_mask_0 = const()[name = tensor("x_39_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_39_cast_fp16 = slice_by_index(begin = x_39_begin_0, end = x_39_end_0, end_mask = x_39_end_mask_0, squeeze_mask = x_39_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("x_39_cast_fp16")]; + tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_19_cast_fp16 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("v_19_cast_fp16")]; + tensor var_2465 = const()[name = tensor("op_2465"), val = tensor([1, 1, 1])]; + tensor position9_to_fp16_dtype_0 = const()[name = tensor("position9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position9_to_fp16 = cast(dtype = position9_to_fp16_dtype_0, x = position9)[name = tensor("cast_505")]; + tensor var_2466_cast_fp16 = reshape(shape = var_2465, x = position9_to_fp16)[name = tensor("op_2466_cast_fp16")]; + tensor angles_37_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_2466_cast_fp16)[name = tensor("angles_37_cast_fp16")]; + tensor rotr_37_cast_fp16 = cos(x = angles_37_cast_fp16)[name = tensor("rotr_37_cast_fp16")]; + tensor roti_37_cast_fp16 = sin(x = angles_37_cast_fp16)[name = tensor("roti_37_cast_fp16")]; + tensor var_2474 = const()[name = tensor("op_2474"), val = tensor([1, 16, 32, 2])]; + tensor pairs_37_cast_fp16 = reshape(shape = var_2474, x = x_37_cast_fp16)[name = tensor("pairs_37_cast_fp16")]; + tensor xr_37_begin_0 = const()[name = tensor("xr_37_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_37_end_0 = const()[name = tensor("xr_37_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_37_end_mask_0 = const()[name = tensor("xr_37_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_37_squeeze_mask_0 = const()[name = tensor("xr_37_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_37_cast_fp16 = slice_by_index(begin = xr_37_begin_0, end = xr_37_end_0, end_mask = xr_37_end_mask_0, squeeze_mask = xr_37_squeeze_mask_0, x = pairs_37_cast_fp16)[name = tensor("xr_37_cast_fp16")]; + tensor xi_37_begin_0 = const()[name = tensor("xi_37_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_37_end_0 = const()[name = tensor("xi_37_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_37_end_mask_0 = const()[name = tensor("xi_37_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_37_squeeze_mask_0 = const()[name = tensor("xi_37_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_37_cast_fp16 = slice_by_index(begin = xi_37_begin_0, end = xi_37_end_0, end_mask = xi_37_end_mask_0, squeeze_mask = xi_37_squeeze_mask_0, x = pairs_37_cast_fp16)[name = tensor("xi_37_cast_fp16")]; + tensor var_2482_cast_fp16 = mul(x = xr_37_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_2482_cast_fp16")]; + tensor var_2483_cast_fp16 = mul(x = xi_37_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_2483_cast_fp16")]; + tensor xor__37_cast_fp16 = sub(x = var_2482_cast_fp16, y = var_2483_cast_fp16)[name = tensor("xor__37_cast_fp16")]; + tensor var_2486_cast_fp16 = mul(x = xr_37_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_2486_cast_fp16")]; + tensor var_2487_cast_fp16 = mul(x = xi_37_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_2487_cast_fp16")]; + tensor xoi_37_cast_fp16 = add(x = var_2486_cast_fp16, y = var_2487_cast_fp16)[name = tensor("xoi_37_cast_fp16")]; + tensor out_37_axis_0 = const()[name = tensor("out_37_axis_0"), val = tensor(-1)]; + tensor out_37_cast_fp16 = stack(axis = out_37_axis_0, values = (xor__37_cast_fp16, xoi_37_cast_fp16))[name = tensor("out_37_cast_fp16")]; + tensor var_2510 = const()[name = tensor("op_2510"), val = tensor([1, 16, 32, 2])]; + tensor pairs_39_cast_fp16 = reshape(shape = var_2510, x = x_39_cast_fp16)[name = tensor("pairs_39_cast_fp16")]; + tensor xr_39_begin_0 = const()[name = tensor("xr_39_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_39_end_0 = const()[name = tensor("xr_39_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_39_end_mask_0 = const()[name = tensor("xr_39_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_39_squeeze_mask_0 = const()[name = tensor("xr_39_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_39_cast_fp16 = slice_by_index(begin = xr_39_begin_0, end = xr_39_end_0, end_mask = xr_39_end_mask_0, squeeze_mask = xr_39_squeeze_mask_0, x = pairs_39_cast_fp16)[name = tensor("xr_39_cast_fp16")]; + tensor xi_39_begin_0 = const()[name = tensor("xi_39_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_39_end_0 = const()[name = tensor("xi_39_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_39_end_mask_0 = const()[name = tensor("xi_39_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_39_squeeze_mask_0 = const()[name = tensor("xi_39_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_39_cast_fp16 = slice_by_index(begin = xi_39_begin_0, end = xi_39_end_0, end_mask = xi_39_end_mask_0, squeeze_mask = xi_39_squeeze_mask_0, x = pairs_39_cast_fp16)[name = tensor("xi_39_cast_fp16")]; + tensor var_2518_cast_fp16 = mul(x = xr_39_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_2518_cast_fp16")]; + tensor var_2519_cast_fp16 = mul(x = xi_39_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_2519_cast_fp16")]; + tensor xor__39_cast_fp16 = sub(x = var_2518_cast_fp16, y = var_2519_cast_fp16)[name = tensor("xor__39_cast_fp16")]; + tensor var_2522_cast_fp16 = mul(x = xr_39_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_2522_cast_fp16")]; + tensor var_2523_cast_fp16 = mul(x = xi_39_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_2523_cast_fp16")]; + tensor xoi_39_cast_fp16 = add(x = var_2522_cast_fp16, y = var_2523_cast_fp16)[name = tensor("xoi_39_cast_fp16")]; + tensor out_39_axis_0 = const()[name = tensor("out_39_axis_0"), val = tensor(-1)]; + tensor out_39_cast_fp16 = stack(axis = out_39_axis_0, values = (xor__39_cast_fp16, xoi_39_cast_fp16))[name = tensor("out_39_cast_fp16")]; + tensor _inversed_2535_y_0_to_fp16 = const()[name = tensor("_inversed_2535_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2535_cast_fp16 = mul(x = position9_to_fp16, y = _inversed_2535_y_0_to_fp16)[name = tensor("_inversed_2535_cast_fp16")]; + tensor var_2536_cast_fp16 = floor(x = _inversed_2535_cast_fp16)[name = tensor("op_2536_cast_fp16")]; + tensor var_2537_to_fp16 = const()[name = tensor("op_2537_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2538_cast_fp16 = mul(x = var_2536_cast_fp16, y = var_2537_to_fp16)[name = tensor("op_2538_cast_fp16")]; + tensor write_idx_19_cast_fp16 = sub(x = position9_to_fp16, y = var_2538_cast_fp16)[name = tensor("write_idx_19_cast_fp16")]; + tensor var_2541_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_19_cast_fp16)[name = tensor("op_2541_cast_fp16")]; + tensor var_2551 = const()[name = tensor("op_2551"), val = tensor([1, 512, 1, 1])]; + tensor onehot_37_to_fp16_dtype_0 = const()[name = tensor("onehot_37_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_2541_cast_fp16_to_fp16 = cast(dtype = onehot_37_to_fp16_dtype_0, x = var_2541_cast_fp16)[name = tensor("cast_504")]; + tensor onehot_39_cast_fp16 = reshape(shape = var_2551, x = var_2541_cast_fp16_to_fp16)[name = tensor("onehot_39_cast_fp16")]; + tensor var_2553_to_fp16 = const()[name = tensor("op_2553_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2555_cast_fp16 = sub(x = var_2553_to_fp16, y = onehot_39_cast_fp16)[name = tensor("op_2555_cast_fp16")]; + tensor k_cache9_to_fp16_dtype_0 = const()[name = tensor("k_cache9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache9_to_fp16 = cast(dtype = k_cache9_to_fp16_dtype_0, x = k_cache9)[name = tensor("cast_503")]; + tensor var_2556_cast_fp16 = mul(x = k_cache9_to_fp16, y = var_2555_cast_fp16)[name = tensor("op_2556_cast_fp16")]; + tensor var_2561 = const()[name = tensor("op_2561"), val = tensor([1, 1, 16, 64])]; + tensor var_2562_cast_fp16 = reshape(shape = var_2561, x = out_39_cast_fp16)[name = tensor("op_2562_cast_fp16")]; + tensor var_2563_cast_fp16 = mul(x = var_2562_cast_fp16, y = onehot_39_cast_fp16)[name = tensor("op_2563_cast_fp16")]; + tensor new_k_cache_19_cast_fp16 = add(x = var_2556_cast_fp16, y = var_2563_cast_fp16)[name = tensor("new_k_cache_19_cast_fp16")]; + tensor new_k_cache_19_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_19_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache9_to_fp16_dtype_0 = const()[name = tensor("v_cache9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache9_to_fp16 = cast(dtype = v_cache9_to_fp16_dtype_0, x = v_cache9)[name = tensor("cast_501")]; + tensor var_2569_cast_fp16 = mul(x = v_cache9_to_fp16, y = var_2555_cast_fp16)[name = tensor("op_2569_cast_fp16")]; + tensor var_2574 = const()[name = tensor("op_2574"), val = tensor([1, 1, 16, 64])]; + tensor var_2575_cast_fp16 = reshape(shape = var_2574, x = v_19_cast_fp16)[name = tensor("op_2575_cast_fp16")]; + tensor var_2576_cast_fp16 = mul(x = var_2575_cast_fp16, y = onehot_39_cast_fp16)[name = tensor("op_2576_cast_fp16")]; + tensor new_v_cache_19_cast_fp16 = add(x = var_2569_cast_fp16, y = var_2576_cast_fp16)[name = tensor("new_v_cache_19_cast_fp16")]; + tensor new_v_cache_19_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_19_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_9_perm_0 = const()[name = tensor("transpose_9_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_2589 = const()[name = tensor("op_2589"), val = tensor([0, 2, 1, 3])]; + tensor var_2595 = const()[name = tensor("op_2595"), val = tensor([1, 16, 1, 64])]; + tensor q4_19_cast_fp16 = reshape(shape = var_2595, x = out_37_cast_fp16)[name = tensor("q4_19_cast_fp16")]; + tensor var_2600_transpose_x_0 = const()[name = tensor("op_2600_transpose_x_0"), val = tensor(false)]; + tensor var_2600_transpose_y_0 = const()[name = tensor("op_2600_transpose_y_0"), val = tensor(false)]; + tensor transpose_9_cast_fp16 = transpose(perm = transpose_9_perm_0, x = new_k_cache_19_cast_fp16)[name = tensor("transpose_68")]; + tensor var_2600_cast_fp16 = matmul(transpose_x = var_2600_transpose_x_0, transpose_y = var_2600_transpose_y_0, x = q4_19_cast_fp16, y = transpose_9_cast_fp16)[name = tensor("op_2600_cast_fp16")]; + tensor var_2601_to_fp16 = const()[name = tensor("op_2601_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_37_cast_fp16 = mul(x = var_2600_cast_fp16, y = var_2601_to_fp16)[name = tensor("scores_37_cast_fp16")]; + tensor var_2605 = const()[name = tensor("op_2605"), val = tensor([1, 1])]; + tensor var_2606_cast_fp16 = reshape(shape = var_2605, x = position9_to_fp16)[name = tensor("op_2606_cast_fp16")]; + tensor var_2607_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_2606_cast_fp16)[name = tensor("op_2607_cast_fp16")]; + tensor mask_19_to_fp16_dtype_0 = const()[name = tensor("mask_19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_2614_to_fp16 = const()[name = tensor("op_2614_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2607_cast_fp16_to_fp16 = cast(dtype = mask_19_to_fp16_dtype_0, x = var_2607_cast_fp16)[name = tensor("cast_499")]; + tensor var_2615_cast_fp16 = sub(x = var_2607_cast_fp16_to_fp16, y = var_2614_to_fp16)[name = tensor("op_2615_cast_fp16")]; + tensor var_2620 = const()[name = tensor("op_2620"), val = tensor([1, 1, 1, 512])]; + tensor var_2621_cast_fp16 = reshape(shape = var_2620, x = var_2615_cast_fp16)[name = tensor("op_2621_cast_fp16")]; + tensor var_2622_to_fp16 = const()[name = tensor("op_2622_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_2623_cast_fp16 = mul(x = var_2621_cast_fp16, y = var_2622_to_fp16)[name = tensor("op_2623_cast_fp16")]; + tensor scores_39_cast_fp16 = add(x = scores_37_cast_fp16, y = var_2623_cast_fp16)[name = tensor("scores_39_cast_fp16")]; + tensor var_2626 = const()[name = tensor("op_2626"), val = tensor(-1)]; + tensor weights_19_cast_fp16 = softmax(axis = var_2626, x = scores_39_cast_fp16)[name = tensor("weights_19_cast_fp16")]; + tensor attn_19_transpose_x_0 = const()[name = tensor("attn_19_transpose_x_0"), val = tensor(false)]; + tensor attn_19_transpose_y_0 = const()[name = tensor("attn_19_transpose_y_0"), val = tensor(false)]; + tensor values_19_cast_fp16 = transpose(perm = var_2589, x = new_v_cache_19_cast_fp16)[name = tensor("transpose_67")]; + tensor attn_19_cast_fp16 = matmul(transpose_x = attn_19_transpose_x_0, transpose_y = attn_19_transpose_y_0, x = weights_19_cast_fp16, y = values_19_cast_fp16)[name = tensor("attn_19_cast_fp16")]; + tensor var_2634 = const()[name = tensor("op_2634"), val = tensor([0, 2, 1, 3])]; + tensor var_2639 = const()[name = tensor("op_2639"), val = tensor([1, 1, 1024])]; + tensor var_2635_cast_fp16 = transpose(perm = var_2634, x = attn_19_cast_fp16)[name = tensor("transpose_66")]; + tensor input_115_cast_fp16 = reshape(shape = var_2639, x = var_2635_cast_fp16)[name = tensor("input_115_cast_fp16")]; + tensor attn9_out_proj_weight_to_fp16 = const()[name = tensor("attn9_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(232949952)))]; + tensor linear_38_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn9_out_proj_weight_to_fp16, x = input_115_cast_fp16)[name = tensor("linear_38_cast_fp16")]; + tensor var_2645_to_fp16 = const()[name = tensor("op_2645_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2646_cast_fp16 = add(x = position9_to_fp16, y = var_2645_to_fp16)[name = tensor("op_2646_cast_fp16")]; + tensor var_2646_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2646_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_117_cast_fp16 = add(x = input_111_cast_fp16, y = linear_38_cast_fp16)[name = tensor("input_117_cast_fp16")]; + tensor input_119_axes_0 = const()[name = tensor("input_119_axes_0"), val = tensor([-1])]; + tensor norm9_2_weight_to_fp16 = const()[name = tensor("norm9_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(235047168)))]; + tensor norm9_2_bias_to_fp16 = const()[name = tensor("norm9_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(235049280)))]; + tensor var_2650_to_fp16 = const()[name = tensor("op_2650_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_119_cast_fp16 = layer_norm(axes = input_119_axes_0, beta = norm9_2_bias_to_fp16, epsilon = var_2650_to_fp16, gamma = norm9_2_weight_to_fp16, x = input_117_cast_fp16)[name = tensor("input_119_cast_fp16")]; + tensor linear9_1_weight_to_fp16 = const()[name = tensor("linear9_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(235051392)))]; + tensor linear_39_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear9_1_weight_to_fp16, x = input_119_cast_fp16)[name = tensor("linear_39_cast_fp16")]; + tensor input_121_mode_0 = const()[name = tensor("input_121_mode_0"), val = tensor("EXACT")]; + tensor input_121_cast_fp16 = gelu(mode = input_121_mode_0, x = linear_39_cast_fp16)[name = tensor("input_121_cast_fp16")]; + tensor linear9_2_weight_to_fp16 = const()[name = tensor("linear9_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(243440064)))]; + tensor linear_40_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear9_2_weight_to_fp16, x = input_121_cast_fp16)[name = tensor("linear_40_cast_fp16")]; + tensor input_123_cast_fp16 = add(x = input_117_cast_fp16, y = linear_40_cast_fp16)[name = tensor("input_123_cast_fp16")]; + tensor input_125_axes_0 = const()[name = tensor("input_125_axes_0"), val = tensor([-1])]; + tensor norm10_1_weight_to_fp16 = const()[name = tensor("norm10_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(251828736)))]; + tensor norm10_1_bias_to_fp16 = const()[name = tensor("norm10_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(251830848)))]; + tensor var_2667_to_fp16 = const()[name = tensor("op_2667_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_125_cast_fp16 = layer_norm(axes = input_125_axes_0, beta = norm10_1_bias_to_fp16, epsilon = var_2667_to_fp16, gamma = norm10_1_weight_to_fp16, x = input_123_cast_fp16)[name = tensor("input_125_cast_fp16")]; + tensor attn10_in_proj_weight_to_fp16 = const()[name = tensor("attn10_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(251832960)))]; + tensor linear_41_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn10_in_proj_weight_to_fp16, x = input_125_cast_fp16)[name = tensor("linear_41_cast_fp16")]; + tensor var_2680 = const()[name = tensor("op_2680"), val = tensor([1, 3, 16, 64])]; + tensor qkv_21_cast_fp16 = reshape(shape = var_2680, x = linear_41_cast_fp16)[name = tensor("qkv_21_cast_fp16")]; + tensor x_41_begin_0 = const()[name = tensor("x_41_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_41_end_0 = const()[name = tensor("x_41_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_41_end_mask_0 = const()[name = tensor("x_41_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_41_squeeze_mask_0 = const()[name = tensor("x_41_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_41_cast_fp16 = slice_by_index(begin = x_41_begin_0, end = x_41_end_0, end_mask = x_41_end_mask_0, squeeze_mask = x_41_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("x_41_cast_fp16")]; + tensor x_43_begin_0 = const()[name = tensor("x_43_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_43_end_0 = const()[name = tensor("x_43_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_43_end_mask_0 = const()[name = tensor("x_43_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_43_squeeze_mask_0 = const()[name = tensor("x_43_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_43_cast_fp16 = slice_by_index(begin = x_43_begin_0, end = x_43_end_0, end_mask = x_43_end_mask_0, squeeze_mask = x_43_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("x_43_cast_fp16")]; + tensor v_21_begin_0 = const()[name = tensor("v_21_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_21_end_0 = const()[name = tensor("v_21_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_21_end_mask_0 = const()[name = tensor("v_21_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_21_squeeze_mask_0 = const()[name = tensor("v_21_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_21_cast_fp16 = slice_by_index(begin = v_21_begin_0, end = v_21_end_0, end_mask = v_21_end_mask_0, squeeze_mask = v_21_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("v_21_cast_fp16")]; + tensor var_2709 = const()[name = tensor("op_2709"), val = tensor([1, 1, 1])]; + tensor position10_to_fp16_dtype_0 = const()[name = tensor("position10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position10_to_fp16 = cast(dtype = position10_to_fp16_dtype_0, x = position10)[name = tensor("cast_497")]; + tensor var_2710_cast_fp16 = reshape(shape = var_2709, x = position10_to_fp16)[name = tensor("op_2710_cast_fp16")]; + tensor angles_41_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_2710_cast_fp16)[name = tensor("angles_41_cast_fp16")]; + tensor rotr_41_cast_fp16 = cos(x = angles_41_cast_fp16)[name = tensor("rotr_41_cast_fp16")]; + tensor roti_41_cast_fp16 = sin(x = angles_41_cast_fp16)[name = tensor("roti_41_cast_fp16")]; + tensor var_2718 = const()[name = tensor("op_2718"), val = tensor([1, 16, 32, 2])]; + tensor pairs_41_cast_fp16 = reshape(shape = var_2718, x = x_41_cast_fp16)[name = tensor("pairs_41_cast_fp16")]; + tensor xr_41_begin_0 = const()[name = tensor("xr_41_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_41_end_0 = const()[name = tensor("xr_41_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_41_end_mask_0 = const()[name = tensor("xr_41_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_41_squeeze_mask_0 = const()[name = tensor("xr_41_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_41_cast_fp16 = slice_by_index(begin = xr_41_begin_0, end = xr_41_end_0, end_mask = xr_41_end_mask_0, squeeze_mask = xr_41_squeeze_mask_0, x = pairs_41_cast_fp16)[name = tensor("xr_41_cast_fp16")]; + tensor xi_41_begin_0 = const()[name = tensor("xi_41_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_41_end_0 = const()[name = tensor("xi_41_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_41_end_mask_0 = const()[name = tensor("xi_41_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_41_squeeze_mask_0 = const()[name = tensor("xi_41_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_41_cast_fp16 = slice_by_index(begin = xi_41_begin_0, end = xi_41_end_0, end_mask = xi_41_end_mask_0, squeeze_mask = xi_41_squeeze_mask_0, x = pairs_41_cast_fp16)[name = tensor("xi_41_cast_fp16")]; + tensor var_2726_cast_fp16 = mul(x = xr_41_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_2726_cast_fp16")]; + tensor var_2727_cast_fp16 = mul(x = xi_41_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_2727_cast_fp16")]; + tensor xor__41_cast_fp16 = sub(x = var_2726_cast_fp16, y = var_2727_cast_fp16)[name = tensor("xor__41_cast_fp16")]; + tensor var_2730_cast_fp16 = mul(x = xr_41_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_2730_cast_fp16")]; + tensor var_2731_cast_fp16 = mul(x = xi_41_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_2731_cast_fp16")]; + tensor xoi_41_cast_fp16 = add(x = var_2730_cast_fp16, y = var_2731_cast_fp16)[name = tensor("xoi_41_cast_fp16")]; + tensor out_41_axis_0 = const()[name = tensor("out_41_axis_0"), val = tensor(-1)]; + tensor out_41_cast_fp16 = stack(axis = out_41_axis_0, values = (xor__41_cast_fp16, xoi_41_cast_fp16))[name = tensor("out_41_cast_fp16")]; + tensor var_2754 = const()[name = tensor("op_2754"), val = tensor([1, 16, 32, 2])]; + tensor pairs_43_cast_fp16 = reshape(shape = var_2754, x = x_43_cast_fp16)[name = tensor("pairs_43_cast_fp16")]; + tensor xr_43_begin_0 = const()[name = tensor("xr_43_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_43_end_0 = const()[name = tensor("xr_43_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_43_end_mask_0 = const()[name = tensor("xr_43_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_43_squeeze_mask_0 = const()[name = tensor("xr_43_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_43_cast_fp16 = slice_by_index(begin = xr_43_begin_0, end = xr_43_end_0, end_mask = xr_43_end_mask_0, squeeze_mask = xr_43_squeeze_mask_0, x = pairs_43_cast_fp16)[name = tensor("xr_43_cast_fp16")]; + tensor xi_43_begin_0 = const()[name = tensor("xi_43_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_43_end_0 = const()[name = tensor("xi_43_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_43_end_mask_0 = const()[name = tensor("xi_43_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_43_squeeze_mask_0 = const()[name = tensor("xi_43_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_43_cast_fp16 = slice_by_index(begin = xi_43_begin_0, end = xi_43_end_0, end_mask = xi_43_end_mask_0, squeeze_mask = xi_43_squeeze_mask_0, x = pairs_43_cast_fp16)[name = tensor("xi_43_cast_fp16")]; + tensor var_2762_cast_fp16 = mul(x = xr_43_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_2762_cast_fp16")]; + tensor var_2763_cast_fp16 = mul(x = xi_43_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_2763_cast_fp16")]; + tensor xor__43_cast_fp16 = sub(x = var_2762_cast_fp16, y = var_2763_cast_fp16)[name = tensor("xor__43_cast_fp16")]; + tensor var_2766_cast_fp16 = mul(x = xr_43_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_2766_cast_fp16")]; + tensor var_2767_cast_fp16 = mul(x = xi_43_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_2767_cast_fp16")]; + tensor xoi_43_cast_fp16 = add(x = var_2766_cast_fp16, y = var_2767_cast_fp16)[name = tensor("xoi_43_cast_fp16")]; + tensor out_43_axis_0 = const()[name = tensor("out_43_axis_0"), val = tensor(-1)]; + tensor out_43_cast_fp16 = stack(axis = out_43_axis_0, values = (xor__43_cast_fp16, xoi_43_cast_fp16))[name = tensor("out_43_cast_fp16")]; + tensor _inversed_2779_y_0_to_fp16 = const()[name = tensor("_inversed_2779_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2779_cast_fp16 = mul(x = position10_to_fp16, y = _inversed_2779_y_0_to_fp16)[name = tensor("_inversed_2779_cast_fp16")]; + tensor var_2780_cast_fp16 = floor(x = _inversed_2779_cast_fp16)[name = tensor("op_2780_cast_fp16")]; + tensor var_2781_to_fp16 = const()[name = tensor("op_2781_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2782_cast_fp16 = mul(x = var_2780_cast_fp16, y = var_2781_to_fp16)[name = tensor("op_2782_cast_fp16")]; + tensor write_idx_21_cast_fp16 = sub(x = position10_to_fp16, y = var_2782_cast_fp16)[name = tensor("write_idx_21_cast_fp16")]; + tensor var_2785_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_21_cast_fp16)[name = tensor("op_2785_cast_fp16")]; + tensor var_2795 = const()[name = tensor("op_2795"), val = tensor([1, 512, 1, 1])]; + tensor onehot_41_to_fp16_dtype_0 = const()[name = tensor("onehot_41_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_2785_cast_fp16_to_fp16 = cast(dtype = onehot_41_to_fp16_dtype_0, x = var_2785_cast_fp16)[name = tensor("cast_496")]; + tensor onehot_43_cast_fp16 = reshape(shape = var_2795, x = var_2785_cast_fp16_to_fp16)[name = tensor("onehot_43_cast_fp16")]; + tensor var_2797_to_fp16 = const()[name = tensor("op_2797_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2799_cast_fp16 = sub(x = var_2797_to_fp16, y = onehot_43_cast_fp16)[name = tensor("op_2799_cast_fp16")]; + tensor k_cache10_to_fp16_dtype_0 = const()[name = tensor("k_cache10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache10_to_fp16 = cast(dtype = k_cache10_to_fp16_dtype_0, x = k_cache10)[name = tensor("cast_495")]; + tensor var_2800_cast_fp16 = mul(x = k_cache10_to_fp16, y = var_2799_cast_fp16)[name = tensor("op_2800_cast_fp16")]; + tensor var_2805 = const()[name = tensor("op_2805"), val = tensor([1, 1, 16, 64])]; + tensor var_2806_cast_fp16 = reshape(shape = var_2805, x = out_43_cast_fp16)[name = tensor("op_2806_cast_fp16")]; + tensor var_2807_cast_fp16 = mul(x = var_2806_cast_fp16, y = onehot_43_cast_fp16)[name = tensor("op_2807_cast_fp16")]; + tensor new_k_cache_21_cast_fp16 = add(x = var_2800_cast_fp16, y = var_2807_cast_fp16)[name = tensor("new_k_cache_21_cast_fp16")]; + tensor new_k_cache_21_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_21_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache10_to_fp16_dtype_0 = const()[name = tensor("v_cache10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache10_to_fp16 = cast(dtype = v_cache10_to_fp16_dtype_0, x = v_cache10)[name = tensor("cast_493")]; + tensor var_2813_cast_fp16 = mul(x = v_cache10_to_fp16, y = var_2799_cast_fp16)[name = tensor("op_2813_cast_fp16")]; + tensor var_2818 = const()[name = tensor("op_2818"), val = tensor([1, 1, 16, 64])]; + tensor var_2819_cast_fp16 = reshape(shape = var_2818, x = v_21_cast_fp16)[name = tensor("op_2819_cast_fp16")]; + tensor var_2820_cast_fp16 = mul(x = var_2819_cast_fp16, y = onehot_43_cast_fp16)[name = tensor("op_2820_cast_fp16")]; + tensor new_v_cache_21_cast_fp16 = add(x = var_2813_cast_fp16, y = var_2820_cast_fp16)[name = tensor("new_v_cache_21_cast_fp16")]; + tensor new_v_cache_21_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_21_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_10_perm_0 = const()[name = tensor("transpose_10_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_2833 = const()[name = tensor("op_2833"), val = tensor([0, 2, 1, 3])]; + tensor var_2839 = const()[name = tensor("op_2839"), val = tensor([1, 16, 1, 64])]; + tensor q4_21_cast_fp16 = reshape(shape = var_2839, x = out_41_cast_fp16)[name = tensor("q4_21_cast_fp16")]; + tensor var_2844_transpose_x_0 = const()[name = tensor("op_2844_transpose_x_0"), val = tensor(false)]; + tensor var_2844_transpose_y_0 = const()[name = tensor("op_2844_transpose_y_0"), val = tensor(false)]; + tensor transpose_10_cast_fp16 = transpose(perm = transpose_10_perm_0, x = new_k_cache_21_cast_fp16)[name = tensor("transpose_65")]; + tensor var_2844_cast_fp16 = matmul(transpose_x = var_2844_transpose_x_0, transpose_y = var_2844_transpose_y_0, x = q4_21_cast_fp16, y = transpose_10_cast_fp16)[name = tensor("op_2844_cast_fp16")]; + tensor var_2845_to_fp16 = const()[name = tensor("op_2845_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_41_cast_fp16 = mul(x = var_2844_cast_fp16, y = var_2845_to_fp16)[name = tensor("scores_41_cast_fp16")]; + tensor var_2849 = const()[name = tensor("op_2849"), val = tensor([1, 1])]; + tensor var_2850_cast_fp16 = reshape(shape = var_2849, x = position10_to_fp16)[name = tensor("op_2850_cast_fp16")]; + tensor var_2851_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_2850_cast_fp16)[name = tensor("op_2851_cast_fp16")]; + tensor mask_21_to_fp16_dtype_0 = const()[name = tensor("mask_21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_2858_to_fp16 = const()[name = tensor("op_2858_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2851_cast_fp16_to_fp16 = cast(dtype = mask_21_to_fp16_dtype_0, x = var_2851_cast_fp16)[name = tensor("cast_491")]; + tensor var_2859_cast_fp16 = sub(x = var_2851_cast_fp16_to_fp16, y = var_2858_to_fp16)[name = tensor("op_2859_cast_fp16")]; + tensor var_2864 = const()[name = tensor("op_2864"), val = tensor([1, 1, 1, 512])]; + tensor var_2865_cast_fp16 = reshape(shape = var_2864, x = var_2859_cast_fp16)[name = tensor("op_2865_cast_fp16")]; + tensor var_2866_to_fp16 = const()[name = tensor("op_2866_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_2867_cast_fp16 = mul(x = var_2865_cast_fp16, y = var_2866_to_fp16)[name = tensor("op_2867_cast_fp16")]; + tensor scores_43_cast_fp16 = add(x = scores_41_cast_fp16, y = var_2867_cast_fp16)[name = tensor("scores_43_cast_fp16")]; + tensor var_2870 = const()[name = tensor("op_2870"), val = tensor(-1)]; + tensor weights_21_cast_fp16 = softmax(axis = var_2870, x = scores_43_cast_fp16)[name = tensor("weights_21_cast_fp16")]; + tensor attn_21_transpose_x_0 = const()[name = tensor("attn_21_transpose_x_0"), val = tensor(false)]; + tensor attn_21_transpose_y_0 = const()[name = tensor("attn_21_transpose_y_0"), val = tensor(false)]; + tensor values_21_cast_fp16 = transpose(perm = var_2833, x = new_v_cache_21_cast_fp16)[name = tensor("transpose_64")]; + tensor attn_21_cast_fp16 = matmul(transpose_x = attn_21_transpose_x_0, transpose_y = attn_21_transpose_y_0, x = weights_21_cast_fp16, y = values_21_cast_fp16)[name = tensor("attn_21_cast_fp16")]; + tensor var_2878 = const()[name = tensor("op_2878"), val = tensor([0, 2, 1, 3])]; + tensor var_2883 = const()[name = tensor("op_2883"), val = tensor([1, 1, 1024])]; + tensor var_2879_cast_fp16 = transpose(perm = var_2878, x = attn_21_cast_fp16)[name = tensor("transpose_63")]; + tensor input_127_cast_fp16 = reshape(shape = var_2883, x = var_2879_cast_fp16)[name = tensor("input_127_cast_fp16")]; + tensor attn10_out_proj_weight_to_fp16 = const()[name = tensor("attn10_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(258124480)))]; + tensor linear_42_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn10_out_proj_weight_to_fp16, x = input_127_cast_fp16)[name = tensor("linear_42_cast_fp16")]; + tensor var_2889_to_fp16 = const()[name = tensor("op_2889_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2890_cast_fp16 = add(x = position10_to_fp16, y = var_2889_to_fp16)[name = tensor("op_2890_cast_fp16")]; + tensor var_2890_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2890_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_129_cast_fp16 = add(x = input_123_cast_fp16, y = linear_42_cast_fp16)[name = tensor("input_129_cast_fp16")]; + tensor input_131_axes_0 = const()[name = tensor("input_131_axes_0"), val = tensor([-1])]; + tensor norm10_2_weight_to_fp16 = const()[name = tensor("norm10_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260221696)))]; + tensor norm10_2_bias_to_fp16 = const()[name = tensor("norm10_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260223808)))]; + tensor var_2894_to_fp16 = const()[name = tensor("op_2894_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_131_cast_fp16 = layer_norm(axes = input_131_axes_0, beta = norm10_2_bias_to_fp16, epsilon = var_2894_to_fp16, gamma = norm10_2_weight_to_fp16, x = input_129_cast_fp16)[name = tensor("input_131_cast_fp16")]; + tensor linear10_1_weight_to_fp16 = const()[name = tensor("linear10_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260225920)))]; + tensor linear_43_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear10_1_weight_to_fp16, x = input_131_cast_fp16)[name = tensor("linear_43_cast_fp16")]; + tensor input_133_mode_0 = const()[name = tensor("input_133_mode_0"), val = tensor("EXACT")]; + tensor input_133_cast_fp16 = gelu(mode = input_133_mode_0, x = linear_43_cast_fp16)[name = tensor("input_133_cast_fp16")]; + tensor linear10_2_weight_to_fp16 = const()[name = tensor("linear10_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(268614592)))]; + tensor linear_44_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear10_2_weight_to_fp16, x = input_133_cast_fp16)[name = tensor("linear_44_cast_fp16")]; + tensor input_135_cast_fp16 = add(x = input_129_cast_fp16, y = linear_44_cast_fp16)[name = tensor("input_135_cast_fp16")]; + tensor input_137_axes_0 = const()[name = tensor("input_137_axes_0"), val = tensor([-1])]; + tensor norm11_1_weight_to_fp16 = const()[name = tensor("norm11_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277003264)))]; + tensor norm11_1_bias_to_fp16 = const()[name = tensor("norm11_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277005376)))]; + tensor var_2911_to_fp16 = const()[name = tensor("op_2911_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_137_cast_fp16 = layer_norm(axes = input_137_axes_0, beta = norm11_1_bias_to_fp16, epsilon = var_2911_to_fp16, gamma = norm11_1_weight_to_fp16, x = input_135_cast_fp16)[name = tensor("input_137_cast_fp16")]; + tensor attn11_in_proj_weight_to_fp16 = const()[name = tensor("attn11_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277007488)))]; + tensor linear_45_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn11_in_proj_weight_to_fp16, x = input_137_cast_fp16)[name = tensor("linear_45_cast_fp16")]; + tensor var_2924 = const()[name = tensor("op_2924"), val = tensor([1, 3, 16, 64])]; + tensor qkv_23_cast_fp16 = reshape(shape = var_2924, x = linear_45_cast_fp16)[name = tensor("qkv_23_cast_fp16")]; + tensor x_45_begin_0 = const()[name = tensor("x_45_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_45_end_0 = const()[name = tensor("x_45_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_45_end_mask_0 = const()[name = tensor("x_45_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_45_squeeze_mask_0 = const()[name = tensor("x_45_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_45_cast_fp16 = slice_by_index(begin = x_45_begin_0, end = x_45_end_0, end_mask = x_45_end_mask_0, squeeze_mask = x_45_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("x_45_cast_fp16")]; + tensor x_47_begin_0 = const()[name = tensor("x_47_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_47_end_0 = const()[name = tensor("x_47_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_47_end_mask_0 = const()[name = tensor("x_47_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_47_squeeze_mask_0 = const()[name = tensor("x_47_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_47_cast_fp16 = slice_by_index(begin = x_47_begin_0, end = x_47_end_0, end_mask = x_47_end_mask_0, squeeze_mask = x_47_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("x_47_cast_fp16")]; + tensor v_23_begin_0 = const()[name = tensor("v_23_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_23_end_0 = const()[name = tensor("v_23_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_23_end_mask_0 = const()[name = tensor("v_23_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_23_squeeze_mask_0 = const()[name = tensor("v_23_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_23_cast_fp16 = slice_by_index(begin = v_23_begin_0, end = v_23_end_0, end_mask = v_23_end_mask_0, squeeze_mask = v_23_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("v_23_cast_fp16")]; + tensor var_2953 = const()[name = tensor("op_2953"), val = tensor([1, 1, 1])]; + tensor position11_to_fp16_dtype_0 = const()[name = tensor("position11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position11_to_fp16 = cast(dtype = position11_to_fp16_dtype_0, x = position11)[name = tensor("cast_489")]; + tensor var_2954_cast_fp16 = reshape(shape = var_2953, x = position11_to_fp16)[name = tensor("op_2954_cast_fp16")]; + tensor angles_45_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_2954_cast_fp16)[name = tensor("angles_45_cast_fp16")]; + tensor rotr_45_cast_fp16 = cos(x = angles_45_cast_fp16)[name = tensor("rotr_45_cast_fp16")]; + tensor roti_45_cast_fp16 = sin(x = angles_45_cast_fp16)[name = tensor("roti_45_cast_fp16")]; + tensor var_2962 = const()[name = tensor("op_2962"), val = tensor([1, 16, 32, 2])]; + tensor pairs_45_cast_fp16 = reshape(shape = var_2962, x = x_45_cast_fp16)[name = tensor("pairs_45_cast_fp16")]; + tensor xr_45_begin_0 = const()[name = tensor("xr_45_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_45_end_0 = const()[name = tensor("xr_45_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_45_end_mask_0 = const()[name = tensor("xr_45_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_45_squeeze_mask_0 = const()[name = tensor("xr_45_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_45_cast_fp16 = slice_by_index(begin = xr_45_begin_0, end = xr_45_end_0, end_mask = xr_45_end_mask_0, squeeze_mask = xr_45_squeeze_mask_0, x = pairs_45_cast_fp16)[name = tensor("xr_45_cast_fp16")]; + tensor xi_45_begin_0 = const()[name = tensor("xi_45_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_45_end_0 = const()[name = tensor("xi_45_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_45_end_mask_0 = const()[name = tensor("xi_45_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_45_squeeze_mask_0 = const()[name = tensor("xi_45_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_45_cast_fp16 = slice_by_index(begin = xi_45_begin_0, end = xi_45_end_0, end_mask = xi_45_end_mask_0, squeeze_mask = xi_45_squeeze_mask_0, x = pairs_45_cast_fp16)[name = tensor("xi_45_cast_fp16")]; + tensor var_2970_cast_fp16 = mul(x = xr_45_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_2970_cast_fp16")]; + tensor var_2971_cast_fp16 = mul(x = xi_45_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_2971_cast_fp16")]; + tensor xor__45_cast_fp16 = sub(x = var_2970_cast_fp16, y = var_2971_cast_fp16)[name = tensor("xor__45_cast_fp16")]; + tensor var_2974_cast_fp16 = mul(x = xr_45_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_2974_cast_fp16")]; + tensor var_2975_cast_fp16 = mul(x = xi_45_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_2975_cast_fp16")]; + tensor xoi_45_cast_fp16 = add(x = var_2974_cast_fp16, y = var_2975_cast_fp16)[name = tensor("xoi_45_cast_fp16")]; + tensor out_45_axis_0 = const()[name = tensor("out_45_axis_0"), val = tensor(-1)]; + tensor out_45_cast_fp16 = stack(axis = out_45_axis_0, values = (xor__45_cast_fp16, xoi_45_cast_fp16))[name = tensor("out_45_cast_fp16")]; + tensor var_2998 = const()[name = tensor("op_2998"), val = tensor([1, 16, 32, 2])]; + tensor pairs_47_cast_fp16 = reshape(shape = var_2998, x = x_47_cast_fp16)[name = tensor("pairs_47_cast_fp16")]; + tensor xr_47_begin_0 = const()[name = tensor("xr_47_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_47_end_0 = const()[name = tensor("xr_47_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_47_end_mask_0 = const()[name = tensor("xr_47_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_47_squeeze_mask_0 = const()[name = tensor("xr_47_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_47_cast_fp16 = slice_by_index(begin = xr_47_begin_0, end = xr_47_end_0, end_mask = xr_47_end_mask_0, squeeze_mask = xr_47_squeeze_mask_0, x = pairs_47_cast_fp16)[name = tensor("xr_47_cast_fp16")]; + tensor xi_47_begin_0 = const()[name = tensor("xi_47_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_47_end_0 = const()[name = tensor("xi_47_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_47_end_mask_0 = const()[name = tensor("xi_47_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_47_squeeze_mask_0 = const()[name = tensor("xi_47_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_47_cast_fp16 = slice_by_index(begin = xi_47_begin_0, end = xi_47_end_0, end_mask = xi_47_end_mask_0, squeeze_mask = xi_47_squeeze_mask_0, x = pairs_47_cast_fp16)[name = tensor("xi_47_cast_fp16")]; + tensor var_3006_cast_fp16 = mul(x = xr_47_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_3006_cast_fp16")]; + tensor var_3007_cast_fp16 = mul(x = xi_47_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_3007_cast_fp16")]; + tensor xor__47_cast_fp16 = sub(x = var_3006_cast_fp16, y = var_3007_cast_fp16)[name = tensor("xor__47_cast_fp16")]; + tensor var_3010_cast_fp16 = mul(x = xr_47_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_3010_cast_fp16")]; + tensor var_3011_cast_fp16 = mul(x = xi_47_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_3011_cast_fp16")]; + tensor xoi_47_cast_fp16 = add(x = var_3010_cast_fp16, y = var_3011_cast_fp16)[name = tensor("xoi_47_cast_fp16")]; + tensor out_47_axis_0 = const()[name = tensor("out_47_axis_0"), val = tensor(-1)]; + tensor out_47_cast_fp16 = stack(axis = out_47_axis_0, values = (xor__47_cast_fp16, xoi_47_cast_fp16))[name = tensor("out_47_cast_fp16")]; + tensor _inversed_3023_y_0_to_fp16 = const()[name = tensor("_inversed_3023_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3023_cast_fp16 = mul(x = position11_to_fp16, y = _inversed_3023_y_0_to_fp16)[name = tensor("_inversed_3023_cast_fp16")]; + tensor var_3024_cast_fp16 = floor(x = _inversed_3023_cast_fp16)[name = tensor("op_3024_cast_fp16")]; + tensor var_3025_to_fp16 = const()[name = tensor("op_3025_to_fp16"), val = tensor(0x1p+9)]; + tensor var_3026_cast_fp16 = mul(x = var_3024_cast_fp16, y = var_3025_to_fp16)[name = tensor("op_3026_cast_fp16")]; + tensor write_idx_23_cast_fp16 = sub(x = position11_to_fp16, y = var_3026_cast_fp16)[name = tensor("write_idx_23_cast_fp16")]; + tensor var_3029_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_23_cast_fp16)[name = tensor("op_3029_cast_fp16")]; + tensor var_3039 = const()[name = tensor("op_3039"), val = tensor([1, 512, 1, 1])]; + tensor onehot_45_to_fp16_dtype_0 = const()[name = tensor("onehot_45_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_3029_cast_fp16_to_fp16 = cast(dtype = onehot_45_to_fp16_dtype_0, x = var_3029_cast_fp16)[name = tensor("cast_488")]; + tensor onehot_47_cast_fp16 = reshape(shape = var_3039, x = var_3029_cast_fp16_to_fp16)[name = tensor("onehot_47_cast_fp16")]; + tensor var_3041_to_fp16 = const()[name = tensor("op_3041_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3043_cast_fp16 = sub(x = var_3041_to_fp16, y = onehot_47_cast_fp16)[name = tensor("op_3043_cast_fp16")]; + tensor k_cache11_to_fp16_dtype_0 = const()[name = tensor("k_cache11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache11_to_fp16 = cast(dtype = k_cache11_to_fp16_dtype_0, x = k_cache11)[name = tensor("cast_487")]; + tensor var_3044_cast_fp16 = mul(x = k_cache11_to_fp16, y = var_3043_cast_fp16)[name = tensor("op_3044_cast_fp16")]; + tensor var_3049 = const()[name = tensor("op_3049"), val = tensor([1, 1, 16, 64])]; + tensor var_3050_cast_fp16 = reshape(shape = var_3049, x = out_47_cast_fp16)[name = tensor("op_3050_cast_fp16")]; + tensor var_3051_cast_fp16 = mul(x = var_3050_cast_fp16, y = onehot_47_cast_fp16)[name = tensor("op_3051_cast_fp16")]; + tensor new_k_cache_23_cast_fp16 = add(x = var_3044_cast_fp16, y = var_3051_cast_fp16)[name = tensor("new_k_cache_23_cast_fp16")]; + tensor new_k_cache_23_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_23_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache11_to_fp16_dtype_0 = const()[name = tensor("v_cache11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache11_to_fp16 = cast(dtype = v_cache11_to_fp16_dtype_0, x = v_cache11)[name = tensor("cast_485")]; + tensor var_3057_cast_fp16 = mul(x = v_cache11_to_fp16, y = var_3043_cast_fp16)[name = tensor("op_3057_cast_fp16")]; + tensor var_3062 = const()[name = tensor("op_3062"), val = tensor([1, 1, 16, 64])]; + tensor var_3063_cast_fp16 = reshape(shape = var_3062, x = v_23_cast_fp16)[name = tensor("op_3063_cast_fp16")]; + tensor var_3064_cast_fp16 = mul(x = var_3063_cast_fp16, y = onehot_47_cast_fp16)[name = tensor("op_3064_cast_fp16")]; + tensor new_v_cache_23_cast_fp16 = add(x = var_3057_cast_fp16, y = var_3064_cast_fp16)[name = tensor("new_v_cache_23_cast_fp16")]; + tensor new_v_cache_23_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_23_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_11_perm_0 = const()[name = tensor("transpose_11_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_3077 = const()[name = tensor("op_3077"), val = tensor([0, 2, 1, 3])]; + tensor var_3083 = const()[name = tensor("op_3083"), val = tensor([1, 16, 1, 64])]; + tensor q4_23_cast_fp16 = reshape(shape = var_3083, x = out_45_cast_fp16)[name = tensor("q4_23_cast_fp16")]; + tensor var_3088_transpose_x_0 = const()[name = tensor("op_3088_transpose_x_0"), val = tensor(false)]; + tensor var_3088_transpose_y_0 = const()[name = tensor("op_3088_transpose_y_0"), val = tensor(false)]; + tensor transpose_11_cast_fp16 = transpose(perm = transpose_11_perm_0, x = new_k_cache_23_cast_fp16)[name = tensor("transpose_62")]; + tensor var_3088_cast_fp16 = matmul(transpose_x = var_3088_transpose_x_0, transpose_y = var_3088_transpose_y_0, x = q4_23_cast_fp16, y = transpose_11_cast_fp16)[name = tensor("op_3088_cast_fp16")]; + tensor var_3089_to_fp16 = const()[name = tensor("op_3089_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_45_cast_fp16 = mul(x = var_3088_cast_fp16, y = var_3089_to_fp16)[name = tensor("scores_45_cast_fp16")]; + tensor var_3093 = const()[name = tensor("op_3093"), val = tensor([1, 1])]; + tensor var_3094_cast_fp16 = reshape(shape = var_3093, x = position11_to_fp16)[name = tensor("op_3094_cast_fp16")]; + tensor var_3095_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_3094_cast_fp16)[name = tensor("op_3095_cast_fp16")]; + tensor mask_23_to_fp16_dtype_0 = const()[name = tensor("mask_23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_3102_to_fp16 = const()[name = tensor("op_3102_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3095_cast_fp16_to_fp16 = cast(dtype = mask_23_to_fp16_dtype_0, x = var_3095_cast_fp16)[name = tensor("cast_483")]; + tensor var_3103_cast_fp16 = sub(x = var_3095_cast_fp16_to_fp16, y = var_3102_to_fp16)[name = tensor("op_3103_cast_fp16")]; + tensor var_3108 = const()[name = tensor("op_3108"), val = tensor([1, 1, 1, 512])]; + tensor var_3109_cast_fp16 = reshape(shape = var_3108, x = var_3103_cast_fp16)[name = tensor("op_3109_cast_fp16")]; + tensor var_3110_to_fp16 = const()[name = tensor("op_3110_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_3111_cast_fp16 = mul(x = var_3109_cast_fp16, y = var_3110_to_fp16)[name = tensor("op_3111_cast_fp16")]; + tensor scores_47_cast_fp16 = add(x = scores_45_cast_fp16, y = var_3111_cast_fp16)[name = tensor("scores_47_cast_fp16")]; + tensor var_3114 = const()[name = tensor("op_3114"), val = tensor(-1)]; + tensor weights_23_cast_fp16 = softmax(axis = var_3114, x = scores_47_cast_fp16)[name = tensor("weights_23_cast_fp16")]; + tensor attn_23_transpose_x_0 = const()[name = tensor("attn_23_transpose_x_0"), val = tensor(false)]; + tensor attn_23_transpose_y_0 = const()[name = tensor("attn_23_transpose_y_0"), val = tensor(false)]; + tensor values_23_cast_fp16 = transpose(perm = var_3077, x = new_v_cache_23_cast_fp16)[name = tensor("transpose_61")]; + tensor attn_23_cast_fp16 = matmul(transpose_x = attn_23_transpose_x_0, transpose_y = attn_23_transpose_y_0, x = weights_23_cast_fp16, y = values_23_cast_fp16)[name = tensor("attn_23_cast_fp16")]; + tensor var_3122 = const()[name = tensor("op_3122"), val = tensor([0, 2, 1, 3])]; + tensor var_3127 = const()[name = tensor("op_3127"), val = tensor([1, 1, 1024])]; + tensor var_3123_cast_fp16 = transpose(perm = var_3122, x = attn_23_cast_fp16)[name = tensor("transpose_60")]; + tensor input_139_cast_fp16 = reshape(shape = var_3127, x = var_3123_cast_fp16)[name = tensor("input_139_cast_fp16")]; + tensor attn11_out_proj_weight_to_fp16 = const()[name = tensor("attn11_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(283299008)))]; + tensor linear_46_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn11_out_proj_weight_to_fp16, x = input_139_cast_fp16)[name = tensor("linear_46_cast_fp16")]; + tensor var_3133_to_fp16 = const()[name = tensor("op_3133_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3134_cast_fp16 = add(x = position11_to_fp16, y = var_3133_to_fp16)[name = tensor("op_3134_cast_fp16")]; + tensor var_3134_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3134_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_141_cast_fp16 = add(x = input_135_cast_fp16, y = linear_46_cast_fp16)[name = tensor("input_141_cast_fp16")]; + tensor input_143_axes_0 = const()[name = tensor("input_143_axes_0"), val = tensor([-1])]; + tensor norm11_2_weight_to_fp16 = const()[name = tensor("norm11_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285396224)))]; + tensor norm11_2_bias_to_fp16 = const()[name = tensor("norm11_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285398336)))]; + tensor var_3138_to_fp16 = const()[name = tensor("op_3138_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_143_cast_fp16 = layer_norm(axes = input_143_axes_0, beta = norm11_2_bias_to_fp16, epsilon = var_3138_to_fp16, gamma = norm11_2_weight_to_fp16, x = input_141_cast_fp16)[name = tensor("input_143_cast_fp16")]; + tensor linear11_1_weight_to_fp16 = const()[name = tensor("linear11_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285400448)))]; + tensor linear_47_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear11_1_weight_to_fp16, x = input_143_cast_fp16)[name = tensor("linear_47_cast_fp16")]; + tensor input_145_mode_0 = const()[name = tensor("input_145_mode_0"), val = tensor("EXACT")]; + tensor input_145_cast_fp16 = gelu(mode = input_145_mode_0, x = linear_47_cast_fp16)[name = tensor("input_145_cast_fp16")]; + tensor linear11_2_weight_to_fp16 = const()[name = tensor("linear11_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293789120)))]; + tensor linear_48_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear11_2_weight_to_fp16, x = input_145_cast_fp16)[name = tensor("linear_48_cast_fp16")]; + tensor input_147_cast_fp16 = add(x = input_141_cast_fp16, y = linear_48_cast_fp16)[name = tensor("input_147_cast_fp16")]; + tensor input_149_axes_0 = const()[name = tensor("input_149_axes_0"), val = tensor([-1])]; + tensor norm12_1_weight_to_fp16 = const()[name = tensor("norm12_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302177792)))]; + tensor norm12_1_bias_to_fp16 = const()[name = tensor("norm12_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302179904)))]; + tensor var_3155_to_fp16 = const()[name = tensor("op_3155_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_149_cast_fp16 = layer_norm(axes = input_149_axes_0, beta = norm12_1_bias_to_fp16, epsilon = var_3155_to_fp16, gamma = norm12_1_weight_to_fp16, x = input_147_cast_fp16)[name = tensor("input_149_cast_fp16")]; + tensor attn12_in_proj_weight_to_fp16 = const()[name = tensor("attn12_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302182016)))]; + tensor linear_49_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn12_in_proj_weight_to_fp16, x = input_149_cast_fp16)[name = tensor("linear_49_cast_fp16")]; + tensor var_3168 = const()[name = tensor("op_3168"), val = tensor([1, 3, 16, 64])]; + tensor qkv_25_cast_fp16 = reshape(shape = var_3168, x = linear_49_cast_fp16)[name = tensor("qkv_25_cast_fp16")]; + tensor x_49_begin_0 = const()[name = tensor("x_49_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_49_end_0 = const()[name = tensor("x_49_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_49_end_mask_0 = const()[name = tensor("x_49_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_49_squeeze_mask_0 = const()[name = tensor("x_49_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_49_cast_fp16 = slice_by_index(begin = x_49_begin_0, end = x_49_end_0, end_mask = x_49_end_mask_0, squeeze_mask = x_49_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("x_49_cast_fp16")]; + tensor x_51_begin_0 = const()[name = tensor("x_51_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_51_end_0 = const()[name = tensor("x_51_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_51_end_mask_0 = const()[name = tensor("x_51_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_51_squeeze_mask_0 = const()[name = tensor("x_51_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_51_cast_fp16 = slice_by_index(begin = x_51_begin_0, end = x_51_end_0, end_mask = x_51_end_mask_0, squeeze_mask = x_51_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("x_51_cast_fp16")]; + tensor v_25_begin_0 = const()[name = tensor("v_25_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_25_end_0 = const()[name = tensor("v_25_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_25_end_mask_0 = const()[name = tensor("v_25_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_25_squeeze_mask_0 = const()[name = tensor("v_25_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_25_cast_fp16 = slice_by_index(begin = v_25_begin_0, end = v_25_end_0, end_mask = v_25_end_mask_0, squeeze_mask = v_25_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("v_25_cast_fp16")]; + tensor var_3197 = const()[name = tensor("op_3197"), val = tensor([1, 1, 1])]; + tensor position12_to_fp16_dtype_0 = const()[name = tensor("position12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position12_to_fp16 = cast(dtype = position12_to_fp16_dtype_0, x = position12)[name = tensor("cast_481")]; + tensor var_3198_cast_fp16 = reshape(shape = var_3197, x = position12_to_fp16)[name = tensor("op_3198_cast_fp16")]; + tensor angles_49_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_3198_cast_fp16)[name = tensor("angles_49_cast_fp16")]; + tensor rotr_49_cast_fp16 = cos(x = angles_49_cast_fp16)[name = tensor("rotr_49_cast_fp16")]; + tensor roti_49_cast_fp16 = sin(x = angles_49_cast_fp16)[name = tensor("roti_49_cast_fp16")]; + tensor var_3206 = const()[name = tensor("op_3206"), val = tensor([1, 16, 32, 2])]; + tensor pairs_49_cast_fp16 = reshape(shape = var_3206, x = x_49_cast_fp16)[name = tensor("pairs_49_cast_fp16")]; + tensor xr_49_begin_0 = const()[name = tensor("xr_49_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_49_end_0 = const()[name = tensor("xr_49_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_49_end_mask_0 = const()[name = tensor("xr_49_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_49_squeeze_mask_0 = const()[name = tensor("xr_49_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_49_cast_fp16 = slice_by_index(begin = xr_49_begin_0, end = xr_49_end_0, end_mask = xr_49_end_mask_0, squeeze_mask = xr_49_squeeze_mask_0, x = pairs_49_cast_fp16)[name = tensor("xr_49_cast_fp16")]; + tensor xi_49_begin_0 = const()[name = tensor("xi_49_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_49_end_0 = const()[name = tensor("xi_49_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_49_end_mask_0 = const()[name = tensor("xi_49_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_49_squeeze_mask_0 = const()[name = tensor("xi_49_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_49_cast_fp16 = slice_by_index(begin = xi_49_begin_0, end = xi_49_end_0, end_mask = xi_49_end_mask_0, squeeze_mask = xi_49_squeeze_mask_0, x = pairs_49_cast_fp16)[name = tensor("xi_49_cast_fp16")]; + tensor var_3214_cast_fp16 = mul(x = xr_49_cast_fp16, y = rotr_49_cast_fp16)[name = tensor("op_3214_cast_fp16")]; + tensor var_3215_cast_fp16 = mul(x = xi_49_cast_fp16, y = roti_49_cast_fp16)[name = tensor("op_3215_cast_fp16")]; + tensor xor__49_cast_fp16 = sub(x = var_3214_cast_fp16, y = var_3215_cast_fp16)[name = tensor("xor__49_cast_fp16")]; + tensor var_3218_cast_fp16 = mul(x = xr_49_cast_fp16, y = roti_49_cast_fp16)[name = tensor("op_3218_cast_fp16")]; + tensor var_3219_cast_fp16 = mul(x = xi_49_cast_fp16, y = rotr_49_cast_fp16)[name = tensor("op_3219_cast_fp16")]; + tensor xoi_49_cast_fp16 = add(x = var_3218_cast_fp16, y = var_3219_cast_fp16)[name = tensor("xoi_49_cast_fp16")]; + tensor out_49_axis_0 = const()[name = tensor("out_49_axis_0"), val = tensor(-1)]; + tensor out_49_cast_fp16 = stack(axis = out_49_axis_0, values = (xor__49_cast_fp16, xoi_49_cast_fp16))[name = tensor("out_49_cast_fp16")]; + tensor var_3242 = const()[name = tensor("op_3242"), val = tensor([1, 16, 32, 2])]; + tensor pairs_51_cast_fp16 = reshape(shape = var_3242, x = x_51_cast_fp16)[name = tensor("pairs_51_cast_fp16")]; + tensor xr_51_begin_0 = const()[name = tensor("xr_51_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_51_end_0 = const()[name = tensor("xr_51_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_51_end_mask_0 = const()[name = tensor("xr_51_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_51_squeeze_mask_0 = const()[name = tensor("xr_51_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_51_cast_fp16 = slice_by_index(begin = xr_51_begin_0, end = xr_51_end_0, end_mask = xr_51_end_mask_0, squeeze_mask = xr_51_squeeze_mask_0, x = pairs_51_cast_fp16)[name = tensor("xr_51_cast_fp16")]; + tensor xi_51_begin_0 = const()[name = tensor("xi_51_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_51_end_0 = const()[name = tensor("xi_51_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_51_end_mask_0 = const()[name = tensor("xi_51_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_51_squeeze_mask_0 = const()[name = tensor("xi_51_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_51_cast_fp16 = slice_by_index(begin = xi_51_begin_0, end = xi_51_end_0, end_mask = xi_51_end_mask_0, squeeze_mask = xi_51_squeeze_mask_0, x = pairs_51_cast_fp16)[name = tensor("xi_51_cast_fp16")]; + tensor var_3250_cast_fp16 = mul(x = xr_51_cast_fp16, y = rotr_49_cast_fp16)[name = tensor("op_3250_cast_fp16")]; + tensor var_3251_cast_fp16 = mul(x = xi_51_cast_fp16, y = roti_49_cast_fp16)[name = tensor("op_3251_cast_fp16")]; + tensor xor__51_cast_fp16 = sub(x = var_3250_cast_fp16, y = var_3251_cast_fp16)[name = tensor("xor__51_cast_fp16")]; + tensor var_3254_cast_fp16 = mul(x = xr_51_cast_fp16, y = roti_49_cast_fp16)[name = tensor("op_3254_cast_fp16")]; + tensor var_3255_cast_fp16 = mul(x = xi_51_cast_fp16, y = rotr_49_cast_fp16)[name = tensor("op_3255_cast_fp16")]; + tensor xoi_51_cast_fp16 = add(x = var_3254_cast_fp16, y = var_3255_cast_fp16)[name = tensor("xoi_51_cast_fp16")]; + tensor out_51_axis_0 = const()[name = tensor("out_51_axis_0"), val = tensor(-1)]; + tensor out_51_cast_fp16 = stack(axis = out_51_axis_0, values = (xor__51_cast_fp16, xoi_51_cast_fp16))[name = tensor("out_51_cast_fp16")]; + tensor _inversed_3267_y_0_to_fp16 = const()[name = tensor("_inversed_3267_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3267_cast_fp16 = mul(x = position12_to_fp16, y = _inversed_3267_y_0_to_fp16)[name = tensor("_inversed_3267_cast_fp16")]; + tensor var_3268_cast_fp16 = floor(x = _inversed_3267_cast_fp16)[name = tensor("op_3268_cast_fp16")]; + tensor var_3269_to_fp16 = const()[name = tensor("op_3269_to_fp16"), val = tensor(0x1p+9)]; + tensor var_3270_cast_fp16 = mul(x = var_3268_cast_fp16, y = var_3269_to_fp16)[name = tensor("op_3270_cast_fp16")]; + tensor write_idx_25_cast_fp16 = sub(x = position12_to_fp16, y = var_3270_cast_fp16)[name = tensor("write_idx_25_cast_fp16")]; + tensor var_3273_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_25_cast_fp16)[name = tensor("op_3273_cast_fp16")]; + tensor var_3283 = const()[name = tensor("op_3283"), val = tensor([1, 512, 1, 1])]; + tensor onehot_49_to_fp16_dtype_0 = const()[name = tensor("onehot_49_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_3273_cast_fp16_to_fp16 = cast(dtype = onehot_49_to_fp16_dtype_0, x = var_3273_cast_fp16)[name = tensor("cast_480")]; + tensor onehot_51_cast_fp16 = reshape(shape = var_3283, x = var_3273_cast_fp16_to_fp16)[name = tensor("onehot_51_cast_fp16")]; + tensor var_3285_to_fp16 = const()[name = tensor("op_3285_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3287_cast_fp16 = sub(x = var_3285_to_fp16, y = onehot_51_cast_fp16)[name = tensor("op_3287_cast_fp16")]; + tensor k_cache12_to_fp16_dtype_0 = const()[name = tensor("k_cache12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache12_to_fp16 = cast(dtype = k_cache12_to_fp16_dtype_0, x = k_cache12)[name = tensor("cast_479")]; + tensor var_3288_cast_fp16 = mul(x = k_cache12_to_fp16, y = var_3287_cast_fp16)[name = tensor("op_3288_cast_fp16")]; + tensor var_3293 = const()[name = tensor("op_3293"), val = tensor([1, 1, 16, 64])]; + tensor var_3294_cast_fp16 = reshape(shape = var_3293, x = out_51_cast_fp16)[name = tensor("op_3294_cast_fp16")]; + tensor var_3295_cast_fp16 = mul(x = var_3294_cast_fp16, y = onehot_51_cast_fp16)[name = tensor("op_3295_cast_fp16")]; + tensor new_k_cache_25_cast_fp16 = add(x = var_3288_cast_fp16, y = var_3295_cast_fp16)[name = tensor("new_k_cache_25_cast_fp16")]; + tensor new_k_cache_25_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_25_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache12_to_fp16_dtype_0 = const()[name = tensor("v_cache12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache12_to_fp16 = cast(dtype = v_cache12_to_fp16_dtype_0, x = v_cache12)[name = tensor("cast_477")]; + tensor var_3301_cast_fp16 = mul(x = v_cache12_to_fp16, y = var_3287_cast_fp16)[name = tensor("op_3301_cast_fp16")]; + tensor var_3306 = const()[name = tensor("op_3306"), val = tensor([1, 1, 16, 64])]; + tensor var_3307_cast_fp16 = reshape(shape = var_3306, x = v_25_cast_fp16)[name = tensor("op_3307_cast_fp16")]; + tensor var_3308_cast_fp16 = mul(x = var_3307_cast_fp16, y = onehot_51_cast_fp16)[name = tensor("op_3308_cast_fp16")]; + tensor new_v_cache_25_cast_fp16 = add(x = var_3301_cast_fp16, y = var_3308_cast_fp16)[name = tensor("new_v_cache_25_cast_fp16")]; + tensor new_v_cache_25_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_25_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_12_perm_0 = const()[name = tensor("transpose_12_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_3321 = const()[name = tensor("op_3321"), val = tensor([0, 2, 1, 3])]; + tensor var_3327 = const()[name = tensor("op_3327"), val = tensor([1, 16, 1, 64])]; + tensor q4_25_cast_fp16 = reshape(shape = var_3327, x = out_49_cast_fp16)[name = tensor("q4_25_cast_fp16")]; + tensor var_3332_transpose_x_0 = const()[name = tensor("op_3332_transpose_x_0"), val = tensor(false)]; + tensor var_3332_transpose_y_0 = const()[name = tensor("op_3332_transpose_y_0"), val = tensor(false)]; + tensor transpose_12_cast_fp16 = transpose(perm = transpose_12_perm_0, x = new_k_cache_25_cast_fp16)[name = tensor("transpose_59")]; + tensor var_3332_cast_fp16 = matmul(transpose_x = var_3332_transpose_x_0, transpose_y = var_3332_transpose_y_0, x = q4_25_cast_fp16, y = transpose_12_cast_fp16)[name = tensor("op_3332_cast_fp16")]; + tensor var_3333_to_fp16 = const()[name = tensor("op_3333_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_49_cast_fp16 = mul(x = var_3332_cast_fp16, y = var_3333_to_fp16)[name = tensor("scores_49_cast_fp16")]; + tensor var_3337 = const()[name = tensor("op_3337"), val = tensor([1, 1])]; + tensor var_3338_cast_fp16 = reshape(shape = var_3337, x = position12_to_fp16)[name = tensor("op_3338_cast_fp16")]; + tensor var_3339_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_3338_cast_fp16)[name = tensor("op_3339_cast_fp16")]; + tensor mask_25_to_fp16_dtype_0 = const()[name = tensor("mask_25_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_3346_to_fp16 = const()[name = tensor("op_3346_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3339_cast_fp16_to_fp16 = cast(dtype = mask_25_to_fp16_dtype_0, x = var_3339_cast_fp16)[name = tensor("cast_475")]; + tensor var_3347_cast_fp16 = sub(x = var_3339_cast_fp16_to_fp16, y = var_3346_to_fp16)[name = tensor("op_3347_cast_fp16")]; + tensor var_3352 = const()[name = tensor("op_3352"), val = tensor([1, 1, 1, 512])]; + tensor var_3353_cast_fp16 = reshape(shape = var_3352, x = var_3347_cast_fp16)[name = tensor("op_3353_cast_fp16")]; + tensor var_3354_to_fp16 = const()[name = tensor("op_3354_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_3355_cast_fp16 = mul(x = var_3353_cast_fp16, y = var_3354_to_fp16)[name = tensor("op_3355_cast_fp16")]; + tensor scores_51_cast_fp16 = add(x = scores_49_cast_fp16, y = var_3355_cast_fp16)[name = tensor("scores_51_cast_fp16")]; + tensor var_3358 = const()[name = tensor("op_3358"), val = tensor(-1)]; + tensor weights_25_cast_fp16 = softmax(axis = var_3358, x = scores_51_cast_fp16)[name = tensor("weights_25_cast_fp16")]; + tensor attn_25_transpose_x_0 = const()[name = tensor("attn_25_transpose_x_0"), val = tensor(false)]; + tensor attn_25_transpose_y_0 = const()[name = tensor("attn_25_transpose_y_0"), val = tensor(false)]; + tensor values_25_cast_fp16 = transpose(perm = var_3321, x = new_v_cache_25_cast_fp16)[name = tensor("transpose_58")]; + tensor attn_25_cast_fp16 = matmul(transpose_x = attn_25_transpose_x_0, transpose_y = attn_25_transpose_y_0, x = weights_25_cast_fp16, y = values_25_cast_fp16)[name = tensor("attn_25_cast_fp16")]; + tensor var_3366 = const()[name = tensor("op_3366"), val = tensor([0, 2, 1, 3])]; + tensor var_3371 = const()[name = tensor("op_3371"), val = tensor([1, 1, 1024])]; + tensor var_3367_cast_fp16 = transpose(perm = var_3366, x = attn_25_cast_fp16)[name = tensor("transpose_57")]; + tensor input_151_cast_fp16 = reshape(shape = var_3371, x = var_3367_cast_fp16)[name = tensor("input_151_cast_fp16")]; + tensor attn12_out_proj_weight_to_fp16 = const()[name = tensor("attn12_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(308473536)))]; + tensor linear_50_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn12_out_proj_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_50_cast_fp16")]; + tensor var_3377_to_fp16 = const()[name = tensor("op_3377_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3378_cast_fp16 = add(x = position12_to_fp16, y = var_3377_to_fp16)[name = tensor("op_3378_cast_fp16")]; + tensor var_3378_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3378_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_153_cast_fp16 = add(x = input_147_cast_fp16, y = linear_50_cast_fp16)[name = tensor("input_153_cast_fp16")]; + tensor input_155_axes_0 = const()[name = tensor("input_155_axes_0"), val = tensor([-1])]; + tensor norm12_2_weight_to_fp16 = const()[name = tensor("norm12_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(310570752)))]; + tensor norm12_2_bias_to_fp16 = const()[name = tensor("norm12_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(310572864)))]; + tensor var_3382_to_fp16 = const()[name = tensor("op_3382_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_155_cast_fp16 = layer_norm(axes = input_155_axes_0, beta = norm12_2_bias_to_fp16, epsilon = var_3382_to_fp16, gamma = norm12_2_weight_to_fp16, x = input_153_cast_fp16)[name = tensor("input_155_cast_fp16")]; + tensor linear12_1_weight_to_fp16 = const()[name = tensor("linear12_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(310574976)))]; + tensor linear_51_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear12_1_weight_to_fp16, x = input_155_cast_fp16)[name = tensor("linear_51_cast_fp16")]; + tensor input_157_mode_0 = const()[name = tensor("input_157_mode_0"), val = tensor("EXACT")]; + tensor input_157_cast_fp16 = gelu(mode = input_157_mode_0, x = linear_51_cast_fp16)[name = tensor("input_157_cast_fp16")]; + tensor linear12_2_weight_to_fp16 = const()[name = tensor("linear12_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(318963648)))]; + tensor linear_52_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear12_2_weight_to_fp16, x = input_157_cast_fp16)[name = tensor("linear_52_cast_fp16")]; + tensor input_159_cast_fp16 = add(x = input_153_cast_fp16, y = linear_52_cast_fp16)[name = tensor("input_159_cast_fp16")]; + tensor input_161_axes_0 = const()[name = tensor("input_161_axes_0"), val = tensor([-1])]; + tensor norm13_1_weight_to_fp16 = const()[name = tensor("norm13_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(327352320)))]; + tensor norm13_1_bias_to_fp16 = const()[name = tensor("norm13_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(327354432)))]; + tensor var_3399_to_fp16 = const()[name = tensor("op_3399_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_161_cast_fp16 = layer_norm(axes = input_161_axes_0, beta = norm13_1_bias_to_fp16, epsilon = var_3399_to_fp16, gamma = norm13_1_weight_to_fp16, x = input_159_cast_fp16)[name = tensor("input_161_cast_fp16")]; + tensor attn13_in_proj_weight_to_fp16 = const()[name = tensor("attn13_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(327356544)))]; + tensor linear_53_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn13_in_proj_weight_to_fp16, x = input_161_cast_fp16)[name = tensor("linear_53_cast_fp16")]; + tensor var_3412 = const()[name = tensor("op_3412"), val = tensor([1, 3, 16, 64])]; + tensor qkv_27_cast_fp16 = reshape(shape = var_3412, x = linear_53_cast_fp16)[name = tensor("qkv_27_cast_fp16")]; + tensor x_53_begin_0 = const()[name = tensor("x_53_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_53_end_0 = const()[name = tensor("x_53_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_53_end_mask_0 = const()[name = tensor("x_53_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_53_squeeze_mask_0 = const()[name = tensor("x_53_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_53_cast_fp16 = slice_by_index(begin = x_53_begin_0, end = x_53_end_0, end_mask = x_53_end_mask_0, squeeze_mask = x_53_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("x_53_cast_fp16")]; + tensor x_55_begin_0 = const()[name = tensor("x_55_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_55_end_0 = const()[name = tensor("x_55_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_55_end_mask_0 = const()[name = tensor("x_55_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_55_squeeze_mask_0 = const()[name = tensor("x_55_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_55_cast_fp16 = slice_by_index(begin = x_55_begin_0, end = x_55_end_0, end_mask = x_55_end_mask_0, squeeze_mask = x_55_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("x_55_cast_fp16")]; + tensor v_27_begin_0 = const()[name = tensor("v_27_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_27_end_0 = const()[name = tensor("v_27_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_27_end_mask_0 = const()[name = tensor("v_27_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_27_squeeze_mask_0 = const()[name = tensor("v_27_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_27_cast_fp16 = slice_by_index(begin = v_27_begin_0, end = v_27_end_0, end_mask = v_27_end_mask_0, squeeze_mask = v_27_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("v_27_cast_fp16")]; + tensor var_3441 = const()[name = tensor("op_3441"), val = tensor([1, 1, 1])]; + tensor position13_to_fp16_dtype_0 = const()[name = tensor("position13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position13_to_fp16 = cast(dtype = position13_to_fp16_dtype_0, x = position13)[name = tensor("cast_473")]; + tensor var_3442_cast_fp16 = reshape(shape = var_3441, x = position13_to_fp16)[name = tensor("op_3442_cast_fp16")]; + tensor angles_53_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_3442_cast_fp16)[name = tensor("angles_53_cast_fp16")]; + tensor rotr_53_cast_fp16 = cos(x = angles_53_cast_fp16)[name = tensor("rotr_53_cast_fp16")]; + tensor roti_53_cast_fp16 = sin(x = angles_53_cast_fp16)[name = tensor("roti_53_cast_fp16")]; + tensor var_3450 = const()[name = tensor("op_3450"), val = tensor([1, 16, 32, 2])]; + tensor pairs_53_cast_fp16 = reshape(shape = var_3450, x = x_53_cast_fp16)[name = tensor("pairs_53_cast_fp16")]; + tensor xr_53_begin_0 = const()[name = tensor("xr_53_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_53_end_0 = const()[name = tensor("xr_53_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_53_end_mask_0 = const()[name = tensor("xr_53_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_53_squeeze_mask_0 = const()[name = tensor("xr_53_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_53_cast_fp16 = slice_by_index(begin = xr_53_begin_0, end = xr_53_end_0, end_mask = xr_53_end_mask_0, squeeze_mask = xr_53_squeeze_mask_0, x = pairs_53_cast_fp16)[name = tensor("xr_53_cast_fp16")]; + tensor xi_53_begin_0 = const()[name = tensor("xi_53_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_53_end_0 = const()[name = tensor("xi_53_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_53_end_mask_0 = const()[name = tensor("xi_53_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_53_squeeze_mask_0 = const()[name = tensor("xi_53_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_53_cast_fp16 = slice_by_index(begin = xi_53_begin_0, end = xi_53_end_0, end_mask = xi_53_end_mask_0, squeeze_mask = xi_53_squeeze_mask_0, x = pairs_53_cast_fp16)[name = tensor("xi_53_cast_fp16")]; + tensor var_3458_cast_fp16 = mul(x = xr_53_cast_fp16, y = rotr_53_cast_fp16)[name = tensor("op_3458_cast_fp16")]; + tensor var_3459_cast_fp16 = mul(x = xi_53_cast_fp16, y = roti_53_cast_fp16)[name = tensor("op_3459_cast_fp16")]; + tensor xor__53_cast_fp16 = sub(x = var_3458_cast_fp16, y = var_3459_cast_fp16)[name = tensor("xor__53_cast_fp16")]; + tensor var_3462_cast_fp16 = mul(x = xr_53_cast_fp16, y = roti_53_cast_fp16)[name = tensor("op_3462_cast_fp16")]; + tensor var_3463_cast_fp16 = mul(x = xi_53_cast_fp16, y = rotr_53_cast_fp16)[name = tensor("op_3463_cast_fp16")]; + tensor xoi_53_cast_fp16 = add(x = var_3462_cast_fp16, y = var_3463_cast_fp16)[name = tensor("xoi_53_cast_fp16")]; + tensor out_53_axis_0 = const()[name = tensor("out_53_axis_0"), val = tensor(-1)]; + tensor out_53_cast_fp16 = stack(axis = out_53_axis_0, values = (xor__53_cast_fp16, xoi_53_cast_fp16))[name = tensor("out_53_cast_fp16")]; + tensor var_3486 = const()[name = tensor("op_3486"), val = tensor([1, 16, 32, 2])]; + tensor pairs_55_cast_fp16 = reshape(shape = var_3486, x = x_55_cast_fp16)[name = tensor("pairs_55_cast_fp16")]; + tensor xr_55_begin_0 = const()[name = tensor("xr_55_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_55_end_0 = const()[name = tensor("xr_55_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_55_end_mask_0 = const()[name = tensor("xr_55_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_55_squeeze_mask_0 = const()[name = tensor("xr_55_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_55_cast_fp16 = slice_by_index(begin = xr_55_begin_0, end = xr_55_end_0, end_mask = xr_55_end_mask_0, squeeze_mask = xr_55_squeeze_mask_0, x = pairs_55_cast_fp16)[name = tensor("xr_55_cast_fp16")]; + tensor xi_55_begin_0 = const()[name = tensor("xi_55_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_55_end_0 = const()[name = tensor("xi_55_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_55_end_mask_0 = const()[name = tensor("xi_55_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_55_squeeze_mask_0 = const()[name = tensor("xi_55_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_55_cast_fp16 = slice_by_index(begin = xi_55_begin_0, end = xi_55_end_0, end_mask = xi_55_end_mask_0, squeeze_mask = xi_55_squeeze_mask_0, x = pairs_55_cast_fp16)[name = tensor("xi_55_cast_fp16")]; + tensor var_3494_cast_fp16 = mul(x = xr_55_cast_fp16, y = rotr_53_cast_fp16)[name = tensor("op_3494_cast_fp16")]; + tensor var_3495_cast_fp16 = mul(x = xi_55_cast_fp16, y = roti_53_cast_fp16)[name = tensor("op_3495_cast_fp16")]; + tensor xor__55_cast_fp16 = sub(x = var_3494_cast_fp16, y = var_3495_cast_fp16)[name = tensor("xor__55_cast_fp16")]; + tensor var_3498_cast_fp16 = mul(x = xr_55_cast_fp16, y = roti_53_cast_fp16)[name = tensor("op_3498_cast_fp16")]; + tensor var_3499_cast_fp16 = mul(x = xi_55_cast_fp16, y = rotr_53_cast_fp16)[name = tensor("op_3499_cast_fp16")]; + tensor xoi_55_cast_fp16 = add(x = var_3498_cast_fp16, y = var_3499_cast_fp16)[name = tensor("xoi_55_cast_fp16")]; + tensor out_55_axis_0 = const()[name = tensor("out_55_axis_0"), val = tensor(-1)]; + tensor out_55_cast_fp16 = stack(axis = out_55_axis_0, values = (xor__55_cast_fp16, xoi_55_cast_fp16))[name = tensor("out_55_cast_fp16")]; + tensor _inversed_3511_y_0_to_fp16 = const()[name = tensor("_inversed_3511_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3511_cast_fp16 = mul(x = position13_to_fp16, y = _inversed_3511_y_0_to_fp16)[name = tensor("_inversed_3511_cast_fp16")]; + tensor var_3512_cast_fp16 = floor(x = _inversed_3511_cast_fp16)[name = tensor("op_3512_cast_fp16")]; + tensor var_3513_to_fp16 = const()[name = tensor("op_3513_to_fp16"), val = tensor(0x1p+9)]; + tensor var_3514_cast_fp16 = mul(x = var_3512_cast_fp16, y = var_3513_to_fp16)[name = tensor("op_3514_cast_fp16")]; + tensor write_idx_27_cast_fp16 = sub(x = position13_to_fp16, y = var_3514_cast_fp16)[name = tensor("write_idx_27_cast_fp16")]; + tensor var_3517_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_27_cast_fp16)[name = tensor("op_3517_cast_fp16")]; + tensor var_3527 = const()[name = tensor("op_3527"), val = tensor([1, 512, 1, 1])]; + tensor onehot_53_to_fp16_dtype_0 = const()[name = tensor("onehot_53_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_3517_cast_fp16_to_fp16 = cast(dtype = onehot_53_to_fp16_dtype_0, x = var_3517_cast_fp16)[name = tensor("cast_472")]; + tensor onehot_55_cast_fp16 = reshape(shape = var_3527, x = var_3517_cast_fp16_to_fp16)[name = tensor("onehot_55_cast_fp16")]; + tensor var_3529_to_fp16 = const()[name = tensor("op_3529_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3531_cast_fp16 = sub(x = var_3529_to_fp16, y = onehot_55_cast_fp16)[name = tensor("op_3531_cast_fp16")]; + tensor k_cache13_to_fp16_dtype_0 = const()[name = tensor("k_cache13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache13_to_fp16 = cast(dtype = k_cache13_to_fp16_dtype_0, x = k_cache13)[name = tensor("cast_471")]; + tensor var_3532_cast_fp16 = mul(x = k_cache13_to_fp16, y = var_3531_cast_fp16)[name = tensor("op_3532_cast_fp16")]; + tensor var_3537 = const()[name = tensor("op_3537"), val = tensor([1, 1, 16, 64])]; + tensor var_3538_cast_fp16 = reshape(shape = var_3537, x = out_55_cast_fp16)[name = tensor("op_3538_cast_fp16")]; + tensor var_3539_cast_fp16 = mul(x = var_3538_cast_fp16, y = onehot_55_cast_fp16)[name = tensor("op_3539_cast_fp16")]; + tensor new_k_cache_27_cast_fp16 = add(x = var_3532_cast_fp16, y = var_3539_cast_fp16)[name = tensor("new_k_cache_27_cast_fp16")]; + tensor new_k_cache_27_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_27_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache13_to_fp16_dtype_0 = const()[name = tensor("v_cache13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache13_to_fp16 = cast(dtype = v_cache13_to_fp16_dtype_0, x = v_cache13)[name = tensor("cast_469")]; + tensor var_3545_cast_fp16 = mul(x = v_cache13_to_fp16, y = var_3531_cast_fp16)[name = tensor("op_3545_cast_fp16")]; + tensor var_3550 = const()[name = tensor("op_3550"), val = tensor([1, 1, 16, 64])]; + tensor var_3551_cast_fp16 = reshape(shape = var_3550, x = v_27_cast_fp16)[name = tensor("op_3551_cast_fp16")]; + tensor var_3552_cast_fp16 = mul(x = var_3551_cast_fp16, y = onehot_55_cast_fp16)[name = tensor("op_3552_cast_fp16")]; + tensor new_v_cache_27_cast_fp16 = add(x = var_3545_cast_fp16, y = var_3552_cast_fp16)[name = tensor("new_v_cache_27_cast_fp16")]; + tensor new_v_cache_27_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_27_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_13_perm_0 = const()[name = tensor("transpose_13_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_3565 = const()[name = tensor("op_3565"), val = tensor([0, 2, 1, 3])]; + tensor var_3571 = const()[name = tensor("op_3571"), val = tensor([1, 16, 1, 64])]; + tensor q4_27_cast_fp16 = reshape(shape = var_3571, x = out_53_cast_fp16)[name = tensor("q4_27_cast_fp16")]; + tensor var_3576_transpose_x_0 = const()[name = tensor("op_3576_transpose_x_0"), val = tensor(false)]; + tensor var_3576_transpose_y_0 = const()[name = tensor("op_3576_transpose_y_0"), val = tensor(false)]; + tensor transpose_13_cast_fp16 = transpose(perm = transpose_13_perm_0, x = new_k_cache_27_cast_fp16)[name = tensor("transpose_56")]; + tensor var_3576_cast_fp16 = matmul(transpose_x = var_3576_transpose_x_0, transpose_y = var_3576_transpose_y_0, x = q4_27_cast_fp16, y = transpose_13_cast_fp16)[name = tensor("op_3576_cast_fp16")]; + tensor var_3577_to_fp16 = const()[name = tensor("op_3577_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_53_cast_fp16 = mul(x = var_3576_cast_fp16, y = var_3577_to_fp16)[name = tensor("scores_53_cast_fp16")]; + tensor var_3581 = const()[name = tensor("op_3581"), val = tensor([1, 1])]; + tensor var_3582_cast_fp16 = reshape(shape = var_3581, x = position13_to_fp16)[name = tensor("op_3582_cast_fp16")]; + tensor var_3583_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_3582_cast_fp16)[name = tensor("op_3583_cast_fp16")]; + tensor mask_27_to_fp16_dtype_0 = const()[name = tensor("mask_27_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_3590_to_fp16 = const()[name = tensor("op_3590_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3583_cast_fp16_to_fp16 = cast(dtype = mask_27_to_fp16_dtype_0, x = var_3583_cast_fp16)[name = tensor("cast_467")]; + tensor var_3591_cast_fp16 = sub(x = var_3583_cast_fp16_to_fp16, y = var_3590_to_fp16)[name = tensor("op_3591_cast_fp16")]; + tensor var_3596 = const()[name = tensor("op_3596"), val = tensor([1, 1, 1, 512])]; + tensor var_3597_cast_fp16 = reshape(shape = var_3596, x = var_3591_cast_fp16)[name = tensor("op_3597_cast_fp16")]; + tensor var_3598_to_fp16 = const()[name = tensor("op_3598_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_3599_cast_fp16 = mul(x = var_3597_cast_fp16, y = var_3598_to_fp16)[name = tensor("op_3599_cast_fp16")]; + tensor scores_55_cast_fp16 = add(x = scores_53_cast_fp16, y = var_3599_cast_fp16)[name = tensor("scores_55_cast_fp16")]; + tensor var_3602 = const()[name = tensor("op_3602"), val = tensor(-1)]; + tensor weights_27_cast_fp16 = softmax(axis = var_3602, x = scores_55_cast_fp16)[name = tensor("weights_27_cast_fp16")]; + tensor attn_27_transpose_x_0 = const()[name = tensor("attn_27_transpose_x_0"), val = tensor(false)]; + tensor attn_27_transpose_y_0 = const()[name = tensor("attn_27_transpose_y_0"), val = tensor(false)]; + tensor values_27_cast_fp16 = transpose(perm = var_3565, x = new_v_cache_27_cast_fp16)[name = tensor("transpose_55")]; + tensor attn_27_cast_fp16 = matmul(transpose_x = attn_27_transpose_x_0, transpose_y = attn_27_transpose_y_0, x = weights_27_cast_fp16, y = values_27_cast_fp16)[name = tensor("attn_27_cast_fp16")]; + tensor var_3610 = const()[name = tensor("op_3610"), val = tensor([0, 2, 1, 3])]; + tensor var_3615 = const()[name = tensor("op_3615"), val = tensor([1, 1, 1024])]; + tensor var_3611_cast_fp16 = transpose(perm = var_3610, x = attn_27_cast_fp16)[name = tensor("transpose_54")]; + tensor input_163_cast_fp16 = reshape(shape = var_3615, x = var_3611_cast_fp16)[name = tensor("input_163_cast_fp16")]; + tensor attn13_out_proj_weight_to_fp16 = const()[name = tensor("attn13_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(333648064)))]; + tensor linear_54_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn13_out_proj_weight_to_fp16, x = input_163_cast_fp16)[name = tensor("linear_54_cast_fp16")]; + tensor var_3621_to_fp16 = const()[name = tensor("op_3621_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3622_cast_fp16 = add(x = position13_to_fp16, y = var_3621_to_fp16)[name = tensor("op_3622_cast_fp16")]; + tensor var_3622_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3622_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_165_cast_fp16 = add(x = input_159_cast_fp16, y = linear_54_cast_fp16)[name = tensor("input_165_cast_fp16")]; + tensor input_167_axes_0 = const()[name = tensor("input_167_axes_0"), val = tensor([-1])]; + tensor norm13_2_weight_to_fp16 = const()[name = tensor("norm13_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335745280)))]; + tensor norm13_2_bias_to_fp16 = const()[name = tensor("norm13_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335747392)))]; + tensor var_3626_to_fp16 = const()[name = tensor("op_3626_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_167_cast_fp16 = layer_norm(axes = input_167_axes_0, beta = norm13_2_bias_to_fp16, epsilon = var_3626_to_fp16, gamma = norm13_2_weight_to_fp16, x = input_165_cast_fp16)[name = tensor("input_167_cast_fp16")]; + tensor linear13_1_weight_to_fp16 = const()[name = tensor("linear13_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335749504)))]; + tensor linear_55_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear13_1_weight_to_fp16, x = input_167_cast_fp16)[name = tensor("linear_55_cast_fp16")]; + tensor input_169_mode_0 = const()[name = tensor("input_169_mode_0"), val = tensor("EXACT")]; + tensor input_169_cast_fp16 = gelu(mode = input_169_mode_0, x = linear_55_cast_fp16)[name = tensor("input_169_cast_fp16")]; + tensor linear13_2_weight_to_fp16 = const()[name = tensor("linear13_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(344138176)))]; + tensor linear_56_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear13_2_weight_to_fp16, x = input_169_cast_fp16)[name = tensor("linear_56_cast_fp16")]; + tensor input_171_cast_fp16 = add(x = input_165_cast_fp16, y = linear_56_cast_fp16)[name = tensor("input_171_cast_fp16")]; + tensor input_173_axes_0 = const()[name = tensor("input_173_axes_0"), val = tensor([-1])]; + tensor norm14_1_weight_to_fp16 = const()[name = tensor("norm14_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(352526848)))]; + tensor norm14_1_bias_to_fp16 = const()[name = tensor("norm14_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(352528960)))]; + tensor var_3643_to_fp16 = const()[name = tensor("op_3643_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_173_cast_fp16 = layer_norm(axes = input_173_axes_0, beta = norm14_1_bias_to_fp16, epsilon = var_3643_to_fp16, gamma = norm14_1_weight_to_fp16, x = input_171_cast_fp16)[name = tensor("input_173_cast_fp16")]; + tensor attn14_in_proj_weight_to_fp16 = const()[name = tensor("attn14_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(352531072)))]; + tensor linear_57_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn14_in_proj_weight_to_fp16, x = input_173_cast_fp16)[name = tensor("linear_57_cast_fp16")]; + tensor var_3656 = const()[name = tensor("op_3656"), val = tensor([1, 3, 16, 64])]; + tensor qkv_29_cast_fp16 = reshape(shape = var_3656, x = linear_57_cast_fp16)[name = tensor("qkv_29_cast_fp16")]; + tensor x_57_begin_0 = const()[name = tensor("x_57_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_57_end_0 = const()[name = tensor("x_57_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_57_end_mask_0 = const()[name = tensor("x_57_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_57_squeeze_mask_0 = const()[name = tensor("x_57_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_57_cast_fp16 = slice_by_index(begin = x_57_begin_0, end = x_57_end_0, end_mask = x_57_end_mask_0, squeeze_mask = x_57_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("x_57_cast_fp16")]; + tensor x_59_begin_0 = const()[name = tensor("x_59_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_59_end_0 = const()[name = tensor("x_59_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_59_end_mask_0 = const()[name = tensor("x_59_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_59_squeeze_mask_0 = const()[name = tensor("x_59_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_59_cast_fp16 = slice_by_index(begin = x_59_begin_0, end = x_59_end_0, end_mask = x_59_end_mask_0, squeeze_mask = x_59_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("x_59_cast_fp16")]; + tensor v_29_begin_0 = const()[name = tensor("v_29_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_29_end_0 = const()[name = tensor("v_29_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_29_end_mask_0 = const()[name = tensor("v_29_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_29_squeeze_mask_0 = const()[name = tensor("v_29_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_29_cast_fp16 = slice_by_index(begin = v_29_begin_0, end = v_29_end_0, end_mask = v_29_end_mask_0, squeeze_mask = v_29_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("v_29_cast_fp16")]; + tensor var_3685 = const()[name = tensor("op_3685"), val = tensor([1, 1, 1])]; + tensor position14_to_fp16_dtype_0 = const()[name = tensor("position14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position14_to_fp16 = cast(dtype = position14_to_fp16_dtype_0, x = position14)[name = tensor("cast_465")]; + tensor var_3686_cast_fp16 = reshape(shape = var_3685, x = position14_to_fp16)[name = tensor("op_3686_cast_fp16")]; + tensor angles_57_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_3686_cast_fp16)[name = tensor("angles_57_cast_fp16")]; + tensor rotr_57_cast_fp16 = cos(x = angles_57_cast_fp16)[name = tensor("rotr_57_cast_fp16")]; + tensor roti_57_cast_fp16 = sin(x = angles_57_cast_fp16)[name = tensor("roti_57_cast_fp16")]; + tensor var_3694 = const()[name = tensor("op_3694"), val = tensor([1, 16, 32, 2])]; + tensor pairs_57_cast_fp16 = reshape(shape = var_3694, x = x_57_cast_fp16)[name = tensor("pairs_57_cast_fp16")]; + tensor xr_57_begin_0 = const()[name = tensor("xr_57_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_57_end_0 = const()[name = tensor("xr_57_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_57_end_mask_0 = const()[name = tensor("xr_57_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_57_squeeze_mask_0 = const()[name = tensor("xr_57_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_57_cast_fp16 = slice_by_index(begin = xr_57_begin_0, end = xr_57_end_0, end_mask = xr_57_end_mask_0, squeeze_mask = xr_57_squeeze_mask_0, x = pairs_57_cast_fp16)[name = tensor("xr_57_cast_fp16")]; + tensor xi_57_begin_0 = const()[name = tensor("xi_57_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_57_end_0 = const()[name = tensor("xi_57_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_57_end_mask_0 = const()[name = tensor("xi_57_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_57_squeeze_mask_0 = const()[name = tensor("xi_57_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_57_cast_fp16 = slice_by_index(begin = xi_57_begin_0, end = xi_57_end_0, end_mask = xi_57_end_mask_0, squeeze_mask = xi_57_squeeze_mask_0, x = pairs_57_cast_fp16)[name = tensor("xi_57_cast_fp16")]; + tensor var_3702_cast_fp16 = mul(x = xr_57_cast_fp16, y = rotr_57_cast_fp16)[name = tensor("op_3702_cast_fp16")]; + tensor var_3703_cast_fp16 = mul(x = xi_57_cast_fp16, y = roti_57_cast_fp16)[name = tensor("op_3703_cast_fp16")]; + tensor xor__57_cast_fp16 = sub(x = var_3702_cast_fp16, y = var_3703_cast_fp16)[name = tensor("xor__57_cast_fp16")]; + tensor var_3706_cast_fp16 = mul(x = xr_57_cast_fp16, y = roti_57_cast_fp16)[name = tensor("op_3706_cast_fp16")]; + tensor var_3707_cast_fp16 = mul(x = xi_57_cast_fp16, y = rotr_57_cast_fp16)[name = tensor("op_3707_cast_fp16")]; + tensor xoi_57_cast_fp16 = add(x = var_3706_cast_fp16, y = var_3707_cast_fp16)[name = tensor("xoi_57_cast_fp16")]; + tensor out_57_axis_0 = const()[name = tensor("out_57_axis_0"), val = tensor(-1)]; + tensor out_57_cast_fp16 = stack(axis = out_57_axis_0, values = (xor__57_cast_fp16, xoi_57_cast_fp16))[name = tensor("out_57_cast_fp16")]; + tensor var_3730 = const()[name = tensor("op_3730"), val = tensor([1, 16, 32, 2])]; + tensor pairs_59_cast_fp16 = reshape(shape = var_3730, x = x_59_cast_fp16)[name = tensor("pairs_59_cast_fp16")]; + tensor xr_59_begin_0 = const()[name = tensor("xr_59_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_59_end_0 = const()[name = tensor("xr_59_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_59_end_mask_0 = const()[name = tensor("xr_59_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_59_squeeze_mask_0 = const()[name = tensor("xr_59_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_59_cast_fp16 = slice_by_index(begin = xr_59_begin_0, end = xr_59_end_0, end_mask = xr_59_end_mask_0, squeeze_mask = xr_59_squeeze_mask_0, x = pairs_59_cast_fp16)[name = tensor("xr_59_cast_fp16")]; + tensor xi_59_begin_0 = const()[name = tensor("xi_59_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_59_end_0 = const()[name = tensor("xi_59_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_59_end_mask_0 = const()[name = tensor("xi_59_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_59_squeeze_mask_0 = const()[name = tensor("xi_59_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_59_cast_fp16 = slice_by_index(begin = xi_59_begin_0, end = xi_59_end_0, end_mask = xi_59_end_mask_0, squeeze_mask = xi_59_squeeze_mask_0, x = pairs_59_cast_fp16)[name = tensor("xi_59_cast_fp16")]; + tensor var_3738_cast_fp16 = mul(x = xr_59_cast_fp16, y = rotr_57_cast_fp16)[name = tensor("op_3738_cast_fp16")]; + tensor var_3739_cast_fp16 = mul(x = xi_59_cast_fp16, y = roti_57_cast_fp16)[name = tensor("op_3739_cast_fp16")]; + tensor xor__59_cast_fp16 = sub(x = var_3738_cast_fp16, y = var_3739_cast_fp16)[name = tensor("xor__59_cast_fp16")]; + tensor var_3742_cast_fp16 = mul(x = xr_59_cast_fp16, y = roti_57_cast_fp16)[name = tensor("op_3742_cast_fp16")]; + tensor var_3743_cast_fp16 = mul(x = xi_59_cast_fp16, y = rotr_57_cast_fp16)[name = tensor("op_3743_cast_fp16")]; + tensor xoi_59_cast_fp16 = add(x = var_3742_cast_fp16, y = var_3743_cast_fp16)[name = tensor("xoi_59_cast_fp16")]; + tensor out_59_axis_0 = const()[name = tensor("out_59_axis_0"), val = tensor(-1)]; + tensor out_59_cast_fp16 = stack(axis = out_59_axis_0, values = (xor__59_cast_fp16, xoi_59_cast_fp16))[name = tensor("out_59_cast_fp16")]; + tensor _inversed_3755_y_0_to_fp16 = const()[name = tensor("_inversed_3755_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3755_cast_fp16 = mul(x = position14_to_fp16, y = _inversed_3755_y_0_to_fp16)[name = tensor("_inversed_3755_cast_fp16")]; + tensor var_3756_cast_fp16 = floor(x = _inversed_3755_cast_fp16)[name = tensor("op_3756_cast_fp16")]; + tensor var_3757_to_fp16 = const()[name = tensor("op_3757_to_fp16"), val = tensor(0x1p+9)]; + tensor var_3758_cast_fp16 = mul(x = var_3756_cast_fp16, y = var_3757_to_fp16)[name = tensor("op_3758_cast_fp16")]; + tensor write_idx_29_cast_fp16 = sub(x = position14_to_fp16, y = var_3758_cast_fp16)[name = tensor("write_idx_29_cast_fp16")]; + tensor var_3761_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_29_cast_fp16)[name = tensor("op_3761_cast_fp16")]; + tensor var_3771 = const()[name = tensor("op_3771"), val = tensor([1, 512, 1, 1])]; + tensor onehot_57_to_fp16_dtype_0 = const()[name = tensor("onehot_57_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_3761_cast_fp16_to_fp16 = cast(dtype = onehot_57_to_fp16_dtype_0, x = var_3761_cast_fp16)[name = tensor("cast_464")]; + tensor onehot_59_cast_fp16 = reshape(shape = var_3771, x = var_3761_cast_fp16_to_fp16)[name = tensor("onehot_59_cast_fp16")]; + tensor var_3773_to_fp16 = const()[name = tensor("op_3773_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3775_cast_fp16 = sub(x = var_3773_to_fp16, y = onehot_59_cast_fp16)[name = tensor("op_3775_cast_fp16")]; + tensor k_cache14_to_fp16_dtype_0 = const()[name = tensor("k_cache14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache14_to_fp16 = cast(dtype = k_cache14_to_fp16_dtype_0, x = k_cache14)[name = tensor("cast_463")]; + tensor var_3776_cast_fp16 = mul(x = k_cache14_to_fp16, y = var_3775_cast_fp16)[name = tensor("op_3776_cast_fp16")]; + tensor var_3781 = const()[name = tensor("op_3781"), val = tensor([1, 1, 16, 64])]; + tensor var_3782_cast_fp16 = reshape(shape = var_3781, x = out_59_cast_fp16)[name = tensor("op_3782_cast_fp16")]; + tensor var_3783_cast_fp16 = mul(x = var_3782_cast_fp16, y = onehot_59_cast_fp16)[name = tensor("op_3783_cast_fp16")]; + tensor new_k_cache_29_cast_fp16 = add(x = var_3776_cast_fp16, y = var_3783_cast_fp16)[name = tensor("new_k_cache_29_cast_fp16")]; + tensor new_k_cache_29_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_29_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache14_to_fp16_dtype_0 = const()[name = tensor("v_cache14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache14_to_fp16 = cast(dtype = v_cache14_to_fp16_dtype_0, x = v_cache14)[name = tensor("cast_461")]; + tensor var_3789_cast_fp16 = mul(x = v_cache14_to_fp16, y = var_3775_cast_fp16)[name = tensor("op_3789_cast_fp16")]; + tensor var_3794 = const()[name = tensor("op_3794"), val = tensor([1, 1, 16, 64])]; + tensor var_3795_cast_fp16 = reshape(shape = var_3794, x = v_29_cast_fp16)[name = tensor("op_3795_cast_fp16")]; + tensor var_3796_cast_fp16 = mul(x = var_3795_cast_fp16, y = onehot_59_cast_fp16)[name = tensor("op_3796_cast_fp16")]; + tensor new_v_cache_29_cast_fp16 = add(x = var_3789_cast_fp16, y = var_3796_cast_fp16)[name = tensor("new_v_cache_29_cast_fp16")]; + tensor new_v_cache_29_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_29_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_14_perm_0 = const()[name = tensor("transpose_14_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_3809 = const()[name = tensor("op_3809"), val = tensor([0, 2, 1, 3])]; + tensor var_3815 = const()[name = tensor("op_3815"), val = tensor([1, 16, 1, 64])]; + tensor q4_29_cast_fp16 = reshape(shape = var_3815, x = out_57_cast_fp16)[name = tensor("q4_29_cast_fp16")]; + tensor var_3820_transpose_x_0 = const()[name = tensor("op_3820_transpose_x_0"), val = tensor(false)]; + tensor var_3820_transpose_y_0 = const()[name = tensor("op_3820_transpose_y_0"), val = tensor(false)]; + tensor transpose_14_cast_fp16 = transpose(perm = transpose_14_perm_0, x = new_k_cache_29_cast_fp16)[name = tensor("transpose_53")]; + tensor var_3820_cast_fp16 = matmul(transpose_x = var_3820_transpose_x_0, transpose_y = var_3820_transpose_y_0, x = q4_29_cast_fp16, y = transpose_14_cast_fp16)[name = tensor("op_3820_cast_fp16")]; + tensor var_3821_to_fp16 = const()[name = tensor("op_3821_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_57_cast_fp16 = mul(x = var_3820_cast_fp16, y = var_3821_to_fp16)[name = tensor("scores_57_cast_fp16")]; + tensor var_3825 = const()[name = tensor("op_3825"), val = tensor([1, 1])]; + tensor var_3826_cast_fp16 = reshape(shape = var_3825, x = position14_to_fp16)[name = tensor("op_3826_cast_fp16")]; + tensor var_3827_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_3826_cast_fp16)[name = tensor("op_3827_cast_fp16")]; + tensor mask_29_to_fp16_dtype_0 = const()[name = tensor("mask_29_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_3834_to_fp16 = const()[name = tensor("op_3834_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3827_cast_fp16_to_fp16 = cast(dtype = mask_29_to_fp16_dtype_0, x = var_3827_cast_fp16)[name = tensor("cast_459")]; + tensor var_3835_cast_fp16 = sub(x = var_3827_cast_fp16_to_fp16, y = var_3834_to_fp16)[name = tensor("op_3835_cast_fp16")]; + tensor var_3840 = const()[name = tensor("op_3840"), val = tensor([1, 1, 1, 512])]; + tensor var_3841_cast_fp16 = reshape(shape = var_3840, x = var_3835_cast_fp16)[name = tensor("op_3841_cast_fp16")]; + tensor var_3842_to_fp16 = const()[name = tensor("op_3842_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_3843_cast_fp16 = mul(x = var_3841_cast_fp16, y = var_3842_to_fp16)[name = tensor("op_3843_cast_fp16")]; + tensor scores_59_cast_fp16 = add(x = scores_57_cast_fp16, y = var_3843_cast_fp16)[name = tensor("scores_59_cast_fp16")]; + tensor var_3846 = const()[name = tensor("op_3846"), val = tensor(-1)]; + tensor weights_29_cast_fp16 = softmax(axis = var_3846, x = scores_59_cast_fp16)[name = tensor("weights_29_cast_fp16")]; + tensor attn_29_transpose_x_0 = const()[name = tensor("attn_29_transpose_x_0"), val = tensor(false)]; + tensor attn_29_transpose_y_0 = const()[name = tensor("attn_29_transpose_y_0"), val = tensor(false)]; + tensor values_29_cast_fp16 = transpose(perm = var_3809, x = new_v_cache_29_cast_fp16)[name = tensor("transpose_52")]; + tensor attn_29_cast_fp16 = matmul(transpose_x = attn_29_transpose_x_0, transpose_y = attn_29_transpose_y_0, x = weights_29_cast_fp16, y = values_29_cast_fp16)[name = tensor("attn_29_cast_fp16")]; + tensor var_3854 = const()[name = tensor("op_3854"), val = tensor([0, 2, 1, 3])]; + tensor var_3859 = const()[name = tensor("op_3859"), val = tensor([1, 1, 1024])]; + tensor var_3855_cast_fp16 = transpose(perm = var_3854, x = attn_29_cast_fp16)[name = tensor("transpose_51")]; + tensor input_175_cast_fp16 = reshape(shape = var_3859, x = var_3855_cast_fp16)[name = tensor("input_175_cast_fp16")]; + tensor attn14_out_proj_weight_to_fp16 = const()[name = tensor("attn14_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(358822592)))]; + tensor linear_58_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn14_out_proj_weight_to_fp16, x = input_175_cast_fp16)[name = tensor("linear_58_cast_fp16")]; + tensor var_3865_to_fp16 = const()[name = tensor("op_3865_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3866_cast_fp16 = add(x = position14_to_fp16, y = var_3865_to_fp16)[name = tensor("op_3866_cast_fp16")]; + tensor var_3866_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3866_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_177_cast_fp16 = add(x = input_171_cast_fp16, y = linear_58_cast_fp16)[name = tensor("input_177_cast_fp16")]; + tensor input_179_axes_0 = const()[name = tensor("input_179_axes_0"), val = tensor([-1])]; + tensor norm14_2_weight_to_fp16 = const()[name = tensor("norm14_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(360919808)))]; + tensor norm14_2_bias_to_fp16 = const()[name = tensor("norm14_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(360921920)))]; + tensor var_3870_to_fp16 = const()[name = tensor("op_3870_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_179_cast_fp16 = layer_norm(axes = input_179_axes_0, beta = norm14_2_bias_to_fp16, epsilon = var_3870_to_fp16, gamma = norm14_2_weight_to_fp16, x = input_177_cast_fp16)[name = tensor("input_179_cast_fp16")]; + tensor linear14_1_weight_to_fp16 = const()[name = tensor("linear14_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(360924032)))]; + tensor linear_59_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear14_1_weight_to_fp16, x = input_179_cast_fp16)[name = tensor("linear_59_cast_fp16")]; + tensor input_181_mode_0 = const()[name = tensor("input_181_mode_0"), val = tensor("EXACT")]; + tensor input_181_cast_fp16 = gelu(mode = input_181_mode_0, x = linear_59_cast_fp16)[name = tensor("input_181_cast_fp16")]; + tensor linear14_2_weight_to_fp16 = const()[name = tensor("linear14_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(369312704)))]; + tensor linear_60_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear14_2_weight_to_fp16, x = input_181_cast_fp16)[name = tensor("linear_60_cast_fp16")]; + tensor input_183_cast_fp16 = add(x = input_177_cast_fp16, y = linear_60_cast_fp16)[name = tensor("input_183_cast_fp16")]; + tensor input_185_axes_0 = const()[name = tensor("input_185_axes_0"), val = tensor([-1])]; + tensor norm15_1_weight_to_fp16 = const()[name = tensor("norm15_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(377701376)))]; + tensor norm15_1_bias_to_fp16 = const()[name = tensor("norm15_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(377703488)))]; + tensor var_3887_to_fp16 = const()[name = tensor("op_3887_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_185_cast_fp16 = layer_norm(axes = input_185_axes_0, beta = norm15_1_bias_to_fp16, epsilon = var_3887_to_fp16, gamma = norm15_1_weight_to_fp16, x = input_183_cast_fp16)[name = tensor("input_185_cast_fp16")]; + tensor attn15_in_proj_weight_to_fp16 = const()[name = tensor("attn15_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(377705600)))]; + tensor linear_61_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn15_in_proj_weight_to_fp16, x = input_185_cast_fp16)[name = tensor("linear_61_cast_fp16")]; + tensor var_3900 = const()[name = tensor("op_3900"), val = tensor([1, 3, 16, 64])]; + tensor qkv_31_cast_fp16 = reshape(shape = var_3900, x = linear_61_cast_fp16)[name = tensor("qkv_31_cast_fp16")]; + tensor x_61_begin_0 = const()[name = tensor("x_61_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_61_end_0 = const()[name = tensor("x_61_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_61_end_mask_0 = const()[name = tensor("x_61_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_61_squeeze_mask_0 = const()[name = tensor("x_61_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_61_cast_fp16 = slice_by_index(begin = x_61_begin_0, end = x_61_end_0, end_mask = x_61_end_mask_0, squeeze_mask = x_61_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("x_61_cast_fp16")]; + tensor x_63_begin_0 = const()[name = tensor("x_63_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_63_end_0 = const()[name = tensor("x_63_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_63_end_mask_0 = const()[name = tensor("x_63_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_63_squeeze_mask_0 = const()[name = tensor("x_63_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_63_cast_fp16 = slice_by_index(begin = x_63_begin_0, end = x_63_end_0, end_mask = x_63_end_mask_0, squeeze_mask = x_63_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("x_63_cast_fp16")]; + tensor v_31_begin_0 = const()[name = tensor("v_31_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_31_end_0 = const()[name = tensor("v_31_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_31_end_mask_0 = const()[name = tensor("v_31_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_31_squeeze_mask_0 = const()[name = tensor("v_31_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_31_cast_fp16 = slice_by_index(begin = v_31_begin_0, end = v_31_end_0, end_mask = v_31_end_mask_0, squeeze_mask = v_31_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("v_31_cast_fp16")]; + tensor var_3929 = const()[name = tensor("op_3929"), val = tensor([1, 1, 1])]; + tensor position15_to_fp16_dtype_0 = const()[name = tensor("position15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position15_to_fp16 = cast(dtype = position15_to_fp16_dtype_0, x = position15)[name = tensor("cast_457")]; + tensor var_3930_cast_fp16 = reshape(shape = var_3929, x = position15_to_fp16)[name = tensor("op_3930_cast_fp16")]; + tensor angles_61_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_3930_cast_fp16)[name = tensor("angles_61_cast_fp16")]; + tensor rotr_61_cast_fp16 = cos(x = angles_61_cast_fp16)[name = tensor("rotr_61_cast_fp16")]; + tensor roti_61_cast_fp16 = sin(x = angles_61_cast_fp16)[name = tensor("roti_61_cast_fp16")]; + tensor var_3938 = const()[name = tensor("op_3938"), val = tensor([1, 16, 32, 2])]; + tensor pairs_61_cast_fp16 = reshape(shape = var_3938, x = x_61_cast_fp16)[name = tensor("pairs_61_cast_fp16")]; + tensor xr_61_begin_0 = const()[name = tensor("xr_61_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_61_end_0 = const()[name = tensor("xr_61_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_61_end_mask_0 = const()[name = tensor("xr_61_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_61_squeeze_mask_0 = const()[name = tensor("xr_61_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_61_cast_fp16 = slice_by_index(begin = xr_61_begin_0, end = xr_61_end_0, end_mask = xr_61_end_mask_0, squeeze_mask = xr_61_squeeze_mask_0, x = pairs_61_cast_fp16)[name = tensor("xr_61_cast_fp16")]; + tensor xi_61_begin_0 = const()[name = tensor("xi_61_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_61_end_0 = const()[name = tensor("xi_61_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_61_end_mask_0 = const()[name = tensor("xi_61_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_61_squeeze_mask_0 = const()[name = tensor("xi_61_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_61_cast_fp16 = slice_by_index(begin = xi_61_begin_0, end = xi_61_end_0, end_mask = xi_61_end_mask_0, squeeze_mask = xi_61_squeeze_mask_0, x = pairs_61_cast_fp16)[name = tensor("xi_61_cast_fp16")]; + tensor var_3946_cast_fp16 = mul(x = xr_61_cast_fp16, y = rotr_61_cast_fp16)[name = tensor("op_3946_cast_fp16")]; + tensor var_3947_cast_fp16 = mul(x = xi_61_cast_fp16, y = roti_61_cast_fp16)[name = tensor("op_3947_cast_fp16")]; + tensor xor__61_cast_fp16 = sub(x = var_3946_cast_fp16, y = var_3947_cast_fp16)[name = tensor("xor__61_cast_fp16")]; + tensor var_3950_cast_fp16 = mul(x = xr_61_cast_fp16, y = roti_61_cast_fp16)[name = tensor("op_3950_cast_fp16")]; + tensor var_3951_cast_fp16 = mul(x = xi_61_cast_fp16, y = rotr_61_cast_fp16)[name = tensor("op_3951_cast_fp16")]; + tensor xoi_61_cast_fp16 = add(x = var_3950_cast_fp16, y = var_3951_cast_fp16)[name = tensor("xoi_61_cast_fp16")]; + tensor out_61_axis_0 = const()[name = tensor("out_61_axis_0"), val = tensor(-1)]; + tensor out_61_cast_fp16 = stack(axis = out_61_axis_0, values = (xor__61_cast_fp16, xoi_61_cast_fp16))[name = tensor("out_61_cast_fp16")]; + tensor var_3974 = const()[name = tensor("op_3974"), val = tensor([1, 16, 32, 2])]; + tensor pairs_63_cast_fp16 = reshape(shape = var_3974, x = x_63_cast_fp16)[name = tensor("pairs_63_cast_fp16")]; + tensor xr_63_begin_0 = const()[name = tensor("xr_63_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_63_end_0 = const()[name = tensor("xr_63_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_63_end_mask_0 = const()[name = tensor("xr_63_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_63_squeeze_mask_0 = const()[name = tensor("xr_63_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_63_cast_fp16 = slice_by_index(begin = xr_63_begin_0, end = xr_63_end_0, end_mask = xr_63_end_mask_0, squeeze_mask = xr_63_squeeze_mask_0, x = pairs_63_cast_fp16)[name = tensor("xr_63_cast_fp16")]; + tensor xi_63_begin_0 = const()[name = tensor("xi_63_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_63_end_0 = const()[name = tensor("xi_63_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_63_end_mask_0 = const()[name = tensor("xi_63_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_63_squeeze_mask_0 = const()[name = tensor("xi_63_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_63_cast_fp16 = slice_by_index(begin = xi_63_begin_0, end = xi_63_end_0, end_mask = xi_63_end_mask_0, squeeze_mask = xi_63_squeeze_mask_0, x = pairs_63_cast_fp16)[name = tensor("xi_63_cast_fp16")]; + tensor var_3982_cast_fp16 = mul(x = xr_63_cast_fp16, y = rotr_61_cast_fp16)[name = tensor("op_3982_cast_fp16")]; + tensor var_3983_cast_fp16 = mul(x = xi_63_cast_fp16, y = roti_61_cast_fp16)[name = tensor("op_3983_cast_fp16")]; + tensor xor__63_cast_fp16 = sub(x = var_3982_cast_fp16, y = var_3983_cast_fp16)[name = tensor("xor__63_cast_fp16")]; + tensor var_3986_cast_fp16 = mul(x = xr_63_cast_fp16, y = roti_61_cast_fp16)[name = tensor("op_3986_cast_fp16")]; + tensor var_3987_cast_fp16 = mul(x = xi_63_cast_fp16, y = rotr_61_cast_fp16)[name = tensor("op_3987_cast_fp16")]; + tensor xoi_63_cast_fp16 = add(x = var_3986_cast_fp16, y = var_3987_cast_fp16)[name = tensor("xoi_63_cast_fp16")]; + tensor out_63_axis_0 = const()[name = tensor("out_63_axis_0"), val = tensor(-1)]; + tensor out_63_cast_fp16 = stack(axis = out_63_axis_0, values = (xor__63_cast_fp16, xoi_63_cast_fp16))[name = tensor("out_63_cast_fp16")]; + tensor _inversed_3999_y_0_to_fp16 = const()[name = tensor("_inversed_3999_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3999_cast_fp16 = mul(x = position15_to_fp16, y = _inversed_3999_y_0_to_fp16)[name = tensor("_inversed_3999_cast_fp16")]; + tensor var_4000_cast_fp16 = floor(x = _inversed_3999_cast_fp16)[name = tensor("op_4000_cast_fp16")]; + tensor var_4001_to_fp16 = const()[name = tensor("op_4001_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4002_cast_fp16 = mul(x = var_4000_cast_fp16, y = var_4001_to_fp16)[name = tensor("op_4002_cast_fp16")]; + tensor write_idx_31_cast_fp16 = sub(x = position15_to_fp16, y = var_4002_cast_fp16)[name = tensor("write_idx_31_cast_fp16")]; + tensor var_4005_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_31_cast_fp16)[name = tensor("op_4005_cast_fp16")]; + tensor var_4015 = const()[name = tensor("op_4015"), val = tensor([1, 512, 1, 1])]; + tensor onehot_61_to_fp16_dtype_0 = const()[name = tensor("onehot_61_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_4005_cast_fp16_to_fp16 = cast(dtype = onehot_61_to_fp16_dtype_0, x = var_4005_cast_fp16)[name = tensor("cast_456")]; + tensor onehot_63_cast_fp16 = reshape(shape = var_4015, x = var_4005_cast_fp16_to_fp16)[name = tensor("onehot_63_cast_fp16")]; + tensor var_4017_to_fp16 = const()[name = tensor("op_4017_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4019_cast_fp16 = sub(x = var_4017_to_fp16, y = onehot_63_cast_fp16)[name = tensor("op_4019_cast_fp16")]; + tensor k_cache15_to_fp16_dtype_0 = const()[name = tensor("k_cache15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache15_to_fp16 = cast(dtype = k_cache15_to_fp16_dtype_0, x = k_cache15)[name = tensor("cast_455")]; + tensor var_4020_cast_fp16 = mul(x = k_cache15_to_fp16, y = var_4019_cast_fp16)[name = tensor("op_4020_cast_fp16")]; + tensor var_4025 = const()[name = tensor("op_4025"), val = tensor([1, 1, 16, 64])]; + tensor var_4026_cast_fp16 = reshape(shape = var_4025, x = out_63_cast_fp16)[name = tensor("op_4026_cast_fp16")]; + tensor var_4027_cast_fp16 = mul(x = var_4026_cast_fp16, y = onehot_63_cast_fp16)[name = tensor("op_4027_cast_fp16")]; + tensor new_k_cache_31_cast_fp16 = add(x = var_4020_cast_fp16, y = var_4027_cast_fp16)[name = tensor("new_k_cache_31_cast_fp16")]; + tensor new_k_cache_31_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_31_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache15_to_fp16_dtype_0 = const()[name = tensor("v_cache15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache15_to_fp16 = cast(dtype = v_cache15_to_fp16_dtype_0, x = v_cache15)[name = tensor("cast_453")]; + tensor var_4033_cast_fp16 = mul(x = v_cache15_to_fp16, y = var_4019_cast_fp16)[name = tensor("op_4033_cast_fp16")]; + tensor var_4038 = const()[name = tensor("op_4038"), val = tensor([1, 1, 16, 64])]; + tensor var_4039_cast_fp16 = reshape(shape = var_4038, x = v_31_cast_fp16)[name = tensor("op_4039_cast_fp16")]; + tensor var_4040_cast_fp16 = mul(x = var_4039_cast_fp16, y = onehot_63_cast_fp16)[name = tensor("op_4040_cast_fp16")]; + tensor new_v_cache_31_cast_fp16 = add(x = var_4033_cast_fp16, y = var_4040_cast_fp16)[name = tensor("new_v_cache_31_cast_fp16")]; + tensor new_v_cache_31_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_31_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_15_perm_0 = const()[name = tensor("transpose_15_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_4053 = const()[name = tensor("op_4053"), val = tensor([0, 2, 1, 3])]; + tensor var_4059 = const()[name = tensor("op_4059"), val = tensor([1, 16, 1, 64])]; + tensor q4_31_cast_fp16 = reshape(shape = var_4059, x = out_61_cast_fp16)[name = tensor("q4_31_cast_fp16")]; + tensor var_4064_transpose_x_0 = const()[name = tensor("op_4064_transpose_x_0"), val = tensor(false)]; + tensor var_4064_transpose_y_0 = const()[name = tensor("op_4064_transpose_y_0"), val = tensor(false)]; + tensor transpose_15_cast_fp16 = transpose(perm = transpose_15_perm_0, x = new_k_cache_31_cast_fp16)[name = tensor("transpose_50")]; + tensor var_4064_cast_fp16 = matmul(transpose_x = var_4064_transpose_x_0, transpose_y = var_4064_transpose_y_0, x = q4_31_cast_fp16, y = transpose_15_cast_fp16)[name = tensor("op_4064_cast_fp16")]; + tensor var_4065_to_fp16 = const()[name = tensor("op_4065_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_61_cast_fp16 = mul(x = var_4064_cast_fp16, y = var_4065_to_fp16)[name = tensor("scores_61_cast_fp16")]; + tensor var_4069 = const()[name = tensor("op_4069"), val = tensor([1, 1])]; + tensor var_4070_cast_fp16 = reshape(shape = var_4069, x = position15_to_fp16)[name = tensor("op_4070_cast_fp16")]; + tensor var_4071_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_4070_cast_fp16)[name = tensor("op_4071_cast_fp16")]; + tensor mask_31_to_fp16_dtype_0 = const()[name = tensor("mask_31_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_4078_to_fp16 = const()[name = tensor("op_4078_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4071_cast_fp16_to_fp16 = cast(dtype = mask_31_to_fp16_dtype_0, x = var_4071_cast_fp16)[name = tensor("cast_451")]; + tensor var_4079_cast_fp16 = sub(x = var_4071_cast_fp16_to_fp16, y = var_4078_to_fp16)[name = tensor("op_4079_cast_fp16")]; + tensor var_4084 = const()[name = tensor("op_4084"), val = tensor([1, 1, 1, 512])]; + tensor var_4085_cast_fp16 = reshape(shape = var_4084, x = var_4079_cast_fp16)[name = tensor("op_4085_cast_fp16")]; + tensor var_4086_to_fp16 = const()[name = tensor("op_4086_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_4087_cast_fp16 = mul(x = var_4085_cast_fp16, y = var_4086_to_fp16)[name = tensor("op_4087_cast_fp16")]; + tensor scores_63_cast_fp16 = add(x = scores_61_cast_fp16, y = var_4087_cast_fp16)[name = tensor("scores_63_cast_fp16")]; + tensor var_4090 = const()[name = tensor("op_4090"), val = tensor(-1)]; + tensor weights_31_cast_fp16 = softmax(axis = var_4090, x = scores_63_cast_fp16)[name = tensor("weights_31_cast_fp16")]; + tensor attn_31_transpose_x_0 = const()[name = tensor("attn_31_transpose_x_0"), val = tensor(false)]; + tensor attn_31_transpose_y_0 = const()[name = tensor("attn_31_transpose_y_0"), val = tensor(false)]; + tensor values_31_cast_fp16 = transpose(perm = var_4053, x = new_v_cache_31_cast_fp16)[name = tensor("transpose_49")]; + tensor attn_31_cast_fp16 = matmul(transpose_x = attn_31_transpose_x_0, transpose_y = attn_31_transpose_y_0, x = weights_31_cast_fp16, y = values_31_cast_fp16)[name = tensor("attn_31_cast_fp16")]; + tensor var_4098 = const()[name = tensor("op_4098"), val = tensor([0, 2, 1, 3])]; + tensor var_4103 = const()[name = tensor("op_4103"), val = tensor([1, 1, 1024])]; + tensor var_4099_cast_fp16 = transpose(perm = var_4098, x = attn_31_cast_fp16)[name = tensor("transpose_48")]; + tensor input_187_cast_fp16 = reshape(shape = var_4103, x = var_4099_cast_fp16)[name = tensor("input_187_cast_fp16")]; + tensor attn15_out_proj_weight_to_fp16 = const()[name = tensor("attn15_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(383997120)))]; + tensor linear_62_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn15_out_proj_weight_to_fp16, x = input_187_cast_fp16)[name = tensor("linear_62_cast_fp16")]; + tensor var_4109_to_fp16 = const()[name = tensor("op_4109_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4110_cast_fp16 = add(x = position15_to_fp16, y = var_4109_to_fp16)[name = tensor("op_4110_cast_fp16")]; + tensor var_4110_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4110_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_189_cast_fp16 = add(x = input_183_cast_fp16, y = linear_62_cast_fp16)[name = tensor("input_189_cast_fp16")]; + tensor input_191_axes_0 = const()[name = tensor("input_191_axes_0"), val = tensor([-1])]; + tensor norm15_2_weight_to_fp16 = const()[name = tensor("norm15_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(386094336)))]; + tensor norm15_2_bias_to_fp16 = const()[name = tensor("norm15_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(386096448)))]; + tensor var_4114_to_fp16 = const()[name = tensor("op_4114_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_191_cast_fp16 = layer_norm(axes = input_191_axes_0, beta = norm15_2_bias_to_fp16, epsilon = var_4114_to_fp16, gamma = norm15_2_weight_to_fp16, x = input_189_cast_fp16)[name = tensor("input_191_cast_fp16")]; + tensor linear15_1_weight_to_fp16 = const()[name = tensor("linear15_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(386098560)))]; + tensor linear_63_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear15_1_weight_to_fp16, x = input_191_cast_fp16)[name = tensor("linear_63_cast_fp16")]; + tensor input_193_mode_0 = const()[name = tensor("input_193_mode_0"), val = tensor("EXACT")]; + tensor input_193_cast_fp16 = gelu(mode = input_193_mode_0, x = linear_63_cast_fp16)[name = tensor("input_193_cast_fp16")]; + tensor linear15_2_weight_to_fp16 = const()[name = tensor("linear15_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(394487232)))]; + tensor linear_64_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear15_2_weight_to_fp16, x = input_193_cast_fp16)[name = tensor("linear_64_cast_fp16")]; + tensor input_195_cast_fp16 = add(x = input_189_cast_fp16, y = linear_64_cast_fp16)[name = tensor("input_195_cast_fp16")]; + tensor input_197_axes_0 = const()[name = tensor("input_197_axes_0"), val = tensor([-1])]; + tensor norm16_1_weight_to_fp16 = const()[name = tensor("norm16_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(402875904)))]; + tensor norm16_1_bias_to_fp16 = const()[name = tensor("norm16_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(402878016)))]; + tensor var_4131_to_fp16 = const()[name = tensor("op_4131_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_197_cast_fp16 = layer_norm(axes = input_197_axes_0, beta = norm16_1_bias_to_fp16, epsilon = var_4131_to_fp16, gamma = norm16_1_weight_to_fp16, x = input_195_cast_fp16)[name = tensor("input_197_cast_fp16")]; + tensor attn16_in_proj_weight_to_fp16 = const()[name = tensor("attn16_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(402880128)))]; + tensor linear_65_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn16_in_proj_weight_to_fp16, x = input_197_cast_fp16)[name = tensor("linear_65_cast_fp16")]; + tensor var_4144 = const()[name = tensor("op_4144"), val = tensor([1, 3, 16, 64])]; + tensor qkv_33_cast_fp16 = reshape(shape = var_4144, x = linear_65_cast_fp16)[name = tensor("qkv_33_cast_fp16")]; + tensor x_65_begin_0 = const()[name = tensor("x_65_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_65_end_0 = const()[name = tensor("x_65_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_65_end_mask_0 = const()[name = tensor("x_65_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_65_squeeze_mask_0 = const()[name = tensor("x_65_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_65_cast_fp16 = slice_by_index(begin = x_65_begin_0, end = x_65_end_0, end_mask = x_65_end_mask_0, squeeze_mask = x_65_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("x_65_cast_fp16")]; + tensor x_67_begin_0 = const()[name = tensor("x_67_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_67_end_0 = const()[name = tensor("x_67_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_67_end_mask_0 = const()[name = tensor("x_67_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_67_squeeze_mask_0 = const()[name = tensor("x_67_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_67_cast_fp16 = slice_by_index(begin = x_67_begin_0, end = x_67_end_0, end_mask = x_67_end_mask_0, squeeze_mask = x_67_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("x_67_cast_fp16")]; + tensor v_33_begin_0 = const()[name = tensor("v_33_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_33_end_0 = const()[name = tensor("v_33_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_33_end_mask_0 = const()[name = tensor("v_33_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_33_squeeze_mask_0 = const()[name = tensor("v_33_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_33_cast_fp16 = slice_by_index(begin = v_33_begin_0, end = v_33_end_0, end_mask = v_33_end_mask_0, squeeze_mask = v_33_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("v_33_cast_fp16")]; + tensor var_4173 = const()[name = tensor("op_4173"), val = tensor([1, 1, 1])]; + tensor position16_to_fp16_dtype_0 = const()[name = tensor("position16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position16_to_fp16 = cast(dtype = position16_to_fp16_dtype_0, x = position16)[name = tensor("cast_449")]; + tensor var_4174_cast_fp16 = reshape(shape = var_4173, x = position16_to_fp16)[name = tensor("op_4174_cast_fp16")]; + tensor angles_65_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_4174_cast_fp16)[name = tensor("angles_65_cast_fp16")]; + tensor rotr_65_cast_fp16 = cos(x = angles_65_cast_fp16)[name = tensor("rotr_65_cast_fp16")]; + tensor roti_65_cast_fp16 = sin(x = angles_65_cast_fp16)[name = tensor("roti_65_cast_fp16")]; + tensor var_4182 = const()[name = tensor("op_4182"), val = tensor([1, 16, 32, 2])]; + tensor pairs_65_cast_fp16 = reshape(shape = var_4182, x = x_65_cast_fp16)[name = tensor("pairs_65_cast_fp16")]; + tensor xr_65_begin_0 = const()[name = tensor("xr_65_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_65_end_0 = const()[name = tensor("xr_65_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_65_end_mask_0 = const()[name = tensor("xr_65_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_65_squeeze_mask_0 = const()[name = tensor("xr_65_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_65_cast_fp16 = slice_by_index(begin = xr_65_begin_0, end = xr_65_end_0, end_mask = xr_65_end_mask_0, squeeze_mask = xr_65_squeeze_mask_0, x = pairs_65_cast_fp16)[name = tensor("xr_65_cast_fp16")]; + tensor xi_65_begin_0 = const()[name = tensor("xi_65_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_65_end_0 = const()[name = tensor("xi_65_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_65_end_mask_0 = const()[name = tensor("xi_65_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_65_squeeze_mask_0 = const()[name = tensor("xi_65_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_65_cast_fp16 = slice_by_index(begin = xi_65_begin_0, end = xi_65_end_0, end_mask = xi_65_end_mask_0, squeeze_mask = xi_65_squeeze_mask_0, x = pairs_65_cast_fp16)[name = tensor("xi_65_cast_fp16")]; + tensor var_4190_cast_fp16 = mul(x = xr_65_cast_fp16, y = rotr_65_cast_fp16)[name = tensor("op_4190_cast_fp16")]; + tensor var_4191_cast_fp16 = mul(x = xi_65_cast_fp16, y = roti_65_cast_fp16)[name = tensor("op_4191_cast_fp16")]; + tensor xor__65_cast_fp16 = sub(x = var_4190_cast_fp16, y = var_4191_cast_fp16)[name = tensor("xor__65_cast_fp16")]; + tensor var_4194_cast_fp16 = mul(x = xr_65_cast_fp16, y = roti_65_cast_fp16)[name = tensor("op_4194_cast_fp16")]; + tensor var_4195_cast_fp16 = mul(x = xi_65_cast_fp16, y = rotr_65_cast_fp16)[name = tensor("op_4195_cast_fp16")]; + tensor xoi_65_cast_fp16 = add(x = var_4194_cast_fp16, y = var_4195_cast_fp16)[name = tensor("xoi_65_cast_fp16")]; + tensor out_65_axis_0 = const()[name = tensor("out_65_axis_0"), val = tensor(-1)]; + tensor out_65_cast_fp16 = stack(axis = out_65_axis_0, values = (xor__65_cast_fp16, xoi_65_cast_fp16))[name = tensor("out_65_cast_fp16")]; + tensor var_4218 = const()[name = tensor("op_4218"), val = tensor([1, 16, 32, 2])]; + tensor pairs_67_cast_fp16 = reshape(shape = var_4218, x = x_67_cast_fp16)[name = tensor("pairs_67_cast_fp16")]; + tensor xr_67_begin_0 = const()[name = tensor("xr_67_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_67_end_0 = const()[name = tensor("xr_67_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_67_end_mask_0 = const()[name = tensor("xr_67_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_67_squeeze_mask_0 = const()[name = tensor("xr_67_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_67_cast_fp16 = slice_by_index(begin = xr_67_begin_0, end = xr_67_end_0, end_mask = xr_67_end_mask_0, squeeze_mask = xr_67_squeeze_mask_0, x = pairs_67_cast_fp16)[name = tensor("xr_67_cast_fp16")]; + tensor xi_67_begin_0 = const()[name = tensor("xi_67_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_67_end_0 = const()[name = tensor("xi_67_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_67_end_mask_0 = const()[name = tensor("xi_67_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_67_squeeze_mask_0 = const()[name = tensor("xi_67_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_67_cast_fp16 = slice_by_index(begin = xi_67_begin_0, end = xi_67_end_0, end_mask = xi_67_end_mask_0, squeeze_mask = xi_67_squeeze_mask_0, x = pairs_67_cast_fp16)[name = tensor("xi_67_cast_fp16")]; + tensor var_4226_cast_fp16 = mul(x = xr_67_cast_fp16, y = rotr_65_cast_fp16)[name = tensor("op_4226_cast_fp16")]; + tensor var_4227_cast_fp16 = mul(x = xi_67_cast_fp16, y = roti_65_cast_fp16)[name = tensor("op_4227_cast_fp16")]; + tensor xor__67_cast_fp16 = sub(x = var_4226_cast_fp16, y = var_4227_cast_fp16)[name = tensor("xor__67_cast_fp16")]; + tensor var_4230_cast_fp16 = mul(x = xr_67_cast_fp16, y = roti_65_cast_fp16)[name = tensor("op_4230_cast_fp16")]; + tensor var_4231_cast_fp16 = mul(x = xi_67_cast_fp16, y = rotr_65_cast_fp16)[name = tensor("op_4231_cast_fp16")]; + tensor xoi_67_cast_fp16 = add(x = var_4230_cast_fp16, y = var_4231_cast_fp16)[name = tensor("xoi_67_cast_fp16")]; + tensor out_67_axis_0 = const()[name = tensor("out_67_axis_0"), val = tensor(-1)]; + tensor out_67_cast_fp16 = stack(axis = out_67_axis_0, values = (xor__67_cast_fp16, xoi_67_cast_fp16))[name = tensor("out_67_cast_fp16")]; + tensor _inversed_4243_y_0_to_fp16 = const()[name = tensor("_inversed_4243_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_4243_cast_fp16 = mul(x = position16_to_fp16, y = _inversed_4243_y_0_to_fp16)[name = tensor("_inversed_4243_cast_fp16")]; + tensor var_4244_cast_fp16 = floor(x = _inversed_4243_cast_fp16)[name = tensor("op_4244_cast_fp16")]; + tensor var_4245_to_fp16 = const()[name = tensor("op_4245_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4246_cast_fp16 = mul(x = var_4244_cast_fp16, y = var_4245_to_fp16)[name = tensor("op_4246_cast_fp16")]; + tensor write_idx_33_cast_fp16 = sub(x = position16_to_fp16, y = var_4246_cast_fp16)[name = tensor("write_idx_33_cast_fp16")]; + tensor var_4249_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_33_cast_fp16)[name = tensor("op_4249_cast_fp16")]; + tensor var_4259 = const()[name = tensor("op_4259"), val = tensor([1, 512, 1, 1])]; + tensor onehot_65_to_fp16_dtype_0 = const()[name = tensor("onehot_65_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_4249_cast_fp16_to_fp16 = cast(dtype = onehot_65_to_fp16_dtype_0, x = var_4249_cast_fp16)[name = tensor("cast_448")]; + tensor onehot_67_cast_fp16 = reshape(shape = var_4259, x = var_4249_cast_fp16_to_fp16)[name = tensor("onehot_67_cast_fp16")]; + tensor var_4261_to_fp16 = const()[name = tensor("op_4261_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4263_cast_fp16 = sub(x = var_4261_to_fp16, y = onehot_67_cast_fp16)[name = tensor("op_4263_cast_fp16")]; + tensor k_cache16_to_fp16_dtype_0 = const()[name = tensor("k_cache16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache16_to_fp16 = cast(dtype = k_cache16_to_fp16_dtype_0, x = k_cache16)[name = tensor("cast_447")]; + tensor var_4264_cast_fp16 = mul(x = k_cache16_to_fp16, y = var_4263_cast_fp16)[name = tensor("op_4264_cast_fp16")]; + tensor var_4269 = const()[name = tensor("op_4269"), val = tensor([1, 1, 16, 64])]; + tensor var_4270_cast_fp16 = reshape(shape = var_4269, x = out_67_cast_fp16)[name = tensor("op_4270_cast_fp16")]; + tensor var_4271_cast_fp16 = mul(x = var_4270_cast_fp16, y = onehot_67_cast_fp16)[name = tensor("op_4271_cast_fp16")]; + tensor new_k_cache_33_cast_fp16 = add(x = var_4264_cast_fp16, y = var_4271_cast_fp16)[name = tensor("new_k_cache_33_cast_fp16")]; + tensor new_k_cache_33_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_33_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache16_to_fp16_dtype_0 = const()[name = tensor("v_cache16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache16_to_fp16 = cast(dtype = v_cache16_to_fp16_dtype_0, x = v_cache16)[name = tensor("cast_445")]; + tensor var_4277_cast_fp16 = mul(x = v_cache16_to_fp16, y = var_4263_cast_fp16)[name = tensor("op_4277_cast_fp16")]; + tensor var_4282 = const()[name = tensor("op_4282"), val = tensor([1, 1, 16, 64])]; + tensor var_4283_cast_fp16 = reshape(shape = var_4282, x = v_33_cast_fp16)[name = tensor("op_4283_cast_fp16")]; + tensor var_4284_cast_fp16 = mul(x = var_4283_cast_fp16, y = onehot_67_cast_fp16)[name = tensor("op_4284_cast_fp16")]; + tensor new_v_cache_33_cast_fp16 = add(x = var_4277_cast_fp16, y = var_4284_cast_fp16)[name = tensor("new_v_cache_33_cast_fp16")]; + tensor new_v_cache_33_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_33_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_16_perm_0 = const()[name = tensor("transpose_16_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_4297 = const()[name = tensor("op_4297"), val = tensor([0, 2, 1, 3])]; + tensor var_4303 = const()[name = tensor("op_4303"), val = tensor([1, 16, 1, 64])]; + tensor q4_33_cast_fp16 = reshape(shape = var_4303, x = out_65_cast_fp16)[name = tensor("q4_33_cast_fp16")]; + tensor var_4308_transpose_x_0 = const()[name = tensor("op_4308_transpose_x_0"), val = tensor(false)]; + tensor var_4308_transpose_y_0 = const()[name = tensor("op_4308_transpose_y_0"), val = tensor(false)]; + tensor transpose_16_cast_fp16 = transpose(perm = transpose_16_perm_0, x = new_k_cache_33_cast_fp16)[name = tensor("transpose_47")]; + tensor var_4308_cast_fp16 = matmul(transpose_x = var_4308_transpose_x_0, transpose_y = var_4308_transpose_y_0, x = q4_33_cast_fp16, y = transpose_16_cast_fp16)[name = tensor("op_4308_cast_fp16")]; + tensor var_4309_to_fp16 = const()[name = tensor("op_4309_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_65_cast_fp16 = mul(x = var_4308_cast_fp16, y = var_4309_to_fp16)[name = tensor("scores_65_cast_fp16")]; + tensor var_4313 = const()[name = tensor("op_4313"), val = tensor([1, 1])]; + tensor var_4314_cast_fp16 = reshape(shape = var_4313, x = position16_to_fp16)[name = tensor("op_4314_cast_fp16")]; + tensor var_4315_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_4314_cast_fp16)[name = tensor("op_4315_cast_fp16")]; + tensor mask_33_to_fp16_dtype_0 = const()[name = tensor("mask_33_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_4322_to_fp16 = const()[name = tensor("op_4322_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4315_cast_fp16_to_fp16 = cast(dtype = mask_33_to_fp16_dtype_0, x = var_4315_cast_fp16)[name = tensor("cast_443")]; + tensor var_4323_cast_fp16 = sub(x = var_4315_cast_fp16_to_fp16, y = var_4322_to_fp16)[name = tensor("op_4323_cast_fp16")]; + tensor var_4328 = const()[name = tensor("op_4328"), val = tensor([1, 1, 1, 512])]; + tensor var_4329_cast_fp16 = reshape(shape = var_4328, x = var_4323_cast_fp16)[name = tensor("op_4329_cast_fp16")]; + tensor var_4330_to_fp16 = const()[name = tensor("op_4330_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_4331_cast_fp16 = mul(x = var_4329_cast_fp16, y = var_4330_to_fp16)[name = tensor("op_4331_cast_fp16")]; + tensor scores_67_cast_fp16 = add(x = scores_65_cast_fp16, y = var_4331_cast_fp16)[name = tensor("scores_67_cast_fp16")]; + tensor var_4334 = const()[name = tensor("op_4334"), val = tensor(-1)]; + tensor weights_33_cast_fp16 = softmax(axis = var_4334, x = scores_67_cast_fp16)[name = tensor("weights_33_cast_fp16")]; + tensor attn_33_transpose_x_0 = const()[name = tensor("attn_33_transpose_x_0"), val = tensor(false)]; + tensor attn_33_transpose_y_0 = const()[name = tensor("attn_33_transpose_y_0"), val = tensor(false)]; + tensor values_33_cast_fp16 = transpose(perm = var_4297, x = new_v_cache_33_cast_fp16)[name = tensor("transpose_46")]; + tensor attn_33_cast_fp16 = matmul(transpose_x = attn_33_transpose_x_0, transpose_y = attn_33_transpose_y_0, x = weights_33_cast_fp16, y = values_33_cast_fp16)[name = tensor("attn_33_cast_fp16")]; + tensor var_4342 = const()[name = tensor("op_4342"), val = tensor([0, 2, 1, 3])]; + tensor var_4347 = const()[name = tensor("op_4347"), val = tensor([1, 1, 1024])]; + tensor var_4343_cast_fp16 = transpose(perm = var_4342, x = attn_33_cast_fp16)[name = tensor("transpose_45")]; + tensor input_199_cast_fp16 = reshape(shape = var_4347, x = var_4343_cast_fp16)[name = tensor("input_199_cast_fp16")]; + tensor attn16_out_proj_weight_to_fp16 = const()[name = tensor("attn16_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(409171648)))]; + tensor linear_66_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn16_out_proj_weight_to_fp16, x = input_199_cast_fp16)[name = tensor("linear_66_cast_fp16")]; + tensor var_4353_to_fp16 = const()[name = tensor("op_4353_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4354_cast_fp16 = add(x = position16_to_fp16, y = var_4353_to_fp16)[name = tensor("op_4354_cast_fp16")]; + tensor var_4354_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4354_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_201_cast_fp16 = add(x = input_195_cast_fp16, y = linear_66_cast_fp16)[name = tensor("input_201_cast_fp16")]; + tensor input_203_axes_0 = const()[name = tensor("input_203_axes_0"), val = tensor([-1])]; + tensor norm16_2_weight_to_fp16 = const()[name = tensor("norm16_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411268864)))]; + tensor norm16_2_bias_to_fp16 = const()[name = tensor("norm16_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411270976)))]; + tensor var_4358_to_fp16 = const()[name = tensor("op_4358_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_203_cast_fp16 = layer_norm(axes = input_203_axes_0, beta = norm16_2_bias_to_fp16, epsilon = var_4358_to_fp16, gamma = norm16_2_weight_to_fp16, x = input_201_cast_fp16)[name = tensor("input_203_cast_fp16")]; + tensor linear16_1_weight_to_fp16 = const()[name = tensor("linear16_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411273088)))]; + tensor linear_67_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear16_1_weight_to_fp16, x = input_203_cast_fp16)[name = tensor("linear_67_cast_fp16")]; + tensor input_205_mode_0 = const()[name = tensor("input_205_mode_0"), val = tensor("EXACT")]; + tensor input_205_cast_fp16 = gelu(mode = input_205_mode_0, x = linear_67_cast_fp16)[name = tensor("input_205_cast_fp16")]; + tensor linear16_2_weight_to_fp16 = const()[name = tensor("linear16_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(419661760)))]; + tensor linear_68_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear16_2_weight_to_fp16, x = input_205_cast_fp16)[name = tensor("linear_68_cast_fp16")]; + tensor input_207_cast_fp16 = add(x = input_201_cast_fp16, y = linear_68_cast_fp16)[name = tensor("input_207_cast_fp16")]; + tensor input_209_axes_0 = const()[name = tensor("input_209_axes_0"), val = tensor([-1])]; + tensor norm17_1_weight_to_fp16 = const()[name = tensor("norm17_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(428050432)))]; + tensor norm17_1_bias_to_fp16 = const()[name = tensor("norm17_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(428052544)))]; + tensor var_4375_to_fp16 = const()[name = tensor("op_4375_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_209_cast_fp16 = layer_norm(axes = input_209_axes_0, beta = norm17_1_bias_to_fp16, epsilon = var_4375_to_fp16, gamma = norm17_1_weight_to_fp16, x = input_207_cast_fp16)[name = tensor("input_209_cast_fp16")]; + tensor attn17_in_proj_weight_to_fp16 = const()[name = tensor("attn17_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(428054656)))]; + tensor linear_69_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn17_in_proj_weight_to_fp16, x = input_209_cast_fp16)[name = tensor("linear_69_cast_fp16")]; + tensor var_4388 = const()[name = tensor("op_4388"), val = tensor([1, 3, 16, 64])]; + tensor qkv_35_cast_fp16 = reshape(shape = var_4388, x = linear_69_cast_fp16)[name = tensor("qkv_35_cast_fp16")]; + tensor x_69_begin_0 = const()[name = tensor("x_69_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_69_end_0 = const()[name = tensor("x_69_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_69_end_mask_0 = const()[name = tensor("x_69_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_69_squeeze_mask_0 = const()[name = tensor("x_69_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_69_cast_fp16 = slice_by_index(begin = x_69_begin_0, end = x_69_end_0, end_mask = x_69_end_mask_0, squeeze_mask = x_69_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("x_69_cast_fp16")]; + tensor x_71_begin_0 = const()[name = tensor("x_71_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_71_end_0 = const()[name = tensor("x_71_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_71_end_mask_0 = const()[name = tensor("x_71_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_71_squeeze_mask_0 = const()[name = tensor("x_71_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_71_cast_fp16 = slice_by_index(begin = x_71_begin_0, end = x_71_end_0, end_mask = x_71_end_mask_0, squeeze_mask = x_71_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("x_71_cast_fp16")]; + tensor v_35_begin_0 = const()[name = tensor("v_35_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_35_end_0 = const()[name = tensor("v_35_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_35_end_mask_0 = const()[name = tensor("v_35_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_35_squeeze_mask_0 = const()[name = tensor("v_35_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_35_cast_fp16 = slice_by_index(begin = v_35_begin_0, end = v_35_end_0, end_mask = v_35_end_mask_0, squeeze_mask = v_35_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("v_35_cast_fp16")]; + tensor var_4417 = const()[name = tensor("op_4417"), val = tensor([1, 1, 1])]; + tensor position17_to_fp16_dtype_0 = const()[name = tensor("position17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position17_to_fp16 = cast(dtype = position17_to_fp16_dtype_0, x = position17)[name = tensor("cast_441")]; + tensor var_4418_cast_fp16 = reshape(shape = var_4417, x = position17_to_fp16)[name = tensor("op_4418_cast_fp16")]; + tensor angles_69_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_4418_cast_fp16)[name = tensor("angles_69_cast_fp16")]; + tensor rotr_69_cast_fp16 = cos(x = angles_69_cast_fp16)[name = tensor("rotr_69_cast_fp16")]; + tensor roti_69_cast_fp16 = sin(x = angles_69_cast_fp16)[name = tensor("roti_69_cast_fp16")]; + tensor var_4426 = const()[name = tensor("op_4426"), val = tensor([1, 16, 32, 2])]; + tensor pairs_69_cast_fp16 = reshape(shape = var_4426, x = x_69_cast_fp16)[name = tensor("pairs_69_cast_fp16")]; + tensor xr_69_begin_0 = const()[name = tensor("xr_69_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_69_end_0 = const()[name = tensor("xr_69_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_69_end_mask_0 = const()[name = tensor("xr_69_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_69_squeeze_mask_0 = const()[name = tensor("xr_69_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_69_cast_fp16 = slice_by_index(begin = xr_69_begin_0, end = xr_69_end_0, end_mask = xr_69_end_mask_0, squeeze_mask = xr_69_squeeze_mask_0, x = pairs_69_cast_fp16)[name = tensor("xr_69_cast_fp16")]; + tensor xi_69_begin_0 = const()[name = tensor("xi_69_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_69_end_0 = const()[name = tensor("xi_69_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_69_end_mask_0 = const()[name = tensor("xi_69_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_69_squeeze_mask_0 = const()[name = tensor("xi_69_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_69_cast_fp16 = slice_by_index(begin = xi_69_begin_0, end = xi_69_end_0, end_mask = xi_69_end_mask_0, squeeze_mask = xi_69_squeeze_mask_0, x = pairs_69_cast_fp16)[name = tensor("xi_69_cast_fp16")]; + tensor var_4434_cast_fp16 = mul(x = xr_69_cast_fp16, y = rotr_69_cast_fp16)[name = tensor("op_4434_cast_fp16")]; + tensor var_4435_cast_fp16 = mul(x = xi_69_cast_fp16, y = roti_69_cast_fp16)[name = tensor("op_4435_cast_fp16")]; + tensor xor__69_cast_fp16 = sub(x = var_4434_cast_fp16, y = var_4435_cast_fp16)[name = tensor("xor__69_cast_fp16")]; + tensor var_4438_cast_fp16 = mul(x = xr_69_cast_fp16, y = roti_69_cast_fp16)[name = tensor("op_4438_cast_fp16")]; + tensor var_4439_cast_fp16 = mul(x = xi_69_cast_fp16, y = rotr_69_cast_fp16)[name = tensor("op_4439_cast_fp16")]; + tensor xoi_69_cast_fp16 = add(x = var_4438_cast_fp16, y = var_4439_cast_fp16)[name = tensor("xoi_69_cast_fp16")]; + tensor out_69_axis_0 = const()[name = tensor("out_69_axis_0"), val = tensor(-1)]; + tensor out_69_cast_fp16 = stack(axis = out_69_axis_0, values = (xor__69_cast_fp16, xoi_69_cast_fp16))[name = tensor("out_69_cast_fp16")]; + tensor var_4462 = const()[name = tensor("op_4462"), val = tensor([1, 16, 32, 2])]; + tensor pairs_71_cast_fp16 = reshape(shape = var_4462, x = x_71_cast_fp16)[name = tensor("pairs_71_cast_fp16")]; + tensor xr_71_begin_0 = const()[name = tensor("xr_71_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_71_end_0 = const()[name = tensor("xr_71_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_71_end_mask_0 = const()[name = tensor("xr_71_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_71_squeeze_mask_0 = const()[name = tensor("xr_71_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_71_cast_fp16 = slice_by_index(begin = xr_71_begin_0, end = xr_71_end_0, end_mask = xr_71_end_mask_0, squeeze_mask = xr_71_squeeze_mask_0, x = pairs_71_cast_fp16)[name = tensor("xr_71_cast_fp16")]; + tensor xi_71_begin_0 = const()[name = tensor("xi_71_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_71_end_0 = const()[name = tensor("xi_71_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_71_end_mask_0 = const()[name = tensor("xi_71_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_71_squeeze_mask_0 = const()[name = tensor("xi_71_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_71_cast_fp16 = slice_by_index(begin = xi_71_begin_0, end = xi_71_end_0, end_mask = xi_71_end_mask_0, squeeze_mask = xi_71_squeeze_mask_0, x = pairs_71_cast_fp16)[name = tensor("xi_71_cast_fp16")]; + tensor var_4470_cast_fp16 = mul(x = xr_71_cast_fp16, y = rotr_69_cast_fp16)[name = tensor("op_4470_cast_fp16")]; + tensor var_4471_cast_fp16 = mul(x = xi_71_cast_fp16, y = roti_69_cast_fp16)[name = tensor("op_4471_cast_fp16")]; + tensor xor__71_cast_fp16 = sub(x = var_4470_cast_fp16, y = var_4471_cast_fp16)[name = tensor("xor__71_cast_fp16")]; + tensor var_4474_cast_fp16 = mul(x = xr_71_cast_fp16, y = roti_69_cast_fp16)[name = tensor("op_4474_cast_fp16")]; + tensor var_4475_cast_fp16 = mul(x = xi_71_cast_fp16, y = rotr_69_cast_fp16)[name = tensor("op_4475_cast_fp16")]; + tensor xoi_71_cast_fp16 = add(x = var_4474_cast_fp16, y = var_4475_cast_fp16)[name = tensor("xoi_71_cast_fp16")]; + tensor out_71_axis_0 = const()[name = tensor("out_71_axis_0"), val = tensor(-1)]; + tensor out_71_cast_fp16 = stack(axis = out_71_axis_0, values = (xor__71_cast_fp16, xoi_71_cast_fp16))[name = tensor("out_71_cast_fp16")]; + tensor _inversed_4487_y_0_to_fp16 = const()[name = tensor("_inversed_4487_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_4487_cast_fp16 = mul(x = position17_to_fp16, y = _inversed_4487_y_0_to_fp16)[name = tensor("_inversed_4487_cast_fp16")]; + tensor var_4488_cast_fp16 = floor(x = _inversed_4487_cast_fp16)[name = tensor("op_4488_cast_fp16")]; + tensor var_4489_to_fp16 = const()[name = tensor("op_4489_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4490_cast_fp16 = mul(x = var_4488_cast_fp16, y = var_4489_to_fp16)[name = tensor("op_4490_cast_fp16")]; + tensor write_idx_35_cast_fp16 = sub(x = position17_to_fp16, y = var_4490_cast_fp16)[name = tensor("write_idx_35_cast_fp16")]; + tensor var_4493_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_35_cast_fp16)[name = tensor("op_4493_cast_fp16")]; + tensor var_4503 = const()[name = tensor("op_4503"), val = tensor([1, 512, 1, 1])]; + tensor onehot_69_to_fp16_dtype_0 = const()[name = tensor("onehot_69_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_4493_cast_fp16_to_fp16 = cast(dtype = onehot_69_to_fp16_dtype_0, x = var_4493_cast_fp16)[name = tensor("cast_440")]; + tensor onehot_71_cast_fp16 = reshape(shape = var_4503, x = var_4493_cast_fp16_to_fp16)[name = tensor("onehot_71_cast_fp16")]; + tensor var_4505_to_fp16 = const()[name = tensor("op_4505_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4507_cast_fp16 = sub(x = var_4505_to_fp16, y = onehot_71_cast_fp16)[name = tensor("op_4507_cast_fp16")]; + tensor k_cache17_to_fp16_dtype_0 = const()[name = tensor("k_cache17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache17_to_fp16 = cast(dtype = k_cache17_to_fp16_dtype_0, x = k_cache17)[name = tensor("cast_439")]; + tensor var_4508_cast_fp16 = mul(x = k_cache17_to_fp16, y = var_4507_cast_fp16)[name = tensor("op_4508_cast_fp16")]; + tensor var_4513 = const()[name = tensor("op_4513"), val = tensor([1, 1, 16, 64])]; + tensor var_4514_cast_fp16 = reshape(shape = var_4513, x = out_71_cast_fp16)[name = tensor("op_4514_cast_fp16")]; + tensor var_4515_cast_fp16 = mul(x = var_4514_cast_fp16, y = onehot_71_cast_fp16)[name = tensor("op_4515_cast_fp16")]; + tensor new_k_cache_35_cast_fp16 = add(x = var_4508_cast_fp16, y = var_4515_cast_fp16)[name = tensor("new_k_cache_35_cast_fp16")]; + tensor new_k_cache_35_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_35_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache17_to_fp16_dtype_0 = const()[name = tensor("v_cache17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache17_to_fp16 = cast(dtype = v_cache17_to_fp16_dtype_0, x = v_cache17)[name = tensor("cast_437")]; + tensor var_4521_cast_fp16 = mul(x = v_cache17_to_fp16, y = var_4507_cast_fp16)[name = tensor("op_4521_cast_fp16")]; + tensor var_4526 = const()[name = tensor("op_4526"), val = tensor([1, 1, 16, 64])]; + tensor var_4527_cast_fp16 = reshape(shape = var_4526, x = v_35_cast_fp16)[name = tensor("op_4527_cast_fp16")]; + tensor var_4528_cast_fp16 = mul(x = var_4527_cast_fp16, y = onehot_71_cast_fp16)[name = tensor("op_4528_cast_fp16")]; + tensor new_v_cache_35_cast_fp16 = add(x = var_4521_cast_fp16, y = var_4528_cast_fp16)[name = tensor("new_v_cache_35_cast_fp16")]; + tensor new_v_cache_35_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_35_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_17_perm_0 = const()[name = tensor("transpose_17_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_4541 = const()[name = tensor("op_4541"), val = tensor([0, 2, 1, 3])]; + tensor var_4547 = const()[name = tensor("op_4547"), val = tensor([1, 16, 1, 64])]; + tensor q4_35_cast_fp16 = reshape(shape = var_4547, x = out_69_cast_fp16)[name = tensor("q4_35_cast_fp16")]; + tensor var_4552_transpose_x_0 = const()[name = tensor("op_4552_transpose_x_0"), val = tensor(false)]; + tensor var_4552_transpose_y_0 = const()[name = tensor("op_4552_transpose_y_0"), val = tensor(false)]; + tensor transpose_17_cast_fp16 = transpose(perm = transpose_17_perm_0, x = new_k_cache_35_cast_fp16)[name = tensor("transpose_44")]; + tensor var_4552_cast_fp16 = matmul(transpose_x = var_4552_transpose_x_0, transpose_y = var_4552_transpose_y_0, x = q4_35_cast_fp16, y = transpose_17_cast_fp16)[name = tensor("op_4552_cast_fp16")]; + tensor var_4553_to_fp16 = const()[name = tensor("op_4553_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_69_cast_fp16 = mul(x = var_4552_cast_fp16, y = var_4553_to_fp16)[name = tensor("scores_69_cast_fp16")]; + tensor var_4557 = const()[name = tensor("op_4557"), val = tensor([1, 1])]; + tensor var_4558_cast_fp16 = reshape(shape = var_4557, x = position17_to_fp16)[name = tensor("op_4558_cast_fp16")]; + tensor var_4559_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_4558_cast_fp16)[name = tensor("op_4559_cast_fp16")]; + tensor mask_35_to_fp16_dtype_0 = const()[name = tensor("mask_35_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_4566_to_fp16 = const()[name = tensor("op_4566_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4559_cast_fp16_to_fp16 = cast(dtype = mask_35_to_fp16_dtype_0, x = var_4559_cast_fp16)[name = tensor("cast_435")]; + tensor var_4567_cast_fp16 = sub(x = var_4559_cast_fp16_to_fp16, y = var_4566_to_fp16)[name = tensor("op_4567_cast_fp16")]; + tensor var_4572 = const()[name = tensor("op_4572"), val = tensor([1, 1, 1, 512])]; + tensor var_4573_cast_fp16 = reshape(shape = var_4572, x = var_4567_cast_fp16)[name = tensor("op_4573_cast_fp16")]; + tensor var_4574_to_fp16 = const()[name = tensor("op_4574_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_4575_cast_fp16 = mul(x = var_4573_cast_fp16, y = var_4574_to_fp16)[name = tensor("op_4575_cast_fp16")]; + tensor scores_71_cast_fp16 = add(x = scores_69_cast_fp16, y = var_4575_cast_fp16)[name = tensor("scores_71_cast_fp16")]; + tensor var_4578 = const()[name = tensor("op_4578"), val = tensor(-1)]; + tensor weights_35_cast_fp16 = softmax(axis = var_4578, x = scores_71_cast_fp16)[name = tensor("weights_35_cast_fp16")]; + tensor attn_35_transpose_x_0 = const()[name = tensor("attn_35_transpose_x_0"), val = tensor(false)]; + tensor attn_35_transpose_y_0 = const()[name = tensor("attn_35_transpose_y_0"), val = tensor(false)]; + tensor values_35_cast_fp16 = transpose(perm = var_4541, x = new_v_cache_35_cast_fp16)[name = tensor("transpose_43")]; + tensor attn_35_cast_fp16 = matmul(transpose_x = attn_35_transpose_x_0, transpose_y = attn_35_transpose_y_0, x = weights_35_cast_fp16, y = values_35_cast_fp16)[name = tensor("attn_35_cast_fp16")]; + tensor var_4586 = const()[name = tensor("op_4586"), val = tensor([0, 2, 1, 3])]; + tensor var_4591 = const()[name = tensor("op_4591"), val = tensor([1, 1, 1024])]; + tensor var_4587_cast_fp16 = transpose(perm = var_4586, x = attn_35_cast_fp16)[name = tensor("transpose_42")]; + tensor input_211_cast_fp16 = reshape(shape = var_4591, x = var_4587_cast_fp16)[name = tensor("input_211_cast_fp16")]; + tensor attn17_out_proj_weight_to_fp16 = const()[name = tensor("attn17_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(434346176)))]; + tensor linear_70_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn17_out_proj_weight_to_fp16, x = input_211_cast_fp16)[name = tensor("linear_70_cast_fp16")]; + tensor var_4597_to_fp16 = const()[name = tensor("op_4597_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4598_cast_fp16 = add(x = position17_to_fp16, y = var_4597_to_fp16)[name = tensor("op_4598_cast_fp16")]; + tensor var_4598_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4598_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_213_cast_fp16 = add(x = input_207_cast_fp16, y = linear_70_cast_fp16)[name = tensor("input_213_cast_fp16")]; + tensor input_215_axes_0 = const()[name = tensor("input_215_axes_0"), val = tensor([-1])]; + tensor norm17_2_weight_to_fp16 = const()[name = tensor("norm17_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(436443392)))]; + tensor norm17_2_bias_to_fp16 = const()[name = tensor("norm17_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(436445504)))]; + tensor var_4602_to_fp16 = const()[name = tensor("op_4602_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_215_cast_fp16 = layer_norm(axes = input_215_axes_0, beta = norm17_2_bias_to_fp16, epsilon = var_4602_to_fp16, gamma = norm17_2_weight_to_fp16, x = input_213_cast_fp16)[name = tensor("input_215_cast_fp16")]; + tensor linear17_1_weight_to_fp16 = const()[name = tensor("linear17_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(436447616)))]; + tensor linear_71_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear17_1_weight_to_fp16, x = input_215_cast_fp16)[name = tensor("linear_71_cast_fp16")]; + tensor input_217_mode_0 = const()[name = tensor("input_217_mode_0"), val = tensor("EXACT")]; + tensor input_217_cast_fp16 = gelu(mode = input_217_mode_0, x = linear_71_cast_fp16)[name = tensor("input_217_cast_fp16")]; + tensor linear17_2_weight_to_fp16 = const()[name = tensor("linear17_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(444836288)))]; + tensor linear_72_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear17_2_weight_to_fp16, x = input_217_cast_fp16)[name = tensor("linear_72_cast_fp16")]; + tensor input_219_cast_fp16 = add(x = input_213_cast_fp16, y = linear_72_cast_fp16)[name = tensor("input_219_cast_fp16")]; + tensor input_221_axes_0 = const()[name = tensor("input_221_axes_0"), val = tensor([-1])]; + tensor norm18_1_weight_to_fp16 = const()[name = tensor("norm18_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(453224960)))]; + tensor norm18_1_bias_to_fp16 = const()[name = tensor("norm18_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(453227072)))]; + tensor var_4619_to_fp16 = const()[name = tensor("op_4619_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_221_cast_fp16 = layer_norm(axes = input_221_axes_0, beta = norm18_1_bias_to_fp16, epsilon = var_4619_to_fp16, gamma = norm18_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("input_221_cast_fp16")]; + tensor attn18_in_proj_weight_to_fp16 = const()[name = tensor("attn18_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(453229184)))]; + tensor linear_73_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn18_in_proj_weight_to_fp16, x = input_221_cast_fp16)[name = tensor("linear_73_cast_fp16")]; + tensor var_4632 = const()[name = tensor("op_4632"), val = tensor([1, 3, 16, 64])]; + tensor qkv_37_cast_fp16 = reshape(shape = var_4632, x = linear_73_cast_fp16)[name = tensor("qkv_37_cast_fp16")]; + tensor x_73_begin_0 = const()[name = tensor("x_73_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_73_end_0 = const()[name = tensor("x_73_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_73_end_mask_0 = const()[name = tensor("x_73_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_73_squeeze_mask_0 = const()[name = tensor("x_73_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_73_cast_fp16 = slice_by_index(begin = x_73_begin_0, end = x_73_end_0, end_mask = x_73_end_mask_0, squeeze_mask = x_73_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("x_73_cast_fp16")]; + tensor x_75_begin_0 = const()[name = tensor("x_75_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_75_end_0 = const()[name = tensor("x_75_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_75_end_mask_0 = const()[name = tensor("x_75_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_75_squeeze_mask_0 = const()[name = tensor("x_75_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_75_cast_fp16 = slice_by_index(begin = x_75_begin_0, end = x_75_end_0, end_mask = x_75_end_mask_0, squeeze_mask = x_75_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("x_75_cast_fp16")]; + tensor v_37_begin_0 = const()[name = tensor("v_37_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_37_end_0 = const()[name = tensor("v_37_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_37_end_mask_0 = const()[name = tensor("v_37_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_37_squeeze_mask_0 = const()[name = tensor("v_37_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_37_cast_fp16 = slice_by_index(begin = v_37_begin_0, end = v_37_end_0, end_mask = v_37_end_mask_0, squeeze_mask = v_37_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("v_37_cast_fp16")]; + tensor var_4661 = const()[name = tensor("op_4661"), val = tensor([1, 1, 1])]; + tensor position18_to_fp16_dtype_0 = const()[name = tensor("position18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position18_to_fp16 = cast(dtype = position18_to_fp16_dtype_0, x = position18)[name = tensor("cast_433")]; + tensor var_4662_cast_fp16 = reshape(shape = var_4661, x = position18_to_fp16)[name = tensor("op_4662_cast_fp16")]; + tensor angles_73_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_4662_cast_fp16)[name = tensor("angles_73_cast_fp16")]; + tensor rotr_73_cast_fp16 = cos(x = angles_73_cast_fp16)[name = tensor("rotr_73_cast_fp16")]; + tensor roti_73_cast_fp16 = sin(x = angles_73_cast_fp16)[name = tensor("roti_73_cast_fp16")]; + tensor var_4670 = const()[name = tensor("op_4670"), val = tensor([1, 16, 32, 2])]; + tensor pairs_73_cast_fp16 = reshape(shape = var_4670, x = x_73_cast_fp16)[name = tensor("pairs_73_cast_fp16")]; + tensor xr_73_begin_0 = const()[name = tensor("xr_73_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_73_end_0 = const()[name = tensor("xr_73_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_73_end_mask_0 = const()[name = tensor("xr_73_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_73_squeeze_mask_0 = const()[name = tensor("xr_73_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_73_cast_fp16 = slice_by_index(begin = xr_73_begin_0, end = xr_73_end_0, end_mask = xr_73_end_mask_0, squeeze_mask = xr_73_squeeze_mask_0, x = pairs_73_cast_fp16)[name = tensor("xr_73_cast_fp16")]; + tensor xi_73_begin_0 = const()[name = tensor("xi_73_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_73_end_0 = const()[name = tensor("xi_73_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_73_end_mask_0 = const()[name = tensor("xi_73_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_73_squeeze_mask_0 = const()[name = tensor("xi_73_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_73_cast_fp16 = slice_by_index(begin = xi_73_begin_0, end = xi_73_end_0, end_mask = xi_73_end_mask_0, squeeze_mask = xi_73_squeeze_mask_0, x = pairs_73_cast_fp16)[name = tensor("xi_73_cast_fp16")]; + tensor var_4678_cast_fp16 = mul(x = xr_73_cast_fp16, y = rotr_73_cast_fp16)[name = tensor("op_4678_cast_fp16")]; + tensor var_4679_cast_fp16 = mul(x = xi_73_cast_fp16, y = roti_73_cast_fp16)[name = tensor("op_4679_cast_fp16")]; + tensor xor__73_cast_fp16 = sub(x = var_4678_cast_fp16, y = var_4679_cast_fp16)[name = tensor("xor__73_cast_fp16")]; + tensor var_4682_cast_fp16 = mul(x = xr_73_cast_fp16, y = roti_73_cast_fp16)[name = tensor("op_4682_cast_fp16")]; + tensor var_4683_cast_fp16 = mul(x = xi_73_cast_fp16, y = rotr_73_cast_fp16)[name = tensor("op_4683_cast_fp16")]; + tensor xoi_73_cast_fp16 = add(x = var_4682_cast_fp16, y = var_4683_cast_fp16)[name = tensor("xoi_73_cast_fp16")]; + tensor out_73_axis_0 = const()[name = tensor("out_73_axis_0"), val = tensor(-1)]; + tensor out_73_cast_fp16 = stack(axis = out_73_axis_0, values = (xor__73_cast_fp16, xoi_73_cast_fp16))[name = tensor("out_73_cast_fp16")]; + tensor var_4706 = const()[name = tensor("op_4706"), val = tensor([1, 16, 32, 2])]; + tensor pairs_75_cast_fp16 = reshape(shape = var_4706, x = x_75_cast_fp16)[name = tensor("pairs_75_cast_fp16")]; + tensor xr_75_begin_0 = const()[name = tensor("xr_75_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_75_end_0 = const()[name = tensor("xr_75_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_75_end_mask_0 = const()[name = tensor("xr_75_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_75_squeeze_mask_0 = const()[name = tensor("xr_75_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_75_cast_fp16 = slice_by_index(begin = xr_75_begin_0, end = xr_75_end_0, end_mask = xr_75_end_mask_0, squeeze_mask = xr_75_squeeze_mask_0, x = pairs_75_cast_fp16)[name = tensor("xr_75_cast_fp16")]; + tensor xi_75_begin_0 = const()[name = tensor("xi_75_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_75_end_0 = const()[name = tensor("xi_75_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_75_end_mask_0 = const()[name = tensor("xi_75_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_75_squeeze_mask_0 = const()[name = tensor("xi_75_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_75_cast_fp16 = slice_by_index(begin = xi_75_begin_0, end = xi_75_end_0, end_mask = xi_75_end_mask_0, squeeze_mask = xi_75_squeeze_mask_0, x = pairs_75_cast_fp16)[name = tensor("xi_75_cast_fp16")]; + tensor var_4714_cast_fp16 = mul(x = xr_75_cast_fp16, y = rotr_73_cast_fp16)[name = tensor("op_4714_cast_fp16")]; + tensor var_4715_cast_fp16 = mul(x = xi_75_cast_fp16, y = roti_73_cast_fp16)[name = tensor("op_4715_cast_fp16")]; + tensor xor__75_cast_fp16 = sub(x = var_4714_cast_fp16, y = var_4715_cast_fp16)[name = tensor("xor__75_cast_fp16")]; + tensor var_4718_cast_fp16 = mul(x = xr_75_cast_fp16, y = roti_73_cast_fp16)[name = tensor("op_4718_cast_fp16")]; + tensor var_4719_cast_fp16 = mul(x = xi_75_cast_fp16, y = rotr_73_cast_fp16)[name = tensor("op_4719_cast_fp16")]; + tensor xoi_75_cast_fp16 = add(x = var_4718_cast_fp16, y = var_4719_cast_fp16)[name = tensor("xoi_75_cast_fp16")]; + tensor out_75_axis_0 = const()[name = tensor("out_75_axis_0"), val = tensor(-1)]; + tensor out_75_cast_fp16 = stack(axis = out_75_axis_0, values = (xor__75_cast_fp16, xoi_75_cast_fp16))[name = tensor("out_75_cast_fp16")]; + tensor _inversed_4731_y_0_to_fp16 = const()[name = tensor("_inversed_4731_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_4731_cast_fp16 = mul(x = position18_to_fp16, y = _inversed_4731_y_0_to_fp16)[name = tensor("_inversed_4731_cast_fp16")]; + tensor var_4732_cast_fp16 = floor(x = _inversed_4731_cast_fp16)[name = tensor("op_4732_cast_fp16")]; + tensor var_4733_to_fp16 = const()[name = tensor("op_4733_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4734_cast_fp16 = mul(x = var_4732_cast_fp16, y = var_4733_to_fp16)[name = tensor("op_4734_cast_fp16")]; + tensor write_idx_37_cast_fp16 = sub(x = position18_to_fp16, y = var_4734_cast_fp16)[name = tensor("write_idx_37_cast_fp16")]; + tensor var_4737_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_37_cast_fp16)[name = tensor("op_4737_cast_fp16")]; + tensor var_4747 = const()[name = tensor("op_4747"), val = tensor([1, 512, 1, 1])]; + tensor onehot_73_to_fp16_dtype_0 = const()[name = tensor("onehot_73_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_4737_cast_fp16_to_fp16 = cast(dtype = onehot_73_to_fp16_dtype_0, x = var_4737_cast_fp16)[name = tensor("cast_432")]; + tensor onehot_75_cast_fp16 = reshape(shape = var_4747, x = var_4737_cast_fp16_to_fp16)[name = tensor("onehot_75_cast_fp16")]; + tensor var_4749_to_fp16 = const()[name = tensor("op_4749_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4751_cast_fp16 = sub(x = var_4749_to_fp16, y = onehot_75_cast_fp16)[name = tensor("op_4751_cast_fp16")]; + tensor k_cache18_to_fp16_dtype_0 = const()[name = tensor("k_cache18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache18_to_fp16 = cast(dtype = k_cache18_to_fp16_dtype_0, x = k_cache18)[name = tensor("cast_431")]; + tensor var_4752_cast_fp16 = mul(x = k_cache18_to_fp16, y = var_4751_cast_fp16)[name = tensor("op_4752_cast_fp16")]; + tensor var_4757 = const()[name = tensor("op_4757"), val = tensor([1, 1, 16, 64])]; + tensor var_4758_cast_fp16 = reshape(shape = var_4757, x = out_75_cast_fp16)[name = tensor("op_4758_cast_fp16")]; + tensor var_4759_cast_fp16 = mul(x = var_4758_cast_fp16, y = onehot_75_cast_fp16)[name = tensor("op_4759_cast_fp16")]; + tensor new_k_cache_37_cast_fp16 = add(x = var_4752_cast_fp16, y = var_4759_cast_fp16)[name = tensor("new_k_cache_37_cast_fp16")]; + tensor new_k_cache_37_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_37_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache18_to_fp16_dtype_0 = const()[name = tensor("v_cache18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache18_to_fp16 = cast(dtype = v_cache18_to_fp16_dtype_0, x = v_cache18)[name = tensor("cast_429")]; + tensor var_4765_cast_fp16 = mul(x = v_cache18_to_fp16, y = var_4751_cast_fp16)[name = tensor("op_4765_cast_fp16")]; + tensor var_4770 = const()[name = tensor("op_4770"), val = tensor([1, 1, 16, 64])]; + tensor var_4771_cast_fp16 = reshape(shape = var_4770, x = v_37_cast_fp16)[name = tensor("op_4771_cast_fp16")]; + tensor var_4772_cast_fp16 = mul(x = var_4771_cast_fp16, y = onehot_75_cast_fp16)[name = tensor("op_4772_cast_fp16")]; + tensor new_v_cache_37_cast_fp16 = add(x = var_4765_cast_fp16, y = var_4772_cast_fp16)[name = tensor("new_v_cache_37_cast_fp16")]; + tensor new_v_cache_37_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_37_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_18_perm_0 = const()[name = tensor("transpose_18_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_4785 = const()[name = tensor("op_4785"), val = tensor([0, 2, 1, 3])]; + tensor var_4791 = const()[name = tensor("op_4791"), val = tensor([1, 16, 1, 64])]; + tensor q4_37_cast_fp16 = reshape(shape = var_4791, x = out_73_cast_fp16)[name = tensor("q4_37_cast_fp16")]; + tensor var_4796_transpose_x_0 = const()[name = tensor("op_4796_transpose_x_0"), val = tensor(false)]; + tensor var_4796_transpose_y_0 = const()[name = tensor("op_4796_transpose_y_0"), val = tensor(false)]; + tensor transpose_18_cast_fp16 = transpose(perm = transpose_18_perm_0, x = new_k_cache_37_cast_fp16)[name = tensor("transpose_41")]; + tensor var_4796_cast_fp16 = matmul(transpose_x = var_4796_transpose_x_0, transpose_y = var_4796_transpose_y_0, x = q4_37_cast_fp16, y = transpose_18_cast_fp16)[name = tensor("op_4796_cast_fp16")]; + tensor var_4797_to_fp16 = const()[name = tensor("op_4797_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_73_cast_fp16 = mul(x = var_4796_cast_fp16, y = var_4797_to_fp16)[name = tensor("scores_73_cast_fp16")]; + tensor var_4801 = const()[name = tensor("op_4801"), val = tensor([1, 1])]; + tensor var_4802_cast_fp16 = reshape(shape = var_4801, x = position18_to_fp16)[name = tensor("op_4802_cast_fp16")]; + tensor var_4803_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_4802_cast_fp16)[name = tensor("op_4803_cast_fp16")]; + tensor mask_37_to_fp16_dtype_0 = const()[name = tensor("mask_37_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_4810_to_fp16 = const()[name = tensor("op_4810_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4803_cast_fp16_to_fp16 = cast(dtype = mask_37_to_fp16_dtype_0, x = var_4803_cast_fp16)[name = tensor("cast_427")]; + tensor var_4811_cast_fp16 = sub(x = var_4803_cast_fp16_to_fp16, y = var_4810_to_fp16)[name = tensor("op_4811_cast_fp16")]; + tensor var_4816 = const()[name = tensor("op_4816"), val = tensor([1, 1, 1, 512])]; + tensor var_4817_cast_fp16 = reshape(shape = var_4816, x = var_4811_cast_fp16)[name = tensor("op_4817_cast_fp16")]; + tensor var_4818_to_fp16 = const()[name = tensor("op_4818_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_4819_cast_fp16 = mul(x = var_4817_cast_fp16, y = var_4818_to_fp16)[name = tensor("op_4819_cast_fp16")]; + tensor scores_75_cast_fp16 = add(x = scores_73_cast_fp16, y = var_4819_cast_fp16)[name = tensor("scores_75_cast_fp16")]; + tensor var_4822 = const()[name = tensor("op_4822"), val = tensor(-1)]; + tensor weights_37_cast_fp16 = softmax(axis = var_4822, x = scores_75_cast_fp16)[name = tensor("weights_37_cast_fp16")]; + tensor attn_37_transpose_x_0 = const()[name = tensor("attn_37_transpose_x_0"), val = tensor(false)]; + tensor attn_37_transpose_y_0 = const()[name = tensor("attn_37_transpose_y_0"), val = tensor(false)]; + tensor values_37_cast_fp16 = transpose(perm = var_4785, x = new_v_cache_37_cast_fp16)[name = tensor("transpose_40")]; + tensor attn_37_cast_fp16 = matmul(transpose_x = attn_37_transpose_x_0, transpose_y = attn_37_transpose_y_0, x = weights_37_cast_fp16, y = values_37_cast_fp16)[name = tensor("attn_37_cast_fp16")]; + tensor var_4830 = const()[name = tensor("op_4830"), val = tensor([0, 2, 1, 3])]; + tensor var_4835 = const()[name = tensor("op_4835"), val = tensor([1, 1, 1024])]; + tensor var_4831_cast_fp16 = transpose(perm = var_4830, x = attn_37_cast_fp16)[name = tensor("transpose_39")]; + tensor input_223_cast_fp16 = reshape(shape = var_4835, x = var_4831_cast_fp16)[name = tensor("input_223_cast_fp16")]; + tensor attn18_out_proj_weight_to_fp16 = const()[name = tensor("attn18_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(459520704)))]; + tensor linear_74_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn18_out_proj_weight_to_fp16, x = input_223_cast_fp16)[name = tensor("linear_74_cast_fp16")]; + tensor var_4841_to_fp16 = const()[name = tensor("op_4841_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4842_cast_fp16 = add(x = position18_to_fp16, y = var_4841_to_fp16)[name = tensor("op_4842_cast_fp16")]; + tensor var_4842_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4842_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_225_cast_fp16 = add(x = input_219_cast_fp16, y = linear_74_cast_fp16)[name = tensor("input_225_cast_fp16")]; + tensor input_227_axes_0 = const()[name = tensor("input_227_axes_0"), val = tensor([-1])]; + tensor norm18_2_weight_to_fp16 = const()[name = tensor("norm18_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(461617920)))]; + tensor norm18_2_bias_to_fp16 = const()[name = tensor("norm18_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(461620032)))]; + tensor var_4846_to_fp16 = const()[name = tensor("op_4846_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_227_cast_fp16 = layer_norm(axes = input_227_axes_0, beta = norm18_2_bias_to_fp16, epsilon = var_4846_to_fp16, gamma = norm18_2_weight_to_fp16, x = input_225_cast_fp16)[name = tensor("input_227_cast_fp16")]; + tensor linear18_1_weight_to_fp16 = const()[name = tensor("linear18_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(461622144)))]; + tensor linear_75_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear18_1_weight_to_fp16, x = input_227_cast_fp16)[name = tensor("linear_75_cast_fp16")]; + tensor input_229_mode_0 = const()[name = tensor("input_229_mode_0"), val = tensor("EXACT")]; + tensor input_229_cast_fp16 = gelu(mode = input_229_mode_0, x = linear_75_cast_fp16)[name = tensor("input_229_cast_fp16")]; + tensor linear18_2_weight_to_fp16 = const()[name = tensor("linear18_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(470010816)))]; + tensor linear_76_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear18_2_weight_to_fp16, x = input_229_cast_fp16)[name = tensor("linear_76_cast_fp16")]; + tensor input_231_cast_fp16 = add(x = input_225_cast_fp16, y = linear_76_cast_fp16)[name = tensor("input_231_cast_fp16")]; + tensor input_233_axes_0 = const()[name = tensor("input_233_axes_0"), val = tensor([-1])]; + tensor norm19_1_weight_to_fp16 = const()[name = tensor("norm19_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(478399488)))]; + tensor norm19_1_bias_to_fp16 = const()[name = tensor("norm19_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(478401600)))]; + tensor var_4863_to_fp16 = const()[name = tensor("op_4863_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_233_cast_fp16 = layer_norm(axes = input_233_axes_0, beta = norm19_1_bias_to_fp16, epsilon = var_4863_to_fp16, gamma = norm19_1_weight_to_fp16, x = input_231_cast_fp16)[name = tensor("input_233_cast_fp16")]; + tensor attn19_in_proj_weight_to_fp16 = const()[name = tensor("attn19_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(478403712)))]; + tensor linear_77_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn19_in_proj_weight_to_fp16, x = input_233_cast_fp16)[name = tensor("linear_77_cast_fp16")]; + tensor var_4876 = const()[name = tensor("op_4876"), val = tensor([1, 3, 16, 64])]; + tensor qkv_39_cast_fp16 = reshape(shape = var_4876, x = linear_77_cast_fp16)[name = tensor("qkv_39_cast_fp16")]; + tensor x_77_begin_0 = const()[name = tensor("x_77_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_77_end_0 = const()[name = tensor("x_77_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_77_end_mask_0 = const()[name = tensor("x_77_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_77_squeeze_mask_0 = const()[name = tensor("x_77_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_77_cast_fp16 = slice_by_index(begin = x_77_begin_0, end = x_77_end_0, end_mask = x_77_end_mask_0, squeeze_mask = x_77_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("x_77_cast_fp16")]; + tensor x_79_begin_0 = const()[name = tensor("x_79_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_79_end_0 = const()[name = tensor("x_79_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_79_end_mask_0 = const()[name = tensor("x_79_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_79_squeeze_mask_0 = const()[name = tensor("x_79_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_79_cast_fp16 = slice_by_index(begin = x_79_begin_0, end = x_79_end_0, end_mask = x_79_end_mask_0, squeeze_mask = x_79_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("x_79_cast_fp16")]; + tensor v_39_begin_0 = const()[name = tensor("v_39_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_39_end_0 = const()[name = tensor("v_39_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_39_end_mask_0 = const()[name = tensor("v_39_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_39_squeeze_mask_0 = const()[name = tensor("v_39_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_39_cast_fp16 = slice_by_index(begin = v_39_begin_0, end = v_39_end_0, end_mask = v_39_end_mask_0, squeeze_mask = v_39_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("v_39_cast_fp16")]; + tensor var_4905 = const()[name = tensor("op_4905"), val = tensor([1, 1, 1])]; + tensor position19_to_fp16_dtype_0 = const()[name = tensor("position19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position19_to_fp16 = cast(dtype = position19_to_fp16_dtype_0, x = position19)[name = tensor("cast_425")]; + tensor var_4906_cast_fp16 = reshape(shape = var_4905, x = position19_to_fp16)[name = tensor("op_4906_cast_fp16")]; + tensor angles_77_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_4906_cast_fp16)[name = tensor("angles_77_cast_fp16")]; + tensor rotr_77_cast_fp16 = cos(x = angles_77_cast_fp16)[name = tensor("rotr_77_cast_fp16")]; + tensor roti_77_cast_fp16 = sin(x = angles_77_cast_fp16)[name = tensor("roti_77_cast_fp16")]; + tensor var_4914 = const()[name = tensor("op_4914"), val = tensor([1, 16, 32, 2])]; + tensor pairs_77_cast_fp16 = reshape(shape = var_4914, x = x_77_cast_fp16)[name = tensor("pairs_77_cast_fp16")]; + tensor xr_77_begin_0 = const()[name = tensor("xr_77_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_77_end_0 = const()[name = tensor("xr_77_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_77_end_mask_0 = const()[name = tensor("xr_77_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_77_squeeze_mask_0 = const()[name = tensor("xr_77_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_77_cast_fp16 = slice_by_index(begin = xr_77_begin_0, end = xr_77_end_0, end_mask = xr_77_end_mask_0, squeeze_mask = xr_77_squeeze_mask_0, x = pairs_77_cast_fp16)[name = tensor("xr_77_cast_fp16")]; + tensor xi_77_begin_0 = const()[name = tensor("xi_77_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_77_end_0 = const()[name = tensor("xi_77_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_77_end_mask_0 = const()[name = tensor("xi_77_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_77_squeeze_mask_0 = const()[name = tensor("xi_77_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_77_cast_fp16 = slice_by_index(begin = xi_77_begin_0, end = xi_77_end_0, end_mask = xi_77_end_mask_0, squeeze_mask = xi_77_squeeze_mask_0, x = pairs_77_cast_fp16)[name = tensor("xi_77_cast_fp16")]; + tensor var_4922_cast_fp16 = mul(x = xr_77_cast_fp16, y = rotr_77_cast_fp16)[name = tensor("op_4922_cast_fp16")]; + tensor var_4923_cast_fp16 = mul(x = xi_77_cast_fp16, y = roti_77_cast_fp16)[name = tensor("op_4923_cast_fp16")]; + tensor xor__77_cast_fp16 = sub(x = var_4922_cast_fp16, y = var_4923_cast_fp16)[name = tensor("xor__77_cast_fp16")]; + tensor var_4926_cast_fp16 = mul(x = xr_77_cast_fp16, y = roti_77_cast_fp16)[name = tensor("op_4926_cast_fp16")]; + tensor var_4927_cast_fp16 = mul(x = xi_77_cast_fp16, y = rotr_77_cast_fp16)[name = tensor("op_4927_cast_fp16")]; + tensor xoi_77_cast_fp16 = add(x = var_4926_cast_fp16, y = var_4927_cast_fp16)[name = tensor("xoi_77_cast_fp16")]; + tensor out_77_axis_0 = const()[name = tensor("out_77_axis_0"), val = tensor(-1)]; + tensor out_77_cast_fp16 = stack(axis = out_77_axis_0, values = (xor__77_cast_fp16, xoi_77_cast_fp16))[name = tensor("out_77_cast_fp16")]; + tensor var_4950 = const()[name = tensor("op_4950"), val = tensor([1, 16, 32, 2])]; + tensor pairs_79_cast_fp16 = reshape(shape = var_4950, x = x_79_cast_fp16)[name = tensor("pairs_79_cast_fp16")]; + tensor xr_79_begin_0 = const()[name = tensor("xr_79_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_79_end_0 = const()[name = tensor("xr_79_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_79_end_mask_0 = const()[name = tensor("xr_79_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_79_squeeze_mask_0 = const()[name = tensor("xr_79_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_79_cast_fp16 = slice_by_index(begin = xr_79_begin_0, end = xr_79_end_0, end_mask = xr_79_end_mask_0, squeeze_mask = xr_79_squeeze_mask_0, x = pairs_79_cast_fp16)[name = tensor("xr_79_cast_fp16")]; + tensor xi_79_begin_0 = const()[name = tensor("xi_79_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_79_end_0 = const()[name = tensor("xi_79_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_79_end_mask_0 = const()[name = tensor("xi_79_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_79_squeeze_mask_0 = const()[name = tensor("xi_79_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_79_cast_fp16 = slice_by_index(begin = xi_79_begin_0, end = xi_79_end_0, end_mask = xi_79_end_mask_0, squeeze_mask = xi_79_squeeze_mask_0, x = pairs_79_cast_fp16)[name = tensor("xi_79_cast_fp16")]; + tensor var_4958_cast_fp16 = mul(x = xr_79_cast_fp16, y = rotr_77_cast_fp16)[name = tensor("op_4958_cast_fp16")]; + tensor var_4959_cast_fp16 = mul(x = xi_79_cast_fp16, y = roti_77_cast_fp16)[name = tensor("op_4959_cast_fp16")]; + tensor xor__79_cast_fp16 = sub(x = var_4958_cast_fp16, y = var_4959_cast_fp16)[name = tensor("xor__79_cast_fp16")]; + tensor var_4962_cast_fp16 = mul(x = xr_79_cast_fp16, y = roti_77_cast_fp16)[name = tensor("op_4962_cast_fp16")]; + tensor var_4963_cast_fp16 = mul(x = xi_79_cast_fp16, y = rotr_77_cast_fp16)[name = tensor("op_4963_cast_fp16")]; + tensor xoi_79_cast_fp16 = add(x = var_4962_cast_fp16, y = var_4963_cast_fp16)[name = tensor("xoi_79_cast_fp16")]; + tensor out_79_axis_0 = const()[name = tensor("out_79_axis_0"), val = tensor(-1)]; + tensor out_79_cast_fp16 = stack(axis = out_79_axis_0, values = (xor__79_cast_fp16, xoi_79_cast_fp16))[name = tensor("out_79_cast_fp16")]; + tensor _inversed_4975_y_0_to_fp16 = const()[name = tensor("_inversed_4975_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_4975_cast_fp16 = mul(x = position19_to_fp16, y = _inversed_4975_y_0_to_fp16)[name = tensor("_inversed_4975_cast_fp16")]; + tensor var_4976_cast_fp16 = floor(x = _inversed_4975_cast_fp16)[name = tensor("op_4976_cast_fp16")]; + tensor var_4977_to_fp16 = const()[name = tensor("op_4977_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4978_cast_fp16 = mul(x = var_4976_cast_fp16, y = var_4977_to_fp16)[name = tensor("op_4978_cast_fp16")]; + tensor write_idx_39_cast_fp16 = sub(x = position19_to_fp16, y = var_4978_cast_fp16)[name = tensor("write_idx_39_cast_fp16")]; + tensor var_4981_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_39_cast_fp16)[name = tensor("op_4981_cast_fp16")]; + tensor var_4991 = const()[name = tensor("op_4991"), val = tensor([1, 512, 1, 1])]; + tensor onehot_77_to_fp16_dtype_0 = const()[name = tensor("onehot_77_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_4981_cast_fp16_to_fp16 = cast(dtype = onehot_77_to_fp16_dtype_0, x = var_4981_cast_fp16)[name = tensor("cast_424")]; + tensor onehot_79_cast_fp16 = reshape(shape = var_4991, x = var_4981_cast_fp16_to_fp16)[name = tensor("onehot_79_cast_fp16")]; + tensor var_4993_to_fp16 = const()[name = tensor("op_4993_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4995_cast_fp16 = sub(x = var_4993_to_fp16, y = onehot_79_cast_fp16)[name = tensor("op_4995_cast_fp16")]; + tensor k_cache19_to_fp16_dtype_0 = const()[name = tensor("k_cache19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache19_to_fp16 = cast(dtype = k_cache19_to_fp16_dtype_0, x = k_cache19)[name = tensor("cast_423")]; + tensor var_4996_cast_fp16 = mul(x = k_cache19_to_fp16, y = var_4995_cast_fp16)[name = tensor("op_4996_cast_fp16")]; + tensor var_5001 = const()[name = tensor("op_5001"), val = tensor([1, 1, 16, 64])]; + tensor var_5002_cast_fp16 = reshape(shape = var_5001, x = out_79_cast_fp16)[name = tensor("op_5002_cast_fp16")]; + tensor var_5003_cast_fp16 = mul(x = var_5002_cast_fp16, y = onehot_79_cast_fp16)[name = tensor("op_5003_cast_fp16")]; + tensor new_k_cache_39_cast_fp16 = add(x = var_4996_cast_fp16, y = var_5003_cast_fp16)[name = tensor("new_k_cache_39_cast_fp16")]; + tensor new_k_cache_39_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_39_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache19_to_fp16_dtype_0 = const()[name = tensor("v_cache19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache19_to_fp16 = cast(dtype = v_cache19_to_fp16_dtype_0, x = v_cache19)[name = tensor("cast_421")]; + tensor var_5009_cast_fp16 = mul(x = v_cache19_to_fp16, y = var_4995_cast_fp16)[name = tensor("op_5009_cast_fp16")]; + tensor var_5014 = const()[name = tensor("op_5014"), val = tensor([1, 1, 16, 64])]; + tensor var_5015_cast_fp16 = reshape(shape = var_5014, x = v_39_cast_fp16)[name = tensor("op_5015_cast_fp16")]; + tensor var_5016_cast_fp16 = mul(x = var_5015_cast_fp16, y = onehot_79_cast_fp16)[name = tensor("op_5016_cast_fp16")]; + tensor new_v_cache_39_cast_fp16 = add(x = var_5009_cast_fp16, y = var_5016_cast_fp16)[name = tensor("new_v_cache_39_cast_fp16")]; + tensor new_v_cache_39_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_39_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_19_perm_0 = const()[name = tensor("transpose_19_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_5029 = const()[name = tensor("op_5029"), val = tensor([0, 2, 1, 3])]; + tensor var_5035 = const()[name = tensor("op_5035"), val = tensor([1, 16, 1, 64])]; + tensor q4_39_cast_fp16 = reshape(shape = var_5035, x = out_77_cast_fp16)[name = tensor("q4_39_cast_fp16")]; + tensor var_5040_transpose_x_0 = const()[name = tensor("op_5040_transpose_x_0"), val = tensor(false)]; + tensor var_5040_transpose_y_0 = const()[name = tensor("op_5040_transpose_y_0"), val = tensor(false)]; + tensor transpose_19_cast_fp16 = transpose(perm = transpose_19_perm_0, x = new_k_cache_39_cast_fp16)[name = tensor("transpose_38")]; + tensor var_5040_cast_fp16 = matmul(transpose_x = var_5040_transpose_x_0, transpose_y = var_5040_transpose_y_0, x = q4_39_cast_fp16, y = transpose_19_cast_fp16)[name = tensor("op_5040_cast_fp16")]; + tensor var_5041_to_fp16 = const()[name = tensor("op_5041_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_77_cast_fp16 = mul(x = var_5040_cast_fp16, y = var_5041_to_fp16)[name = tensor("scores_77_cast_fp16")]; + tensor var_5045 = const()[name = tensor("op_5045"), val = tensor([1, 1])]; + tensor var_5046_cast_fp16 = reshape(shape = var_5045, x = position19_to_fp16)[name = tensor("op_5046_cast_fp16")]; + tensor var_5047_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_5046_cast_fp16)[name = tensor("op_5047_cast_fp16")]; + tensor mask_39_to_fp16_dtype_0 = const()[name = tensor("mask_39_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_5054_to_fp16 = const()[name = tensor("op_5054_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5047_cast_fp16_to_fp16 = cast(dtype = mask_39_to_fp16_dtype_0, x = var_5047_cast_fp16)[name = tensor("cast_419")]; + tensor var_5055_cast_fp16 = sub(x = var_5047_cast_fp16_to_fp16, y = var_5054_to_fp16)[name = tensor("op_5055_cast_fp16")]; + tensor var_5060 = const()[name = tensor("op_5060"), val = tensor([1, 1, 1, 512])]; + tensor var_5061_cast_fp16 = reshape(shape = var_5060, x = var_5055_cast_fp16)[name = tensor("op_5061_cast_fp16")]; + tensor var_5062_to_fp16 = const()[name = tensor("op_5062_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_5063_cast_fp16 = mul(x = var_5061_cast_fp16, y = var_5062_to_fp16)[name = tensor("op_5063_cast_fp16")]; + tensor scores_79_cast_fp16 = add(x = scores_77_cast_fp16, y = var_5063_cast_fp16)[name = tensor("scores_79_cast_fp16")]; + tensor var_5066 = const()[name = tensor("op_5066"), val = tensor(-1)]; + tensor weights_39_cast_fp16 = softmax(axis = var_5066, x = scores_79_cast_fp16)[name = tensor("weights_39_cast_fp16")]; + tensor attn_39_transpose_x_0 = const()[name = tensor("attn_39_transpose_x_0"), val = tensor(false)]; + tensor attn_39_transpose_y_0 = const()[name = tensor("attn_39_transpose_y_0"), val = tensor(false)]; + tensor values_39_cast_fp16 = transpose(perm = var_5029, x = new_v_cache_39_cast_fp16)[name = tensor("transpose_37")]; + tensor attn_39_cast_fp16 = matmul(transpose_x = attn_39_transpose_x_0, transpose_y = attn_39_transpose_y_0, x = weights_39_cast_fp16, y = values_39_cast_fp16)[name = tensor("attn_39_cast_fp16")]; + tensor var_5074 = const()[name = tensor("op_5074"), val = tensor([0, 2, 1, 3])]; + tensor var_5079 = const()[name = tensor("op_5079"), val = tensor([1, 1, 1024])]; + tensor var_5075_cast_fp16 = transpose(perm = var_5074, x = attn_39_cast_fp16)[name = tensor("transpose_36")]; + tensor input_235_cast_fp16 = reshape(shape = var_5079, x = var_5075_cast_fp16)[name = tensor("input_235_cast_fp16")]; + tensor attn19_out_proj_weight_to_fp16 = const()[name = tensor("attn19_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(484695232)))]; + tensor linear_78_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn19_out_proj_weight_to_fp16, x = input_235_cast_fp16)[name = tensor("linear_78_cast_fp16")]; + tensor var_5085_to_fp16 = const()[name = tensor("op_5085_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5086_cast_fp16 = add(x = position19_to_fp16, y = var_5085_to_fp16)[name = tensor("op_5086_cast_fp16")]; + tensor var_5086_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5086_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_237_cast_fp16 = add(x = input_231_cast_fp16, y = linear_78_cast_fp16)[name = tensor("input_237_cast_fp16")]; + tensor input_239_axes_0 = const()[name = tensor("input_239_axes_0"), val = tensor([-1])]; + tensor norm19_2_weight_to_fp16 = const()[name = tensor("norm19_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486792448)))]; + tensor norm19_2_bias_to_fp16 = const()[name = tensor("norm19_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486794560)))]; + tensor var_5090_to_fp16 = const()[name = tensor("op_5090_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_239_cast_fp16 = layer_norm(axes = input_239_axes_0, beta = norm19_2_bias_to_fp16, epsilon = var_5090_to_fp16, gamma = norm19_2_weight_to_fp16, x = input_237_cast_fp16)[name = tensor("input_239_cast_fp16")]; + tensor linear19_1_weight_to_fp16 = const()[name = tensor("linear19_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486796672)))]; + tensor linear_79_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear19_1_weight_to_fp16, x = input_239_cast_fp16)[name = tensor("linear_79_cast_fp16")]; + tensor input_241_mode_0 = const()[name = tensor("input_241_mode_0"), val = tensor("EXACT")]; + tensor input_241_cast_fp16 = gelu(mode = input_241_mode_0, x = linear_79_cast_fp16)[name = tensor("input_241_cast_fp16")]; + tensor linear19_2_weight_to_fp16 = const()[name = tensor("linear19_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(495185344)))]; + tensor linear_80_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear19_2_weight_to_fp16, x = input_241_cast_fp16)[name = tensor("linear_80_cast_fp16")]; + tensor input_243_cast_fp16 = add(x = input_237_cast_fp16, y = linear_80_cast_fp16)[name = tensor("input_243_cast_fp16")]; + tensor input_245_axes_0 = const()[name = tensor("input_245_axes_0"), val = tensor([-1])]; + tensor norm20_1_weight_to_fp16 = const()[name = tensor("norm20_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(503574016)))]; + tensor norm20_1_bias_to_fp16 = const()[name = tensor("norm20_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(503576128)))]; + tensor var_5107_to_fp16 = const()[name = tensor("op_5107_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_245_cast_fp16 = layer_norm(axes = input_245_axes_0, beta = norm20_1_bias_to_fp16, epsilon = var_5107_to_fp16, gamma = norm20_1_weight_to_fp16, x = input_243_cast_fp16)[name = tensor("input_245_cast_fp16")]; + tensor attn20_in_proj_weight_to_fp16 = const()[name = tensor("attn20_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(503578240)))]; + tensor linear_81_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn20_in_proj_weight_to_fp16, x = input_245_cast_fp16)[name = tensor("linear_81_cast_fp16")]; + tensor var_5120 = const()[name = tensor("op_5120"), val = tensor([1, 3, 16, 64])]; + tensor qkv_41_cast_fp16 = reshape(shape = var_5120, x = linear_81_cast_fp16)[name = tensor("qkv_41_cast_fp16")]; + tensor x_81_begin_0 = const()[name = tensor("x_81_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_81_end_0 = const()[name = tensor("x_81_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_81_end_mask_0 = const()[name = tensor("x_81_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_81_squeeze_mask_0 = const()[name = tensor("x_81_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_81_cast_fp16 = slice_by_index(begin = x_81_begin_0, end = x_81_end_0, end_mask = x_81_end_mask_0, squeeze_mask = x_81_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("x_81_cast_fp16")]; + tensor x_83_begin_0 = const()[name = tensor("x_83_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_83_end_0 = const()[name = tensor("x_83_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_83_end_mask_0 = const()[name = tensor("x_83_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_83_squeeze_mask_0 = const()[name = tensor("x_83_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_83_cast_fp16 = slice_by_index(begin = x_83_begin_0, end = x_83_end_0, end_mask = x_83_end_mask_0, squeeze_mask = x_83_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("x_83_cast_fp16")]; + tensor v_41_begin_0 = const()[name = tensor("v_41_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_41_end_0 = const()[name = tensor("v_41_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_41_end_mask_0 = const()[name = tensor("v_41_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_41_squeeze_mask_0 = const()[name = tensor("v_41_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_41_cast_fp16 = slice_by_index(begin = v_41_begin_0, end = v_41_end_0, end_mask = v_41_end_mask_0, squeeze_mask = v_41_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("v_41_cast_fp16")]; + tensor var_5149 = const()[name = tensor("op_5149"), val = tensor([1, 1, 1])]; + tensor position20_to_fp16_dtype_0 = const()[name = tensor("position20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position20_to_fp16 = cast(dtype = position20_to_fp16_dtype_0, x = position20)[name = tensor("cast_417")]; + tensor var_5150_cast_fp16 = reshape(shape = var_5149, x = position20_to_fp16)[name = tensor("op_5150_cast_fp16")]; + tensor angles_81_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_5150_cast_fp16)[name = tensor("angles_81_cast_fp16")]; + tensor rotr_81_cast_fp16 = cos(x = angles_81_cast_fp16)[name = tensor("rotr_81_cast_fp16")]; + tensor roti_81_cast_fp16 = sin(x = angles_81_cast_fp16)[name = tensor("roti_81_cast_fp16")]; + tensor var_5158 = const()[name = tensor("op_5158"), val = tensor([1, 16, 32, 2])]; + tensor pairs_81_cast_fp16 = reshape(shape = var_5158, x = x_81_cast_fp16)[name = tensor("pairs_81_cast_fp16")]; + tensor xr_81_begin_0 = const()[name = tensor("xr_81_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_81_end_0 = const()[name = tensor("xr_81_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_81_end_mask_0 = const()[name = tensor("xr_81_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_81_squeeze_mask_0 = const()[name = tensor("xr_81_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_81_cast_fp16 = slice_by_index(begin = xr_81_begin_0, end = xr_81_end_0, end_mask = xr_81_end_mask_0, squeeze_mask = xr_81_squeeze_mask_0, x = pairs_81_cast_fp16)[name = tensor("xr_81_cast_fp16")]; + tensor xi_81_begin_0 = const()[name = tensor("xi_81_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_81_end_0 = const()[name = tensor("xi_81_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_81_end_mask_0 = const()[name = tensor("xi_81_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_81_squeeze_mask_0 = const()[name = tensor("xi_81_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_81_cast_fp16 = slice_by_index(begin = xi_81_begin_0, end = xi_81_end_0, end_mask = xi_81_end_mask_0, squeeze_mask = xi_81_squeeze_mask_0, x = pairs_81_cast_fp16)[name = tensor("xi_81_cast_fp16")]; + tensor var_5166_cast_fp16 = mul(x = xr_81_cast_fp16, y = rotr_81_cast_fp16)[name = tensor("op_5166_cast_fp16")]; + tensor var_5167_cast_fp16 = mul(x = xi_81_cast_fp16, y = roti_81_cast_fp16)[name = tensor("op_5167_cast_fp16")]; + tensor xor__81_cast_fp16 = sub(x = var_5166_cast_fp16, y = var_5167_cast_fp16)[name = tensor("xor__81_cast_fp16")]; + tensor var_5170_cast_fp16 = mul(x = xr_81_cast_fp16, y = roti_81_cast_fp16)[name = tensor("op_5170_cast_fp16")]; + tensor var_5171_cast_fp16 = mul(x = xi_81_cast_fp16, y = rotr_81_cast_fp16)[name = tensor("op_5171_cast_fp16")]; + tensor xoi_81_cast_fp16 = add(x = var_5170_cast_fp16, y = var_5171_cast_fp16)[name = tensor("xoi_81_cast_fp16")]; + tensor out_81_axis_0 = const()[name = tensor("out_81_axis_0"), val = tensor(-1)]; + tensor out_81_cast_fp16 = stack(axis = out_81_axis_0, values = (xor__81_cast_fp16, xoi_81_cast_fp16))[name = tensor("out_81_cast_fp16")]; + tensor var_5194 = const()[name = tensor("op_5194"), val = tensor([1, 16, 32, 2])]; + tensor pairs_83_cast_fp16 = reshape(shape = var_5194, x = x_83_cast_fp16)[name = tensor("pairs_83_cast_fp16")]; + tensor xr_83_begin_0 = const()[name = tensor("xr_83_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_83_end_0 = const()[name = tensor("xr_83_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_83_end_mask_0 = const()[name = tensor("xr_83_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_83_squeeze_mask_0 = const()[name = tensor("xr_83_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_83_cast_fp16 = slice_by_index(begin = xr_83_begin_0, end = xr_83_end_0, end_mask = xr_83_end_mask_0, squeeze_mask = xr_83_squeeze_mask_0, x = pairs_83_cast_fp16)[name = tensor("xr_83_cast_fp16")]; + tensor xi_83_begin_0 = const()[name = tensor("xi_83_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_83_end_0 = const()[name = tensor("xi_83_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_83_end_mask_0 = const()[name = tensor("xi_83_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_83_squeeze_mask_0 = const()[name = tensor("xi_83_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_83_cast_fp16 = slice_by_index(begin = xi_83_begin_0, end = xi_83_end_0, end_mask = xi_83_end_mask_0, squeeze_mask = xi_83_squeeze_mask_0, x = pairs_83_cast_fp16)[name = tensor("xi_83_cast_fp16")]; + tensor var_5202_cast_fp16 = mul(x = xr_83_cast_fp16, y = rotr_81_cast_fp16)[name = tensor("op_5202_cast_fp16")]; + tensor var_5203_cast_fp16 = mul(x = xi_83_cast_fp16, y = roti_81_cast_fp16)[name = tensor("op_5203_cast_fp16")]; + tensor xor__83_cast_fp16 = sub(x = var_5202_cast_fp16, y = var_5203_cast_fp16)[name = tensor("xor__83_cast_fp16")]; + tensor var_5206_cast_fp16 = mul(x = xr_83_cast_fp16, y = roti_81_cast_fp16)[name = tensor("op_5206_cast_fp16")]; + tensor var_5207_cast_fp16 = mul(x = xi_83_cast_fp16, y = rotr_81_cast_fp16)[name = tensor("op_5207_cast_fp16")]; + tensor xoi_83_cast_fp16 = add(x = var_5206_cast_fp16, y = var_5207_cast_fp16)[name = tensor("xoi_83_cast_fp16")]; + tensor out_83_axis_0 = const()[name = tensor("out_83_axis_0"), val = tensor(-1)]; + tensor out_83_cast_fp16 = stack(axis = out_83_axis_0, values = (xor__83_cast_fp16, xoi_83_cast_fp16))[name = tensor("out_83_cast_fp16")]; + tensor _inversed_5219_y_0_to_fp16 = const()[name = tensor("_inversed_5219_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5219_cast_fp16 = mul(x = position20_to_fp16, y = _inversed_5219_y_0_to_fp16)[name = tensor("_inversed_5219_cast_fp16")]; + tensor var_5220_cast_fp16 = floor(x = _inversed_5219_cast_fp16)[name = tensor("op_5220_cast_fp16")]; + tensor var_5221_to_fp16 = const()[name = tensor("op_5221_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5222_cast_fp16 = mul(x = var_5220_cast_fp16, y = var_5221_to_fp16)[name = tensor("op_5222_cast_fp16")]; + tensor write_idx_41_cast_fp16 = sub(x = position20_to_fp16, y = var_5222_cast_fp16)[name = tensor("write_idx_41_cast_fp16")]; + tensor var_5225_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_41_cast_fp16)[name = tensor("op_5225_cast_fp16")]; + tensor var_5235 = const()[name = tensor("op_5235"), val = tensor([1, 512, 1, 1])]; + tensor onehot_81_to_fp16_dtype_0 = const()[name = tensor("onehot_81_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_5225_cast_fp16_to_fp16 = cast(dtype = onehot_81_to_fp16_dtype_0, x = var_5225_cast_fp16)[name = tensor("cast_416")]; + tensor onehot_83_cast_fp16 = reshape(shape = var_5235, x = var_5225_cast_fp16_to_fp16)[name = tensor("onehot_83_cast_fp16")]; + tensor var_5237_to_fp16 = const()[name = tensor("op_5237_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5239_cast_fp16 = sub(x = var_5237_to_fp16, y = onehot_83_cast_fp16)[name = tensor("op_5239_cast_fp16")]; + tensor k_cache20_to_fp16_dtype_0 = const()[name = tensor("k_cache20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache20_to_fp16 = cast(dtype = k_cache20_to_fp16_dtype_0, x = k_cache20)[name = tensor("cast_415")]; + tensor var_5240_cast_fp16 = mul(x = k_cache20_to_fp16, y = var_5239_cast_fp16)[name = tensor("op_5240_cast_fp16")]; + tensor var_5245 = const()[name = tensor("op_5245"), val = tensor([1, 1, 16, 64])]; + tensor var_5246_cast_fp16 = reshape(shape = var_5245, x = out_83_cast_fp16)[name = tensor("op_5246_cast_fp16")]; + tensor var_5247_cast_fp16 = mul(x = var_5246_cast_fp16, y = onehot_83_cast_fp16)[name = tensor("op_5247_cast_fp16")]; + tensor new_k_cache_41_cast_fp16 = add(x = var_5240_cast_fp16, y = var_5247_cast_fp16)[name = tensor("new_k_cache_41_cast_fp16")]; + tensor new_k_cache_41_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_41_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache20_to_fp16_dtype_0 = const()[name = tensor("v_cache20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache20_to_fp16 = cast(dtype = v_cache20_to_fp16_dtype_0, x = v_cache20)[name = tensor("cast_413")]; + tensor var_5253_cast_fp16 = mul(x = v_cache20_to_fp16, y = var_5239_cast_fp16)[name = tensor("op_5253_cast_fp16")]; + tensor var_5258 = const()[name = tensor("op_5258"), val = tensor([1, 1, 16, 64])]; + tensor var_5259_cast_fp16 = reshape(shape = var_5258, x = v_41_cast_fp16)[name = tensor("op_5259_cast_fp16")]; + tensor var_5260_cast_fp16 = mul(x = var_5259_cast_fp16, y = onehot_83_cast_fp16)[name = tensor("op_5260_cast_fp16")]; + tensor new_v_cache_41_cast_fp16 = add(x = var_5253_cast_fp16, y = var_5260_cast_fp16)[name = tensor("new_v_cache_41_cast_fp16")]; + tensor new_v_cache_41_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_41_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_20_perm_0 = const()[name = tensor("transpose_20_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_5273 = const()[name = tensor("op_5273"), val = tensor([0, 2, 1, 3])]; + tensor var_5279 = const()[name = tensor("op_5279"), val = tensor([1, 16, 1, 64])]; + tensor q4_41_cast_fp16 = reshape(shape = var_5279, x = out_81_cast_fp16)[name = tensor("q4_41_cast_fp16")]; + tensor var_5284_transpose_x_0 = const()[name = tensor("op_5284_transpose_x_0"), val = tensor(false)]; + tensor var_5284_transpose_y_0 = const()[name = tensor("op_5284_transpose_y_0"), val = tensor(false)]; + tensor transpose_20_cast_fp16 = transpose(perm = transpose_20_perm_0, x = new_k_cache_41_cast_fp16)[name = tensor("transpose_35")]; + tensor var_5284_cast_fp16 = matmul(transpose_x = var_5284_transpose_x_0, transpose_y = var_5284_transpose_y_0, x = q4_41_cast_fp16, y = transpose_20_cast_fp16)[name = tensor("op_5284_cast_fp16")]; + tensor var_5285_to_fp16 = const()[name = tensor("op_5285_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_81_cast_fp16 = mul(x = var_5284_cast_fp16, y = var_5285_to_fp16)[name = tensor("scores_81_cast_fp16")]; + tensor var_5289 = const()[name = tensor("op_5289"), val = tensor([1, 1])]; + tensor var_5290_cast_fp16 = reshape(shape = var_5289, x = position20_to_fp16)[name = tensor("op_5290_cast_fp16")]; + tensor var_5291_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_5290_cast_fp16)[name = tensor("op_5291_cast_fp16")]; + tensor mask_41_to_fp16_dtype_0 = const()[name = tensor("mask_41_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_5298_to_fp16 = const()[name = tensor("op_5298_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5291_cast_fp16_to_fp16 = cast(dtype = mask_41_to_fp16_dtype_0, x = var_5291_cast_fp16)[name = tensor("cast_411")]; + tensor var_5299_cast_fp16 = sub(x = var_5291_cast_fp16_to_fp16, y = var_5298_to_fp16)[name = tensor("op_5299_cast_fp16")]; + tensor var_5304 = const()[name = tensor("op_5304"), val = tensor([1, 1, 1, 512])]; + tensor var_5305_cast_fp16 = reshape(shape = var_5304, x = var_5299_cast_fp16)[name = tensor("op_5305_cast_fp16")]; + tensor var_5306_to_fp16 = const()[name = tensor("op_5306_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_5307_cast_fp16 = mul(x = var_5305_cast_fp16, y = var_5306_to_fp16)[name = tensor("op_5307_cast_fp16")]; + tensor scores_83_cast_fp16 = add(x = scores_81_cast_fp16, y = var_5307_cast_fp16)[name = tensor("scores_83_cast_fp16")]; + tensor var_5310 = const()[name = tensor("op_5310"), val = tensor(-1)]; + tensor weights_41_cast_fp16 = softmax(axis = var_5310, x = scores_83_cast_fp16)[name = tensor("weights_41_cast_fp16")]; + tensor attn_41_transpose_x_0 = const()[name = tensor("attn_41_transpose_x_0"), val = tensor(false)]; + tensor attn_41_transpose_y_0 = const()[name = tensor("attn_41_transpose_y_0"), val = tensor(false)]; + tensor values_41_cast_fp16 = transpose(perm = var_5273, x = new_v_cache_41_cast_fp16)[name = tensor("transpose_34")]; + tensor attn_41_cast_fp16 = matmul(transpose_x = attn_41_transpose_x_0, transpose_y = attn_41_transpose_y_0, x = weights_41_cast_fp16, y = values_41_cast_fp16)[name = tensor("attn_41_cast_fp16")]; + tensor var_5318 = const()[name = tensor("op_5318"), val = tensor([0, 2, 1, 3])]; + tensor var_5323 = const()[name = tensor("op_5323"), val = tensor([1, 1, 1024])]; + tensor var_5319_cast_fp16 = transpose(perm = var_5318, x = attn_41_cast_fp16)[name = tensor("transpose_33")]; + tensor input_247_cast_fp16 = reshape(shape = var_5323, x = var_5319_cast_fp16)[name = tensor("input_247_cast_fp16")]; + tensor attn20_out_proj_weight_to_fp16 = const()[name = tensor("attn20_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(509869760)))]; + tensor linear_82_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn20_out_proj_weight_to_fp16, x = input_247_cast_fp16)[name = tensor("linear_82_cast_fp16")]; + tensor var_5329_to_fp16 = const()[name = tensor("op_5329_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5330_cast_fp16 = add(x = position20_to_fp16, y = var_5329_to_fp16)[name = tensor("op_5330_cast_fp16")]; + tensor var_5330_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5330_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_249_cast_fp16 = add(x = input_243_cast_fp16, y = linear_82_cast_fp16)[name = tensor("input_249_cast_fp16")]; + tensor input_251_axes_0 = const()[name = tensor("input_251_axes_0"), val = tensor([-1])]; + tensor norm20_2_weight_to_fp16 = const()[name = tensor("norm20_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(511966976)))]; + tensor norm20_2_bias_to_fp16 = const()[name = tensor("norm20_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(511969088)))]; + tensor var_5334_to_fp16 = const()[name = tensor("op_5334_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_251_cast_fp16 = layer_norm(axes = input_251_axes_0, beta = norm20_2_bias_to_fp16, epsilon = var_5334_to_fp16, gamma = norm20_2_weight_to_fp16, x = input_249_cast_fp16)[name = tensor("input_251_cast_fp16")]; + tensor linear20_1_weight_to_fp16 = const()[name = tensor("linear20_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(511971200)))]; + tensor linear_83_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear20_1_weight_to_fp16, x = input_251_cast_fp16)[name = tensor("linear_83_cast_fp16")]; + tensor input_253_mode_0 = const()[name = tensor("input_253_mode_0"), val = tensor("EXACT")]; + tensor input_253_cast_fp16 = gelu(mode = input_253_mode_0, x = linear_83_cast_fp16)[name = tensor("input_253_cast_fp16")]; + tensor linear20_2_weight_to_fp16 = const()[name = tensor("linear20_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(520359872)))]; + tensor linear_84_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear20_2_weight_to_fp16, x = input_253_cast_fp16)[name = tensor("linear_84_cast_fp16")]; + tensor input_255_cast_fp16 = add(x = input_249_cast_fp16, y = linear_84_cast_fp16)[name = tensor("input_255_cast_fp16")]; + tensor input_257_axes_0 = const()[name = tensor("input_257_axes_0"), val = tensor([-1])]; + tensor norm21_1_weight_to_fp16 = const()[name = tensor("norm21_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(528748544)))]; + tensor norm21_1_bias_to_fp16 = const()[name = tensor("norm21_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(528750656)))]; + tensor var_5351_to_fp16 = const()[name = tensor("op_5351_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_257_cast_fp16 = layer_norm(axes = input_257_axes_0, beta = norm21_1_bias_to_fp16, epsilon = var_5351_to_fp16, gamma = norm21_1_weight_to_fp16, x = input_255_cast_fp16)[name = tensor("input_257_cast_fp16")]; + tensor attn21_in_proj_weight_to_fp16 = const()[name = tensor("attn21_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(528752768)))]; + tensor linear_85_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn21_in_proj_weight_to_fp16, x = input_257_cast_fp16)[name = tensor("linear_85_cast_fp16")]; + tensor var_5364 = const()[name = tensor("op_5364"), val = tensor([1, 3, 16, 64])]; + tensor qkv_43_cast_fp16 = reshape(shape = var_5364, x = linear_85_cast_fp16)[name = tensor("qkv_43_cast_fp16")]; + tensor x_85_begin_0 = const()[name = tensor("x_85_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_85_end_0 = const()[name = tensor("x_85_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_85_end_mask_0 = const()[name = tensor("x_85_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_85_squeeze_mask_0 = const()[name = tensor("x_85_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_85_cast_fp16 = slice_by_index(begin = x_85_begin_0, end = x_85_end_0, end_mask = x_85_end_mask_0, squeeze_mask = x_85_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("x_85_cast_fp16")]; + tensor x_87_begin_0 = const()[name = tensor("x_87_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_87_end_0 = const()[name = tensor("x_87_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_87_end_mask_0 = const()[name = tensor("x_87_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_87_squeeze_mask_0 = const()[name = tensor("x_87_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_87_cast_fp16 = slice_by_index(begin = x_87_begin_0, end = x_87_end_0, end_mask = x_87_end_mask_0, squeeze_mask = x_87_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("x_87_cast_fp16")]; + tensor v_43_begin_0 = const()[name = tensor("v_43_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_43_end_0 = const()[name = tensor("v_43_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_43_end_mask_0 = const()[name = tensor("v_43_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_43_squeeze_mask_0 = const()[name = tensor("v_43_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_43_cast_fp16 = slice_by_index(begin = v_43_begin_0, end = v_43_end_0, end_mask = v_43_end_mask_0, squeeze_mask = v_43_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("v_43_cast_fp16")]; + tensor var_5393 = const()[name = tensor("op_5393"), val = tensor([1, 1, 1])]; + tensor position21_to_fp16_dtype_0 = const()[name = tensor("position21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position21_to_fp16 = cast(dtype = position21_to_fp16_dtype_0, x = position21)[name = tensor("cast_409")]; + tensor var_5394_cast_fp16 = reshape(shape = var_5393, x = position21_to_fp16)[name = tensor("op_5394_cast_fp16")]; + tensor angles_85_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_5394_cast_fp16)[name = tensor("angles_85_cast_fp16")]; + tensor rotr_85_cast_fp16 = cos(x = angles_85_cast_fp16)[name = tensor("rotr_85_cast_fp16")]; + tensor roti_85_cast_fp16 = sin(x = angles_85_cast_fp16)[name = tensor("roti_85_cast_fp16")]; + tensor var_5402 = const()[name = tensor("op_5402"), val = tensor([1, 16, 32, 2])]; + tensor pairs_85_cast_fp16 = reshape(shape = var_5402, x = x_85_cast_fp16)[name = tensor("pairs_85_cast_fp16")]; + tensor xr_85_begin_0 = const()[name = tensor("xr_85_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_85_end_0 = const()[name = tensor("xr_85_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_85_end_mask_0 = const()[name = tensor("xr_85_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_85_squeeze_mask_0 = const()[name = tensor("xr_85_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_85_cast_fp16 = slice_by_index(begin = xr_85_begin_0, end = xr_85_end_0, end_mask = xr_85_end_mask_0, squeeze_mask = xr_85_squeeze_mask_0, x = pairs_85_cast_fp16)[name = tensor("xr_85_cast_fp16")]; + tensor xi_85_begin_0 = const()[name = tensor("xi_85_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_85_end_0 = const()[name = tensor("xi_85_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_85_end_mask_0 = const()[name = tensor("xi_85_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_85_squeeze_mask_0 = const()[name = tensor("xi_85_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_85_cast_fp16 = slice_by_index(begin = xi_85_begin_0, end = xi_85_end_0, end_mask = xi_85_end_mask_0, squeeze_mask = xi_85_squeeze_mask_0, x = pairs_85_cast_fp16)[name = tensor("xi_85_cast_fp16")]; + tensor var_5410_cast_fp16 = mul(x = xr_85_cast_fp16, y = rotr_85_cast_fp16)[name = tensor("op_5410_cast_fp16")]; + tensor var_5411_cast_fp16 = mul(x = xi_85_cast_fp16, y = roti_85_cast_fp16)[name = tensor("op_5411_cast_fp16")]; + tensor xor__85_cast_fp16 = sub(x = var_5410_cast_fp16, y = var_5411_cast_fp16)[name = tensor("xor__85_cast_fp16")]; + tensor var_5414_cast_fp16 = mul(x = xr_85_cast_fp16, y = roti_85_cast_fp16)[name = tensor("op_5414_cast_fp16")]; + tensor var_5415_cast_fp16 = mul(x = xi_85_cast_fp16, y = rotr_85_cast_fp16)[name = tensor("op_5415_cast_fp16")]; + tensor xoi_85_cast_fp16 = add(x = var_5414_cast_fp16, y = var_5415_cast_fp16)[name = tensor("xoi_85_cast_fp16")]; + tensor out_85_axis_0 = const()[name = tensor("out_85_axis_0"), val = tensor(-1)]; + tensor out_85_cast_fp16 = stack(axis = out_85_axis_0, values = (xor__85_cast_fp16, xoi_85_cast_fp16))[name = tensor("out_85_cast_fp16")]; + tensor var_5438 = const()[name = tensor("op_5438"), val = tensor([1, 16, 32, 2])]; + tensor pairs_87_cast_fp16 = reshape(shape = var_5438, x = x_87_cast_fp16)[name = tensor("pairs_87_cast_fp16")]; + tensor xr_87_begin_0 = const()[name = tensor("xr_87_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_87_end_0 = const()[name = tensor("xr_87_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_87_end_mask_0 = const()[name = tensor("xr_87_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_87_squeeze_mask_0 = const()[name = tensor("xr_87_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_87_cast_fp16 = slice_by_index(begin = xr_87_begin_0, end = xr_87_end_0, end_mask = xr_87_end_mask_0, squeeze_mask = xr_87_squeeze_mask_0, x = pairs_87_cast_fp16)[name = tensor("xr_87_cast_fp16")]; + tensor xi_87_begin_0 = const()[name = tensor("xi_87_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_87_end_0 = const()[name = tensor("xi_87_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_87_end_mask_0 = const()[name = tensor("xi_87_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_87_squeeze_mask_0 = const()[name = tensor("xi_87_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_87_cast_fp16 = slice_by_index(begin = xi_87_begin_0, end = xi_87_end_0, end_mask = xi_87_end_mask_0, squeeze_mask = xi_87_squeeze_mask_0, x = pairs_87_cast_fp16)[name = tensor("xi_87_cast_fp16")]; + tensor var_5446_cast_fp16 = mul(x = xr_87_cast_fp16, y = rotr_85_cast_fp16)[name = tensor("op_5446_cast_fp16")]; + tensor var_5447_cast_fp16 = mul(x = xi_87_cast_fp16, y = roti_85_cast_fp16)[name = tensor("op_5447_cast_fp16")]; + tensor xor__87_cast_fp16 = sub(x = var_5446_cast_fp16, y = var_5447_cast_fp16)[name = tensor("xor__87_cast_fp16")]; + tensor var_5450_cast_fp16 = mul(x = xr_87_cast_fp16, y = roti_85_cast_fp16)[name = tensor("op_5450_cast_fp16")]; + tensor var_5451_cast_fp16 = mul(x = xi_87_cast_fp16, y = rotr_85_cast_fp16)[name = tensor("op_5451_cast_fp16")]; + tensor xoi_87_cast_fp16 = add(x = var_5450_cast_fp16, y = var_5451_cast_fp16)[name = tensor("xoi_87_cast_fp16")]; + tensor out_87_axis_0 = const()[name = tensor("out_87_axis_0"), val = tensor(-1)]; + tensor out_87_cast_fp16 = stack(axis = out_87_axis_0, values = (xor__87_cast_fp16, xoi_87_cast_fp16))[name = tensor("out_87_cast_fp16")]; + tensor _inversed_5463_y_0_to_fp16 = const()[name = tensor("_inversed_5463_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5463_cast_fp16 = mul(x = position21_to_fp16, y = _inversed_5463_y_0_to_fp16)[name = tensor("_inversed_5463_cast_fp16")]; + tensor var_5464_cast_fp16 = floor(x = _inversed_5463_cast_fp16)[name = tensor("op_5464_cast_fp16")]; + tensor var_5465_to_fp16 = const()[name = tensor("op_5465_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5466_cast_fp16 = mul(x = var_5464_cast_fp16, y = var_5465_to_fp16)[name = tensor("op_5466_cast_fp16")]; + tensor write_idx_43_cast_fp16 = sub(x = position21_to_fp16, y = var_5466_cast_fp16)[name = tensor("write_idx_43_cast_fp16")]; + tensor var_5469_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_43_cast_fp16)[name = tensor("op_5469_cast_fp16")]; + tensor var_5479 = const()[name = tensor("op_5479"), val = tensor([1, 512, 1, 1])]; + tensor onehot_85_to_fp16_dtype_0 = const()[name = tensor("onehot_85_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_5469_cast_fp16_to_fp16 = cast(dtype = onehot_85_to_fp16_dtype_0, x = var_5469_cast_fp16)[name = tensor("cast_408")]; + tensor onehot_87_cast_fp16 = reshape(shape = var_5479, x = var_5469_cast_fp16_to_fp16)[name = tensor("onehot_87_cast_fp16")]; + tensor var_5481_to_fp16 = const()[name = tensor("op_5481_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5483_cast_fp16 = sub(x = var_5481_to_fp16, y = onehot_87_cast_fp16)[name = tensor("op_5483_cast_fp16")]; + tensor k_cache21_to_fp16_dtype_0 = const()[name = tensor("k_cache21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache21_to_fp16 = cast(dtype = k_cache21_to_fp16_dtype_0, x = k_cache21)[name = tensor("cast_407")]; + tensor var_5484_cast_fp16 = mul(x = k_cache21_to_fp16, y = var_5483_cast_fp16)[name = tensor("op_5484_cast_fp16")]; + tensor var_5489 = const()[name = tensor("op_5489"), val = tensor([1, 1, 16, 64])]; + tensor var_5490_cast_fp16 = reshape(shape = var_5489, x = out_87_cast_fp16)[name = tensor("op_5490_cast_fp16")]; + tensor var_5491_cast_fp16 = mul(x = var_5490_cast_fp16, y = onehot_87_cast_fp16)[name = tensor("op_5491_cast_fp16")]; + tensor new_k_cache_43_cast_fp16 = add(x = var_5484_cast_fp16, y = var_5491_cast_fp16)[name = tensor("new_k_cache_43_cast_fp16")]; + tensor new_k_cache_43_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_43_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache21_to_fp16_dtype_0 = const()[name = tensor("v_cache21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache21_to_fp16 = cast(dtype = v_cache21_to_fp16_dtype_0, x = v_cache21)[name = tensor("cast_405")]; + tensor var_5497_cast_fp16 = mul(x = v_cache21_to_fp16, y = var_5483_cast_fp16)[name = tensor("op_5497_cast_fp16")]; + tensor var_5502 = const()[name = tensor("op_5502"), val = tensor([1, 1, 16, 64])]; + tensor var_5503_cast_fp16 = reshape(shape = var_5502, x = v_43_cast_fp16)[name = tensor("op_5503_cast_fp16")]; + tensor var_5504_cast_fp16 = mul(x = var_5503_cast_fp16, y = onehot_87_cast_fp16)[name = tensor("op_5504_cast_fp16")]; + tensor new_v_cache_43_cast_fp16 = add(x = var_5497_cast_fp16, y = var_5504_cast_fp16)[name = tensor("new_v_cache_43_cast_fp16")]; + tensor new_v_cache_43_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_43_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_21_perm_0 = const()[name = tensor("transpose_21_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_5517 = const()[name = tensor("op_5517"), val = tensor([0, 2, 1, 3])]; + tensor var_5523 = const()[name = tensor("op_5523"), val = tensor([1, 16, 1, 64])]; + tensor q4_43_cast_fp16 = reshape(shape = var_5523, x = out_85_cast_fp16)[name = tensor("q4_43_cast_fp16")]; + tensor var_5528_transpose_x_0 = const()[name = tensor("op_5528_transpose_x_0"), val = tensor(false)]; + tensor var_5528_transpose_y_0 = const()[name = tensor("op_5528_transpose_y_0"), val = tensor(false)]; + tensor transpose_21_cast_fp16 = transpose(perm = transpose_21_perm_0, x = new_k_cache_43_cast_fp16)[name = tensor("transpose_32")]; + tensor var_5528_cast_fp16 = matmul(transpose_x = var_5528_transpose_x_0, transpose_y = var_5528_transpose_y_0, x = q4_43_cast_fp16, y = transpose_21_cast_fp16)[name = tensor("op_5528_cast_fp16")]; + tensor var_5529_to_fp16 = const()[name = tensor("op_5529_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_85_cast_fp16 = mul(x = var_5528_cast_fp16, y = var_5529_to_fp16)[name = tensor("scores_85_cast_fp16")]; + tensor var_5533 = const()[name = tensor("op_5533"), val = tensor([1, 1])]; + tensor var_5534_cast_fp16 = reshape(shape = var_5533, x = position21_to_fp16)[name = tensor("op_5534_cast_fp16")]; + tensor var_5535_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_5534_cast_fp16)[name = tensor("op_5535_cast_fp16")]; + tensor mask_43_to_fp16_dtype_0 = const()[name = tensor("mask_43_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_5542_to_fp16 = const()[name = tensor("op_5542_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5535_cast_fp16_to_fp16 = cast(dtype = mask_43_to_fp16_dtype_0, x = var_5535_cast_fp16)[name = tensor("cast_403")]; + tensor var_5543_cast_fp16 = sub(x = var_5535_cast_fp16_to_fp16, y = var_5542_to_fp16)[name = tensor("op_5543_cast_fp16")]; + tensor var_5548 = const()[name = tensor("op_5548"), val = tensor([1, 1, 1, 512])]; + tensor var_5549_cast_fp16 = reshape(shape = var_5548, x = var_5543_cast_fp16)[name = tensor("op_5549_cast_fp16")]; + tensor var_5550_to_fp16 = const()[name = tensor("op_5550_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_5551_cast_fp16 = mul(x = var_5549_cast_fp16, y = var_5550_to_fp16)[name = tensor("op_5551_cast_fp16")]; + tensor scores_87_cast_fp16 = add(x = scores_85_cast_fp16, y = var_5551_cast_fp16)[name = tensor("scores_87_cast_fp16")]; + tensor var_5554 = const()[name = tensor("op_5554"), val = tensor(-1)]; + tensor weights_43_cast_fp16 = softmax(axis = var_5554, x = scores_87_cast_fp16)[name = tensor("weights_43_cast_fp16")]; + tensor attn_43_transpose_x_0 = const()[name = tensor("attn_43_transpose_x_0"), val = tensor(false)]; + tensor attn_43_transpose_y_0 = const()[name = tensor("attn_43_transpose_y_0"), val = tensor(false)]; + tensor values_43_cast_fp16 = transpose(perm = var_5517, x = new_v_cache_43_cast_fp16)[name = tensor("transpose_31")]; + tensor attn_43_cast_fp16 = matmul(transpose_x = attn_43_transpose_x_0, transpose_y = attn_43_transpose_y_0, x = weights_43_cast_fp16, y = values_43_cast_fp16)[name = tensor("attn_43_cast_fp16")]; + tensor var_5562 = const()[name = tensor("op_5562"), val = tensor([0, 2, 1, 3])]; + tensor var_5567 = const()[name = tensor("op_5567"), val = tensor([1, 1, 1024])]; + tensor var_5563_cast_fp16 = transpose(perm = var_5562, x = attn_43_cast_fp16)[name = tensor("transpose_30")]; + tensor input_259_cast_fp16 = reshape(shape = var_5567, x = var_5563_cast_fp16)[name = tensor("input_259_cast_fp16")]; + tensor attn21_out_proj_weight_to_fp16 = const()[name = tensor("attn21_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(535044288)))]; + tensor linear_86_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn21_out_proj_weight_to_fp16, x = input_259_cast_fp16)[name = tensor("linear_86_cast_fp16")]; + tensor var_5573_to_fp16 = const()[name = tensor("op_5573_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5574_cast_fp16 = add(x = position21_to_fp16, y = var_5573_to_fp16)[name = tensor("op_5574_cast_fp16")]; + tensor var_5574_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5574_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_261_cast_fp16 = add(x = input_255_cast_fp16, y = linear_86_cast_fp16)[name = tensor("input_261_cast_fp16")]; + tensor input_263_axes_0 = const()[name = tensor("input_263_axes_0"), val = tensor([-1])]; + tensor norm21_2_weight_to_fp16 = const()[name = tensor("norm21_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(537141504)))]; + tensor norm21_2_bias_to_fp16 = const()[name = tensor("norm21_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(537143616)))]; + tensor var_5578_to_fp16 = const()[name = tensor("op_5578_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_263_cast_fp16 = layer_norm(axes = input_263_axes_0, beta = norm21_2_bias_to_fp16, epsilon = var_5578_to_fp16, gamma = norm21_2_weight_to_fp16, x = input_261_cast_fp16)[name = tensor("input_263_cast_fp16")]; + tensor linear21_1_weight_to_fp16 = const()[name = tensor("linear21_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(537145728)))]; + tensor linear_87_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear21_1_weight_to_fp16, x = input_263_cast_fp16)[name = tensor("linear_87_cast_fp16")]; + tensor input_265_mode_0 = const()[name = tensor("input_265_mode_0"), val = tensor("EXACT")]; + tensor input_265_cast_fp16 = gelu(mode = input_265_mode_0, x = linear_87_cast_fp16)[name = tensor("input_265_cast_fp16")]; + tensor linear21_2_weight_to_fp16 = const()[name = tensor("linear21_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(545534400)))]; + tensor linear_88_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear21_2_weight_to_fp16, x = input_265_cast_fp16)[name = tensor("linear_88_cast_fp16")]; + tensor input_267_cast_fp16 = add(x = input_261_cast_fp16, y = linear_88_cast_fp16)[name = tensor("input_267_cast_fp16")]; + tensor input_269_axes_0 = const()[name = tensor("input_269_axes_0"), val = tensor([-1])]; + tensor norm22_1_weight_to_fp16 = const()[name = tensor("norm22_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(553923072)))]; + tensor norm22_1_bias_to_fp16 = const()[name = tensor("norm22_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(553925184)))]; + tensor var_5595_to_fp16 = const()[name = tensor("op_5595_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_269_cast_fp16 = layer_norm(axes = input_269_axes_0, beta = norm22_1_bias_to_fp16, epsilon = var_5595_to_fp16, gamma = norm22_1_weight_to_fp16, x = input_267_cast_fp16)[name = tensor("input_269_cast_fp16")]; + tensor attn22_in_proj_weight_to_fp16 = const()[name = tensor("attn22_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(553927296)))]; + tensor linear_89_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn22_in_proj_weight_to_fp16, x = input_269_cast_fp16)[name = tensor("linear_89_cast_fp16")]; + tensor var_5608 = const()[name = tensor("op_5608"), val = tensor([1, 3, 16, 64])]; + tensor qkv_45_cast_fp16 = reshape(shape = var_5608, x = linear_89_cast_fp16)[name = tensor("qkv_45_cast_fp16")]; + tensor x_89_begin_0 = const()[name = tensor("x_89_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_89_end_0 = const()[name = tensor("x_89_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_89_end_mask_0 = const()[name = tensor("x_89_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_89_squeeze_mask_0 = const()[name = tensor("x_89_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_89_cast_fp16 = slice_by_index(begin = x_89_begin_0, end = x_89_end_0, end_mask = x_89_end_mask_0, squeeze_mask = x_89_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("x_89_cast_fp16")]; + tensor x_91_begin_0 = const()[name = tensor("x_91_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_91_end_0 = const()[name = tensor("x_91_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_91_end_mask_0 = const()[name = tensor("x_91_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_91_squeeze_mask_0 = const()[name = tensor("x_91_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_91_cast_fp16 = slice_by_index(begin = x_91_begin_0, end = x_91_end_0, end_mask = x_91_end_mask_0, squeeze_mask = x_91_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("x_91_cast_fp16")]; + tensor v_45_begin_0 = const()[name = tensor("v_45_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_45_end_0 = const()[name = tensor("v_45_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_45_end_mask_0 = const()[name = tensor("v_45_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_45_squeeze_mask_0 = const()[name = tensor("v_45_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_45_cast_fp16 = slice_by_index(begin = v_45_begin_0, end = v_45_end_0, end_mask = v_45_end_mask_0, squeeze_mask = v_45_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("v_45_cast_fp16")]; + tensor var_5637 = const()[name = tensor("op_5637"), val = tensor([1, 1, 1])]; + tensor position22_to_fp16_dtype_0 = const()[name = tensor("position22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position22_to_fp16 = cast(dtype = position22_to_fp16_dtype_0, x = position22)[name = tensor("cast_401")]; + tensor var_5638_cast_fp16 = reshape(shape = var_5637, x = position22_to_fp16)[name = tensor("op_5638_cast_fp16")]; + tensor angles_89_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_5638_cast_fp16)[name = tensor("angles_89_cast_fp16")]; + tensor rotr_89_cast_fp16 = cos(x = angles_89_cast_fp16)[name = tensor("rotr_89_cast_fp16")]; + tensor roti_89_cast_fp16 = sin(x = angles_89_cast_fp16)[name = tensor("roti_89_cast_fp16")]; + tensor var_5646 = const()[name = tensor("op_5646"), val = tensor([1, 16, 32, 2])]; + tensor pairs_89_cast_fp16 = reshape(shape = var_5646, x = x_89_cast_fp16)[name = tensor("pairs_89_cast_fp16")]; + tensor xr_89_begin_0 = const()[name = tensor("xr_89_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_89_end_0 = const()[name = tensor("xr_89_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_89_end_mask_0 = const()[name = tensor("xr_89_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_89_squeeze_mask_0 = const()[name = tensor("xr_89_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_89_cast_fp16 = slice_by_index(begin = xr_89_begin_0, end = xr_89_end_0, end_mask = xr_89_end_mask_0, squeeze_mask = xr_89_squeeze_mask_0, x = pairs_89_cast_fp16)[name = tensor("xr_89_cast_fp16")]; + tensor xi_89_begin_0 = const()[name = tensor("xi_89_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_89_end_0 = const()[name = tensor("xi_89_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_89_end_mask_0 = const()[name = tensor("xi_89_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_89_squeeze_mask_0 = const()[name = tensor("xi_89_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_89_cast_fp16 = slice_by_index(begin = xi_89_begin_0, end = xi_89_end_0, end_mask = xi_89_end_mask_0, squeeze_mask = xi_89_squeeze_mask_0, x = pairs_89_cast_fp16)[name = tensor("xi_89_cast_fp16")]; + tensor var_5654_cast_fp16 = mul(x = xr_89_cast_fp16, y = rotr_89_cast_fp16)[name = tensor("op_5654_cast_fp16")]; + tensor var_5655_cast_fp16 = mul(x = xi_89_cast_fp16, y = roti_89_cast_fp16)[name = tensor("op_5655_cast_fp16")]; + tensor xor__89_cast_fp16 = sub(x = var_5654_cast_fp16, y = var_5655_cast_fp16)[name = tensor("xor__89_cast_fp16")]; + tensor var_5658_cast_fp16 = mul(x = xr_89_cast_fp16, y = roti_89_cast_fp16)[name = tensor("op_5658_cast_fp16")]; + tensor var_5659_cast_fp16 = mul(x = xi_89_cast_fp16, y = rotr_89_cast_fp16)[name = tensor("op_5659_cast_fp16")]; + tensor xoi_89_cast_fp16 = add(x = var_5658_cast_fp16, y = var_5659_cast_fp16)[name = tensor("xoi_89_cast_fp16")]; + tensor out_89_axis_0 = const()[name = tensor("out_89_axis_0"), val = tensor(-1)]; + tensor out_89_cast_fp16 = stack(axis = out_89_axis_0, values = (xor__89_cast_fp16, xoi_89_cast_fp16))[name = tensor("out_89_cast_fp16")]; + tensor var_5682 = const()[name = tensor("op_5682"), val = tensor([1, 16, 32, 2])]; + tensor pairs_91_cast_fp16 = reshape(shape = var_5682, x = x_91_cast_fp16)[name = tensor("pairs_91_cast_fp16")]; + tensor xr_91_begin_0 = const()[name = tensor("xr_91_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_91_end_0 = const()[name = tensor("xr_91_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_91_end_mask_0 = const()[name = tensor("xr_91_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_91_squeeze_mask_0 = const()[name = tensor("xr_91_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_91_cast_fp16 = slice_by_index(begin = xr_91_begin_0, end = xr_91_end_0, end_mask = xr_91_end_mask_0, squeeze_mask = xr_91_squeeze_mask_0, x = pairs_91_cast_fp16)[name = tensor("xr_91_cast_fp16")]; + tensor xi_91_begin_0 = const()[name = tensor("xi_91_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_91_end_0 = const()[name = tensor("xi_91_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_91_end_mask_0 = const()[name = tensor("xi_91_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_91_squeeze_mask_0 = const()[name = tensor("xi_91_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_91_cast_fp16 = slice_by_index(begin = xi_91_begin_0, end = xi_91_end_0, end_mask = xi_91_end_mask_0, squeeze_mask = xi_91_squeeze_mask_0, x = pairs_91_cast_fp16)[name = tensor("xi_91_cast_fp16")]; + tensor var_5690_cast_fp16 = mul(x = xr_91_cast_fp16, y = rotr_89_cast_fp16)[name = tensor("op_5690_cast_fp16")]; + tensor var_5691_cast_fp16 = mul(x = xi_91_cast_fp16, y = roti_89_cast_fp16)[name = tensor("op_5691_cast_fp16")]; + tensor xor__91_cast_fp16 = sub(x = var_5690_cast_fp16, y = var_5691_cast_fp16)[name = tensor("xor__91_cast_fp16")]; + tensor var_5694_cast_fp16 = mul(x = xr_91_cast_fp16, y = roti_89_cast_fp16)[name = tensor("op_5694_cast_fp16")]; + tensor var_5695_cast_fp16 = mul(x = xi_91_cast_fp16, y = rotr_89_cast_fp16)[name = tensor("op_5695_cast_fp16")]; + tensor xoi_91_cast_fp16 = add(x = var_5694_cast_fp16, y = var_5695_cast_fp16)[name = tensor("xoi_91_cast_fp16")]; + tensor out_91_axis_0 = const()[name = tensor("out_91_axis_0"), val = tensor(-1)]; + tensor out_91_cast_fp16 = stack(axis = out_91_axis_0, values = (xor__91_cast_fp16, xoi_91_cast_fp16))[name = tensor("out_91_cast_fp16")]; + tensor _inversed_5707_y_0_to_fp16 = const()[name = tensor("_inversed_5707_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5707_cast_fp16 = mul(x = position22_to_fp16, y = _inversed_5707_y_0_to_fp16)[name = tensor("_inversed_5707_cast_fp16")]; + tensor var_5708_cast_fp16 = floor(x = _inversed_5707_cast_fp16)[name = tensor("op_5708_cast_fp16")]; + tensor var_5709_to_fp16 = const()[name = tensor("op_5709_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5710_cast_fp16 = mul(x = var_5708_cast_fp16, y = var_5709_to_fp16)[name = tensor("op_5710_cast_fp16")]; + tensor write_idx_45_cast_fp16 = sub(x = position22_to_fp16, y = var_5710_cast_fp16)[name = tensor("write_idx_45_cast_fp16")]; + tensor var_5713_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_45_cast_fp16)[name = tensor("op_5713_cast_fp16")]; + tensor var_5723 = const()[name = tensor("op_5723"), val = tensor([1, 512, 1, 1])]; + tensor onehot_89_to_fp16_dtype_0 = const()[name = tensor("onehot_89_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_5713_cast_fp16_to_fp16 = cast(dtype = onehot_89_to_fp16_dtype_0, x = var_5713_cast_fp16)[name = tensor("cast_400")]; + tensor onehot_91_cast_fp16 = reshape(shape = var_5723, x = var_5713_cast_fp16_to_fp16)[name = tensor("onehot_91_cast_fp16")]; + tensor var_5725_to_fp16 = const()[name = tensor("op_5725_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5727_cast_fp16 = sub(x = var_5725_to_fp16, y = onehot_91_cast_fp16)[name = tensor("op_5727_cast_fp16")]; + tensor k_cache22_to_fp16_dtype_0 = const()[name = tensor("k_cache22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache22_to_fp16 = cast(dtype = k_cache22_to_fp16_dtype_0, x = k_cache22)[name = tensor("cast_399")]; + tensor var_5728_cast_fp16 = mul(x = k_cache22_to_fp16, y = var_5727_cast_fp16)[name = tensor("op_5728_cast_fp16")]; + tensor var_5733 = const()[name = tensor("op_5733"), val = tensor([1, 1, 16, 64])]; + tensor var_5734_cast_fp16 = reshape(shape = var_5733, x = out_91_cast_fp16)[name = tensor("op_5734_cast_fp16")]; + tensor var_5735_cast_fp16 = mul(x = var_5734_cast_fp16, y = onehot_91_cast_fp16)[name = tensor("op_5735_cast_fp16")]; + tensor new_k_cache_45_cast_fp16 = add(x = var_5728_cast_fp16, y = var_5735_cast_fp16)[name = tensor("new_k_cache_45_cast_fp16")]; + tensor new_k_cache_45_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_45_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache22_to_fp16_dtype_0 = const()[name = tensor("v_cache22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache22_to_fp16 = cast(dtype = v_cache22_to_fp16_dtype_0, x = v_cache22)[name = tensor("cast_397")]; + tensor var_5741_cast_fp16 = mul(x = v_cache22_to_fp16, y = var_5727_cast_fp16)[name = tensor("op_5741_cast_fp16")]; + tensor var_5746 = const()[name = tensor("op_5746"), val = tensor([1, 1, 16, 64])]; + tensor var_5747_cast_fp16 = reshape(shape = var_5746, x = v_45_cast_fp16)[name = tensor("op_5747_cast_fp16")]; + tensor var_5748_cast_fp16 = mul(x = var_5747_cast_fp16, y = onehot_91_cast_fp16)[name = tensor("op_5748_cast_fp16")]; + tensor new_v_cache_45_cast_fp16 = add(x = var_5741_cast_fp16, y = var_5748_cast_fp16)[name = tensor("new_v_cache_45_cast_fp16")]; + tensor new_v_cache_45_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_45_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_22_perm_0 = const()[name = tensor("transpose_22_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_5761 = const()[name = tensor("op_5761"), val = tensor([0, 2, 1, 3])]; + tensor var_5767 = const()[name = tensor("op_5767"), val = tensor([1, 16, 1, 64])]; + tensor q4_45_cast_fp16 = reshape(shape = var_5767, x = out_89_cast_fp16)[name = tensor("q4_45_cast_fp16")]; + tensor var_5772_transpose_x_0 = const()[name = tensor("op_5772_transpose_x_0"), val = tensor(false)]; + tensor var_5772_transpose_y_0 = const()[name = tensor("op_5772_transpose_y_0"), val = tensor(false)]; + tensor transpose_22_cast_fp16 = transpose(perm = transpose_22_perm_0, x = new_k_cache_45_cast_fp16)[name = tensor("transpose_29")]; + tensor var_5772_cast_fp16 = matmul(transpose_x = var_5772_transpose_x_0, transpose_y = var_5772_transpose_y_0, x = q4_45_cast_fp16, y = transpose_22_cast_fp16)[name = tensor("op_5772_cast_fp16")]; + tensor var_5773_to_fp16 = const()[name = tensor("op_5773_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_89_cast_fp16 = mul(x = var_5772_cast_fp16, y = var_5773_to_fp16)[name = tensor("scores_89_cast_fp16")]; + tensor var_5777 = const()[name = tensor("op_5777"), val = tensor([1, 1])]; + tensor var_5778_cast_fp16 = reshape(shape = var_5777, x = position22_to_fp16)[name = tensor("op_5778_cast_fp16")]; + tensor var_5779_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_5778_cast_fp16)[name = tensor("op_5779_cast_fp16")]; + tensor mask_45_to_fp16_dtype_0 = const()[name = tensor("mask_45_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_5786_to_fp16 = const()[name = tensor("op_5786_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5779_cast_fp16_to_fp16 = cast(dtype = mask_45_to_fp16_dtype_0, x = var_5779_cast_fp16)[name = tensor("cast_395")]; + tensor var_5787_cast_fp16 = sub(x = var_5779_cast_fp16_to_fp16, y = var_5786_to_fp16)[name = tensor("op_5787_cast_fp16")]; + tensor var_5792 = const()[name = tensor("op_5792"), val = tensor([1, 1, 1, 512])]; + tensor var_5793_cast_fp16 = reshape(shape = var_5792, x = var_5787_cast_fp16)[name = tensor("op_5793_cast_fp16")]; + tensor var_5794_to_fp16 = const()[name = tensor("op_5794_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_5795_cast_fp16 = mul(x = var_5793_cast_fp16, y = var_5794_to_fp16)[name = tensor("op_5795_cast_fp16")]; + tensor scores_91_cast_fp16 = add(x = scores_89_cast_fp16, y = var_5795_cast_fp16)[name = tensor("scores_91_cast_fp16")]; + tensor var_5798 = const()[name = tensor("op_5798"), val = tensor(-1)]; + tensor weights_45_cast_fp16 = softmax(axis = var_5798, x = scores_91_cast_fp16)[name = tensor("weights_45_cast_fp16")]; + tensor attn_45_transpose_x_0 = const()[name = tensor("attn_45_transpose_x_0"), val = tensor(false)]; + tensor attn_45_transpose_y_0 = const()[name = tensor("attn_45_transpose_y_0"), val = tensor(false)]; + tensor values_45_cast_fp16 = transpose(perm = var_5761, x = new_v_cache_45_cast_fp16)[name = tensor("transpose_28")]; + tensor attn_45_cast_fp16 = matmul(transpose_x = attn_45_transpose_x_0, transpose_y = attn_45_transpose_y_0, x = weights_45_cast_fp16, y = values_45_cast_fp16)[name = tensor("attn_45_cast_fp16")]; + tensor var_5806 = const()[name = tensor("op_5806"), val = tensor([0, 2, 1, 3])]; + tensor var_5811 = const()[name = tensor("op_5811"), val = tensor([1, 1, 1024])]; + tensor var_5807_cast_fp16 = transpose(perm = var_5806, x = attn_45_cast_fp16)[name = tensor("transpose_27")]; + tensor input_271_cast_fp16 = reshape(shape = var_5811, x = var_5807_cast_fp16)[name = tensor("input_271_cast_fp16")]; + tensor attn22_out_proj_weight_to_fp16 = const()[name = tensor("attn22_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(560218816)))]; + tensor linear_90_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn22_out_proj_weight_to_fp16, x = input_271_cast_fp16)[name = tensor("linear_90_cast_fp16")]; + tensor var_5817_to_fp16 = const()[name = tensor("op_5817_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5818_cast_fp16 = add(x = position22_to_fp16, y = var_5817_to_fp16)[name = tensor("op_5818_cast_fp16")]; + tensor var_5818_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5818_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_273_cast_fp16 = add(x = input_267_cast_fp16, y = linear_90_cast_fp16)[name = tensor("input_273_cast_fp16")]; + tensor input_275_axes_0 = const()[name = tensor("input_275_axes_0"), val = tensor([-1])]; + tensor norm22_2_weight_to_fp16 = const()[name = tensor("norm22_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(562316032)))]; + tensor norm22_2_bias_to_fp16 = const()[name = tensor("norm22_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(562318144)))]; + tensor var_5822_to_fp16 = const()[name = tensor("op_5822_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_275_cast_fp16 = layer_norm(axes = input_275_axes_0, beta = norm22_2_bias_to_fp16, epsilon = var_5822_to_fp16, gamma = norm22_2_weight_to_fp16, x = input_273_cast_fp16)[name = tensor("input_275_cast_fp16")]; + tensor linear22_1_weight_to_fp16 = const()[name = tensor("linear22_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(562320256)))]; + tensor linear_91_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear22_1_weight_to_fp16, x = input_275_cast_fp16)[name = tensor("linear_91_cast_fp16")]; + tensor input_277_mode_0 = const()[name = tensor("input_277_mode_0"), val = tensor("EXACT")]; + tensor input_277_cast_fp16 = gelu(mode = input_277_mode_0, x = linear_91_cast_fp16)[name = tensor("input_277_cast_fp16")]; + tensor linear22_2_weight_to_fp16 = const()[name = tensor("linear22_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(570708928)))]; + tensor linear_92_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear22_2_weight_to_fp16, x = input_277_cast_fp16)[name = tensor("linear_92_cast_fp16")]; + tensor input_279_cast_fp16 = add(x = input_273_cast_fp16, y = linear_92_cast_fp16)[name = tensor("input_279_cast_fp16")]; + tensor input_281_axes_0 = const()[name = tensor("input_281_axes_0"), val = tensor([-1])]; + tensor norm23_1_weight_to_fp16 = const()[name = tensor("norm23_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(579097600)))]; + tensor norm23_1_bias_to_fp16 = const()[name = tensor("norm23_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(579099712)))]; + tensor var_5839_to_fp16 = const()[name = tensor("op_5839_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_281_cast_fp16 = layer_norm(axes = input_281_axes_0, beta = norm23_1_bias_to_fp16, epsilon = var_5839_to_fp16, gamma = norm23_1_weight_to_fp16, x = input_279_cast_fp16)[name = tensor("input_281_cast_fp16")]; + tensor attn23_in_proj_weight_to_fp16 = const()[name = tensor("attn23_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(579101824)))]; + tensor linear_93_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn23_in_proj_weight_to_fp16, x = input_281_cast_fp16)[name = tensor("linear_93_cast_fp16")]; + tensor var_5852 = const()[name = tensor("op_5852"), val = tensor([1, 3, 16, 64])]; + tensor qkv_cast_fp16 = reshape(shape = var_5852, x = linear_93_cast_fp16)[name = tensor("qkv_cast_fp16")]; + tensor x_93_begin_0 = const()[name = tensor("x_93_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_93_end_0 = const()[name = tensor("x_93_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_93_end_mask_0 = const()[name = tensor("x_93_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_93_squeeze_mask_0 = const()[name = tensor("x_93_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_93_cast_fp16 = slice_by_index(begin = x_93_begin_0, end = x_93_end_0, end_mask = x_93_end_mask_0, squeeze_mask = x_93_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("x_93_cast_fp16")]; + tensor x_begin_0 = const()[name = tensor("x_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_end_0 = const()[name = tensor("x_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_end_mask_0 = const()[name = tensor("x_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_squeeze_mask_0 = const()[name = tensor("x_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_cast_fp16 = slice_by_index(begin = x_begin_0, end = x_end_0, end_mask = x_end_mask_0, squeeze_mask = x_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("x_cast_fp16")]; + tensor v_begin_0 = const()[name = tensor("v_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_end_0 = const()[name = tensor("v_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_end_mask_0 = const()[name = tensor("v_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_squeeze_mask_0 = const()[name = tensor("v_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_cast_fp16 = slice_by_index(begin = v_begin_0, end = v_end_0, end_mask = v_end_mask_0, squeeze_mask = v_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("v_cast_fp16")]; + tensor var_5881 = const()[name = tensor("op_5881"), val = tensor([1, 1, 1])]; + tensor position23_to_fp16_dtype_0 = const()[name = tensor("position23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position23_to_fp16 = cast(dtype = position23_to_fp16_dtype_0, x = position23)[name = tensor("cast_393")]; + tensor var_5882_cast_fp16 = reshape(shape = var_5881, x = position23_to_fp16)[name = tensor("op_5882_cast_fp16")]; + tensor angles_93_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_5882_cast_fp16)[name = tensor("angles_93_cast_fp16")]; + tensor rotr_93_cast_fp16 = cos(x = angles_93_cast_fp16)[name = tensor("rotr_93_cast_fp16")]; + tensor roti_93_cast_fp16 = sin(x = angles_93_cast_fp16)[name = tensor("roti_93_cast_fp16")]; + tensor var_5890 = const()[name = tensor("op_5890"), val = tensor([1, 16, 32, 2])]; + tensor pairs_93_cast_fp16 = reshape(shape = var_5890, x = x_93_cast_fp16)[name = tensor("pairs_93_cast_fp16")]; + tensor xr_93_begin_0 = const()[name = tensor("xr_93_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_93_end_0 = const()[name = tensor("xr_93_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_93_end_mask_0 = const()[name = tensor("xr_93_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_93_squeeze_mask_0 = const()[name = tensor("xr_93_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_93_cast_fp16 = slice_by_index(begin = xr_93_begin_0, end = xr_93_end_0, end_mask = xr_93_end_mask_0, squeeze_mask = xr_93_squeeze_mask_0, x = pairs_93_cast_fp16)[name = tensor("xr_93_cast_fp16")]; + tensor xi_93_begin_0 = const()[name = tensor("xi_93_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_93_end_0 = const()[name = tensor("xi_93_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_93_end_mask_0 = const()[name = tensor("xi_93_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_93_squeeze_mask_0 = const()[name = tensor("xi_93_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_93_cast_fp16 = slice_by_index(begin = xi_93_begin_0, end = xi_93_end_0, end_mask = xi_93_end_mask_0, squeeze_mask = xi_93_squeeze_mask_0, x = pairs_93_cast_fp16)[name = tensor("xi_93_cast_fp16")]; + tensor var_5898_cast_fp16 = mul(x = xr_93_cast_fp16, y = rotr_93_cast_fp16)[name = tensor("op_5898_cast_fp16")]; + tensor var_5899_cast_fp16 = mul(x = xi_93_cast_fp16, y = roti_93_cast_fp16)[name = tensor("op_5899_cast_fp16")]; + tensor xor__93_cast_fp16 = sub(x = var_5898_cast_fp16, y = var_5899_cast_fp16)[name = tensor("xor__93_cast_fp16")]; + tensor var_5902_cast_fp16 = mul(x = xr_93_cast_fp16, y = roti_93_cast_fp16)[name = tensor("op_5902_cast_fp16")]; + tensor var_5903_cast_fp16 = mul(x = xi_93_cast_fp16, y = rotr_93_cast_fp16)[name = tensor("op_5903_cast_fp16")]; + tensor xoi_93_cast_fp16 = add(x = var_5902_cast_fp16, y = var_5903_cast_fp16)[name = tensor("xoi_93_cast_fp16")]; + tensor out_93_axis_0 = const()[name = tensor("out_93_axis_0"), val = tensor(-1)]; + tensor out_93_cast_fp16 = stack(axis = out_93_axis_0, values = (xor__93_cast_fp16, xoi_93_cast_fp16))[name = tensor("out_93_cast_fp16")]; + tensor var_5926 = const()[name = tensor("op_5926"), val = tensor([1, 16, 32, 2])]; + tensor pairs_cast_fp16 = reshape(shape = var_5926, x = x_cast_fp16)[name = tensor("pairs_cast_fp16")]; + tensor xr_begin_0 = const()[name = tensor("xr_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_end_0 = const()[name = tensor("xr_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_end_mask_0 = const()[name = tensor("xr_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_squeeze_mask_0 = const()[name = tensor("xr_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_cast_fp16 = slice_by_index(begin = xr_begin_0, end = xr_end_0, end_mask = xr_end_mask_0, squeeze_mask = xr_squeeze_mask_0, x = pairs_cast_fp16)[name = tensor("xr_cast_fp16")]; + tensor xi_begin_0 = const()[name = tensor("xi_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_end_0 = const()[name = tensor("xi_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_end_mask_0 = const()[name = tensor("xi_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_squeeze_mask_0 = const()[name = tensor("xi_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_cast_fp16 = slice_by_index(begin = xi_begin_0, end = xi_end_0, end_mask = xi_end_mask_0, squeeze_mask = xi_squeeze_mask_0, x = pairs_cast_fp16)[name = tensor("xi_cast_fp16")]; + tensor var_5934_cast_fp16 = mul(x = xr_cast_fp16, y = rotr_93_cast_fp16)[name = tensor("op_5934_cast_fp16")]; + tensor var_5935_cast_fp16 = mul(x = xi_cast_fp16, y = roti_93_cast_fp16)[name = tensor("op_5935_cast_fp16")]; + tensor xor__cast_fp16 = sub(x = var_5934_cast_fp16, y = var_5935_cast_fp16)[name = tensor("xor__cast_fp16")]; + tensor var_5938_cast_fp16 = mul(x = xr_cast_fp16, y = roti_93_cast_fp16)[name = tensor("op_5938_cast_fp16")]; + tensor var_5939_cast_fp16 = mul(x = xi_cast_fp16, y = rotr_93_cast_fp16)[name = tensor("op_5939_cast_fp16")]; + tensor xoi_cast_fp16 = add(x = var_5938_cast_fp16, y = var_5939_cast_fp16)[name = tensor("xoi_cast_fp16")]; + tensor out_axis_0 = const()[name = tensor("out_axis_0"), val = tensor(-1)]; + tensor out_cast_fp16 = stack(axis = out_axis_0, values = (xor__cast_fp16, xoi_cast_fp16))[name = tensor("out_cast_fp16")]; + tensor _inversed_5951_y_0_to_fp16 = const()[name = tensor("_inversed_5951_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5951_cast_fp16 = mul(x = position23_to_fp16, y = _inversed_5951_y_0_to_fp16)[name = tensor("_inversed_5951_cast_fp16")]; + tensor var_5952_cast_fp16 = floor(x = _inversed_5951_cast_fp16)[name = tensor("op_5952_cast_fp16")]; + tensor var_5953_to_fp16 = const()[name = tensor("op_5953_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5954_cast_fp16 = mul(x = var_5952_cast_fp16, y = var_5953_to_fp16)[name = tensor("op_5954_cast_fp16")]; + tensor write_idx_cast_fp16 = sub(x = position23_to_fp16, y = var_5954_cast_fp16)[name = tensor("write_idx_cast_fp16")]; + tensor var_5957_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_cast_fp16)[name = tensor("op_5957_cast_fp16")]; + tensor var_5967 = const()[name = tensor("op_5967"), val = tensor([1, 512, 1, 1])]; + tensor onehot_93_to_fp16_dtype_0 = const()[name = tensor("onehot_93_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_5957_cast_fp16_to_fp16 = cast(dtype = onehot_93_to_fp16_dtype_0, x = var_5957_cast_fp16)[name = tensor("cast_392")]; + tensor onehot_cast_fp16 = reshape(shape = var_5967, x = var_5957_cast_fp16_to_fp16)[name = tensor("onehot_cast_fp16")]; + tensor var_5969_to_fp16 = const()[name = tensor("op_5969_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5971_cast_fp16 = sub(x = var_5969_to_fp16, y = onehot_cast_fp16)[name = tensor("op_5971_cast_fp16")]; + tensor k_cache23_to_fp16_dtype_0 = const()[name = tensor("k_cache23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache23_to_fp16 = cast(dtype = k_cache23_to_fp16_dtype_0, x = k_cache23)[name = tensor("cast_391")]; + tensor var_5972_cast_fp16 = mul(x = k_cache23_to_fp16, y = var_5971_cast_fp16)[name = tensor("op_5972_cast_fp16")]; + tensor var_5977 = const()[name = tensor("op_5977"), val = tensor([1, 1, 16, 64])]; + tensor var_5978_cast_fp16 = reshape(shape = var_5977, x = out_cast_fp16)[name = tensor("op_5978_cast_fp16")]; + tensor var_5979_cast_fp16 = mul(x = var_5978_cast_fp16, y = onehot_cast_fp16)[name = tensor("op_5979_cast_fp16")]; + tensor new_k_cache_cast_fp16 = add(x = var_5972_cast_fp16, y = var_5979_cast_fp16)[name = tensor("new_k_cache_cast_fp16")]; + tensor new_k_cache_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache23_to_fp16_dtype_0 = const()[name = tensor("v_cache23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache23_to_fp16 = cast(dtype = v_cache23_to_fp16_dtype_0, x = v_cache23)[name = tensor("cast_389")]; + tensor var_5985_cast_fp16 = mul(x = v_cache23_to_fp16, y = var_5971_cast_fp16)[name = tensor("op_5985_cast_fp16")]; + tensor var_5990 = const()[name = tensor("op_5990"), val = tensor([1, 1, 16, 64])]; + tensor var_5991_cast_fp16 = reshape(shape = var_5990, x = v_cast_fp16)[name = tensor("op_5991_cast_fp16")]; + tensor var_5992_cast_fp16 = mul(x = var_5991_cast_fp16, y = onehot_cast_fp16)[name = tensor("op_5992_cast_fp16")]; + tensor new_v_cache_cast_fp16 = add(x = var_5985_cast_fp16, y = var_5992_cast_fp16)[name = tensor("new_v_cache_cast_fp16")]; + tensor new_v_cache_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_23_perm_0 = const()[name = tensor("transpose_23_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_6005 = const()[name = tensor("op_6005"), val = tensor([0, 2, 1, 3])]; + tensor var_6011 = const()[name = tensor("op_6011"), val = tensor([1, 16, 1, 64])]; + tensor q4_cast_fp16 = reshape(shape = var_6011, x = out_93_cast_fp16)[name = tensor("q4_cast_fp16")]; + tensor var_6016_transpose_x_0 = const()[name = tensor("op_6016_transpose_x_0"), val = tensor(false)]; + tensor var_6016_transpose_y_0 = const()[name = tensor("op_6016_transpose_y_0"), val = tensor(false)]; + tensor transpose_23_cast_fp16 = transpose(perm = transpose_23_perm_0, x = new_k_cache_cast_fp16)[name = tensor("transpose_26")]; + tensor var_6016_cast_fp16 = matmul(transpose_x = var_6016_transpose_x_0, transpose_y = var_6016_transpose_y_0, x = q4_cast_fp16, y = transpose_23_cast_fp16)[name = tensor("op_6016_cast_fp16")]; + tensor var_6017_to_fp16 = const()[name = tensor("op_6017_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_93_cast_fp16 = mul(x = var_6016_cast_fp16, y = var_6017_to_fp16)[name = tensor("scores_93_cast_fp16")]; + tensor var_6021 = const()[name = tensor("op_6021"), val = tensor([1, 1])]; + tensor var_6022_cast_fp16 = reshape(shape = var_6021, x = position23_to_fp16)[name = tensor("op_6022_cast_fp16")]; + tensor var_6023_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_6022_cast_fp16)[name = tensor("op_6023_cast_fp16")]; + tensor mask_to_fp16_dtype_0 = const()[name = tensor("mask_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_6030_to_fp16 = const()[name = tensor("op_6030_to_fp16"), val = tensor(0x1p+0)]; + tensor var_6023_cast_fp16_to_fp16 = cast(dtype = mask_to_fp16_dtype_0, x = var_6023_cast_fp16)[name = tensor("cast_387")]; + tensor var_6031_cast_fp16 = sub(x = var_6023_cast_fp16_to_fp16, y = var_6030_to_fp16)[name = tensor("op_6031_cast_fp16")]; + tensor var_6036 = const()[name = tensor("op_6036"), val = tensor([1, 1, 1, 512])]; + tensor var_6037_cast_fp16 = reshape(shape = var_6036, x = var_6031_cast_fp16)[name = tensor("op_6037_cast_fp16")]; + tensor var_6038_to_fp16 = const()[name = tensor("op_6038_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_6039_cast_fp16 = mul(x = var_6037_cast_fp16, y = var_6038_to_fp16)[name = tensor("op_6039_cast_fp16")]; + tensor scores_cast_fp16 = add(x = scores_93_cast_fp16, y = var_6039_cast_fp16)[name = tensor("scores_cast_fp16")]; + tensor var_6042 = const()[name = tensor("op_6042"), val = tensor(-1)]; + tensor weights_cast_fp16 = softmax(axis = var_6042, x = scores_cast_fp16)[name = tensor("weights_cast_fp16")]; + tensor attn_transpose_x_0 = const()[name = tensor("attn_transpose_x_0"), val = tensor(false)]; + tensor attn_transpose_y_0 = const()[name = tensor("attn_transpose_y_0"), val = tensor(false)]; + tensor values_cast_fp16 = transpose(perm = var_6005, x = new_v_cache_cast_fp16)[name = tensor("transpose_25")]; + tensor attn_cast_fp16 = matmul(transpose_x = attn_transpose_x_0, transpose_y = attn_transpose_y_0, x = weights_cast_fp16, y = values_cast_fp16)[name = tensor("attn_cast_fp16")]; + tensor var_6050 = const()[name = tensor("op_6050"), val = tensor([0, 2, 1, 3])]; + tensor var_6055 = const()[name = tensor("op_6055"), val = tensor([1, 1, 1024])]; + tensor var_6051_cast_fp16 = transpose(perm = var_6050, x = attn_cast_fp16)[name = tensor("transpose_24")]; + tensor input_283_cast_fp16 = reshape(shape = var_6055, x = var_6051_cast_fp16)[name = tensor("input_283_cast_fp16")]; + tensor attn23_out_proj_weight_to_fp16 = const()[name = tensor("attn23_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(585393344)))]; + tensor linear_94_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn23_out_proj_weight_to_fp16, x = input_283_cast_fp16)[name = tensor("linear_94_cast_fp16")]; + tensor var_6061_to_fp16 = const()[name = tensor("op_6061_to_fp16"), val = tensor(0x1p+0)]; + tensor var_6062_cast_fp16 = add(x = position23_to_fp16, y = var_6061_to_fp16)[name = tensor("op_6062_cast_fp16")]; + tensor var_6062_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_6062_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_285_cast_fp16 = add(x = input_279_cast_fp16, y = linear_94_cast_fp16)[name = tensor("input_285_cast_fp16")]; + tensor input_287_axes_0 = const()[name = tensor("input_287_axes_0"), val = tensor([-1])]; + tensor norm23_2_weight_to_fp16 = const()[name = tensor("norm23_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(587490560)))]; + tensor norm23_2_bias_to_fp16 = const()[name = tensor("norm23_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(587492672)))]; + tensor var_6066_to_fp16 = const()[name = tensor("op_6066_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_287_cast_fp16 = layer_norm(axes = input_287_axes_0, beta = norm23_2_bias_to_fp16, epsilon = var_6066_to_fp16, gamma = norm23_2_weight_to_fp16, x = input_285_cast_fp16)[name = tensor("input_287_cast_fp16")]; + tensor linear23_1_weight_to_fp16 = const()[name = tensor("linear23_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(587494784)))]; + tensor linear_95_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear23_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_95_cast_fp16")]; + tensor input_289_mode_0 = const()[name = tensor("input_289_mode_0"), val = tensor("EXACT")]; + tensor input_289_cast_fp16 = gelu(mode = input_289_mode_0, x = linear_95_cast_fp16)[name = tensor("input_289_cast_fp16")]; + tensor linear23_2_weight_to_fp16 = const()[name = tensor("linear23_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(595883456)))]; + tensor linear_96_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear23_2_weight_to_fp16, x = input_289_cast_fp16)[name = tensor("linear_96_cast_fp16")]; + tensor input_291_cast_fp16 = add(x = input_285_cast_fp16, y = linear_96_cast_fp16)[name = tensor("input_291_cast_fp16")]; + tensor input_axes_0 = const()[name = tensor("input_axes_0"), val = tensor([-1])]; + tensor out_norm_weight_to_fp16 = const()[name = tensor("out_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(604272128)))]; + tensor out_norm_bias_to_fp16 = const()[name = tensor("out_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(604274240)))]; + tensor var_6083_to_fp16 = const()[name = tensor("op_6083_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_cast_fp16 = layer_norm(axes = input_axes_0, beta = out_norm_bias_to_fp16, epsilon = var_6083_to_fp16, gamma = out_norm_weight_to_fp16, x = input_291_cast_fp16)[name = tensor("input_cast_fp16")]; + tensor input_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("input_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor out_eos_weight_to_fp16 = const()[name = tensor("out_eos_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(604276352)))]; + tensor out_eos_bias_to_fp16 = const()[name = tensor("out_eos_bias_to_fp16"), val = tensor([-0x1.64p-3])]; + tensor linear_97_cast_fp16 = linear(bias = out_eos_bias_to_fp16, weight = out_eos_weight_to_fp16, x = input_cast_fp16)[name = tensor("linear_97_cast_fp16")]; + tensor linear_97_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("linear_97_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor is_eos = cast(dtype = linear_97_cast_fp16_to_fp32_dtype_0, x = linear_97_cast_fp16)[name = tensor("cast_384")]; + tensor transformer_out = cast(dtype = input_cast_fp16_to_fp32_dtype_0, x = input_cast_fp16)[name = tensor("cast_385")]; + tensor new_position23 = cast(dtype = var_6062_cast_fp16_to_fp32_dtype_0, x = var_6062_cast_fp16)[name = tensor("cast_386")]; + tensor new_v_cache23 = cast(dtype = new_v_cache_cast_fp16_to_fp32_dtype_0, x = new_v_cache_cast_fp16)[name = tensor("cast_388")]; + tensor new_k_cache23 = cast(dtype = new_k_cache_cast_fp16_to_fp32_dtype_0, x = new_k_cache_cast_fp16)[name = tensor("cast_390")]; + tensor new_position22 = cast(dtype = var_5818_cast_fp16_to_fp32_dtype_0, x = var_5818_cast_fp16)[name = tensor("cast_394")]; + tensor new_v_cache22 = cast(dtype = new_v_cache_45_cast_fp16_to_fp32_dtype_0, x = new_v_cache_45_cast_fp16)[name = tensor("cast_396")]; + tensor new_k_cache22 = cast(dtype = new_k_cache_45_cast_fp16_to_fp32_dtype_0, x = new_k_cache_45_cast_fp16)[name = tensor("cast_398")]; + tensor new_position21 = cast(dtype = var_5574_cast_fp16_to_fp32_dtype_0, x = var_5574_cast_fp16)[name = tensor("cast_402")]; + tensor new_v_cache21 = cast(dtype = new_v_cache_43_cast_fp16_to_fp32_dtype_0, x = new_v_cache_43_cast_fp16)[name = tensor("cast_404")]; + tensor new_k_cache21 = cast(dtype = new_k_cache_43_cast_fp16_to_fp32_dtype_0, x = new_k_cache_43_cast_fp16)[name = tensor("cast_406")]; + tensor new_position20 = cast(dtype = var_5330_cast_fp16_to_fp32_dtype_0, x = var_5330_cast_fp16)[name = tensor("cast_410")]; + tensor new_v_cache20 = cast(dtype = new_v_cache_41_cast_fp16_to_fp32_dtype_0, x = new_v_cache_41_cast_fp16)[name = tensor("cast_412")]; + tensor new_k_cache20 = cast(dtype = new_k_cache_41_cast_fp16_to_fp32_dtype_0, x = new_k_cache_41_cast_fp16)[name = tensor("cast_414")]; + tensor new_position19 = cast(dtype = var_5086_cast_fp16_to_fp32_dtype_0, x = var_5086_cast_fp16)[name = tensor("cast_418")]; + tensor new_v_cache19 = cast(dtype = new_v_cache_39_cast_fp16_to_fp32_dtype_0, x = new_v_cache_39_cast_fp16)[name = tensor("cast_420")]; + tensor new_k_cache19 = cast(dtype = new_k_cache_39_cast_fp16_to_fp32_dtype_0, x = new_k_cache_39_cast_fp16)[name = tensor("cast_422")]; + tensor new_position18 = cast(dtype = var_4842_cast_fp16_to_fp32_dtype_0, x = var_4842_cast_fp16)[name = tensor("cast_426")]; + tensor new_v_cache18 = cast(dtype = new_v_cache_37_cast_fp16_to_fp32_dtype_0, x = new_v_cache_37_cast_fp16)[name = tensor("cast_428")]; + tensor new_k_cache18 = cast(dtype = new_k_cache_37_cast_fp16_to_fp32_dtype_0, x = new_k_cache_37_cast_fp16)[name = tensor("cast_430")]; + tensor new_position17 = cast(dtype = var_4598_cast_fp16_to_fp32_dtype_0, x = var_4598_cast_fp16)[name = tensor("cast_434")]; + tensor new_v_cache17 = cast(dtype = new_v_cache_35_cast_fp16_to_fp32_dtype_0, x = new_v_cache_35_cast_fp16)[name = tensor("cast_436")]; + tensor new_k_cache17 = cast(dtype = new_k_cache_35_cast_fp16_to_fp32_dtype_0, x = new_k_cache_35_cast_fp16)[name = tensor("cast_438")]; + tensor new_position16 = cast(dtype = var_4354_cast_fp16_to_fp32_dtype_0, x = var_4354_cast_fp16)[name = tensor("cast_442")]; + tensor new_v_cache16 = cast(dtype = new_v_cache_33_cast_fp16_to_fp32_dtype_0, x = new_v_cache_33_cast_fp16)[name = tensor("cast_444")]; + tensor new_k_cache16 = cast(dtype = new_k_cache_33_cast_fp16_to_fp32_dtype_0, x = new_k_cache_33_cast_fp16)[name = tensor("cast_446")]; + tensor new_position15 = cast(dtype = var_4110_cast_fp16_to_fp32_dtype_0, x = var_4110_cast_fp16)[name = tensor("cast_450")]; + tensor new_v_cache15 = cast(dtype = new_v_cache_31_cast_fp16_to_fp32_dtype_0, x = new_v_cache_31_cast_fp16)[name = tensor("cast_452")]; + tensor new_k_cache15 = cast(dtype = new_k_cache_31_cast_fp16_to_fp32_dtype_0, x = new_k_cache_31_cast_fp16)[name = tensor("cast_454")]; + tensor new_position14 = cast(dtype = var_3866_cast_fp16_to_fp32_dtype_0, x = var_3866_cast_fp16)[name = tensor("cast_458")]; + tensor new_v_cache14 = cast(dtype = new_v_cache_29_cast_fp16_to_fp32_dtype_0, x = new_v_cache_29_cast_fp16)[name = tensor("cast_460")]; + tensor new_k_cache14 = cast(dtype = new_k_cache_29_cast_fp16_to_fp32_dtype_0, x = new_k_cache_29_cast_fp16)[name = tensor("cast_462")]; + tensor new_position13 = cast(dtype = var_3622_cast_fp16_to_fp32_dtype_0, x = var_3622_cast_fp16)[name = tensor("cast_466")]; + tensor new_v_cache13 = cast(dtype = new_v_cache_27_cast_fp16_to_fp32_dtype_0, x = new_v_cache_27_cast_fp16)[name = tensor("cast_468")]; + tensor new_k_cache13 = cast(dtype = new_k_cache_27_cast_fp16_to_fp32_dtype_0, x = new_k_cache_27_cast_fp16)[name = tensor("cast_470")]; + tensor new_position12 = cast(dtype = var_3378_cast_fp16_to_fp32_dtype_0, x = var_3378_cast_fp16)[name = tensor("cast_474")]; + tensor new_v_cache12 = cast(dtype = new_v_cache_25_cast_fp16_to_fp32_dtype_0, x = new_v_cache_25_cast_fp16)[name = tensor("cast_476")]; + tensor new_k_cache12 = cast(dtype = new_k_cache_25_cast_fp16_to_fp32_dtype_0, x = new_k_cache_25_cast_fp16)[name = tensor("cast_478")]; + tensor new_position11 = cast(dtype = var_3134_cast_fp16_to_fp32_dtype_0, x = var_3134_cast_fp16)[name = tensor("cast_482")]; + tensor new_v_cache11 = cast(dtype = new_v_cache_23_cast_fp16_to_fp32_dtype_0, x = new_v_cache_23_cast_fp16)[name = tensor("cast_484")]; + tensor new_k_cache11 = cast(dtype = new_k_cache_23_cast_fp16_to_fp32_dtype_0, x = new_k_cache_23_cast_fp16)[name = tensor("cast_486")]; + tensor new_position10 = cast(dtype = var_2890_cast_fp16_to_fp32_dtype_0, x = var_2890_cast_fp16)[name = tensor("cast_490")]; + tensor new_v_cache10 = cast(dtype = new_v_cache_21_cast_fp16_to_fp32_dtype_0, x = new_v_cache_21_cast_fp16)[name = tensor("cast_492")]; + tensor new_k_cache10 = cast(dtype = new_k_cache_21_cast_fp16_to_fp32_dtype_0, x = new_k_cache_21_cast_fp16)[name = tensor("cast_494")]; + tensor new_position9 = cast(dtype = var_2646_cast_fp16_to_fp32_dtype_0, x = var_2646_cast_fp16)[name = tensor("cast_498")]; + tensor new_v_cache9 = cast(dtype = new_v_cache_19_cast_fp16_to_fp32_dtype_0, x = new_v_cache_19_cast_fp16)[name = tensor("cast_500")]; + tensor new_k_cache9 = cast(dtype = new_k_cache_19_cast_fp16_to_fp32_dtype_0, x = new_k_cache_19_cast_fp16)[name = tensor("cast_502")]; + tensor new_position8 = cast(dtype = var_2402_cast_fp16_to_fp32_dtype_0, x = var_2402_cast_fp16)[name = tensor("cast_506")]; + tensor new_v_cache8 = cast(dtype = new_v_cache_17_cast_fp16_to_fp32_dtype_0, x = new_v_cache_17_cast_fp16)[name = tensor("cast_508")]; + tensor new_k_cache8 = cast(dtype = new_k_cache_17_cast_fp16_to_fp32_dtype_0, x = new_k_cache_17_cast_fp16)[name = tensor("cast_510")]; + tensor new_position7 = cast(dtype = var_2158_cast_fp16_to_fp32_dtype_0, x = var_2158_cast_fp16)[name = tensor("cast_514")]; + tensor new_v_cache7 = cast(dtype = new_v_cache_15_cast_fp16_to_fp32_dtype_0, x = new_v_cache_15_cast_fp16)[name = tensor("cast_516")]; + tensor new_k_cache7 = cast(dtype = new_k_cache_15_cast_fp16_to_fp32_dtype_0, x = new_k_cache_15_cast_fp16)[name = tensor("cast_518")]; + tensor new_position6 = cast(dtype = var_1914_cast_fp16_to_fp32_dtype_0, x = var_1914_cast_fp16)[name = tensor("cast_522")]; + tensor new_v_cache6 = cast(dtype = new_v_cache_13_cast_fp16_to_fp32_dtype_0, x = new_v_cache_13_cast_fp16)[name = tensor("cast_524")]; + tensor new_k_cache6 = cast(dtype = new_k_cache_13_cast_fp16_to_fp32_dtype_0, x = new_k_cache_13_cast_fp16)[name = tensor("cast_526")]; + tensor new_position5 = cast(dtype = var_1670_cast_fp16_to_fp32_dtype_0, x = var_1670_cast_fp16)[name = tensor("cast_530")]; + tensor new_v_cache5 = cast(dtype = new_v_cache_11_cast_fp16_to_fp32_dtype_0, x = new_v_cache_11_cast_fp16)[name = tensor("cast_532")]; + tensor new_k_cache5 = cast(dtype = new_k_cache_11_cast_fp16_to_fp32_dtype_0, x = new_k_cache_11_cast_fp16)[name = tensor("cast_534")]; + tensor new_position4 = cast(dtype = var_1426_cast_fp16_to_fp32_dtype_0, x = var_1426_cast_fp16)[name = tensor("cast_538")]; + tensor new_v_cache4 = cast(dtype = new_v_cache_9_cast_fp16_to_fp32_dtype_0, x = new_v_cache_9_cast_fp16)[name = tensor("cast_540")]; + tensor new_k_cache4 = cast(dtype = new_k_cache_9_cast_fp16_to_fp32_dtype_0, x = new_k_cache_9_cast_fp16)[name = tensor("cast_542")]; + tensor new_position3 = cast(dtype = var_1182_cast_fp16_to_fp32_dtype_0, x = var_1182_cast_fp16)[name = tensor("cast_546")]; + tensor new_v_cache3 = cast(dtype = new_v_cache_7_cast_fp16_to_fp32_dtype_0, x = new_v_cache_7_cast_fp16)[name = tensor("cast_548")]; + tensor new_k_cache3 = cast(dtype = new_k_cache_7_cast_fp16_to_fp32_dtype_0, x = new_k_cache_7_cast_fp16)[name = tensor("cast_550")]; + tensor new_position2 = cast(dtype = var_938_cast_fp16_to_fp32_dtype_0, x = var_938_cast_fp16)[name = tensor("cast_554")]; + tensor new_v_cache2 = cast(dtype = new_v_cache_5_cast_fp16_to_fp32_dtype_0, x = new_v_cache_5_cast_fp16)[name = tensor("cast_556")]; + tensor new_k_cache2 = cast(dtype = new_k_cache_5_cast_fp16_to_fp32_dtype_0, x = new_k_cache_5_cast_fp16)[name = tensor("cast_558")]; + tensor new_position1 = cast(dtype = var_694_cast_fp16_to_fp32_dtype_0, x = var_694_cast_fp16)[name = tensor("cast_562")]; + tensor new_v_cache1 = cast(dtype = new_v_cache_3_cast_fp16_to_fp32_dtype_0, x = new_v_cache_3_cast_fp16)[name = tensor("cast_564")]; + tensor new_k_cache1 = cast(dtype = new_k_cache_3_cast_fp16_to_fp32_dtype_0, x = new_k_cache_3_cast_fp16)[name = tensor("cast_566")]; + tensor new_position0 = cast(dtype = var_450_cast_fp16_to_fp32_dtype_0, x = var_450_cast_fp16)[name = tensor("cast_570")]; + tensor new_v_cache0 = cast(dtype = new_v_cache_1_cast_fp16_to_fp32_dtype_0, x = new_v_cache_1_cast_fp16)[name = tensor("cast_572")]; + tensor new_k_cache0 = cast(dtype = new_k_cache_1_cast_fp16_to_fp32_dtype_0, x = new_k_cache_1_cast_fp16)[name = tensor("cast_574")]; + } -> (transformer_out, is_eos, new_k_cache0, new_v_cache0, new_position0, new_k_cache1, new_v_cache1, new_position1, new_k_cache2, new_v_cache2, new_position2, new_k_cache3, new_v_cache3, new_position3, new_k_cache4, new_v_cache4, new_position4, new_k_cache5, new_v_cache5, new_position5, new_k_cache6, new_v_cache6, new_position6, new_k_cache7, new_v_cache7, new_position7, new_k_cache8, new_v_cache8, new_position8, new_k_cache9, new_v_cache9, new_position9, new_k_cache10, new_v_cache10, new_position10, new_k_cache11, new_v_cache11, new_position11, new_k_cache12, new_v_cache12, new_position12, new_k_cache13, new_v_cache13, new_position13, new_k_cache14, new_v_cache14, new_position14, new_k_cache15, new_v_cache15, new_position15, new_k_cache16, new_v_cache16, new_position16, new_k_cache17, new_v_cache17, new_position17, new_k_cache18, new_v_cache18, new_position18, new_k_cache19, new_v_cache19, new_position19, new_k_cache20, new_v_cache20, new_position20, new_k_cache21, new_v_cache21, new_position21, new_k_cache22, new_v_cache22, new_position22, new_k_cache23, new_v_cache23, new_position23); +} \ No newline at end of file diff --git a/v2.1/italian_24l/flowlm_step_ane.mlmodelc/weights/weight.bin b/v2.1/italian_24l/flowlm_step_ane.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..2a3d6cc0c17fc3cb96c2778c8cf6b8600b01a8fc --- /dev/null +++ b/v2.1/italian_24l/flowlm_step_ane.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d68ff442fad20d6ec483dda9dc5816fc0ffce8e1ffb9d9e8fdf805edbaf1f82a +size 604278464 diff --git a/v2.1/italian_24l/flowlm_step_ane.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/italian_24l/flowlm_step_ane.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..479ed772e45febb152b839011a9d4172ee7d61c9 --- /dev/null +++ b/v2.1/italian_24l/flowlm_step_ane.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:90e9182dfa523ad27b5b16061f4af1a0fe3a9e621b69807cd7ccf341774617fc +size 545925 diff --git a/v2.1/italian_24l/flowlm_step_ane.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/italian_24l/flowlm_step_ane.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..2a3d6cc0c17fc3cb96c2778c8cf6b8600b01a8fc --- /dev/null +++ b/v2.1/italian_24l/flowlm_step_ane.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d68ff442fad20d6ec483dda9dc5816fc0ffce8e1ffb9d9e8fdf805edbaf1f82a +size 604278464 diff --git a/v2.1/italian_24l/flowlm_step_ane.mlpackage/Manifest.json b/v2.1/italian_24l/flowlm_step_ane.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..a6fef383dddb1878d11504f2b1590aa16830273a --- /dev/null +++ b/v2.1/italian_24l/flowlm_step_ane.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "9CE006C9-8A1A-40B8-B6B4-58220D462E98": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "C342707E-652C-4E59-83C4-2A61FB9BA255": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "9CE006C9-8A1A-40B8-B6B4-58220D462E98" +} diff --git a/v2.1/portuguese/cond_prefill_ane.mlmodelc/analytics/coremldata.bin b/v2.1/portuguese/cond_prefill_ane.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..90eeb4b835b9693dc5e2ec06589b6e07979549e3 --- /dev/null +++ b/v2.1/portuguese/cond_prefill_ane.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ae26b20655b1f9e81aa32f949b189ddc4abb507401484df853af6102c38afe2 +size 243 diff --git a/v2.1/portuguese/cond_prefill_ane.mlmodelc/coremldata.bin b/v2.1/portuguese/cond_prefill_ane.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..b52168fe32dccc1f3abfad5127d7383fc049bc3a --- /dev/null +++ b/v2.1/portuguese/cond_prefill_ane.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8221e35581faaa4f54675dfc1680a1ab35b48102bd8a9876444b1d78581c7cde +size 1388 diff --git a/v2.1/portuguese/cond_prefill_ane.mlmodelc/model.mil b/v2.1/portuguese/cond_prefill_ane.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..2aef44ccd7c895876ef314e46a742909df52fe28 --- /dev/null +++ b/v2.1/portuguese/cond_prefill_ane.mlmodelc/model.mil @@ -0,0 +1,804 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.9.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor conditioning, tensor k_cache0, tensor k_cache1, tensor k_cache2, tensor k_cache3, tensor k_cache4, tensor k_cache5, tensor position0, tensor position1, tensor position2, tensor position3, tensor position4, tensor position5, tensor v_cache0, tensor v_cache1, tensor v_cache2, tensor v_cache3, tensor v_cache4, tensor v_cache5, tensor valid_len) { + tensor token_indices_to_fp16 = const()[name = tensor("token_indices_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor position0_to_fp16_dtype_0 = const()[name = tensor("position0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position0_to_fp16 = cast(dtype = position0_to_fp16_dtype_0, x = position0)[name = tensor("cast_128")]; + tensor ts_cast_fp16 = add(x = token_indices_to_fp16, y = position0_to_fp16)[name = tensor("ts_cast_fp16")]; + tensor var_69 = const()[name = tensor("op_69"), val = tensor([256, 1, 1])]; + tensor var_70_cast_fp16 = reshape(shape = var_69, x = ts_cast_fp16)[name = tensor("op_70_cast_fp16")]; + tensor var_65_to_fp16 = const()[name = tensor("op_65_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(640)))]; + tensor angles_cast_fp16 = mul(x = var_65_to_fp16, y = var_70_cast_fp16)[name = tensor("angles_cast_fp16")]; + tensor rotr_cast_fp16 = cos(x = angles_cast_fp16)[name = tensor("rotr_cast_fp16")]; + tensor roti_cast_fp16 = sin(x = angles_cast_fp16)[name = tensor("roti_cast_fp16")]; + tensor _inversed_75_y_0_to_fp16 = const()[name = tensor("_inversed_75_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_75_cast_fp16 = mul(x = ts_cast_fp16, y = _inversed_75_y_0_to_fp16)[name = tensor("_inversed_75_cast_fp16")]; + tensor var_76_cast_fp16 = floor(x = _inversed_75_cast_fp16)[name = tensor("op_76_cast_fp16")]; + tensor var_77_to_fp16 = const()[name = tensor("op_77_to_fp16"), val = tensor(0x1p+9)]; + tensor var_78_cast_fp16 = mul(x = var_76_cast_fp16, y = var_77_to_fp16)[name = tensor("op_78_cast_fp16")]; + tensor write_idx_1_cast_fp16 = sub(x = ts_cast_fp16, y = var_78_cast_fp16)[name = tensor("write_idx_1_cast_fp16")]; + tensor valid_len_to_fp16_dtype_0 = const()[name = tensor("valid_len_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor valid_len_to_fp16 = cast(dtype = valid_len_to_fp16_dtype_0, x = valid_len)[name = tensor("cast_127")]; + tensor var_84_cast_fp16 = less(x = token_indices_to_fp16, y = valid_len_to_fp16)[name = tensor("op_84_cast_fp16")]; + tensor token_valid_to_fp16_dtype_0 = const()[name = tensor("token_valid_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_84_cast_fp16_to_fp16 = cast(dtype = token_valid_to_fp16_dtype_0, x = var_84_cast_fp16)[name = tensor("cast_126")]; + tensor var_90_cast_fp16 = mul(x = write_idx_1_cast_fp16, y = var_84_cast_fp16_to_fp16)[name = tensor("op_90_cast_fp16")]; + tensor var_91_to_fp16 = const()[name = tensor("op_91_to_fp16"), val = tensor(0x1p+0)]; + tensor var_93_cast_fp16 = sub(x = var_91_to_fp16, y = var_84_cast_fp16_to_fp16)[name = tensor("op_93_cast_fp16")]; + tensor var_94_to_fp16 = const()[name = tensor("op_94_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_95_cast_fp16 = mul(x = var_93_cast_fp16, y = var_94_to_fp16)[name = tensor("op_95_cast_fp16")]; + tensor write_idx_cast_fp16 = add(x = var_90_cast_fp16, y = var_95_cast_fp16)[name = tensor("write_idx_cast_fp16")]; + tensor var_100 = const()[name = tensor("op_100"), val = tensor([256, 1])]; + tensor var_101_cast_fp16 = reshape(shape = var_100, x = write_idx_cast_fp16)[name = tensor("op_101_cast_fp16")]; + tensor var_105_to_fp16 = const()[name = tensor("op_105_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(768)))]; + tensor var_106_cast_fp16 = equal(x = var_101_cast_fp16, y = var_105_to_fp16)[name = tensor("op_106_cast_fp16")]; + tensor var_116_axes_0 = const()[name = tensor("op_116_axes_0"), val = tensor([0])]; + tensor var_116_keep_dims_0 = const()[name = tensor("op_116_keep_dims_0"), val = tensor(false)]; + tensor assign_to_fp16_dtype_0 = const()[name = tensor("assign_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_106_cast_fp16_to_fp16 = cast(dtype = assign_to_fp16_dtype_0, x = var_106_cast_fp16)[name = tensor("cast_125")]; + tensor var_116_cast_fp16 = reduce_sum(axes = var_116_axes_0, keep_dims = var_116_keep_dims_0, x = var_106_cast_fp16_to_fp16)[name = tensor("op_116_cast_fp16")]; + tensor const_0_to_fp16 = const()[name = tensor("const_0_to_fp16"), val = tensor(-inf)]; + tensor var_118_to_fp16 = const()[name = tensor("op_118_to_fp16"), val = tensor(0x1p+0)]; + tensor clip_0_cast_fp16 = clip(alpha = const_0_to_fp16, beta = var_118_to_fp16, x = var_116_cast_fp16)[name = tensor("clip_0_cast_fp16")]; + tensor var_122 = const()[name = tensor("op_122"), val = tensor([1, 1])]; + tensor var_123_cast_fp16 = reshape(shape = var_122, x = position0_to_fp16)[name = tensor("op_123_cast_fp16")]; + tensor var_127_to_fp16 = const()[name = tensor("op_127_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1856)))]; + tensor q_positions_cast_fp16 = add(x = var_123_cast_fp16, y = var_127_to_fp16)[name = tensor("q_positions_cast_fp16")]; + tensor var_135_cast_fp16 = add(x = position0_to_fp16, y = valid_len_to_fp16)[name = tensor("op_135_cast_fp16")]; + tensor var_138 = const()[name = tensor("op_138"), val = tensor([1, 1])]; + tensor valid_end_cast_fp16 = reshape(shape = var_138, x = var_135_cast_fp16)[name = tensor("valid_end_cast_fp16")]; + tensor var_140_cast_fp16 = less(x = var_105_to_fp16, y = valid_end_cast_fp16)[name = tensor("op_140_cast_fp16")]; + tensor var_141_cast_fp16 = less_equal(x = var_105_to_fp16, y = q_positions_cast_fp16)[name = tensor("op_141_cast_fp16")]; + tensor var_142 = logical_and(x = var_140_cast_fp16, y = var_141_cast_fp16)[name = tensor("op_142")]; + tensor mask_to_fp16_dtype_0 = const()[name = tensor("mask_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_149_to_fp16 = const()[name = tensor("op_149_to_fp16"), val = tensor(0x1p+0)]; + tensor var_142_to_fp16 = cast(dtype = mask_to_fp16_dtype_0, x = var_142)[name = tensor("cast_124")]; + tensor var_150_cast_fp16 = sub(x = var_142_to_fp16, y = var_149_to_fp16)[name = tensor("op_150_cast_fp16")]; + tensor var_151_to_fp16 = const()[name = tensor("op_151_to_fp16"), val = tensor(0x1.388p+13)]; + tensor mask_bias_cast_fp16 = mul(x = var_150_cast_fp16, y = var_151_to_fp16)[name = tensor("mask_bias_cast_fp16")]; + tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; + tensor conditioning_to_fp16_dtype_0 = const()[name = tensor("conditioning_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor norm0_1_weight_to_fp16 = const()[name = tensor("norm0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2432)))]; + tensor norm0_1_bias_to_fp16 = const()[name = tensor("norm0_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4544)))]; + tensor var_154_to_fp16 = const()[name = tensor("op_154_to_fp16"), val = tensor(0x1.5p-17)]; + tensor conditioning_to_fp16 = cast(dtype = conditioning_to_fp16_dtype_0, x = conditioning)[name = tensor("cast_123")]; + tensor input_3_cast_fp16 = layer_norm(axes = input_3_axes_0, beta = norm0_1_bias_to_fp16, epsilon = var_154_to_fp16, gamma = norm0_1_weight_to_fp16, x = conditioning_to_fp16)[name = tensor("input_3_cast_fp16")]; + tensor attn0_in_proj_weight_to_fp16 = const()[name = tensor("attn0_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6656)))]; + tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6298176)))]; + tensor linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn0_in_proj_weight_to_fp16, x = input_3_cast_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor var_168 = const()[name = tensor("op_168"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_1_cast_fp16 = reshape(shape = var_168, x = linear_0_cast_fp16)[name = tensor("qkv_1_cast_fp16")]; + tensor x_1_begin_0 = const()[name = tensor("x_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_1_end_0 = const()[name = tensor("x_1_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_1_end_mask_0 = const()[name = tensor("x_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_1_squeeze_mask_0 = const()[name = tensor("x_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_1_cast_fp16 = slice_by_index(begin = x_1_begin_0, end = x_1_end_0, end_mask = x_1_end_mask_0, squeeze_mask = x_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("x_1_cast_fp16")]; + tensor x_3_begin_0 = const()[name = tensor("x_3_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_3_end_0 = const()[name = tensor("x_3_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_3_end_mask_0 = const()[name = tensor("x_3_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_3_squeeze_mask_0 = const()[name = tensor("x_3_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_3_cast_fp16 = slice_by_index(begin = x_3_begin_0, end = x_3_end_0, end_mask = x_3_end_mask_0, squeeze_mask = x_3_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("x_3_cast_fp16")]; + tensor v_1_begin_0 = const()[name = tensor("v_1_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_1_end_0 = const()[name = tensor("v_1_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_1_end_mask_0 = const()[name = tensor("v_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_1_squeeze_mask_0 = const()[name = tensor("v_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_1_cast_fp16 = slice_by_index(begin = v_1_begin_0, end = v_1_end_0, end_mask = v_1_end_mask_0, squeeze_mask = v_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("v_1_cast_fp16")]; + tensor var_213 = const()[name = tensor("op_213"), val = tensor([256, 16, 32, 2])]; + tensor pairs_1_cast_fp16 = reshape(shape = var_213, x = x_1_cast_fp16)[name = tensor("pairs_1_cast_fp16")]; + tensor xr_1_begin_0 = const()[name = tensor("xr_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_1_end_0 = const()[name = tensor("xr_1_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_1_end_mask_0 = const()[name = tensor("xr_1_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_1_squeeze_mask_0 = const()[name = tensor("xr_1_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_1_cast_fp16 = slice_by_index(begin = xr_1_begin_0, end = xr_1_end_0, end_mask = xr_1_end_mask_0, squeeze_mask = xr_1_squeeze_mask_0, x = pairs_1_cast_fp16)[name = tensor("xr_1_cast_fp16")]; + tensor xi_1_begin_0 = const()[name = tensor("xi_1_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_1_end_0 = const()[name = tensor("xi_1_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_1_end_mask_0 = const()[name = tensor("xi_1_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_1_squeeze_mask_0 = const()[name = tensor("xi_1_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_1_cast_fp16 = slice_by_index(begin = xi_1_begin_0, end = xi_1_end_0, end_mask = xi_1_end_mask_0, squeeze_mask = xi_1_squeeze_mask_0, x = pairs_1_cast_fp16)[name = tensor("xi_1_cast_fp16")]; + tensor var_221_cast_fp16 = mul(x = xr_1_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_221_cast_fp16")]; + tensor var_222_cast_fp16 = mul(x = xi_1_cast_fp16, y = roti_cast_fp16)[name = tensor("op_222_cast_fp16")]; + tensor xor__1_cast_fp16 = sub(x = var_221_cast_fp16, y = var_222_cast_fp16)[name = tensor("xor__1_cast_fp16")]; + tensor var_225_cast_fp16 = mul(x = xr_1_cast_fp16, y = roti_cast_fp16)[name = tensor("op_225_cast_fp16")]; + tensor var_226_cast_fp16 = mul(x = xi_1_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_226_cast_fp16")]; + tensor xoi_1_cast_fp16 = add(x = var_225_cast_fp16, y = var_226_cast_fp16)[name = tensor("xoi_1_cast_fp16")]; + tensor out_1_axis_0 = const()[name = tensor("out_1_axis_0"), val = tensor(-1)]; + tensor out_1_cast_fp16 = stack(axis = out_1_axis_0, values = (xor__1_cast_fp16, xoi_1_cast_fp16))[name = tensor("out_1_cast_fp16")]; + tensor var_236 = const()[name = tensor("op_236"), val = tensor([1, 256, 16, 64])]; + tensor q_1_cast_fp16 = reshape(shape = var_236, x = out_1_cast_fp16)[name = tensor("q_1_cast_fp16")]; + tensor var_242 = const()[name = tensor("op_242"), val = tensor([256, 16, 32, 2])]; + tensor pairs_3_cast_fp16 = reshape(shape = var_242, x = x_3_cast_fp16)[name = tensor("pairs_3_cast_fp16")]; + tensor xr_3_begin_0 = const()[name = tensor("xr_3_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_3_end_0 = const()[name = tensor("xr_3_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_3_end_mask_0 = const()[name = tensor("xr_3_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_3_squeeze_mask_0 = const()[name = tensor("xr_3_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_3_cast_fp16 = slice_by_index(begin = xr_3_begin_0, end = xr_3_end_0, end_mask = xr_3_end_mask_0, squeeze_mask = xr_3_squeeze_mask_0, x = pairs_3_cast_fp16)[name = tensor("xr_3_cast_fp16")]; + tensor xi_3_begin_0 = const()[name = tensor("xi_3_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_3_end_0 = const()[name = tensor("xi_3_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_3_end_mask_0 = const()[name = tensor("xi_3_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_3_squeeze_mask_0 = const()[name = tensor("xi_3_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_3_cast_fp16 = slice_by_index(begin = xi_3_begin_0, end = xi_3_end_0, end_mask = xi_3_end_mask_0, squeeze_mask = xi_3_squeeze_mask_0, x = pairs_3_cast_fp16)[name = tensor("xi_3_cast_fp16")]; + tensor var_250_cast_fp16 = mul(x = xr_3_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_250_cast_fp16")]; + tensor var_251_cast_fp16 = mul(x = xi_3_cast_fp16, y = roti_cast_fp16)[name = tensor("op_251_cast_fp16")]; + tensor xor__3_cast_fp16 = sub(x = var_250_cast_fp16, y = var_251_cast_fp16)[name = tensor("xor__3_cast_fp16")]; + tensor var_254_cast_fp16 = mul(x = xr_3_cast_fp16, y = roti_cast_fp16)[name = tensor("op_254_cast_fp16")]; + tensor var_255_cast_fp16 = mul(x = xi_3_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_255_cast_fp16")]; + tensor xoi_3_cast_fp16 = add(x = var_254_cast_fp16, y = var_255_cast_fp16)[name = tensor("xoi_3_cast_fp16")]; + tensor out_3_axis_0 = const()[name = tensor("out_3_axis_0"), val = tensor(-1)]; + tensor out_3_cast_fp16 = stack(axis = out_3_axis_0, values = (xor__3_cast_fp16, xoi_3_cast_fp16))[name = tensor("out_3_cast_fp16")]; + tensor var_272 = const()[name = tensor("op_272"), val = tensor([256, 1024])]; + tensor var_273_cast_fp16 = reshape(shape = var_272, x = out_3_cast_fp16)[name = tensor("op_273_cast_fp16")]; + tensor written_k_1_transpose_x_1 = const()[name = tensor("written_k_1_transpose_x_1"), val = tensor(true)]; + tensor written_k_1_transpose_y_1 = const()[name = tensor("written_k_1_transpose_y_1"), val = tensor(false)]; + tensor written_k_1_cast_fp16 = matmul(transpose_x = written_k_1_transpose_x_1, transpose_y = written_k_1_transpose_y_1, x = var_106_cast_fp16_to_fp16, y = var_273_cast_fp16)[name = tensor("written_k_1_cast_fp16")]; + tensor var_280 = const()[name = tensor("op_280"), val = tensor([256, 1024])]; + tensor var_281_cast_fp16 = reshape(shape = var_280, x = v_1_cast_fp16)[name = tensor("op_281_cast_fp16")]; + tensor written_v_1_transpose_x_1 = const()[name = tensor("written_v_1_transpose_x_1"), val = tensor(true)]; + tensor written_v_1_transpose_y_1 = const()[name = tensor("written_v_1_transpose_y_1"), val = tensor(false)]; + tensor written_v_1_cast_fp16 = matmul(transpose_x = written_v_1_transpose_x_1, transpose_y = written_v_1_transpose_y_1, x = var_106_cast_fp16_to_fp16, y = var_281_cast_fp16)[name = tensor("written_v_1_cast_fp16")]; + tensor var_283_to_fp16 = const()[name = tensor("op_283_to_fp16"), val = tensor(0x1p+0)]; + tensor var_285_cast_fp16 = sub(x = var_283_to_fp16, y = clip_0_cast_fp16)[name = tensor("op_285_cast_fp16")]; + tensor var_290 = const()[name = tensor("op_290"), val = tensor([1, 512, 1, 1])]; + tensor keep_1_cast_fp16 = reshape(shape = var_290, x = var_285_cast_fp16)[name = tensor("keep_1_cast_fp16")]; + tensor k_cache0_to_fp16_dtype_0 = const()[name = tensor("k_cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache0_to_fp16 = cast(dtype = k_cache0_to_fp16_dtype_0, x = k_cache0)[name = tensor("cast_122")]; + tensor var_292_cast_fp16 = mul(x = k_cache0_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_292_cast_fp16")]; + tensor var_297 = const()[name = tensor("op_297"), val = tensor([1, 512, 16, 64])]; + tensor var_298_cast_fp16 = reshape(shape = var_297, x = written_k_1_cast_fp16)[name = tensor("op_298_cast_fp16")]; + tensor new_k_cache_1_cast_fp16 = add(x = var_292_cast_fp16, y = var_298_cast_fp16)[name = tensor("new_k_cache_1_cast_fp16")]; + tensor new_k_cache_1_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_1_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache0_to_fp16_dtype_0 = const()[name = tensor("v_cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache0_to_fp16 = cast(dtype = v_cache0_to_fp16_dtype_0, x = v_cache0)[name = tensor("cast_120")]; + tensor var_301_cast_fp16 = mul(x = v_cache0_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_301_cast_fp16")]; + tensor var_306 = const()[name = tensor("op_306"), val = tensor([1, 512, 16, 64])]; + tensor var_307_cast_fp16 = reshape(shape = var_306, x = written_v_1_cast_fp16)[name = tensor("op_307_cast_fp16")]; + tensor new_v_cache_1_cast_fp16 = add(x = var_301_cast_fp16, y = var_307_cast_fp16)[name = tensor("new_v_cache_1_cast_fp16")]; + tensor new_v_cache_1_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_1_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_320 = const()[name = tensor("op_320"), val = tensor([0, 2, 1, 3])]; + tensor var_331_transpose_x_0 = const()[name = tensor("op_331_transpose_x_0"), val = tensor(false)]; + tensor var_331_transpose_y_0 = const()[name = tensor("op_331_transpose_y_0"), val = tensor(false)]; + tensor transpose_15_perm_0 = const()[name = tensor("transpose_15_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_16_perm_0 = const()[name = tensor("transpose_16_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_16 = transpose(perm = transpose_16_perm_0, x = new_k_cache_1_cast_fp16)[name = tensor("transpose_42")]; + tensor transpose_15 = transpose(perm = transpose_15_perm_0, x = q_1_cast_fp16)[name = tensor("transpose_43")]; + tensor var_331_cast_fp16 = matmul(transpose_x = var_331_transpose_x_0, transpose_y = var_331_transpose_y_0, x = transpose_15, y = transpose_16)[name = tensor("op_331_cast_fp16")]; + tensor var_332_to_fp16 = const()[name = tensor("op_332_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_1_cast_fp16 = mul(x = var_331_cast_fp16, y = var_332_to_fp16)[name = tensor("scores_1_cast_fp16")]; + tensor var_338 = const()[name = tensor("op_338"), val = tensor([1, 1, 256, 512])]; + tensor var_339_cast_fp16 = reshape(shape = var_338, x = mask_bias_cast_fp16)[name = tensor("op_339_cast_fp16")]; + tensor scores_3_cast_fp16 = add(x = scores_1_cast_fp16, y = var_339_cast_fp16)[name = tensor("scores_3_cast_fp16")]; + tensor var_342 = const()[name = tensor("op_342"), val = tensor(-1)]; + tensor weights_1_cast_fp16 = softmax(axis = var_342, x = scores_3_cast_fp16)[name = tensor("weights_1_cast_fp16")]; + tensor attn_1_transpose_x_0 = const()[name = tensor("attn_1_transpose_x_0"), val = tensor(false)]; + tensor attn_1_transpose_y_0 = const()[name = tensor("attn_1_transpose_y_0"), val = tensor(false)]; + tensor values_1_cast_fp16 = transpose(perm = var_320, x = new_v_cache_1_cast_fp16)[name = tensor("transpose_44")]; + tensor attn_1_cast_fp16 = matmul(transpose_x = attn_1_transpose_x_0, transpose_y = attn_1_transpose_y_0, x = weights_1_cast_fp16, y = values_1_cast_fp16)[name = tensor("attn_1_cast_fp16")]; + tensor var_350 = const()[name = tensor("op_350"), val = tensor([0, 2, 1, 3])]; + tensor var_355 = const()[name = tensor("op_355"), val = tensor([1, 256, 1024])]; + tensor var_351_cast_fp16 = transpose(perm = var_350, x = attn_1_cast_fp16)[name = tensor("transpose_41")]; + tensor input_5_cast_fp16 = reshape(shape = var_355, x = var_351_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor attn0_out_proj_weight_to_fp16 = const()[name = tensor("attn0_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6304384)))]; + tensor linear_1_bias_0_to_fp16 = const()[name = tensor("linear_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401600)))]; + tensor linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn0_out_proj_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("linear_1_cast_fp16")]; + tensor var_364_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_364_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_7_cast_fp16 = add(x = conditioning_to_fp16, y = linear_1_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor input_9_axes_0 = const()[name = tensor("input_9_axes_0"), val = tensor([-1])]; + tensor norm0_2_weight_to_fp16 = const()[name = tensor("norm0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8403712)))]; + tensor norm0_2_bias_to_fp16 = const()[name = tensor("norm0_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8405824)))]; + tensor var_368_to_fp16 = const()[name = tensor("op_368_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_9_cast_fp16 = layer_norm(axes = input_9_axes_0, beta = norm0_2_bias_to_fp16, epsilon = var_368_to_fp16, gamma = norm0_2_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor linear0_1_weight_to_fp16 = const()[name = tensor("linear0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8407936)))]; + tensor linear_2_bias_0_to_fp16 = const()[name = tensor("linear_2_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16796608)))]; + tensor linear_2_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear0_1_weight_to_fp16, x = input_9_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor input_11_mode_0 = const()[name = tensor("input_11_mode_0"), val = tensor("EXACT")]; + tensor input_11_cast_fp16 = gelu(mode = input_11_mode_0, x = linear_2_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor linear0_2_weight_to_fp16 = const()[name = tensor("linear0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16804864)))]; + tensor linear_3_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear0_2_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("linear_3_cast_fp16")]; + tensor input_13_cast_fp16 = add(x = input_7_cast_fp16, y = linear_3_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor input_15_axes_0 = const()[name = tensor("input_15_axes_0"), val = tensor([-1])]; + tensor norm1_1_weight_to_fp16 = const()[name = tensor("norm1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25193536)))]; + tensor norm1_1_bias_to_fp16 = const()[name = tensor("norm1_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25195648)))]; + tensor var_385_to_fp16 = const()[name = tensor("op_385_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_15_cast_fp16 = layer_norm(axes = input_15_axes_0, beta = norm1_1_bias_to_fp16, epsilon = var_385_to_fp16, gamma = norm1_1_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor attn1_in_proj_weight_to_fp16 = const()[name = tensor("attn1_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25197760)))]; + tensor linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn1_in_proj_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor var_399 = const()[name = tensor("op_399"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_3_cast_fp16 = reshape(shape = var_399, x = linear_4_cast_fp16)[name = tensor("qkv_3_cast_fp16")]; + tensor x_5_begin_0 = const()[name = tensor("x_5_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_5_end_0 = const()[name = tensor("x_5_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_5_end_mask_0 = const()[name = tensor("x_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_5_squeeze_mask_0 = const()[name = tensor("x_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_5_cast_fp16 = slice_by_index(begin = x_5_begin_0, end = x_5_end_0, end_mask = x_5_end_mask_0, squeeze_mask = x_5_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("x_5_cast_fp16")]; + tensor x_7_begin_0 = const()[name = tensor("x_7_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_7_end_0 = const()[name = tensor("x_7_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_7_end_mask_0 = const()[name = tensor("x_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_7_squeeze_mask_0 = const()[name = tensor("x_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_7_cast_fp16 = slice_by_index(begin = x_7_begin_0, end = x_7_end_0, end_mask = x_7_end_mask_0, squeeze_mask = x_7_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("x_7_cast_fp16")]; + tensor v_3_begin_0 = const()[name = tensor("v_3_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_3_end_0 = const()[name = tensor("v_3_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_3_end_mask_0 = const()[name = tensor("v_3_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_3_squeeze_mask_0 = const()[name = tensor("v_3_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_3_cast_fp16 = slice_by_index(begin = v_3_begin_0, end = v_3_end_0, end_mask = v_3_end_mask_0, squeeze_mask = v_3_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("v_3_cast_fp16")]; + tensor var_444 = const()[name = tensor("op_444"), val = tensor([256, 16, 32, 2])]; + tensor pairs_5_cast_fp16 = reshape(shape = var_444, x = x_5_cast_fp16)[name = tensor("pairs_5_cast_fp16")]; + tensor xr_5_begin_0 = const()[name = tensor("xr_5_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_5_end_0 = const()[name = tensor("xr_5_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_5_end_mask_0 = const()[name = tensor("xr_5_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_5_squeeze_mask_0 = const()[name = tensor("xr_5_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_5_cast_fp16 = slice_by_index(begin = xr_5_begin_0, end = xr_5_end_0, end_mask = xr_5_end_mask_0, squeeze_mask = xr_5_squeeze_mask_0, x = pairs_5_cast_fp16)[name = tensor("xr_5_cast_fp16")]; + tensor xi_5_begin_0 = const()[name = tensor("xi_5_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_5_end_0 = const()[name = tensor("xi_5_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_5_end_mask_0 = const()[name = tensor("xi_5_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_5_squeeze_mask_0 = const()[name = tensor("xi_5_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_5_cast_fp16 = slice_by_index(begin = xi_5_begin_0, end = xi_5_end_0, end_mask = xi_5_end_mask_0, squeeze_mask = xi_5_squeeze_mask_0, x = pairs_5_cast_fp16)[name = tensor("xi_5_cast_fp16")]; + tensor var_452_cast_fp16 = mul(x = xr_5_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_452_cast_fp16")]; + tensor var_453_cast_fp16 = mul(x = xi_5_cast_fp16, y = roti_cast_fp16)[name = tensor("op_453_cast_fp16")]; + tensor xor__5_cast_fp16 = sub(x = var_452_cast_fp16, y = var_453_cast_fp16)[name = tensor("xor__5_cast_fp16")]; + tensor var_456_cast_fp16 = mul(x = xr_5_cast_fp16, y = roti_cast_fp16)[name = tensor("op_456_cast_fp16")]; + tensor var_457_cast_fp16 = mul(x = xi_5_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_457_cast_fp16")]; + tensor xoi_5_cast_fp16 = add(x = var_456_cast_fp16, y = var_457_cast_fp16)[name = tensor("xoi_5_cast_fp16")]; + tensor out_5_axis_0 = const()[name = tensor("out_5_axis_0"), val = tensor(-1)]; + tensor out_5_cast_fp16 = stack(axis = out_5_axis_0, values = (xor__5_cast_fp16, xoi_5_cast_fp16))[name = tensor("out_5_cast_fp16")]; + tensor var_467 = const()[name = tensor("op_467"), val = tensor([1, 256, 16, 64])]; + tensor q_3_cast_fp16 = reshape(shape = var_467, x = out_5_cast_fp16)[name = tensor("q_3_cast_fp16")]; + tensor var_473 = const()[name = tensor("op_473"), val = tensor([256, 16, 32, 2])]; + tensor pairs_7_cast_fp16 = reshape(shape = var_473, x = x_7_cast_fp16)[name = tensor("pairs_7_cast_fp16")]; + tensor xr_7_begin_0 = const()[name = tensor("xr_7_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_7_end_0 = const()[name = tensor("xr_7_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_7_end_mask_0 = const()[name = tensor("xr_7_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_7_squeeze_mask_0 = const()[name = tensor("xr_7_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_7_cast_fp16 = slice_by_index(begin = xr_7_begin_0, end = xr_7_end_0, end_mask = xr_7_end_mask_0, squeeze_mask = xr_7_squeeze_mask_0, x = pairs_7_cast_fp16)[name = tensor("xr_7_cast_fp16")]; + tensor xi_7_begin_0 = const()[name = tensor("xi_7_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_7_end_0 = const()[name = tensor("xi_7_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_7_end_mask_0 = const()[name = tensor("xi_7_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_7_squeeze_mask_0 = const()[name = tensor("xi_7_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_7_cast_fp16 = slice_by_index(begin = xi_7_begin_0, end = xi_7_end_0, end_mask = xi_7_end_mask_0, squeeze_mask = xi_7_squeeze_mask_0, x = pairs_7_cast_fp16)[name = tensor("xi_7_cast_fp16")]; + tensor var_481_cast_fp16 = mul(x = xr_7_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_481_cast_fp16")]; + tensor var_482_cast_fp16 = mul(x = xi_7_cast_fp16, y = roti_cast_fp16)[name = tensor("op_482_cast_fp16")]; + tensor xor__7_cast_fp16 = sub(x = var_481_cast_fp16, y = var_482_cast_fp16)[name = tensor("xor__7_cast_fp16")]; + tensor var_485_cast_fp16 = mul(x = xr_7_cast_fp16, y = roti_cast_fp16)[name = tensor("op_485_cast_fp16")]; + tensor var_486_cast_fp16 = mul(x = xi_7_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_486_cast_fp16")]; + tensor xoi_7_cast_fp16 = add(x = var_485_cast_fp16, y = var_486_cast_fp16)[name = tensor("xoi_7_cast_fp16")]; + tensor out_7_axis_0 = const()[name = tensor("out_7_axis_0"), val = tensor(-1)]; + tensor out_7_cast_fp16 = stack(axis = out_7_axis_0, values = (xor__7_cast_fp16, xoi_7_cast_fp16))[name = tensor("out_7_cast_fp16")]; + tensor var_503 = const()[name = tensor("op_503"), val = tensor([256, 1024])]; + tensor var_504_cast_fp16 = reshape(shape = var_503, x = out_7_cast_fp16)[name = tensor("op_504_cast_fp16")]; + tensor written_k_3_transpose_x_1 = const()[name = tensor("written_k_3_transpose_x_1"), val = tensor(true)]; + tensor written_k_3_transpose_y_1 = const()[name = tensor("written_k_3_transpose_y_1"), val = tensor(false)]; + tensor written_k_3_cast_fp16 = matmul(transpose_x = written_k_3_transpose_x_1, transpose_y = written_k_3_transpose_y_1, x = var_106_cast_fp16_to_fp16, y = var_504_cast_fp16)[name = tensor("written_k_3_cast_fp16")]; + tensor var_511 = const()[name = tensor("op_511"), val = tensor([256, 1024])]; + tensor var_512_cast_fp16 = reshape(shape = var_511, x = v_3_cast_fp16)[name = tensor("op_512_cast_fp16")]; + tensor written_v_3_transpose_x_1 = const()[name = tensor("written_v_3_transpose_x_1"), val = tensor(true)]; + tensor written_v_3_transpose_y_1 = const()[name = tensor("written_v_3_transpose_y_1"), val = tensor(false)]; + tensor written_v_3_cast_fp16 = matmul(transpose_x = written_v_3_transpose_x_1, transpose_y = written_v_3_transpose_y_1, x = var_106_cast_fp16_to_fp16, y = var_512_cast_fp16)[name = tensor("written_v_3_cast_fp16")]; + tensor k_cache1_to_fp16_dtype_0 = const()[name = tensor("k_cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache1_to_fp16 = cast(dtype = k_cache1_to_fp16_dtype_0, x = k_cache1)[name = tensor("cast_117")]; + tensor var_523_cast_fp16 = mul(x = k_cache1_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_523_cast_fp16")]; + tensor var_528 = const()[name = tensor("op_528"), val = tensor([1, 512, 16, 64])]; + tensor var_529_cast_fp16 = reshape(shape = var_528, x = written_k_3_cast_fp16)[name = tensor("op_529_cast_fp16")]; + tensor new_k_cache_3_cast_fp16 = add(x = var_523_cast_fp16, y = var_529_cast_fp16)[name = tensor("new_k_cache_3_cast_fp16")]; + tensor new_k_cache_3_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_3_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache1_to_fp16_dtype_0 = const()[name = tensor("v_cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache1_to_fp16 = cast(dtype = v_cache1_to_fp16_dtype_0, x = v_cache1)[name = tensor("cast_115")]; + tensor var_532_cast_fp16 = mul(x = v_cache1_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_532_cast_fp16")]; + tensor var_537 = const()[name = tensor("op_537"), val = tensor([1, 512, 16, 64])]; + tensor var_538_cast_fp16 = reshape(shape = var_537, x = written_v_3_cast_fp16)[name = tensor("op_538_cast_fp16")]; + tensor new_v_cache_3_cast_fp16 = add(x = var_532_cast_fp16, y = var_538_cast_fp16)[name = tensor("new_v_cache_3_cast_fp16")]; + tensor new_v_cache_3_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_3_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_551 = const()[name = tensor("op_551"), val = tensor([0, 2, 1, 3])]; + tensor var_562_transpose_x_0 = const()[name = tensor("op_562_transpose_x_0"), val = tensor(false)]; + tensor var_562_transpose_y_0 = const()[name = tensor("op_562_transpose_y_0"), val = tensor(false)]; + tensor transpose_17_perm_0 = const()[name = tensor("transpose_17_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_18_perm_0 = const()[name = tensor("transpose_18_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_18 = transpose(perm = transpose_18_perm_0, x = new_k_cache_3_cast_fp16)[name = tensor("transpose_38")]; + tensor transpose_17 = transpose(perm = transpose_17_perm_0, x = q_3_cast_fp16)[name = tensor("transpose_39")]; + tensor var_562_cast_fp16 = matmul(transpose_x = var_562_transpose_x_0, transpose_y = var_562_transpose_y_0, x = transpose_17, y = transpose_18)[name = tensor("op_562_cast_fp16")]; + tensor var_563_to_fp16 = const()[name = tensor("op_563_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_5_cast_fp16 = mul(x = var_562_cast_fp16, y = var_563_to_fp16)[name = tensor("scores_5_cast_fp16")]; + tensor scores_7_cast_fp16 = add(x = scores_5_cast_fp16, y = var_339_cast_fp16)[name = tensor("scores_7_cast_fp16")]; + tensor var_573 = const()[name = tensor("op_573"), val = tensor(-1)]; + tensor weights_3_cast_fp16 = softmax(axis = var_573, x = scores_7_cast_fp16)[name = tensor("weights_3_cast_fp16")]; + tensor attn_3_transpose_x_0 = const()[name = tensor("attn_3_transpose_x_0"), val = tensor(false)]; + tensor attn_3_transpose_y_0 = const()[name = tensor("attn_3_transpose_y_0"), val = tensor(false)]; + tensor values_3_cast_fp16 = transpose(perm = var_551, x = new_v_cache_3_cast_fp16)[name = tensor("transpose_40")]; + tensor attn_3_cast_fp16 = matmul(transpose_x = attn_3_transpose_x_0, transpose_y = attn_3_transpose_y_0, x = weights_3_cast_fp16, y = values_3_cast_fp16)[name = tensor("attn_3_cast_fp16")]; + tensor var_581 = const()[name = tensor("op_581"), val = tensor([0, 2, 1, 3])]; + tensor var_586 = const()[name = tensor("op_586"), val = tensor([1, 256, 1024])]; + tensor var_582_cast_fp16 = transpose(perm = var_581, x = attn_3_cast_fp16)[name = tensor("transpose_37")]; + tensor input_17_cast_fp16 = reshape(shape = var_586, x = var_582_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor attn1_out_proj_weight_to_fp16 = const()[name = tensor("attn1_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31489280)))]; + tensor linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn1_out_proj_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor position1_to_fp16_dtype_0 = const()[name = tensor("position1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position1_to_fp16 = cast(dtype = position1_to_fp16_dtype_0, x = position1)[name = tensor("cast_113")]; + tensor var_595_cast_fp16 = add(x = position1_to_fp16, y = valid_len_to_fp16)[name = tensor("op_595_cast_fp16")]; + tensor var_595_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_595_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_19_cast_fp16 = add(x = input_13_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor input_21_axes_0 = const()[name = tensor("input_21_axes_0"), val = tensor([-1])]; + tensor norm1_2_weight_to_fp16 = const()[name = tensor("norm1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33586496)))]; + tensor norm1_2_bias_to_fp16 = const()[name = tensor("norm1_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33588608)))]; + tensor var_599_to_fp16 = const()[name = tensor("op_599_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_21_cast_fp16 = layer_norm(axes = input_21_axes_0, beta = norm1_2_bias_to_fp16, epsilon = var_599_to_fp16, gamma = norm1_2_weight_to_fp16, x = input_19_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor linear1_1_weight_to_fp16 = const()[name = tensor("linear1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33590720)))]; + tensor linear_6_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear1_1_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor input_23_mode_0 = const()[name = tensor("input_23_mode_0"), val = tensor("EXACT")]; + tensor input_23_cast_fp16 = gelu(mode = input_23_mode_0, x = linear_6_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor linear1_2_weight_to_fp16 = const()[name = tensor("linear1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41979392)))]; + tensor linear_7_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear1_2_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_25_cast_fp16 = add(x = input_19_cast_fp16, y = linear_7_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor input_27_axes_0 = const()[name = tensor("input_27_axes_0"), val = tensor([-1])]; + tensor norm2_1_weight_to_fp16 = const()[name = tensor("norm2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50368064)))]; + tensor norm2_1_bias_to_fp16 = const()[name = tensor("norm2_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50370176)))]; + tensor var_616_to_fp16 = const()[name = tensor("op_616_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = norm2_1_bias_to_fp16, epsilon = var_616_to_fp16, gamma = norm2_1_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor attn2_in_proj_weight_to_fp16 = const()[name = tensor("attn2_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50372288)))]; + tensor linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn2_in_proj_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor var_630 = const()[name = tensor("op_630"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_5_cast_fp16 = reshape(shape = var_630, x = linear_8_cast_fp16)[name = tensor("qkv_5_cast_fp16")]; + tensor x_9_begin_0 = const()[name = tensor("x_9_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_9_end_0 = const()[name = tensor("x_9_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_9_end_mask_0 = const()[name = tensor("x_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_9_squeeze_mask_0 = const()[name = tensor("x_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_9_cast_fp16 = slice_by_index(begin = x_9_begin_0, end = x_9_end_0, end_mask = x_9_end_mask_0, squeeze_mask = x_9_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("x_9_cast_fp16")]; + tensor x_11_begin_0 = const()[name = tensor("x_11_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_11_end_0 = const()[name = tensor("x_11_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_11_end_mask_0 = const()[name = tensor("x_11_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_11_squeeze_mask_0 = const()[name = tensor("x_11_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_11_cast_fp16 = slice_by_index(begin = x_11_begin_0, end = x_11_end_0, end_mask = x_11_end_mask_0, squeeze_mask = x_11_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("x_11_cast_fp16")]; + tensor v_5_begin_0 = const()[name = tensor("v_5_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_5_end_0 = const()[name = tensor("v_5_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_5_end_mask_0 = const()[name = tensor("v_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_5_squeeze_mask_0 = const()[name = tensor("v_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_5_cast_fp16 = slice_by_index(begin = v_5_begin_0, end = v_5_end_0, end_mask = v_5_end_mask_0, squeeze_mask = v_5_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("v_5_cast_fp16")]; + tensor var_675 = const()[name = tensor("op_675"), val = tensor([256, 16, 32, 2])]; + tensor pairs_9_cast_fp16 = reshape(shape = var_675, x = x_9_cast_fp16)[name = tensor("pairs_9_cast_fp16")]; + tensor xr_9_begin_0 = const()[name = tensor("xr_9_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_9_end_0 = const()[name = tensor("xr_9_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_9_end_mask_0 = const()[name = tensor("xr_9_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_9_squeeze_mask_0 = const()[name = tensor("xr_9_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_9_cast_fp16 = slice_by_index(begin = xr_9_begin_0, end = xr_9_end_0, end_mask = xr_9_end_mask_0, squeeze_mask = xr_9_squeeze_mask_0, x = pairs_9_cast_fp16)[name = tensor("xr_9_cast_fp16")]; + tensor xi_9_begin_0 = const()[name = tensor("xi_9_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_9_end_0 = const()[name = tensor("xi_9_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_9_end_mask_0 = const()[name = tensor("xi_9_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_9_squeeze_mask_0 = const()[name = tensor("xi_9_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_9_cast_fp16 = slice_by_index(begin = xi_9_begin_0, end = xi_9_end_0, end_mask = xi_9_end_mask_0, squeeze_mask = xi_9_squeeze_mask_0, x = pairs_9_cast_fp16)[name = tensor("xi_9_cast_fp16")]; + tensor var_683_cast_fp16 = mul(x = xr_9_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_683_cast_fp16")]; + tensor var_684_cast_fp16 = mul(x = xi_9_cast_fp16, y = roti_cast_fp16)[name = tensor("op_684_cast_fp16")]; + tensor xor__9_cast_fp16 = sub(x = var_683_cast_fp16, y = var_684_cast_fp16)[name = tensor("xor__9_cast_fp16")]; + tensor var_687_cast_fp16 = mul(x = xr_9_cast_fp16, y = roti_cast_fp16)[name = tensor("op_687_cast_fp16")]; + tensor var_688_cast_fp16 = mul(x = xi_9_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_688_cast_fp16")]; + tensor xoi_9_cast_fp16 = add(x = var_687_cast_fp16, y = var_688_cast_fp16)[name = tensor("xoi_9_cast_fp16")]; + tensor out_9_axis_0 = const()[name = tensor("out_9_axis_0"), val = tensor(-1)]; + tensor out_9_cast_fp16 = stack(axis = out_9_axis_0, values = (xor__9_cast_fp16, xoi_9_cast_fp16))[name = tensor("out_9_cast_fp16")]; + tensor var_698 = const()[name = tensor("op_698"), val = tensor([1, 256, 16, 64])]; + tensor q_5_cast_fp16 = reshape(shape = var_698, x = out_9_cast_fp16)[name = tensor("q_5_cast_fp16")]; + tensor var_704 = const()[name = tensor("op_704"), val = tensor([256, 16, 32, 2])]; + tensor pairs_11_cast_fp16 = reshape(shape = var_704, x = x_11_cast_fp16)[name = tensor("pairs_11_cast_fp16")]; + tensor xr_11_begin_0 = const()[name = tensor("xr_11_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_11_end_0 = const()[name = tensor("xr_11_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_11_end_mask_0 = const()[name = tensor("xr_11_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_11_squeeze_mask_0 = const()[name = tensor("xr_11_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_11_cast_fp16 = slice_by_index(begin = xr_11_begin_0, end = xr_11_end_0, end_mask = xr_11_end_mask_0, squeeze_mask = xr_11_squeeze_mask_0, x = pairs_11_cast_fp16)[name = tensor("xr_11_cast_fp16")]; + tensor xi_11_begin_0 = const()[name = tensor("xi_11_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_11_end_0 = const()[name = tensor("xi_11_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_11_end_mask_0 = const()[name = tensor("xi_11_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_11_squeeze_mask_0 = const()[name = tensor("xi_11_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_11_cast_fp16 = slice_by_index(begin = xi_11_begin_0, end = xi_11_end_0, end_mask = xi_11_end_mask_0, squeeze_mask = xi_11_squeeze_mask_0, x = pairs_11_cast_fp16)[name = tensor("xi_11_cast_fp16")]; + tensor var_712_cast_fp16 = mul(x = xr_11_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_712_cast_fp16")]; + tensor var_713_cast_fp16 = mul(x = xi_11_cast_fp16, y = roti_cast_fp16)[name = tensor("op_713_cast_fp16")]; + tensor xor__11_cast_fp16 = sub(x = var_712_cast_fp16, y = var_713_cast_fp16)[name = tensor("xor__11_cast_fp16")]; + tensor var_716_cast_fp16 = mul(x = xr_11_cast_fp16, y = roti_cast_fp16)[name = tensor("op_716_cast_fp16")]; + tensor var_717_cast_fp16 = mul(x = xi_11_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_717_cast_fp16")]; + tensor xoi_11_cast_fp16 = add(x = var_716_cast_fp16, y = var_717_cast_fp16)[name = tensor("xoi_11_cast_fp16")]; + tensor out_11_axis_0 = const()[name = tensor("out_11_axis_0"), val = tensor(-1)]; + tensor out_11_cast_fp16 = stack(axis = out_11_axis_0, values = (xor__11_cast_fp16, xoi_11_cast_fp16))[name = tensor("out_11_cast_fp16")]; + tensor var_734 = const()[name = tensor("op_734"), val = tensor([256, 1024])]; + tensor var_735_cast_fp16 = reshape(shape = var_734, x = out_11_cast_fp16)[name = tensor("op_735_cast_fp16")]; + tensor written_k_5_transpose_x_1 = const()[name = tensor("written_k_5_transpose_x_1"), val = tensor(true)]; + tensor written_k_5_transpose_y_1 = const()[name = tensor("written_k_5_transpose_y_1"), val = tensor(false)]; + tensor written_k_5_cast_fp16 = matmul(transpose_x = written_k_5_transpose_x_1, transpose_y = written_k_5_transpose_y_1, x = var_106_cast_fp16_to_fp16, y = var_735_cast_fp16)[name = tensor("written_k_5_cast_fp16")]; + tensor var_742 = const()[name = tensor("op_742"), val = tensor([256, 1024])]; + tensor var_743_cast_fp16 = reshape(shape = var_742, x = v_5_cast_fp16)[name = tensor("op_743_cast_fp16")]; + tensor written_v_5_transpose_x_1 = const()[name = tensor("written_v_5_transpose_x_1"), val = tensor(true)]; + tensor written_v_5_transpose_y_1 = const()[name = tensor("written_v_5_transpose_y_1"), val = tensor(false)]; + tensor written_v_5_cast_fp16 = matmul(transpose_x = written_v_5_transpose_x_1, transpose_y = written_v_5_transpose_y_1, x = var_106_cast_fp16_to_fp16, y = var_743_cast_fp16)[name = tensor("written_v_5_cast_fp16")]; + tensor k_cache2_to_fp16_dtype_0 = const()[name = tensor("k_cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache2_to_fp16 = cast(dtype = k_cache2_to_fp16_dtype_0, x = k_cache2)[name = tensor("cast_111")]; + tensor var_754_cast_fp16 = mul(x = k_cache2_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_754_cast_fp16")]; + tensor var_759 = const()[name = tensor("op_759"), val = tensor([1, 512, 16, 64])]; + tensor var_760_cast_fp16 = reshape(shape = var_759, x = written_k_5_cast_fp16)[name = tensor("op_760_cast_fp16")]; + tensor new_k_cache_5_cast_fp16 = add(x = var_754_cast_fp16, y = var_760_cast_fp16)[name = tensor("new_k_cache_5_cast_fp16")]; + tensor new_k_cache_5_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_5_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache2_to_fp16_dtype_0 = const()[name = tensor("v_cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache2_to_fp16 = cast(dtype = v_cache2_to_fp16_dtype_0, x = v_cache2)[name = tensor("cast_109")]; + tensor var_763_cast_fp16 = mul(x = v_cache2_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_763_cast_fp16")]; + tensor var_768 = const()[name = tensor("op_768"), val = tensor([1, 512, 16, 64])]; + tensor var_769_cast_fp16 = reshape(shape = var_768, x = written_v_5_cast_fp16)[name = tensor("op_769_cast_fp16")]; + tensor new_v_cache_5_cast_fp16 = add(x = var_763_cast_fp16, y = var_769_cast_fp16)[name = tensor("new_v_cache_5_cast_fp16")]; + tensor new_v_cache_5_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_5_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_782 = const()[name = tensor("op_782"), val = tensor([0, 2, 1, 3])]; + tensor var_793_transpose_x_0 = const()[name = tensor("op_793_transpose_x_0"), val = tensor(false)]; + tensor var_793_transpose_y_0 = const()[name = tensor("op_793_transpose_y_0"), val = tensor(false)]; + tensor transpose_19_perm_0 = const()[name = tensor("transpose_19_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_20_perm_0 = const()[name = tensor("transpose_20_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_20 = transpose(perm = transpose_20_perm_0, x = new_k_cache_5_cast_fp16)[name = tensor("transpose_34")]; + tensor transpose_19 = transpose(perm = transpose_19_perm_0, x = q_5_cast_fp16)[name = tensor("transpose_35")]; + tensor var_793_cast_fp16 = matmul(transpose_x = var_793_transpose_x_0, transpose_y = var_793_transpose_y_0, x = transpose_19, y = transpose_20)[name = tensor("op_793_cast_fp16")]; + tensor var_794_to_fp16 = const()[name = tensor("op_794_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_9_cast_fp16 = mul(x = var_793_cast_fp16, y = var_794_to_fp16)[name = tensor("scores_9_cast_fp16")]; + tensor scores_11_cast_fp16 = add(x = scores_9_cast_fp16, y = var_339_cast_fp16)[name = tensor("scores_11_cast_fp16")]; + tensor var_804 = const()[name = tensor("op_804"), val = tensor(-1)]; + tensor weights_5_cast_fp16 = softmax(axis = var_804, x = scores_11_cast_fp16)[name = tensor("weights_5_cast_fp16")]; + tensor attn_5_transpose_x_0 = const()[name = tensor("attn_5_transpose_x_0"), val = tensor(false)]; + tensor attn_5_transpose_y_0 = const()[name = tensor("attn_5_transpose_y_0"), val = tensor(false)]; + tensor values_5_cast_fp16 = transpose(perm = var_782, x = new_v_cache_5_cast_fp16)[name = tensor("transpose_36")]; + tensor attn_5_cast_fp16 = matmul(transpose_x = attn_5_transpose_x_0, transpose_y = attn_5_transpose_y_0, x = weights_5_cast_fp16, y = values_5_cast_fp16)[name = tensor("attn_5_cast_fp16")]; + tensor var_812 = const()[name = tensor("op_812"), val = tensor([0, 2, 1, 3])]; + tensor var_817 = const()[name = tensor("op_817"), val = tensor([1, 256, 1024])]; + tensor var_813_cast_fp16 = transpose(perm = var_812, x = attn_5_cast_fp16)[name = tensor("transpose_33")]; + tensor input_29_cast_fp16 = reshape(shape = var_817, x = var_813_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor attn2_out_proj_weight_to_fp16 = const()[name = tensor("attn2_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(56663808)))]; + tensor linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn2_out_proj_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor position2_to_fp16_dtype_0 = const()[name = tensor("position2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position2_to_fp16 = cast(dtype = position2_to_fp16_dtype_0, x = position2)[name = tensor("cast_107")]; + tensor var_826_cast_fp16 = add(x = position2_to_fp16, y = valid_len_to_fp16)[name = tensor("op_826_cast_fp16")]; + tensor var_826_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_826_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_31_cast_fp16 = add(x = input_25_cast_fp16, y = linear_9_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; + tensor norm2_2_weight_to_fp16 = const()[name = tensor("norm2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58761024)))]; + tensor norm2_2_bias_to_fp16 = const()[name = tensor("norm2_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58763136)))]; + tensor var_830_to_fp16 = const()[name = tensor("op_830_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_33_cast_fp16 = layer_norm(axes = input_33_axes_0, beta = norm2_2_bias_to_fp16, epsilon = var_830_to_fp16, gamma = norm2_2_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor linear2_1_weight_to_fp16 = const()[name = tensor("linear2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58765248)))]; + tensor linear_10_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear2_1_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor input_35_mode_0 = const()[name = tensor("input_35_mode_0"), val = tensor("EXACT")]; + tensor input_35_cast_fp16 = gelu(mode = input_35_mode_0, x = linear_10_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor linear2_2_weight_to_fp16 = const()[name = tensor("linear2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67153920)))]; + tensor linear_11_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear2_2_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor input_37_cast_fp16 = add(x = input_31_cast_fp16, y = linear_11_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor input_39_axes_0 = const()[name = tensor("input_39_axes_0"), val = tensor([-1])]; + tensor norm3_1_weight_to_fp16 = const()[name = tensor("norm3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75542592)))]; + tensor norm3_1_bias_to_fp16 = const()[name = tensor("norm3_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75544704)))]; + tensor var_847_to_fp16 = const()[name = tensor("op_847_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_39_cast_fp16 = layer_norm(axes = input_39_axes_0, beta = norm3_1_bias_to_fp16, epsilon = var_847_to_fp16, gamma = norm3_1_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor attn3_in_proj_weight_to_fp16 = const()[name = tensor("attn3_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75546816)))]; + tensor linear_12_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn3_in_proj_weight_to_fp16, x = input_39_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor var_861 = const()[name = tensor("op_861"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_7_cast_fp16 = reshape(shape = var_861, x = linear_12_cast_fp16)[name = tensor("qkv_7_cast_fp16")]; + tensor x_13_begin_0 = const()[name = tensor("x_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_13_end_0 = const()[name = tensor("x_13_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_13_end_mask_0 = const()[name = tensor("x_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_13_squeeze_mask_0 = const()[name = tensor("x_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_13_cast_fp16 = slice_by_index(begin = x_13_begin_0, end = x_13_end_0, end_mask = x_13_end_mask_0, squeeze_mask = x_13_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("x_13_cast_fp16")]; + tensor x_15_begin_0 = const()[name = tensor("x_15_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_15_end_0 = const()[name = tensor("x_15_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_15_end_mask_0 = const()[name = tensor("x_15_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_15_squeeze_mask_0 = const()[name = tensor("x_15_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_15_cast_fp16 = slice_by_index(begin = x_15_begin_0, end = x_15_end_0, end_mask = x_15_end_mask_0, squeeze_mask = x_15_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("x_15_cast_fp16")]; + tensor v_7_begin_0 = const()[name = tensor("v_7_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_7_end_0 = const()[name = tensor("v_7_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_7_end_mask_0 = const()[name = tensor("v_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_7_squeeze_mask_0 = const()[name = tensor("v_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_7_cast_fp16 = slice_by_index(begin = v_7_begin_0, end = v_7_end_0, end_mask = v_7_end_mask_0, squeeze_mask = v_7_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("v_7_cast_fp16")]; + tensor var_906 = const()[name = tensor("op_906"), val = tensor([256, 16, 32, 2])]; + tensor pairs_13_cast_fp16 = reshape(shape = var_906, x = x_13_cast_fp16)[name = tensor("pairs_13_cast_fp16")]; + tensor xr_13_begin_0 = const()[name = tensor("xr_13_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_13_end_0 = const()[name = tensor("xr_13_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_13_end_mask_0 = const()[name = tensor("xr_13_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_13_squeeze_mask_0 = const()[name = tensor("xr_13_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_13_cast_fp16 = slice_by_index(begin = xr_13_begin_0, end = xr_13_end_0, end_mask = xr_13_end_mask_0, squeeze_mask = xr_13_squeeze_mask_0, x = pairs_13_cast_fp16)[name = tensor("xr_13_cast_fp16")]; + tensor xi_13_begin_0 = const()[name = tensor("xi_13_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_13_end_0 = const()[name = tensor("xi_13_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_13_end_mask_0 = const()[name = tensor("xi_13_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_13_squeeze_mask_0 = const()[name = tensor("xi_13_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_13_cast_fp16 = slice_by_index(begin = xi_13_begin_0, end = xi_13_end_0, end_mask = xi_13_end_mask_0, squeeze_mask = xi_13_squeeze_mask_0, x = pairs_13_cast_fp16)[name = tensor("xi_13_cast_fp16")]; + tensor var_914_cast_fp16 = mul(x = xr_13_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_914_cast_fp16")]; + tensor var_915_cast_fp16 = mul(x = xi_13_cast_fp16, y = roti_cast_fp16)[name = tensor("op_915_cast_fp16")]; + tensor xor__13_cast_fp16 = sub(x = var_914_cast_fp16, y = var_915_cast_fp16)[name = tensor("xor__13_cast_fp16")]; + tensor var_918_cast_fp16 = mul(x = xr_13_cast_fp16, y = roti_cast_fp16)[name = tensor("op_918_cast_fp16")]; + tensor var_919_cast_fp16 = mul(x = xi_13_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_919_cast_fp16")]; + tensor xoi_13_cast_fp16 = add(x = var_918_cast_fp16, y = var_919_cast_fp16)[name = tensor("xoi_13_cast_fp16")]; + tensor out_13_axis_0 = const()[name = tensor("out_13_axis_0"), val = tensor(-1)]; + tensor out_13_cast_fp16 = stack(axis = out_13_axis_0, values = (xor__13_cast_fp16, xoi_13_cast_fp16))[name = tensor("out_13_cast_fp16")]; + tensor var_929 = const()[name = tensor("op_929"), val = tensor([1, 256, 16, 64])]; + tensor q_7_cast_fp16 = reshape(shape = var_929, x = out_13_cast_fp16)[name = tensor("q_7_cast_fp16")]; + tensor var_935 = const()[name = tensor("op_935"), val = tensor([256, 16, 32, 2])]; + tensor pairs_15_cast_fp16 = reshape(shape = var_935, x = x_15_cast_fp16)[name = tensor("pairs_15_cast_fp16")]; + tensor xr_15_begin_0 = const()[name = tensor("xr_15_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_15_end_0 = const()[name = tensor("xr_15_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_15_end_mask_0 = const()[name = tensor("xr_15_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_15_squeeze_mask_0 = const()[name = tensor("xr_15_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_15_cast_fp16 = slice_by_index(begin = xr_15_begin_0, end = xr_15_end_0, end_mask = xr_15_end_mask_0, squeeze_mask = xr_15_squeeze_mask_0, x = pairs_15_cast_fp16)[name = tensor("xr_15_cast_fp16")]; + tensor xi_15_begin_0 = const()[name = tensor("xi_15_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_15_end_0 = const()[name = tensor("xi_15_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_15_end_mask_0 = const()[name = tensor("xi_15_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_15_squeeze_mask_0 = const()[name = tensor("xi_15_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_15_cast_fp16 = slice_by_index(begin = xi_15_begin_0, end = xi_15_end_0, end_mask = xi_15_end_mask_0, squeeze_mask = xi_15_squeeze_mask_0, x = pairs_15_cast_fp16)[name = tensor("xi_15_cast_fp16")]; + tensor var_943_cast_fp16 = mul(x = xr_15_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_943_cast_fp16")]; + tensor var_944_cast_fp16 = mul(x = xi_15_cast_fp16, y = roti_cast_fp16)[name = tensor("op_944_cast_fp16")]; + tensor xor__15_cast_fp16 = sub(x = var_943_cast_fp16, y = var_944_cast_fp16)[name = tensor("xor__15_cast_fp16")]; + tensor var_947_cast_fp16 = mul(x = xr_15_cast_fp16, y = roti_cast_fp16)[name = tensor("op_947_cast_fp16")]; + tensor var_948_cast_fp16 = mul(x = xi_15_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_948_cast_fp16")]; + tensor xoi_15_cast_fp16 = add(x = var_947_cast_fp16, y = var_948_cast_fp16)[name = tensor("xoi_15_cast_fp16")]; + tensor out_15_axis_0 = const()[name = tensor("out_15_axis_0"), val = tensor(-1)]; + tensor out_15_cast_fp16 = stack(axis = out_15_axis_0, values = (xor__15_cast_fp16, xoi_15_cast_fp16))[name = tensor("out_15_cast_fp16")]; + tensor var_965 = const()[name = tensor("op_965"), val = tensor([256, 1024])]; + tensor var_966_cast_fp16 = reshape(shape = var_965, x = out_15_cast_fp16)[name = tensor("op_966_cast_fp16")]; + tensor written_k_7_transpose_x_1 = const()[name = tensor("written_k_7_transpose_x_1"), val = tensor(true)]; + tensor written_k_7_transpose_y_1 = const()[name = tensor("written_k_7_transpose_y_1"), val = tensor(false)]; + tensor written_k_7_cast_fp16 = matmul(transpose_x = written_k_7_transpose_x_1, transpose_y = written_k_7_transpose_y_1, x = var_106_cast_fp16_to_fp16, y = var_966_cast_fp16)[name = tensor("written_k_7_cast_fp16")]; + tensor var_973 = const()[name = tensor("op_973"), val = tensor([256, 1024])]; + tensor var_974_cast_fp16 = reshape(shape = var_973, x = v_7_cast_fp16)[name = tensor("op_974_cast_fp16")]; + tensor written_v_7_transpose_x_1 = const()[name = tensor("written_v_7_transpose_x_1"), val = tensor(true)]; + tensor written_v_7_transpose_y_1 = const()[name = tensor("written_v_7_transpose_y_1"), val = tensor(false)]; + tensor written_v_7_cast_fp16 = matmul(transpose_x = written_v_7_transpose_x_1, transpose_y = written_v_7_transpose_y_1, x = var_106_cast_fp16_to_fp16, y = var_974_cast_fp16)[name = tensor("written_v_7_cast_fp16")]; + tensor k_cache3_to_fp16_dtype_0 = const()[name = tensor("k_cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache3_to_fp16 = cast(dtype = k_cache3_to_fp16_dtype_0, x = k_cache3)[name = tensor("cast_105")]; + tensor var_985_cast_fp16 = mul(x = k_cache3_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_985_cast_fp16")]; + tensor var_990 = const()[name = tensor("op_990"), val = tensor([1, 512, 16, 64])]; + tensor var_991_cast_fp16 = reshape(shape = var_990, x = written_k_7_cast_fp16)[name = tensor("op_991_cast_fp16")]; + tensor new_k_cache_7_cast_fp16 = add(x = var_985_cast_fp16, y = var_991_cast_fp16)[name = tensor("new_k_cache_7_cast_fp16")]; + tensor new_k_cache_7_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_7_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache3_to_fp16_dtype_0 = const()[name = tensor("v_cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache3_to_fp16 = cast(dtype = v_cache3_to_fp16_dtype_0, x = v_cache3)[name = tensor("cast_103")]; + tensor var_994_cast_fp16 = mul(x = v_cache3_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_994_cast_fp16")]; + tensor var_999 = const()[name = tensor("op_999"), val = tensor([1, 512, 16, 64])]; + tensor var_1000_cast_fp16 = reshape(shape = var_999, x = written_v_7_cast_fp16)[name = tensor("op_1000_cast_fp16")]; + tensor new_v_cache_7_cast_fp16 = add(x = var_994_cast_fp16, y = var_1000_cast_fp16)[name = tensor("new_v_cache_7_cast_fp16")]; + tensor new_v_cache_7_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_7_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_1013 = const()[name = tensor("op_1013"), val = tensor([0, 2, 1, 3])]; + tensor var_1024_transpose_x_0 = const()[name = tensor("op_1024_transpose_x_0"), val = tensor(false)]; + tensor var_1024_transpose_y_0 = const()[name = tensor("op_1024_transpose_y_0"), val = tensor(false)]; + tensor transpose_21_perm_0 = const()[name = tensor("transpose_21_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_22_perm_0 = const()[name = tensor("transpose_22_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_22 = transpose(perm = transpose_22_perm_0, x = new_k_cache_7_cast_fp16)[name = tensor("transpose_30")]; + tensor transpose_21 = transpose(perm = transpose_21_perm_0, x = q_7_cast_fp16)[name = tensor("transpose_31")]; + tensor var_1024_cast_fp16 = matmul(transpose_x = var_1024_transpose_x_0, transpose_y = var_1024_transpose_y_0, x = transpose_21, y = transpose_22)[name = tensor("op_1024_cast_fp16")]; + tensor var_1025_to_fp16 = const()[name = tensor("op_1025_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_13_cast_fp16 = mul(x = var_1024_cast_fp16, y = var_1025_to_fp16)[name = tensor("scores_13_cast_fp16")]; + tensor scores_15_cast_fp16 = add(x = scores_13_cast_fp16, y = var_339_cast_fp16)[name = tensor("scores_15_cast_fp16")]; + tensor var_1035 = const()[name = tensor("op_1035"), val = tensor(-1)]; + tensor weights_7_cast_fp16 = softmax(axis = var_1035, x = scores_15_cast_fp16)[name = tensor("weights_7_cast_fp16")]; + tensor attn_7_transpose_x_0 = const()[name = tensor("attn_7_transpose_x_0"), val = tensor(false)]; + tensor attn_7_transpose_y_0 = const()[name = tensor("attn_7_transpose_y_0"), val = tensor(false)]; + tensor values_7_cast_fp16 = transpose(perm = var_1013, x = new_v_cache_7_cast_fp16)[name = tensor("transpose_32")]; + tensor attn_7_cast_fp16 = matmul(transpose_x = attn_7_transpose_x_0, transpose_y = attn_7_transpose_y_0, x = weights_7_cast_fp16, y = values_7_cast_fp16)[name = tensor("attn_7_cast_fp16")]; + tensor var_1043 = const()[name = tensor("op_1043"), val = tensor([0, 2, 1, 3])]; + tensor var_1048 = const()[name = tensor("op_1048"), val = tensor([1, 256, 1024])]; + tensor var_1044_cast_fp16 = transpose(perm = var_1043, x = attn_7_cast_fp16)[name = tensor("transpose_29")]; + tensor input_41_cast_fp16 = reshape(shape = var_1048, x = var_1044_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor attn3_out_proj_weight_to_fp16 = const()[name = tensor("attn3_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(81838336)))]; + tensor linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn3_out_proj_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor position3_to_fp16_dtype_0 = const()[name = tensor("position3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position3_to_fp16 = cast(dtype = position3_to_fp16_dtype_0, x = position3)[name = tensor("cast_101")]; + tensor var_1057_cast_fp16 = add(x = position3_to_fp16, y = valid_len_to_fp16)[name = tensor("op_1057_cast_fp16")]; + tensor var_1057_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1057_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_43_cast_fp16 = add(x = input_37_cast_fp16, y = linear_13_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; + tensor norm3_2_weight_to_fp16 = const()[name = tensor("norm3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(83935552)))]; + tensor norm3_2_bias_to_fp16 = const()[name = tensor("norm3_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(83937664)))]; + tensor var_1061_to_fp16 = const()[name = tensor("op_1061_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_45_cast_fp16 = layer_norm(axes = input_45_axes_0, beta = norm3_2_bias_to_fp16, epsilon = var_1061_to_fp16, gamma = norm3_2_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor linear3_1_weight_to_fp16 = const()[name = tensor("linear3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(83939776)))]; + tensor linear_14_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear3_1_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor input_47_mode_0 = const()[name = tensor("input_47_mode_0"), val = tensor("EXACT")]; + tensor input_47_cast_fp16 = gelu(mode = input_47_mode_0, x = linear_14_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor linear3_2_weight_to_fp16 = const()[name = tensor("linear3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(92328448)))]; + tensor linear_15_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear3_2_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor input_49_cast_fp16 = add(x = input_43_cast_fp16, y = linear_15_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor input_51_axes_0 = const()[name = tensor("input_51_axes_0"), val = tensor([-1])]; + tensor norm4_1_weight_to_fp16 = const()[name = tensor("norm4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100717120)))]; + tensor norm4_1_bias_to_fp16 = const()[name = tensor("norm4_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100719232)))]; + tensor var_1078_to_fp16 = const()[name = tensor("op_1078_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_51_cast_fp16 = layer_norm(axes = input_51_axes_0, beta = norm4_1_bias_to_fp16, epsilon = var_1078_to_fp16, gamma = norm4_1_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor attn4_in_proj_weight_to_fp16 = const()[name = tensor("attn4_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100721344)))]; + tensor linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn4_in_proj_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor var_1092 = const()[name = tensor("op_1092"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_9_cast_fp16 = reshape(shape = var_1092, x = linear_16_cast_fp16)[name = tensor("qkv_9_cast_fp16")]; + tensor x_17_begin_0 = const()[name = tensor("x_17_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_17_end_0 = const()[name = tensor("x_17_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_17_end_mask_0 = const()[name = tensor("x_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_17_squeeze_mask_0 = const()[name = tensor("x_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_17_cast_fp16 = slice_by_index(begin = x_17_begin_0, end = x_17_end_0, end_mask = x_17_end_mask_0, squeeze_mask = x_17_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("x_17_cast_fp16")]; + tensor x_19_begin_0 = const()[name = tensor("x_19_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_19_end_0 = const()[name = tensor("x_19_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_19_end_mask_0 = const()[name = tensor("x_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_19_squeeze_mask_0 = const()[name = tensor("x_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_19_cast_fp16 = slice_by_index(begin = x_19_begin_0, end = x_19_end_0, end_mask = x_19_end_mask_0, squeeze_mask = x_19_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("x_19_cast_fp16")]; + tensor v_9_begin_0 = const()[name = tensor("v_9_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_9_end_0 = const()[name = tensor("v_9_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_9_end_mask_0 = const()[name = tensor("v_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_9_squeeze_mask_0 = const()[name = tensor("v_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_9_cast_fp16 = slice_by_index(begin = v_9_begin_0, end = v_9_end_0, end_mask = v_9_end_mask_0, squeeze_mask = v_9_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("v_9_cast_fp16")]; + tensor var_1137 = const()[name = tensor("op_1137"), val = tensor([256, 16, 32, 2])]; + tensor pairs_17_cast_fp16 = reshape(shape = var_1137, x = x_17_cast_fp16)[name = tensor("pairs_17_cast_fp16")]; + tensor xr_17_begin_0 = const()[name = tensor("xr_17_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_17_end_0 = const()[name = tensor("xr_17_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_17_end_mask_0 = const()[name = tensor("xr_17_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_17_squeeze_mask_0 = const()[name = tensor("xr_17_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_17_cast_fp16 = slice_by_index(begin = xr_17_begin_0, end = xr_17_end_0, end_mask = xr_17_end_mask_0, squeeze_mask = xr_17_squeeze_mask_0, x = pairs_17_cast_fp16)[name = tensor("xr_17_cast_fp16")]; + tensor xi_17_begin_0 = const()[name = tensor("xi_17_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_17_end_0 = const()[name = tensor("xi_17_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_17_end_mask_0 = const()[name = tensor("xi_17_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_17_squeeze_mask_0 = const()[name = tensor("xi_17_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_17_cast_fp16 = slice_by_index(begin = xi_17_begin_0, end = xi_17_end_0, end_mask = xi_17_end_mask_0, squeeze_mask = xi_17_squeeze_mask_0, x = pairs_17_cast_fp16)[name = tensor("xi_17_cast_fp16")]; + tensor var_1145_cast_fp16 = mul(x = xr_17_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1145_cast_fp16")]; + tensor var_1146_cast_fp16 = mul(x = xi_17_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1146_cast_fp16")]; + tensor xor__17_cast_fp16 = sub(x = var_1145_cast_fp16, y = var_1146_cast_fp16)[name = tensor("xor__17_cast_fp16")]; + tensor var_1149_cast_fp16 = mul(x = xr_17_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1149_cast_fp16")]; + tensor var_1150_cast_fp16 = mul(x = xi_17_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1150_cast_fp16")]; + tensor xoi_17_cast_fp16 = add(x = var_1149_cast_fp16, y = var_1150_cast_fp16)[name = tensor("xoi_17_cast_fp16")]; + tensor out_17_axis_0 = const()[name = tensor("out_17_axis_0"), val = tensor(-1)]; + tensor out_17_cast_fp16 = stack(axis = out_17_axis_0, values = (xor__17_cast_fp16, xoi_17_cast_fp16))[name = tensor("out_17_cast_fp16")]; + tensor var_1160 = const()[name = tensor("op_1160"), val = tensor([1, 256, 16, 64])]; + tensor q_9_cast_fp16 = reshape(shape = var_1160, x = out_17_cast_fp16)[name = tensor("q_9_cast_fp16")]; + tensor var_1166 = const()[name = tensor("op_1166"), val = tensor([256, 16, 32, 2])]; + tensor pairs_19_cast_fp16 = reshape(shape = var_1166, x = x_19_cast_fp16)[name = tensor("pairs_19_cast_fp16")]; + tensor xr_19_begin_0 = const()[name = tensor("xr_19_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_19_end_0 = const()[name = tensor("xr_19_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_19_end_mask_0 = const()[name = tensor("xr_19_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_19_squeeze_mask_0 = const()[name = tensor("xr_19_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_19_cast_fp16 = slice_by_index(begin = xr_19_begin_0, end = xr_19_end_0, end_mask = xr_19_end_mask_0, squeeze_mask = xr_19_squeeze_mask_0, x = pairs_19_cast_fp16)[name = tensor("xr_19_cast_fp16")]; + tensor xi_19_begin_0 = const()[name = tensor("xi_19_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_19_end_0 = const()[name = tensor("xi_19_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_19_end_mask_0 = const()[name = tensor("xi_19_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_19_squeeze_mask_0 = const()[name = tensor("xi_19_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_19_cast_fp16 = slice_by_index(begin = xi_19_begin_0, end = xi_19_end_0, end_mask = xi_19_end_mask_0, squeeze_mask = xi_19_squeeze_mask_0, x = pairs_19_cast_fp16)[name = tensor("xi_19_cast_fp16")]; + tensor var_1174_cast_fp16 = mul(x = xr_19_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1174_cast_fp16")]; + tensor var_1175_cast_fp16 = mul(x = xi_19_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1175_cast_fp16")]; + tensor xor__19_cast_fp16 = sub(x = var_1174_cast_fp16, y = var_1175_cast_fp16)[name = tensor("xor__19_cast_fp16")]; + tensor var_1178_cast_fp16 = mul(x = xr_19_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1178_cast_fp16")]; + tensor var_1179_cast_fp16 = mul(x = xi_19_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1179_cast_fp16")]; + tensor xoi_19_cast_fp16 = add(x = var_1178_cast_fp16, y = var_1179_cast_fp16)[name = tensor("xoi_19_cast_fp16")]; + tensor out_19_axis_0 = const()[name = tensor("out_19_axis_0"), val = tensor(-1)]; + tensor out_19_cast_fp16 = stack(axis = out_19_axis_0, values = (xor__19_cast_fp16, xoi_19_cast_fp16))[name = tensor("out_19_cast_fp16")]; + tensor var_1196 = const()[name = tensor("op_1196"), val = tensor([256, 1024])]; + tensor var_1197_cast_fp16 = reshape(shape = var_1196, x = out_19_cast_fp16)[name = tensor("op_1197_cast_fp16")]; + tensor written_k_9_transpose_x_1 = const()[name = tensor("written_k_9_transpose_x_1"), val = tensor(true)]; + tensor written_k_9_transpose_y_1 = const()[name = tensor("written_k_9_transpose_y_1"), val = tensor(false)]; + tensor written_k_9_cast_fp16 = matmul(transpose_x = written_k_9_transpose_x_1, transpose_y = written_k_9_transpose_y_1, x = var_106_cast_fp16_to_fp16, y = var_1197_cast_fp16)[name = tensor("written_k_9_cast_fp16")]; + tensor var_1204 = const()[name = tensor("op_1204"), val = tensor([256, 1024])]; + tensor var_1205_cast_fp16 = reshape(shape = var_1204, x = v_9_cast_fp16)[name = tensor("op_1205_cast_fp16")]; + tensor written_v_9_transpose_x_1 = const()[name = tensor("written_v_9_transpose_x_1"), val = tensor(true)]; + tensor written_v_9_transpose_y_1 = const()[name = tensor("written_v_9_transpose_y_1"), val = tensor(false)]; + tensor written_v_9_cast_fp16 = matmul(transpose_x = written_v_9_transpose_x_1, transpose_y = written_v_9_transpose_y_1, x = var_106_cast_fp16_to_fp16, y = var_1205_cast_fp16)[name = tensor("written_v_9_cast_fp16")]; + tensor k_cache4_to_fp16_dtype_0 = const()[name = tensor("k_cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache4_to_fp16 = cast(dtype = k_cache4_to_fp16_dtype_0, x = k_cache4)[name = tensor("cast_99")]; + tensor var_1216_cast_fp16 = mul(x = k_cache4_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_1216_cast_fp16")]; + tensor var_1221 = const()[name = tensor("op_1221"), val = tensor([1, 512, 16, 64])]; + tensor var_1222_cast_fp16 = reshape(shape = var_1221, x = written_k_9_cast_fp16)[name = tensor("op_1222_cast_fp16")]; + tensor new_k_cache_9_cast_fp16 = add(x = var_1216_cast_fp16, y = var_1222_cast_fp16)[name = tensor("new_k_cache_9_cast_fp16")]; + tensor new_k_cache_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache4_to_fp16_dtype_0 = const()[name = tensor("v_cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache4_to_fp16 = cast(dtype = v_cache4_to_fp16_dtype_0, x = v_cache4)[name = tensor("cast_97")]; + tensor var_1225_cast_fp16 = mul(x = v_cache4_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_1225_cast_fp16")]; + tensor var_1230 = const()[name = tensor("op_1230"), val = tensor([1, 512, 16, 64])]; + tensor var_1231_cast_fp16 = reshape(shape = var_1230, x = written_v_9_cast_fp16)[name = tensor("op_1231_cast_fp16")]; + tensor new_v_cache_9_cast_fp16 = add(x = var_1225_cast_fp16, y = var_1231_cast_fp16)[name = tensor("new_v_cache_9_cast_fp16")]; + tensor new_v_cache_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_1244 = const()[name = tensor("op_1244"), val = tensor([0, 2, 1, 3])]; + tensor var_1255_transpose_x_0 = const()[name = tensor("op_1255_transpose_x_0"), val = tensor(false)]; + tensor var_1255_transpose_y_0 = const()[name = tensor("op_1255_transpose_y_0"), val = tensor(false)]; + tensor transpose_23_perm_0 = const()[name = tensor("transpose_23_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_24_perm_0 = const()[name = tensor("transpose_24_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_24 = transpose(perm = transpose_24_perm_0, x = new_k_cache_9_cast_fp16)[name = tensor("transpose_26")]; + tensor transpose_23 = transpose(perm = transpose_23_perm_0, x = q_9_cast_fp16)[name = tensor("transpose_27")]; + tensor var_1255_cast_fp16 = matmul(transpose_x = var_1255_transpose_x_0, transpose_y = var_1255_transpose_y_0, x = transpose_23, y = transpose_24)[name = tensor("op_1255_cast_fp16")]; + tensor var_1256_to_fp16 = const()[name = tensor("op_1256_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_17_cast_fp16 = mul(x = var_1255_cast_fp16, y = var_1256_to_fp16)[name = tensor("scores_17_cast_fp16")]; + tensor scores_19_cast_fp16 = add(x = scores_17_cast_fp16, y = var_339_cast_fp16)[name = tensor("scores_19_cast_fp16")]; + tensor var_1266 = const()[name = tensor("op_1266"), val = tensor(-1)]; + tensor weights_9_cast_fp16 = softmax(axis = var_1266, x = scores_19_cast_fp16)[name = tensor("weights_9_cast_fp16")]; + tensor attn_9_transpose_x_0 = const()[name = tensor("attn_9_transpose_x_0"), val = tensor(false)]; + tensor attn_9_transpose_y_0 = const()[name = tensor("attn_9_transpose_y_0"), val = tensor(false)]; + tensor values_9_cast_fp16 = transpose(perm = var_1244, x = new_v_cache_9_cast_fp16)[name = tensor("transpose_28")]; + tensor attn_9_cast_fp16 = matmul(transpose_x = attn_9_transpose_x_0, transpose_y = attn_9_transpose_y_0, x = weights_9_cast_fp16, y = values_9_cast_fp16)[name = tensor("attn_9_cast_fp16")]; + tensor var_1274 = const()[name = tensor("op_1274"), val = tensor([0, 2, 1, 3])]; + tensor var_1279 = const()[name = tensor("op_1279"), val = tensor([1, 256, 1024])]; + tensor var_1275_cast_fp16 = transpose(perm = var_1274, x = attn_9_cast_fp16)[name = tensor("transpose_25")]; + tensor input_53_cast_fp16 = reshape(shape = var_1279, x = var_1275_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor attn4_out_proj_weight_to_fp16 = const()[name = tensor("attn4_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(107012864)))]; + tensor linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn4_out_proj_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor position4_to_fp16_dtype_0 = const()[name = tensor("position4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position4_to_fp16 = cast(dtype = position4_to_fp16_dtype_0, x = position4)[name = tensor("cast_95")]; + tensor var_1288_cast_fp16 = add(x = position4_to_fp16, y = valid_len_to_fp16)[name = tensor("op_1288_cast_fp16")]; + tensor var_1288_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1288_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_55_cast_fp16 = add(x = input_49_cast_fp16, y = linear_17_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor input_57_axes_0 = const()[name = tensor("input_57_axes_0"), val = tensor([-1])]; + tensor norm4_2_weight_to_fp16 = const()[name = tensor("norm4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109110080)))]; + tensor norm4_2_bias_to_fp16 = const()[name = tensor("norm4_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109112192)))]; + tensor var_1292_to_fp16 = const()[name = tensor("op_1292_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_57_cast_fp16 = layer_norm(axes = input_57_axes_0, beta = norm4_2_bias_to_fp16, epsilon = var_1292_to_fp16, gamma = norm4_2_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor linear4_1_weight_to_fp16 = const()[name = tensor("linear4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109114304)))]; + tensor linear_18_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear4_1_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor input_59_mode_0 = const()[name = tensor("input_59_mode_0"), val = tensor("EXACT")]; + tensor input_59_cast_fp16 = gelu(mode = input_59_mode_0, x = linear_18_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor linear4_2_weight_to_fp16 = const()[name = tensor("linear4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(117502976)))]; + tensor linear_19_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear4_2_weight_to_fp16, x = input_59_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_61_cast_fp16 = add(x = input_55_cast_fp16, y = linear_19_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor input_63_axes_0 = const()[name = tensor("input_63_axes_0"), val = tensor([-1])]; + tensor norm5_1_weight_to_fp16 = const()[name = tensor("norm5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125891648)))]; + tensor norm5_1_bias_to_fp16 = const()[name = tensor("norm5_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125893760)))]; + tensor var_1309_to_fp16 = const()[name = tensor("op_1309_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_63_cast_fp16 = layer_norm(axes = input_63_axes_0, beta = norm5_1_bias_to_fp16, epsilon = var_1309_to_fp16, gamma = norm5_1_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor attn5_in_proj_weight_to_fp16 = const()[name = tensor("attn5_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125895872)))]; + tensor linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn5_in_proj_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor var_1323 = const()[name = tensor("op_1323"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_cast_fp16 = reshape(shape = var_1323, x = linear_20_cast_fp16)[name = tensor("qkv_cast_fp16")]; + tensor x_begin_0 = const()[name = tensor("x_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_end_0 = const()[name = tensor("x_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_end_mask_0 = const()[name = tensor("x_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_squeeze_mask_0 = const()[name = tensor("x_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_cast_fp16 = slice_by_index(begin = x_begin_0, end = x_end_0, end_mask = x_end_mask_0, squeeze_mask = x_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("x_cast_fp16")]; + tensor v_begin_0 = const()[name = tensor("v_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_end_0 = const()[name = tensor("v_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_end_mask_0 = const()[name = tensor("v_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_squeeze_mask_0 = const()[name = tensor("v_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_cast_fp16 = slice_by_index(begin = v_begin_0, end = v_end_0, end_mask = v_end_mask_0, squeeze_mask = v_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("v_cast_fp16")]; + tensor var_1355 = const()[name = tensor("op_1355"), val = tensor([256, 16, 32, 2])]; + tensor pairs_cast_fp16 = reshape(shape = var_1355, x = x_cast_fp16)[name = tensor("pairs_cast_fp16")]; + tensor xr_begin_0 = const()[name = tensor("xr_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_end_0 = const()[name = tensor("xr_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_end_mask_0 = const()[name = tensor("xr_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_squeeze_mask_0 = const()[name = tensor("xr_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_cast_fp16 = slice_by_index(begin = xr_begin_0, end = xr_end_0, end_mask = xr_end_mask_0, squeeze_mask = xr_squeeze_mask_0, x = pairs_cast_fp16)[name = tensor("xr_cast_fp16")]; + tensor xi_begin_0 = const()[name = tensor("xi_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_end_0 = const()[name = tensor("xi_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_end_mask_0 = const()[name = tensor("xi_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_squeeze_mask_0 = const()[name = tensor("xi_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_cast_fp16 = slice_by_index(begin = xi_begin_0, end = xi_end_0, end_mask = xi_end_mask_0, squeeze_mask = xi_squeeze_mask_0, x = pairs_cast_fp16)[name = tensor("xi_cast_fp16")]; + tensor var_1363_cast_fp16 = mul(x = xr_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1363_cast_fp16")]; + tensor var_1364_cast_fp16 = mul(x = xi_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1364_cast_fp16")]; + tensor xor__cast_fp16 = sub(x = var_1363_cast_fp16, y = var_1364_cast_fp16)[name = tensor("xor__cast_fp16")]; + tensor var_1367_cast_fp16 = mul(x = xr_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1367_cast_fp16")]; + tensor var_1368_cast_fp16 = mul(x = xi_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1368_cast_fp16")]; + tensor xoi_cast_fp16 = add(x = var_1367_cast_fp16, y = var_1368_cast_fp16)[name = tensor("xoi_cast_fp16")]; + tensor out_axis_0 = const()[name = tensor("out_axis_0"), val = tensor(-1)]; + tensor out_cast_fp16 = stack(axis = out_axis_0, values = (xor__cast_fp16, xoi_cast_fp16))[name = tensor("out_cast_fp16")]; + tensor var_1385 = const()[name = tensor("op_1385"), val = tensor([256, 1024])]; + tensor var_1386_cast_fp16 = reshape(shape = var_1385, x = out_cast_fp16)[name = tensor("op_1386_cast_fp16")]; + tensor written_k_transpose_x_1 = const()[name = tensor("written_k_transpose_x_1"), val = tensor(true)]; + tensor written_k_transpose_y_1 = const()[name = tensor("written_k_transpose_y_1"), val = tensor(false)]; + tensor written_k_cast_fp16 = matmul(transpose_x = written_k_transpose_x_1, transpose_y = written_k_transpose_y_1, x = var_106_cast_fp16_to_fp16, y = var_1386_cast_fp16)[name = tensor("written_k_cast_fp16")]; + tensor var_1393 = const()[name = tensor("op_1393"), val = tensor([256, 1024])]; + tensor var_1394_cast_fp16 = reshape(shape = var_1393, x = v_cast_fp16)[name = tensor("op_1394_cast_fp16")]; + tensor written_v_transpose_x_1 = const()[name = tensor("written_v_transpose_x_1"), val = tensor(true)]; + tensor written_v_transpose_y_1 = const()[name = tensor("written_v_transpose_y_1"), val = tensor(false)]; + tensor written_v_cast_fp16 = matmul(transpose_x = written_v_transpose_x_1, transpose_y = written_v_transpose_y_1, x = var_106_cast_fp16_to_fp16, y = var_1394_cast_fp16)[name = tensor("written_v_cast_fp16")]; + tensor k_cache5_to_fp16_dtype_0 = const()[name = tensor("k_cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache5_to_fp16 = cast(dtype = k_cache5_to_fp16_dtype_0, x = k_cache5)[name = tensor("cast_93")]; + tensor var_1405_cast_fp16 = mul(x = k_cache5_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_1405_cast_fp16")]; + tensor var_1410 = const()[name = tensor("op_1410"), val = tensor([1, 512, 16, 64])]; + tensor var_1411_cast_fp16 = reshape(shape = var_1410, x = written_k_cast_fp16)[name = tensor("op_1411_cast_fp16")]; + tensor new_k_cache_cast_fp16 = add(x = var_1405_cast_fp16, y = var_1411_cast_fp16)[name = tensor("new_k_cache_cast_fp16")]; + tensor new_k_cache_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache5_to_fp16_dtype_0 = const()[name = tensor("v_cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache5_to_fp16 = cast(dtype = v_cache5_to_fp16_dtype_0, x = v_cache5)[name = tensor("cast_91")]; + tensor var_1414_cast_fp16 = mul(x = v_cache5_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_1414_cast_fp16")]; + tensor var_1419 = const()[name = tensor("op_1419"), val = tensor([1, 512, 16, 64])]; + tensor var_1420_cast_fp16 = reshape(shape = var_1419, x = written_v_cast_fp16)[name = tensor("op_1420_cast_fp16")]; + tensor new_v_cache_cast_fp16 = add(x = var_1414_cast_fp16, y = var_1420_cast_fp16)[name = tensor("new_v_cache_cast_fp16")]; + tensor new_v_cache_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor position5_to_fp16_dtype_0 = const()[name = tensor("position5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position5_to_fp16 = cast(dtype = position5_to_fp16_dtype_0, x = position5)[name = tensor("cast_89")]; + tensor var_1427_cast_fp16 = add(x = position5_to_fp16, y = valid_len_to_fp16)[name = tensor("op_1427_cast_fp16")]; + tensor var_1427_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1427_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor new_position5 = cast(dtype = var_1427_cast_fp16_to_fp32_dtype_0, x = var_1427_cast_fp16)[name = tensor("cast_88")]; + tensor new_v_cache5 = cast(dtype = new_v_cache_cast_fp16_to_fp32_dtype_0, x = new_v_cache_cast_fp16)[name = tensor("cast_90")]; + tensor new_k_cache5 = cast(dtype = new_k_cache_cast_fp16_to_fp32_dtype_0, x = new_k_cache_cast_fp16)[name = tensor("cast_92")]; + tensor new_position4 = cast(dtype = var_1288_cast_fp16_to_fp32_dtype_0, x = var_1288_cast_fp16)[name = tensor("cast_94")]; + tensor new_v_cache4 = cast(dtype = new_v_cache_9_cast_fp16_to_fp32_dtype_0, x = new_v_cache_9_cast_fp16)[name = tensor("cast_96")]; + tensor new_k_cache4 = cast(dtype = new_k_cache_9_cast_fp16_to_fp32_dtype_0, x = new_k_cache_9_cast_fp16)[name = tensor("cast_98")]; + tensor new_position3 = cast(dtype = var_1057_cast_fp16_to_fp32_dtype_0, x = var_1057_cast_fp16)[name = tensor("cast_100")]; + tensor new_v_cache3 = cast(dtype = new_v_cache_7_cast_fp16_to_fp32_dtype_0, x = new_v_cache_7_cast_fp16)[name = tensor("cast_102")]; + tensor new_k_cache3 = cast(dtype = new_k_cache_7_cast_fp16_to_fp32_dtype_0, x = new_k_cache_7_cast_fp16)[name = tensor("cast_104")]; + tensor new_position2 = cast(dtype = var_826_cast_fp16_to_fp32_dtype_0, x = var_826_cast_fp16)[name = tensor("cast_106")]; + tensor new_v_cache2 = cast(dtype = new_v_cache_5_cast_fp16_to_fp32_dtype_0, x = new_v_cache_5_cast_fp16)[name = tensor("cast_108")]; + tensor new_k_cache2 = cast(dtype = new_k_cache_5_cast_fp16_to_fp32_dtype_0, x = new_k_cache_5_cast_fp16)[name = tensor("cast_110")]; + tensor new_position1 = cast(dtype = var_595_cast_fp16_to_fp32_dtype_0, x = var_595_cast_fp16)[name = tensor("cast_112")]; + tensor new_v_cache1 = cast(dtype = new_v_cache_3_cast_fp16_to_fp32_dtype_0, x = new_v_cache_3_cast_fp16)[name = tensor("cast_114")]; + tensor new_k_cache1 = cast(dtype = new_k_cache_3_cast_fp16_to_fp32_dtype_0, x = new_k_cache_3_cast_fp16)[name = tensor("cast_116")]; + tensor new_position0 = cast(dtype = var_364_cast_fp16_to_fp32_dtype_0, x = var_135_cast_fp16)[name = tensor("cast_118")]; + tensor new_v_cache0 = cast(dtype = new_v_cache_1_cast_fp16_to_fp32_dtype_0, x = new_v_cache_1_cast_fp16)[name = tensor("cast_119")]; + tensor new_k_cache0 = cast(dtype = new_k_cache_1_cast_fp16_to_fp32_dtype_0, x = new_k_cache_1_cast_fp16)[name = tensor("cast_121")]; + } -> (new_k_cache0, new_v_cache0, new_position0, new_k_cache1, new_v_cache1, new_position1, new_k_cache2, new_v_cache2, new_position2, new_k_cache3, new_v_cache3, new_position3, new_k_cache4, new_v_cache4, new_position4, new_k_cache5, new_v_cache5, new_position5); +} \ No newline at end of file diff --git a/v2.1/portuguese/cond_prefill_ane.mlmodelc/weights/weight.bin b/v2.1/portuguese/cond_prefill_ane.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..3af36c48a29a5ce05e49bbd125eeb844caf0d0a8 --- /dev/null +++ b/v2.1/portuguese/cond_prefill_ane.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0f70cf8e384d088c29e1cb2962593a43a622f25e4b1f2198adb0332503d920b0 +size 132187392 diff --git a/v2.1/portuguese/cond_prefill_ane.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/portuguese/cond_prefill_ane.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..2da27e37217066d95cf0f58a56343c73c52bacbf --- /dev/null +++ b/v2.1/portuguese/cond_prefill_ane.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:93327d50d4285a02206e3a302f28280f2c8f3d3bf1b9880e0686a691b056dbd6 +size 119086 diff --git a/v2.1/portuguese/cond_prefill_ane.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/portuguese/cond_prefill_ane.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..3af36c48a29a5ce05e49bbd125eeb844caf0d0a8 --- /dev/null +++ b/v2.1/portuguese/cond_prefill_ane.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0f70cf8e384d088c29e1cb2962593a43a622f25e4b1f2198adb0332503d920b0 +size 132187392 diff --git a/v2.1/portuguese/cond_prefill_ane.mlpackage/Manifest.json b/v2.1/portuguese/cond_prefill_ane.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..e358388edb6fa8601caede8fb704188c0ddb8c1a --- /dev/null +++ b/v2.1/portuguese/cond_prefill_ane.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "66AEC651-8EFE-4104-ADA5-84B4B731310E": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "B5EA83D8-01C4-4EDF-95F8-A9F1F018AC4A": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "66AEC651-8EFE-4104-ADA5-84B4B731310E" +} diff --git a/v2.1/portuguese/flowlm_step_ane.mlmodelc/analytics/coremldata.bin b/v2.1/portuguese/flowlm_step_ane.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..b083beddba6ae9d3e3ec46b5c54ca1c1e799e7fb --- /dev/null +++ b/v2.1/portuguese/flowlm_step_ane.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e40fda940b03580d26bc27cf185e67d7fede40d3be59430f38baebaef1ff83c8 +size 243 diff --git a/v2.1/portuguese/flowlm_step_ane.mlmodelc/coremldata.bin b/v2.1/portuguese/flowlm_step_ane.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..a48b8bd83053805276a370fd316948827dbc602f --- /dev/null +++ b/v2.1/portuguese/flowlm_step_ane.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a4cfa942154903778b6d4dbc0c301cc966a60c0904978335740d1880e7b09569 +size 1414 diff --git a/v2.1/portuguese/flowlm_step_ane.mlmodelc/model.mil b/v2.1/portuguese/flowlm_step_ane.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..bce41c8a0bac7043d2c997eb77ff01df0c237aad --- /dev/null +++ b/v2.1/portuguese/flowlm_step_ane.mlmodelc/model.mil @@ -0,0 +1,945 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.9.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor k_cache0, tensor k_cache1, tensor k_cache2, tensor k_cache3, tensor k_cache4, tensor k_cache5, tensor position0, tensor position1, tensor position2, tensor position3, tensor position4, tensor position5, tensor sequence, tensor v_cache0, tensor v_cache1, tensor v_cache2, tensor v_cache3, tensor v_cache4, tensor v_cache5) { + tensor sequence_to_fp16_dtype_0 = const()[name = tensor("sequence_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor input_linear_weight_to_fp16 = const()[name = tensor("input_linear_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(65664)))]; + tensor sequence_to_fp16 = cast(dtype = sequence_to_fp16_dtype_0, x = sequence)[name = tensor("cast_146")]; + tensor linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = input_linear_weight_to_fp16, x = sequence_to_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; + tensor norm0_1_weight_to_fp16 = const()[name = tensor("norm0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67776)))]; + tensor norm0_1_bias_to_fp16 = const()[name = tensor("norm0_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(69888)))]; + tensor var_65_to_fp16 = const()[name = tensor("op_65_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_5_cast_fp16 = layer_norm(axes = input_5_axes_0, beta = norm0_1_bias_to_fp16, epsilon = var_65_to_fp16, gamma = norm0_1_weight_to_fp16, x = linear_0_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor attn0_in_proj_weight_to_fp16 = const()[name = tensor("attn0_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(72000)))]; + tensor linear_1_bias_0_to_fp16 = const()[name = tensor("linear_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6363520)))]; + tensor linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn0_in_proj_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("linear_1_cast_fp16")]; + tensor var_78 = const()[name = tensor("op_78"), val = tensor([1, 3, 16, 64])]; + tensor qkv_1_cast_fp16 = reshape(shape = var_78, x = linear_1_cast_fp16)[name = tensor("qkv_1_cast_fp16")]; + tensor x_1_begin_0 = const()[name = tensor("x_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_1_end_0 = const()[name = tensor("x_1_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_1_end_mask_0 = const()[name = tensor("x_1_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_1_squeeze_mask_0 = const()[name = tensor("x_1_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_1_cast_fp16 = slice_by_index(begin = x_1_begin_0, end = x_1_end_0, end_mask = x_1_end_mask_0, squeeze_mask = x_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("x_1_cast_fp16")]; + tensor x_3_begin_0 = const()[name = tensor("x_3_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_3_end_0 = const()[name = tensor("x_3_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_3_end_mask_0 = const()[name = tensor("x_3_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_3_squeeze_mask_0 = const()[name = tensor("x_3_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_3_cast_fp16 = slice_by_index(begin = x_3_begin_0, end = x_3_end_0, end_mask = x_3_end_mask_0, squeeze_mask = x_3_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("x_3_cast_fp16")]; + tensor v_1_begin_0 = const()[name = tensor("v_1_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_1_end_0 = const()[name = tensor("v_1_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_1_end_mask_0 = const()[name = tensor("v_1_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_1_squeeze_mask_0 = const()[name = tensor("v_1_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_1_cast_fp16 = slice_by_index(begin = v_1_begin_0, end = v_1_end_0, end_mask = v_1_end_mask_0, squeeze_mask = v_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("v_1_cast_fp16")]; + tensor var_107 = const()[name = tensor("op_107"), val = tensor([1, 1, 1])]; + tensor position0_to_fp16_dtype_0 = const()[name = tensor("position0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position0_to_fp16 = cast(dtype = position0_to_fp16_dtype_0, x = position0)[name = tensor("cast_145")]; + tensor var_108_cast_fp16 = reshape(shape = var_107, x = position0_to_fp16)[name = tensor("op_108_cast_fp16")]; + tensor rope_freqs_to_fp16 = const()[name = tensor("rope_freqs_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6369728)))]; + tensor angles_1_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_108_cast_fp16)[name = tensor("angles_1_cast_fp16")]; + tensor rotr_1_cast_fp16 = cos(x = angles_1_cast_fp16)[name = tensor("rotr_1_cast_fp16")]; + tensor roti_1_cast_fp16 = sin(x = angles_1_cast_fp16)[name = tensor("roti_1_cast_fp16")]; + tensor var_116 = const()[name = tensor("op_116"), val = tensor([1, 16, 32, 2])]; + tensor pairs_1_cast_fp16 = reshape(shape = var_116, x = x_1_cast_fp16)[name = tensor("pairs_1_cast_fp16")]; + tensor xr_1_begin_0 = const()[name = tensor("xr_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_1_end_0 = const()[name = tensor("xr_1_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_1_end_mask_0 = const()[name = tensor("xr_1_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_1_squeeze_mask_0 = const()[name = tensor("xr_1_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_1_cast_fp16 = slice_by_index(begin = xr_1_begin_0, end = xr_1_end_0, end_mask = xr_1_end_mask_0, squeeze_mask = xr_1_squeeze_mask_0, x = pairs_1_cast_fp16)[name = tensor("xr_1_cast_fp16")]; + tensor xi_1_begin_0 = const()[name = tensor("xi_1_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_1_end_0 = const()[name = tensor("xi_1_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_1_end_mask_0 = const()[name = tensor("xi_1_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_1_squeeze_mask_0 = const()[name = tensor("xi_1_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_1_cast_fp16 = slice_by_index(begin = xi_1_begin_0, end = xi_1_end_0, end_mask = xi_1_end_mask_0, squeeze_mask = xi_1_squeeze_mask_0, x = pairs_1_cast_fp16)[name = tensor("xi_1_cast_fp16")]; + tensor var_124_cast_fp16 = mul(x = xr_1_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_124_cast_fp16")]; + tensor var_125_cast_fp16 = mul(x = xi_1_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_125_cast_fp16")]; + tensor xor__1_cast_fp16 = sub(x = var_124_cast_fp16, y = var_125_cast_fp16)[name = tensor("xor__1_cast_fp16")]; + tensor var_128_cast_fp16 = mul(x = xr_1_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_128_cast_fp16")]; + tensor var_129_cast_fp16 = mul(x = xi_1_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_129_cast_fp16")]; + tensor xoi_1_cast_fp16 = add(x = var_128_cast_fp16, y = var_129_cast_fp16)[name = tensor("xoi_1_cast_fp16")]; + tensor out_1_axis_0 = const()[name = tensor("out_1_axis_0"), val = tensor(-1)]; + tensor out_1_cast_fp16 = stack(axis = out_1_axis_0, values = (xor__1_cast_fp16, xoi_1_cast_fp16))[name = tensor("out_1_cast_fp16")]; + tensor var_152 = const()[name = tensor("op_152"), val = tensor([1, 16, 32, 2])]; + tensor pairs_3_cast_fp16 = reshape(shape = var_152, x = x_3_cast_fp16)[name = tensor("pairs_3_cast_fp16")]; + tensor xr_3_begin_0 = const()[name = tensor("xr_3_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_3_end_0 = const()[name = tensor("xr_3_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_3_end_mask_0 = const()[name = tensor("xr_3_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_3_squeeze_mask_0 = const()[name = tensor("xr_3_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_3_cast_fp16 = slice_by_index(begin = xr_3_begin_0, end = xr_3_end_0, end_mask = xr_3_end_mask_0, squeeze_mask = xr_3_squeeze_mask_0, x = pairs_3_cast_fp16)[name = tensor("xr_3_cast_fp16")]; + tensor xi_3_begin_0 = const()[name = tensor("xi_3_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_3_end_0 = const()[name = tensor("xi_3_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_3_end_mask_0 = const()[name = tensor("xi_3_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_3_squeeze_mask_0 = const()[name = tensor("xi_3_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_3_cast_fp16 = slice_by_index(begin = xi_3_begin_0, end = xi_3_end_0, end_mask = xi_3_end_mask_0, squeeze_mask = xi_3_squeeze_mask_0, x = pairs_3_cast_fp16)[name = tensor("xi_3_cast_fp16")]; + tensor var_160_cast_fp16 = mul(x = xr_3_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_160_cast_fp16")]; + tensor var_161_cast_fp16 = mul(x = xi_3_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_161_cast_fp16")]; + tensor xor__3_cast_fp16 = sub(x = var_160_cast_fp16, y = var_161_cast_fp16)[name = tensor("xor__3_cast_fp16")]; + tensor var_164_cast_fp16 = mul(x = xr_3_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_164_cast_fp16")]; + tensor var_165_cast_fp16 = mul(x = xi_3_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_165_cast_fp16")]; + tensor xoi_3_cast_fp16 = add(x = var_164_cast_fp16, y = var_165_cast_fp16)[name = tensor("xoi_3_cast_fp16")]; + tensor out_3_axis_0 = const()[name = tensor("out_3_axis_0"), val = tensor(-1)]; + tensor out_3_cast_fp16 = stack(axis = out_3_axis_0, values = (xor__3_cast_fp16, xoi_3_cast_fp16))[name = tensor("out_3_cast_fp16")]; + tensor _inversed_177_y_0_to_fp16 = const()[name = tensor("_inversed_177_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_177_cast_fp16 = mul(x = position0_to_fp16, y = _inversed_177_y_0_to_fp16)[name = tensor("_inversed_177_cast_fp16")]; + tensor var_178_cast_fp16 = floor(x = _inversed_177_cast_fp16)[name = tensor("op_178_cast_fp16")]; + tensor var_179_to_fp16 = const()[name = tensor("op_179_to_fp16"), val = tensor(0x1p+9)]; + tensor var_180_cast_fp16 = mul(x = var_178_cast_fp16, y = var_179_to_fp16)[name = tensor("op_180_cast_fp16")]; + tensor write_idx_1_cast_fp16 = sub(x = position0_to_fp16, y = var_180_cast_fp16)[name = tensor("write_idx_1_cast_fp16")]; + tensor slot_indices_to_fp16 = const()[name = tensor("slot_indices_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6369856)))]; + tensor var_183_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_1_cast_fp16)[name = tensor("op_183_cast_fp16")]; + tensor var_193 = const()[name = tensor("op_193"), val = tensor([1, 512, 1, 1])]; + tensor onehot_1_to_fp16_dtype_0 = const()[name = tensor("onehot_1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_183_cast_fp16_to_fp16 = cast(dtype = onehot_1_to_fp16_dtype_0, x = var_183_cast_fp16)[name = tensor("cast_144")]; + tensor onehot_3_cast_fp16 = reshape(shape = var_193, x = var_183_cast_fp16_to_fp16)[name = tensor("onehot_3_cast_fp16")]; + tensor var_195_to_fp16 = const()[name = tensor("op_195_to_fp16"), val = tensor(0x1p+0)]; + tensor var_197_cast_fp16 = sub(x = var_195_to_fp16, y = onehot_3_cast_fp16)[name = tensor("op_197_cast_fp16")]; + tensor k_cache0_to_fp16_dtype_0 = const()[name = tensor("k_cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache0_to_fp16 = cast(dtype = k_cache0_to_fp16_dtype_0, x = k_cache0)[name = tensor("cast_143")]; + tensor var_198_cast_fp16 = mul(x = k_cache0_to_fp16, y = var_197_cast_fp16)[name = tensor("op_198_cast_fp16")]; + tensor var_203 = const()[name = tensor("op_203"), val = tensor([1, 1, 16, 64])]; + tensor var_204_cast_fp16 = reshape(shape = var_203, x = out_3_cast_fp16)[name = tensor("op_204_cast_fp16")]; + tensor var_205_cast_fp16 = mul(x = var_204_cast_fp16, y = onehot_3_cast_fp16)[name = tensor("op_205_cast_fp16")]; + tensor new_k_cache_1_cast_fp16 = add(x = var_198_cast_fp16, y = var_205_cast_fp16)[name = tensor("new_k_cache_1_cast_fp16")]; + tensor new_k_cache_1_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_1_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache0_to_fp16_dtype_0 = const()[name = tensor("v_cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache0_to_fp16 = cast(dtype = v_cache0_to_fp16_dtype_0, x = v_cache0)[name = tensor("cast_141")]; + tensor var_211_cast_fp16 = mul(x = v_cache0_to_fp16, y = var_197_cast_fp16)[name = tensor("op_211_cast_fp16")]; + tensor var_216 = const()[name = tensor("op_216"), val = tensor([1, 1, 16, 64])]; + tensor var_217_cast_fp16 = reshape(shape = var_216, x = v_1_cast_fp16)[name = tensor("op_217_cast_fp16")]; + tensor var_218_cast_fp16 = mul(x = var_217_cast_fp16, y = onehot_3_cast_fp16)[name = tensor("op_218_cast_fp16")]; + tensor new_v_cache_1_cast_fp16 = add(x = var_211_cast_fp16, y = var_218_cast_fp16)[name = tensor("new_v_cache_1_cast_fp16")]; + tensor new_v_cache_1_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_1_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_0_perm_0 = const()[name = tensor("transpose_0_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_231 = const()[name = tensor("op_231"), val = tensor([0, 2, 1, 3])]; + tensor var_237 = const()[name = tensor("op_237"), val = tensor([1, 16, 1, 64])]; + tensor q4_1_cast_fp16 = reshape(shape = var_237, x = out_1_cast_fp16)[name = tensor("q4_1_cast_fp16")]; + tensor var_242_transpose_x_0 = const()[name = tensor("op_242_transpose_x_0"), val = tensor(false)]; + tensor var_242_transpose_y_0 = const()[name = tensor("op_242_transpose_y_0"), val = tensor(false)]; + tensor transpose_0_cast_fp16 = transpose(perm = transpose_0_perm_0, x = new_k_cache_1_cast_fp16)[name = tensor("transpose_23")]; + tensor var_242_cast_fp16 = matmul(transpose_x = var_242_transpose_x_0, transpose_y = var_242_transpose_y_0, x = q4_1_cast_fp16, y = transpose_0_cast_fp16)[name = tensor("op_242_cast_fp16")]; + tensor var_243_to_fp16 = const()[name = tensor("op_243_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_1_cast_fp16 = mul(x = var_242_cast_fp16, y = var_243_to_fp16)[name = tensor("scores_1_cast_fp16")]; + tensor var_247 = const()[name = tensor("op_247"), val = tensor([1, 1])]; + tensor var_248_cast_fp16 = reshape(shape = var_247, x = position0_to_fp16)[name = tensor("op_248_cast_fp16")]; + tensor var_249_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_248_cast_fp16)[name = tensor("op_249_cast_fp16")]; + tensor mask_1_to_fp16_dtype_0 = const()[name = tensor("mask_1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_256_to_fp16 = const()[name = tensor("op_256_to_fp16"), val = tensor(0x1p+0)]; + tensor var_249_cast_fp16_to_fp16 = cast(dtype = mask_1_to_fp16_dtype_0, x = var_249_cast_fp16)[name = tensor("cast_139")]; + tensor var_257_cast_fp16 = sub(x = var_249_cast_fp16_to_fp16, y = var_256_to_fp16)[name = tensor("op_257_cast_fp16")]; + tensor var_262 = const()[name = tensor("op_262"), val = tensor([1, 1, 1, 512])]; + tensor var_263_cast_fp16 = reshape(shape = var_262, x = var_257_cast_fp16)[name = tensor("op_263_cast_fp16")]; + tensor var_264_to_fp16 = const()[name = tensor("op_264_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_265_cast_fp16 = mul(x = var_263_cast_fp16, y = var_264_to_fp16)[name = tensor("op_265_cast_fp16")]; + tensor scores_3_cast_fp16 = add(x = scores_1_cast_fp16, y = var_265_cast_fp16)[name = tensor("scores_3_cast_fp16")]; + tensor var_268 = const()[name = tensor("op_268"), val = tensor(-1)]; + tensor weights_1_cast_fp16 = softmax(axis = var_268, x = scores_3_cast_fp16)[name = tensor("weights_1_cast_fp16")]; + tensor attn_1_transpose_x_0 = const()[name = tensor("attn_1_transpose_x_0"), val = tensor(false)]; + tensor attn_1_transpose_y_0 = const()[name = tensor("attn_1_transpose_y_0"), val = tensor(false)]; + tensor values_1_cast_fp16 = transpose(perm = var_231, x = new_v_cache_1_cast_fp16)[name = tensor("transpose_22")]; + tensor attn_1_cast_fp16 = matmul(transpose_x = attn_1_transpose_x_0, transpose_y = attn_1_transpose_y_0, x = weights_1_cast_fp16, y = values_1_cast_fp16)[name = tensor("attn_1_cast_fp16")]; + tensor var_276 = const()[name = tensor("op_276"), val = tensor([0, 2, 1, 3])]; + tensor var_281 = const()[name = tensor("op_281"), val = tensor([1, 1, 1024])]; + tensor var_277_cast_fp16 = transpose(perm = var_276, x = attn_1_cast_fp16)[name = tensor("transpose_21")]; + tensor input_7_cast_fp16 = reshape(shape = var_281, x = var_277_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor attn0_out_proj_weight_to_fp16 = const()[name = tensor("attn0_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6370944)))]; + tensor linear_2_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn0_out_proj_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor var_287_to_fp16 = const()[name = tensor("op_287_to_fp16"), val = tensor(0x1p+0)]; + tensor var_288_cast_fp16 = add(x = position0_to_fp16, y = var_287_to_fp16)[name = tensor("op_288_cast_fp16")]; + tensor var_288_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_288_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_9_cast_fp16 = add(x = linear_0_cast_fp16, y = linear_2_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor input_11_axes_0 = const()[name = tensor("input_11_axes_0"), val = tensor([-1])]; + tensor norm0_2_weight_to_fp16 = const()[name = tensor("norm0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8468160)))]; + tensor norm0_2_bias_to_fp16 = const()[name = tensor("norm0_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8470272)))]; + tensor var_292_to_fp16 = const()[name = tensor("op_292_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_11_cast_fp16 = layer_norm(axes = input_11_axes_0, beta = norm0_2_bias_to_fp16, epsilon = var_292_to_fp16, gamma = norm0_2_weight_to_fp16, x = input_9_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor linear0_1_weight_to_fp16 = const()[name = tensor("linear0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8472384)))]; + tensor linear_3_bias_0_to_fp16 = const()[name = tensor("linear_3_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16861056)))]; + tensor linear_3_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear0_1_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("linear_3_cast_fp16")]; + tensor input_13_mode_0 = const()[name = tensor("input_13_mode_0"), val = tensor("EXACT")]; + tensor input_13_cast_fp16 = gelu(mode = input_13_mode_0, x = linear_3_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor linear0_2_weight_to_fp16 = const()[name = tensor("linear0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16869312)))]; + tensor linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear0_2_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor input_15_cast_fp16 = add(x = input_9_cast_fp16, y = linear_4_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor input_17_axes_0 = const()[name = tensor("input_17_axes_0"), val = tensor([-1])]; + tensor norm1_1_weight_to_fp16 = const()[name = tensor("norm1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25257984)))]; + tensor norm1_1_bias_to_fp16 = const()[name = tensor("norm1_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25260096)))]; + tensor var_309_to_fp16 = const()[name = tensor("op_309_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_17_cast_fp16 = layer_norm(axes = input_17_axes_0, beta = norm1_1_bias_to_fp16, epsilon = var_309_to_fp16, gamma = norm1_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor attn1_in_proj_weight_to_fp16 = const()[name = tensor("attn1_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25262208)))]; + tensor linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn1_in_proj_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor var_322 = const()[name = tensor("op_322"), val = tensor([1, 3, 16, 64])]; + tensor qkv_3_cast_fp16 = reshape(shape = var_322, x = linear_5_cast_fp16)[name = tensor("qkv_3_cast_fp16")]; + tensor x_5_begin_0 = const()[name = tensor("x_5_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_5_end_0 = const()[name = tensor("x_5_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_5_end_mask_0 = const()[name = tensor("x_5_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_5_squeeze_mask_0 = const()[name = tensor("x_5_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_5_cast_fp16 = slice_by_index(begin = x_5_begin_0, end = x_5_end_0, end_mask = x_5_end_mask_0, squeeze_mask = x_5_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("x_5_cast_fp16")]; + tensor x_7_begin_0 = const()[name = tensor("x_7_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_7_end_0 = const()[name = tensor("x_7_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_7_end_mask_0 = const()[name = tensor("x_7_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_7_squeeze_mask_0 = const()[name = tensor("x_7_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_7_cast_fp16 = slice_by_index(begin = x_7_begin_0, end = x_7_end_0, end_mask = x_7_end_mask_0, squeeze_mask = x_7_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("x_7_cast_fp16")]; + tensor v_3_begin_0 = const()[name = tensor("v_3_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_3_end_0 = const()[name = tensor("v_3_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_3_end_mask_0 = const()[name = tensor("v_3_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_3_squeeze_mask_0 = const()[name = tensor("v_3_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_3_cast_fp16 = slice_by_index(begin = v_3_begin_0, end = v_3_end_0, end_mask = v_3_end_mask_0, squeeze_mask = v_3_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("v_3_cast_fp16")]; + tensor var_351 = const()[name = tensor("op_351"), val = tensor([1, 1, 1])]; + tensor position1_to_fp16_dtype_0 = const()[name = tensor("position1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position1_to_fp16 = cast(dtype = position1_to_fp16_dtype_0, x = position1)[name = tensor("cast_137")]; + tensor var_352_cast_fp16 = reshape(shape = var_351, x = position1_to_fp16)[name = tensor("op_352_cast_fp16")]; + tensor angles_5_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_352_cast_fp16)[name = tensor("angles_5_cast_fp16")]; + tensor rotr_5_cast_fp16 = cos(x = angles_5_cast_fp16)[name = tensor("rotr_5_cast_fp16")]; + tensor roti_5_cast_fp16 = sin(x = angles_5_cast_fp16)[name = tensor("roti_5_cast_fp16")]; + tensor var_360 = const()[name = tensor("op_360"), val = tensor([1, 16, 32, 2])]; + tensor pairs_5_cast_fp16 = reshape(shape = var_360, x = x_5_cast_fp16)[name = tensor("pairs_5_cast_fp16")]; + tensor xr_5_begin_0 = const()[name = tensor("xr_5_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_5_end_0 = const()[name = tensor("xr_5_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_5_end_mask_0 = const()[name = tensor("xr_5_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_5_squeeze_mask_0 = const()[name = tensor("xr_5_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_5_cast_fp16 = slice_by_index(begin = xr_5_begin_0, end = xr_5_end_0, end_mask = xr_5_end_mask_0, squeeze_mask = xr_5_squeeze_mask_0, x = pairs_5_cast_fp16)[name = tensor("xr_5_cast_fp16")]; + tensor xi_5_begin_0 = const()[name = tensor("xi_5_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_5_end_0 = const()[name = tensor("xi_5_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_5_end_mask_0 = const()[name = tensor("xi_5_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_5_squeeze_mask_0 = const()[name = tensor("xi_5_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_5_cast_fp16 = slice_by_index(begin = xi_5_begin_0, end = xi_5_end_0, end_mask = xi_5_end_mask_0, squeeze_mask = xi_5_squeeze_mask_0, x = pairs_5_cast_fp16)[name = tensor("xi_5_cast_fp16")]; + tensor var_368_cast_fp16 = mul(x = xr_5_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_368_cast_fp16")]; + tensor var_369_cast_fp16 = mul(x = xi_5_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_369_cast_fp16")]; + tensor xor__5_cast_fp16 = sub(x = var_368_cast_fp16, y = var_369_cast_fp16)[name = tensor("xor__5_cast_fp16")]; + tensor var_372_cast_fp16 = mul(x = xr_5_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_372_cast_fp16")]; + tensor var_373_cast_fp16 = mul(x = xi_5_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_373_cast_fp16")]; + tensor xoi_5_cast_fp16 = add(x = var_372_cast_fp16, y = var_373_cast_fp16)[name = tensor("xoi_5_cast_fp16")]; + tensor out_5_axis_0 = const()[name = tensor("out_5_axis_0"), val = tensor(-1)]; + tensor out_5_cast_fp16 = stack(axis = out_5_axis_0, values = (xor__5_cast_fp16, xoi_5_cast_fp16))[name = tensor("out_5_cast_fp16")]; + tensor var_396 = const()[name = tensor("op_396"), val = tensor([1, 16, 32, 2])]; + tensor pairs_7_cast_fp16 = reshape(shape = var_396, x = x_7_cast_fp16)[name = tensor("pairs_7_cast_fp16")]; + tensor xr_7_begin_0 = const()[name = tensor("xr_7_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_7_end_0 = const()[name = tensor("xr_7_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_7_end_mask_0 = const()[name = tensor("xr_7_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_7_squeeze_mask_0 = const()[name = tensor("xr_7_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_7_cast_fp16 = slice_by_index(begin = xr_7_begin_0, end = xr_7_end_0, end_mask = xr_7_end_mask_0, squeeze_mask = xr_7_squeeze_mask_0, x = pairs_7_cast_fp16)[name = tensor("xr_7_cast_fp16")]; + tensor xi_7_begin_0 = const()[name = tensor("xi_7_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_7_end_0 = const()[name = tensor("xi_7_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_7_end_mask_0 = const()[name = tensor("xi_7_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_7_squeeze_mask_0 = const()[name = tensor("xi_7_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_7_cast_fp16 = slice_by_index(begin = xi_7_begin_0, end = xi_7_end_0, end_mask = xi_7_end_mask_0, squeeze_mask = xi_7_squeeze_mask_0, x = pairs_7_cast_fp16)[name = tensor("xi_7_cast_fp16")]; + tensor var_404_cast_fp16 = mul(x = xr_7_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_404_cast_fp16")]; + tensor var_405_cast_fp16 = mul(x = xi_7_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_405_cast_fp16")]; + tensor xor__7_cast_fp16 = sub(x = var_404_cast_fp16, y = var_405_cast_fp16)[name = tensor("xor__7_cast_fp16")]; + tensor var_408_cast_fp16 = mul(x = xr_7_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_408_cast_fp16")]; + tensor var_409_cast_fp16 = mul(x = xi_7_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_409_cast_fp16")]; + tensor xoi_7_cast_fp16 = add(x = var_408_cast_fp16, y = var_409_cast_fp16)[name = tensor("xoi_7_cast_fp16")]; + tensor out_7_axis_0 = const()[name = tensor("out_7_axis_0"), val = tensor(-1)]; + tensor out_7_cast_fp16 = stack(axis = out_7_axis_0, values = (xor__7_cast_fp16, xoi_7_cast_fp16))[name = tensor("out_7_cast_fp16")]; + tensor _inversed_421_y_0_to_fp16 = const()[name = tensor("_inversed_421_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_421_cast_fp16 = mul(x = position1_to_fp16, y = _inversed_421_y_0_to_fp16)[name = tensor("_inversed_421_cast_fp16")]; + tensor var_422_cast_fp16 = floor(x = _inversed_421_cast_fp16)[name = tensor("op_422_cast_fp16")]; + tensor var_423_to_fp16 = const()[name = tensor("op_423_to_fp16"), val = tensor(0x1p+9)]; + tensor var_424_cast_fp16 = mul(x = var_422_cast_fp16, y = var_423_to_fp16)[name = tensor("op_424_cast_fp16")]; + tensor write_idx_3_cast_fp16 = sub(x = position1_to_fp16, y = var_424_cast_fp16)[name = tensor("write_idx_3_cast_fp16")]; + tensor var_427_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_3_cast_fp16)[name = tensor("op_427_cast_fp16")]; + tensor var_437 = const()[name = tensor("op_437"), val = tensor([1, 512, 1, 1])]; + tensor onehot_5_to_fp16_dtype_0 = const()[name = tensor("onehot_5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_427_cast_fp16_to_fp16 = cast(dtype = onehot_5_to_fp16_dtype_0, x = var_427_cast_fp16)[name = tensor("cast_136")]; + tensor onehot_7_cast_fp16 = reshape(shape = var_437, x = var_427_cast_fp16_to_fp16)[name = tensor("onehot_7_cast_fp16")]; + tensor var_439_to_fp16 = const()[name = tensor("op_439_to_fp16"), val = tensor(0x1p+0)]; + tensor var_441_cast_fp16 = sub(x = var_439_to_fp16, y = onehot_7_cast_fp16)[name = tensor("op_441_cast_fp16")]; + tensor k_cache1_to_fp16_dtype_0 = const()[name = tensor("k_cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache1_to_fp16 = cast(dtype = k_cache1_to_fp16_dtype_0, x = k_cache1)[name = tensor("cast_135")]; + tensor var_442_cast_fp16 = mul(x = k_cache1_to_fp16, y = var_441_cast_fp16)[name = tensor("op_442_cast_fp16")]; + tensor var_447 = const()[name = tensor("op_447"), val = tensor([1, 1, 16, 64])]; + tensor var_448_cast_fp16 = reshape(shape = var_447, x = out_7_cast_fp16)[name = tensor("op_448_cast_fp16")]; + tensor var_449_cast_fp16 = mul(x = var_448_cast_fp16, y = onehot_7_cast_fp16)[name = tensor("op_449_cast_fp16")]; + tensor new_k_cache_3_cast_fp16 = add(x = var_442_cast_fp16, y = var_449_cast_fp16)[name = tensor("new_k_cache_3_cast_fp16")]; + tensor new_k_cache_3_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_3_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache1_to_fp16_dtype_0 = const()[name = tensor("v_cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache1_to_fp16 = cast(dtype = v_cache1_to_fp16_dtype_0, x = v_cache1)[name = tensor("cast_133")]; + tensor var_455_cast_fp16 = mul(x = v_cache1_to_fp16, y = var_441_cast_fp16)[name = tensor("op_455_cast_fp16")]; + tensor var_460 = const()[name = tensor("op_460"), val = tensor([1, 1, 16, 64])]; + tensor var_461_cast_fp16 = reshape(shape = var_460, x = v_3_cast_fp16)[name = tensor("op_461_cast_fp16")]; + tensor var_462_cast_fp16 = mul(x = var_461_cast_fp16, y = onehot_7_cast_fp16)[name = tensor("op_462_cast_fp16")]; + tensor new_v_cache_3_cast_fp16 = add(x = var_455_cast_fp16, y = var_462_cast_fp16)[name = tensor("new_v_cache_3_cast_fp16")]; + tensor new_v_cache_3_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_3_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_1_perm_0 = const()[name = tensor("transpose_1_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_475 = const()[name = tensor("op_475"), val = tensor([0, 2, 1, 3])]; + tensor var_481 = const()[name = tensor("op_481"), val = tensor([1, 16, 1, 64])]; + tensor q4_3_cast_fp16 = reshape(shape = var_481, x = out_5_cast_fp16)[name = tensor("q4_3_cast_fp16")]; + tensor var_486_transpose_x_0 = const()[name = tensor("op_486_transpose_x_0"), val = tensor(false)]; + tensor var_486_transpose_y_0 = const()[name = tensor("op_486_transpose_y_0"), val = tensor(false)]; + tensor transpose_1_cast_fp16 = transpose(perm = transpose_1_perm_0, x = new_k_cache_3_cast_fp16)[name = tensor("transpose_20")]; + tensor var_486_cast_fp16 = matmul(transpose_x = var_486_transpose_x_0, transpose_y = var_486_transpose_y_0, x = q4_3_cast_fp16, y = transpose_1_cast_fp16)[name = tensor("op_486_cast_fp16")]; + tensor var_487_to_fp16 = const()[name = tensor("op_487_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_5_cast_fp16 = mul(x = var_486_cast_fp16, y = var_487_to_fp16)[name = tensor("scores_5_cast_fp16")]; + tensor var_491 = const()[name = tensor("op_491"), val = tensor([1, 1])]; + tensor var_492_cast_fp16 = reshape(shape = var_491, x = position1_to_fp16)[name = tensor("op_492_cast_fp16")]; + tensor var_493_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_492_cast_fp16)[name = tensor("op_493_cast_fp16")]; + tensor mask_3_to_fp16_dtype_0 = const()[name = tensor("mask_3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_500_to_fp16 = const()[name = tensor("op_500_to_fp16"), val = tensor(0x1p+0)]; + tensor var_493_cast_fp16_to_fp16 = cast(dtype = mask_3_to_fp16_dtype_0, x = var_493_cast_fp16)[name = tensor("cast_131")]; + tensor var_501_cast_fp16 = sub(x = var_493_cast_fp16_to_fp16, y = var_500_to_fp16)[name = tensor("op_501_cast_fp16")]; + tensor var_506 = const()[name = tensor("op_506"), val = tensor([1, 1, 1, 512])]; + tensor var_507_cast_fp16 = reshape(shape = var_506, x = var_501_cast_fp16)[name = tensor("op_507_cast_fp16")]; + tensor var_508_to_fp16 = const()[name = tensor("op_508_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_509_cast_fp16 = mul(x = var_507_cast_fp16, y = var_508_to_fp16)[name = tensor("op_509_cast_fp16")]; + tensor scores_7_cast_fp16 = add(x = scores_5_cast_fp16, y = var_509_cast_fp16)[name = tensor("scores_7_cast_fp16")]; + tensor var_512 = const()[name = tensor("op_512"), val = tensor(-1)]; + tensor weights_3_cast_fp16 = softmax(axis = var_512, x = scores_7_cast_fp16)[name = tensor("weights_3_cast_fp16")]; + tensor attn_3_transpose_x_0 = const()[name = tensor("attn_3_transpose_x_0"), val = tensor(false)]; + tensor attn_3_transpose_y_0 = const()[name = tensor("attn_3_transpose_y_0"), val = tensor(false)]; + tensor values_3_cast_fp16 = transpose(perm = var_475, x = new_v_cache_3_cast_fp16)[name = tensor("transpose_19")]; + tensor attn_3_cast_fp16 = matmul(transpose_x = attn_3_transpose_x_0, transpose_y = attn_3_transpose_y_0, x = weights_3_cast_fp16, y = values_3_cast_fp16)[name = tensor("attn_3_cast_fp16")]; + tensor var_520 = const()[name = tensor("op_520"), val = tensor([0, 2, 1, 3])]; + tensor var_525 = const()[name = tensor("op_525"), val = tensor([1, 1, 1024])]; + tensor var_521_cast_fp16 = transpose(perm = var_520, x = attn_3_cast_fp16)[name = tensor("transpose_18")]; + tensor input_19_cast_fp16 = reshape(shape = var_525, x = var_521_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor attn1_out_proj_weight_to_fp16 = const()[name = tensor("attn1_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31553728)))]; + tensor linear_6_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn1_out_proj_weight_to_fp16, x = input_19_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor var_531_to_fp16 = const()[name = tensor("op_531_to_fp16"), val = tensor(0x1p+0)]; + tensor var_532_cast_fp16 = add(x = position1_to_fp16, y = var_531_to_fp16)[name = tensor("op_532_cast_fp16")]; + tensor var_532_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_532_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_21_cast_fp16 = add(x = input_15_cast_fp16, y = linear_6_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor input_23_axes_0 = const()[name = tensor("input_23_axes_0"), val = tensor([-1])]; + tensor norm1_2_weight_to_fp16 = const()[name = tensor("norm1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33650944)))]; + tensor norm1_2_bias_to_fp16 = const()[name = tensor("norm1_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33653056)))]; + tensor var_536_to_fp16 = const()[name = tensor("op_536_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_23_cast_fp16 = layer_norm(axes = input_23_axes_0, beta = norm1_2_bias_to_fp16, epsilon = var_536_to_fp16, gamma = norm1_2_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor linear1_1_weight_to_fp16 = const()[name = tensor("linear1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33655168)))]; + tensor linear_7_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear1_1_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_25_mode_0 = const()[name = tensor("input_25_mode_0"), val = tensor("EXACT")]; + tensor input_25_cast_fp16 = gelu(mode = input_25_mode_0, x = linear_7_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor linear1_2_weight_to_fp16 = const()[name = tensor("linear1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42043840)))]; + tensor linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear1_2_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor input_27_cast_fp16 = add(x = input_21_cast_fp16, y = linear_8_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor input_29_axes_0 = const()[name = tensor("input_29_axes_0"), val = tensor([-1])]; + tensor norm2_1_weight_to_fp16 = const()[name = tensor("norm2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50432512)))]; + tensor norm2_1_bias_to_fp16 = const()[name = tensor("norm2_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50434624)))]; + tensor var_553_to_fp16 = const()[name = tensor("op_553_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_29_cast_fp16 = layer_norm(axes = input_29_axes_0, beta = norm2_1_bias_to_fp16, epsilon = var_553_to_fp16, gamma = norm2_1_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor attn2_in_proj_weight_to_fp16 = const()[name = tensor("attn2_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50436736)))]; + tensor linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn2_in_proj_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor var_566 = const()[name = tensor("op_566"), val = tensor([1, 3, 16, 64])]; + tensor qkv_5_cast_fp16 = reshape(shape = var_566, x = linear_9_cast_fp16)[name = tensor("qkv_5_cast_fp16")]; + tensor x_9_begin_0 = const()[name = tensor("x_9_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_9_end_0 = const()[name = tensor("x_9_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_9_end_mask_0 = const()[name = tensor("x_9_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_9_squeeze_mask_0 = const()[name = tensor("x_9_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_9_cast_fp16 = slice_by_index(begin = x_9_begin_0, end = x_9_end_0, end_mask = x_9_end_mask_0, squeeze_mask = x_9_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("x_9_cast_fp16")]; + tensor x_11_begin_0 = const()[name = tensor("x_11_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_11_end_0 = const()[name = tensor("x_11_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_11_end_mask_0 = const()[name = tensor("x_11_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_11_squeeze_mask_0 = const()[name = tensor("x_11_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_11_cast_fp16 = slice_by_index(begin = x_11_begin_0, end = x_11_end_0, end_mask = x_11_end_mask_0, squeeze_mask = x_11_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("x_11_cast_fp16")]; + tensor v_5_begin_0 = const()[name = tensor("v_5_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_5_end_0 = const()[name = tensor("v_5_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_5_end_mask_0 = const()[name = tensor("v_5_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_5_squeeze_mask_0 = const()[name = tensor("v_5_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_5_cast_fp16 = slice_by_index(begin = v_5_begin_0, end = v_5_end_0, end_mask = v_5_end_mask_0, squeeze_mask = v_5_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("v_5_cast_fp16")]; + tensor var_595 = const()[name = tensor("op_595"), val = tensor([1, 1, 1])]; + tensor position2_to_fp16_dtype_0 = const()[name = tensor("position2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position2_to_fp16 = cast(dtype = position2_to_fp16_dtype_0, x = position2)[name = tensor("cast_129")]; + tensor var_596_cast_fp16 = reshape(shape = var_595, x = position2_to_fp16)[name = tensor("op_596_cast_fp16")]; + tensor angles_9_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_596_cast_fp16)[name = tensor("angles_9_cast_fp16")]; + tensor rotr_9_cast_fp16 = cos(x = angles_9_cast_fp16)[name = tensor("rotr_9_cast_fp16")]; + tensor roti_9_cast_fp16 = sin(x = angles_9_cast_fp16)[name = tensor("roti_9_cast_fp16")]; + tensor var_604 = const()[name = tensor("op_604"), val = tensor([1, 16, 32, 2])]; + tensor pairs_9_cast_fp16 = reshape(shape = var_604, x = x_9_cast_fp16)[name = tensor("pairs_9_cast_fp16")]; + tensor xr_9_begin_0 = const()[name = tensor("xr_9_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_9_end_0 = const()[name = tensor("xr_9_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_9_end_mask_0 = const()[name = tensor("xr_9_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_9_squeeze_mask_0 = const()[name = tensor("xr_9_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_9_cast_fp16 = slice_by_index(begin = xr_9_begin_0, end = xr_9_end_0, end_mask = xr_9_end_mask_0, squeeze_mask = xr_9_squeeze_mask_0, x = pairs_9_cast_fp16)[name = tensor("xr_9_cast_fp16")]; + tensor xi_9_begin_0 = const()[name = tensor("xi_9_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_9_end_0 = const()[name = tensor("xi_9_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_9_end_mask_0 = const()[name = tensor("xi_9_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_9_squeeze_mask_0 = const()[name = tensor("xi_9_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_9_cast_fp16 = slice_by_index(begin = xi_9_begin_0, end = xi_9_end_0, end_mask = xi_9_end_mask_0, squeeze_mask = xi_9_squeeze_mask_0, x = pairs_9_cast_fp16)[name = tensor("xi_9_cast_fp16")]; + tensor var_612_cast_fp16 = mul(x = xr_9_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_612_cast_fp16")]; + tensor var_613_cast_fp16 = mul(x = xi_9_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_613_cast_fp16")]; + tensor xor__9_cast_fp16 = sub(x = var_612_cast_fp16, y = var_613_cast_fp16)[name = tensor("xor__9_cast_fp16")]; + tensor var_616_cast_fp16 = mul(x = xr_9_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_616_cast_fp16")]; + tensor var_617_cast_fp16 = mul(x = xi_9_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_617_cast_fp16")]; + tensor xoi_9_cast_fp16 = add(x = var_616_cast_fp16, y = var_617_cast_fp16)[name = tensor("xoi_9_cast_fp16")]; + tensor out_9_axis_0 = const()[name = tensor("out_9_axis_0"), val = tensor(-1)]; + tensor out_9_cast_fp16 = stack(axis = out_9_axis_0, values = (xor__9_cast_fp16, xoi_9_cast_fp16))[name = tensor("out_9_cast_fp16")]; + tensor var_640 = const()[name = tensor("op_640"), val = tensor([1, 16, 32, 2])]; + tensor pairs_11_cast_fp16 = reshape(shape = var_640, x = x_11_cast_fp16)[name = tensor("pairs_11_cast_fp16")]; + tensor xr_11_begin_0 = const()[name = tensor("xr_11_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_11_end_0 = const()[name = tensor("xr_11_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_11_end_mask_0 = const()[name = tensor("xr_11_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_11_squeeze_mask_0 = const()[name = tensor("xr_11_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_11_cast_fp16 = slice_by_index(begin = xr_11_begin_0, end = xr_11_end_0, end_mask = xr_11_end_mask_0, squeeze_mask = xr_11_squeeze_mask_0, x = pairs_11_cast_fp16)[name = tensor("xr_11_cast_fp16")]; + tensor xi_11_begin_0 = const()[name = tensor("xi_11_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_11_end_0 = const()[name = tensor("xi_11_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_11_end_mask_0 = const()[name = tensor("xi_11_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_11_squeeze_mask_0 = const()[name = tensor("xi_11_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_11_cast_fp16 = slice_by_index(begin = xi_11_begin_0, end = xi_11_end_0, end_mask = xi_11_end_mask_0, squeeze_mask = xi_11_squeeze_mask_0, x = pairs_11_cast_fp16)[name = tensor("xi_11_cast_fp16")]; + tensor var_648_cast_fp16 = mul(x = xr_11_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_648_cast_fp16")]; + tensor var_649_cast_fp16 = mul(x = xi_11_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_649_cast_fp16")]; + tensor xor__11_cast_fp16 = sub(x = var_648_cast_fp16, y = var_649_cast_fp16)[name = tensor("xor__11_cast_fp16")]; + tensor var_652_cast_fp16 = mul(x = xr_11_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_652_cast_fp16")]; + tensor var_653_cast_fp16 = mul(x = xi_11_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_653_cast_fp16")]; + tensor xoi_11_cast_fp16 = add(x = var_652_cast_fp16, y = var_653_cast_fp16)[name = tensor("xoi_11_cast_fp16")]; + tensor out_11_axis_0 = const()[name = tensor("out_11_axis_0"), val = tensor(-1)]; + tensor out_11_cast_fp16 = stack(axis = out_11_axis_0, values = (xor__11_cast_fp16, xoi_11_cast_fp16))[name = tensor("out_11_cast_fp16")]; + tensor _inversed_665_y_0_to_fp16 = const()[name = tensor("_inversed_665_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_665_cast_fp16 = mul(x = position2_to_fp16, y = _inversed_665_y_0_to_fp16)[name = tensor("_inversed_665_cast_fp16")]; + tensor var_666_cast_fp16 = floor(x = _inversed_665_cast_fp16)[name = tensor("op_666_cast_fp16")]; + tensor var_667_to_fp16 = const()[name = tensor("op_667_to_fp16"), val = tensor(0x1p+9)]; + tensor var_668_cast_fp16 = mul(x = var_666_cast_fp16, y = var_667_to_fp16)[name = tensor("op_668_cast_fp16")]; + tensor write_idx_5_cast_fp16 = sub(x = position2_to_fp16, y = var_668_cast_fp16)[name = tensor("write_idx_5_cast_fp16")]; + tensor var_671_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_5_cast_fp16)[name = tensor("op_671_cast_fp16")]; + tensor var_681 = const()[name = tensor("op_681"), val = tensor([1, 512, 1, 1])]; + tensor onehot_9_to_fp16_dtype_0 = const()[name = tensor("onehot_9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_671_cast_fp16_to_fp16 = cast(dtype = onehot_9_to_fp16_dtype_0, x = var_671_cast_fp16)[name = tensor("cast_128")]; + tensor onehot_11_cast_fp16 = reshape(shape = var_681, x = var_671_cast_fp16_to_fp16)[name = tensor("onehot_11_cast_fp16")]; + tensor var_683_to_fp16 = const()[name = tensor("op_683_to_fp16"), val = tensor(0x1p+0)]; + tensor var_685_cast_fp16 = sub(x = var_683_to_fp16, y = onehot_11_cast_fp16)[name = tensor("op_685_cast_fp16")]; + tensor k_cache2_to_fp16_dtype_0 = const()[name = tensor("k_cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache2_to_fp16 = cast(dtype = k_cache2_to_fp16_dtype_0, x = k_cache2)[name = tensor("cast_127")]; + tensor var_686_cast_fp16 = mul(x = k_cache2_to_fp16, y = var_685_cast_fp16)[name = tensor("op_686_cast_fp16")]; + tensor var_691 = const()[name = tensor("op_691"), val = tensor([1, 1, 16, 64])]; + tensor var_692_cast_fp16 = reshape(shape = var_691, x = out_11_cast_fp16)[name = tensor("op_692_cast_fp16")]; + tensor var_693_cast_fp16 = mul(x = var_692_cast_fp16, y = onehot_11_cast_fp16)[name = tensor("op_693_cast_fp16")]; + tensor new_k_cache_5_cast_fp16 = add(x = var_686_cast_fp16, y = var_693_cast_fp16)[name = tensor("new_k_cache_5_cast_fp16")]; + tensor new_k_cache_5_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_5_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache2_to_fp16_dtype_0 = const()[name = tensor("v_cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache2_to_fp16 = cast(dtype = v_cache2_to_fp16_dtype_0, x = v_cache2)[name = tensor("cast_125")]; + tensor var_699_cast_fp16 = mul(x = v_cache2_to_fp16, y = var_685_cast_fp16)[name = tensor("op_699_cast_fp16")]; + tensor var_704 = const()[name = tensor("op_704"), val = tensor([1, 1, 16, 64])]; + tensor var_705_cast_fp16 = reshape(shape = var_704, x = v_5_cast_fp16)[name = tensor("op_705_cast_fp16")]; + tensor var_706_cast_fp16 = mul(x = var_705_cast_fp16, y = onehot_11_cast_fp16)[name = tensor("op_706_cast_fp16")]; + tensor new_v_cache_5_cast_fp16 = add(x = var_699_cast_fp16, y = var_706_cast_fp16)[name = tensor("new_v_cache_5_cast_fp16")]; + tensor new_v_cache_5_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_5_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_2_perm_0 = const()[name = tensor("transpose_2_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_719 = const()[name = tensor("op_719"), val = tensor([0, 2, 1, 3])]; + tensor var_725 = const()[name = tensor("op_725"), val = tensor([1, 16, 1, 64])]; + tensor q4_5_cast_fp16 = reshape(shape = var_725, x = out_9_cast_fp16)[name = tensor("q4_5_cast_fp16")]; + tensor var_730_transpose_x_0 = const()[name = tensor("op_730_transpose_x_0"), val = tensor(false)]; + tensor var_730_transpose_y_0 = const()[name = tensor("op_730_transpose_y_0"), val = tensor(false)]; + tensor transpose_2_cast_fp16 = transpose(perm = transpose_2_perm_0, x = new_k_cache_5_cast_fp16)[name = tensor("transpose_17")]; + tensor var_730_cast_fp16 = matmul(transpose_x = var_730_transpose_x_0, transpose_y = var_730_transpose_y_0, x = q4_5_cast_fp16, y = transpose_2_cast_fp16)[name = tensor("op_730_cast_fp16")]; + tensor var_731_to_fp16 = const()[name = tensor("op_731_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_9_cast_fp16 = mul(x = var_730_cast_fp16, y = var_731_to_fp16)[name = tensor("scores_9_cast_fp16")]; + tensor var_735 = const()[name = tensor("op_735"), val = tensor([1, 1])]; + tensor var_736_cast_fp16 = reshape(shape = var_735, x = position2_to_fp16)[name = tensor("op_736_cast_fp16")]; + tensor var_737_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_736_cast_fp16)[name = tensor("op_737_cast_fp16")]; + tensor mask_5_to_fp16_dtype_0 = const()[name = tensor("mask_5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_744_to_fp16 = const()[name = tensor("op_744_to_fp16"), val = tensor(0x1p+0)]; + tensor var_737_cast_fp16_to_fp16 = cast(dtype = mask_5_to_fp16_dtype_0, x = var_737_cast_fp16)[name = tensor("cast_123")]; + tensor var_745_cast_fp16 = sub(x = var_737_cast_fp16_to_fp16, y = var_744_to_fp16)[name = tensor("op_745_cast_fp16")]; + tensor var_750 = const()[name = tensor("op_750"), val = tensor([1, 1, 1, 512])]; + tensor var_751_cast_fp16 = reshape(shape = var_750, x = var_745_cast_fp16)[name = tensor("op_751_cast_fp16")]; + tensor var_752_to_fp16 = const()[name = tensor("op_752_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_753_cast_fp16 = mul(x = var_751_cast_fp16, y = var_752_to_fp16)[name = tensor("op_753_cast_fp16")]; + tensor scores_11_cast_fp16 = add(x = scores_9_cast_fp16, y = var_753_cast_fp16)[name = tensor("scores_11_cast_fp16")]; + tensor var_756 = const()[name = tensor("op_756"), val = tensor(-1)]; + tensor weights_5_cast_fp16 = softmax(axis = var_756, x = scores_11_cast_fp16)[name = tensor("weights_5_cast_fp16")]; + tensor attn_5_transpose_x_0 = const()[name = tensor("attn_5_transpose_x_0"), val = tensor(false)]; + tensor attn_5_transpose_y_0 = const()[name = tensor("attn_5_transpose_y_0"), val = tensor(false)]; + tensor values_5_cast_fp16 = transpose(perm = var_719, x = new_v_cache_5_cast_fp16)[name = tensor("transpose_16")]; + tensor attn_5_cast_fp16 = matmul(transpose_x = attn_5_transpose_x_0, transpose_y = attn_5_transpose_y_0, x = weights_5_cast_fp16, y = values_5_cast_fp16)[name = tensor("attn_5_cast_fp16")]; + tensor var_764 = const()[name = tensor("op_764"), val = tensor([0, 2, 1, 3])]; + tensor var_769 = const()[name = tensor("op_769"), val = tensor([1, 1, 1024])]; + tensor var_765_cast_fp16 = transpose(perm = var_764, x = attn_5_cast_fp16)[name = tensor("transpose_15")]; + tensor input_31_cast_fp16 = reshape(shape = var_769, x = var_765_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor attn2_out_proj_weight_to_fp16 = const()[name = tensor("attn2_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(56728256)))]; + tensor linear_10_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn2_out_proj_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor var_775_to_fp16 = const()[name = tensor("op_775_to_fp16"), val = tensor(0x1p+0)]; + tensor var_776_cast_fp16 = add(x = position2_to_fp16, y = var_775_to_fp16)[name = tensor("op_776_cast_fp16")]; + tensor var_776_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_776_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_33_cast_fp16 = add(x = input_27_cast_fp16, y = linear_10_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor input_35_axes_0 = const()[name = tensor("input_35_axes_0"), val = tensor([-1])]; + tensor norm2_2_weight_to_fp16 = const()[name = tensor("norm2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58825472)))]; + tensor norm2_2_bias_to_fp16 = const()[name = tensor("norm2_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58827584)))]; + tensor var_780_to_fp16 = const()[name = tensor("op_780_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_35_cast_fp16 = layer_norm(axes = input_35_axes_0, beta = norm2_2_bias_to_fp16, epsilon = var_780_to_fp16, gamma = norm2_2_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor linear2_1_weight_to_fp16 = const()[name = tensor("linear2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58829696)))]; + tensor linear_11_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear2_1_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor input_37_mode_0 = const()[name = tensor("input_37_mode_0"), val = tensor("EXACT")]; + tensor input_37_cast_fp16 = gelu(mode = input_37_mode_0, x = linear_11_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor linear2_2_weight_to_fp16 = const()[name = tensor("linear2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67218368)))]; + tensor linear_12_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear2_2_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor input_39_cast_fp16 = add(x = input_33_cast_fp16, y = linear_12_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor input_41_axes_0 = const()[name = tensor("input_41_axes_0"), val = tensor([-1])]; + tensor norm3_1_weight_to_fp16 = const()[name = tensor("norm3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75607040)))]; + tensor norm3_1_bias_to_fp16 = const()[name = tensor("norm3_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75609152)))]; + tensor var_797_to_fp16 = const()[name = tensor("op_797_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_41_cast_fp16 = layer_norm(axes = input_41_axes_0, beta = norm3_1_bias_to_fp16, epsilon = var_797_to_fp16, gamma = norm3_1_weight_to_fp16, x = input_39_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor attn3_in_proj_weight_to_fp16 = const()[name = tensor("attn3_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75611264)))]; + tensor linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn3_in_proj_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor var_810 = const()[name = tensor("op_810"), val = tensor([1, 3, 16, 64])]; + tensor qkv_7_cast_fp16 = reshape(shape = var_810, x = linear_13_cast_fp16)[name = tensor("qkv_7_cast_fp16")]; + tensor x_13_begin_0 = const()[name = tensor("x_13_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_13_end_0 = const()[name = tensor("x_13_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_13_end_mask_0 = const()[name = tensor("x_13_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_13_squeeze_mask_0 = const()[name = tensor("x_13_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_13_cast_fp16 = slice_by_index(begin = x_13_begin_0, end = x_13_end_0, end_mask = x_13_end_mask_0, squeeze_mask = x_13_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("x_13_cast_fp16")]; + tensor x_15_begin_0 = const()[name = tensor("x_15_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_15_end_0 = const()[name = tensor("x_15_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_15_end_mask_0 = const()[name = tensor("x_15_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_15_squeeze_mask_0 = const()[name = tensor("x_15_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_15_cast_fp16 = slice_by_index(begin = x_15_begin_0, end = x_15_end_0, end_mask = x_15_end_mask_0, squeeze_mask = x_15_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("x_15_cast_fp16")]; + tensor v_7_begin_0 = const()[name = tensor("v_7_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_7_end_0 = const()[name = tensor("v_7_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_7_end_mask_0 = const()[name = tensor("v_7_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_7_squeeze_mask_0 = const()[name = tensor("v_7_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_7_cast_fp16 = slice_by_index(begin = v_7_begin_0, end = v_7_end_0, end_mask = v_7_end_mask_0, squeeze_mask = v_7_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("v_7_cast_fp16")]; + tensor var_839 = const()[name = tensor("op_839"), val = tensor([1, 1, 1])]; + tensor position3_to_fp16_dtype_0 = const()[name = tensor("position3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position3_to_fp16 = cast(dtype = position3_to_fp16_dtype_0, x = position3)[name = tensor("cast_121")]; + tensor var_840_cast_fp16 = reshape(shape = var_839, x = position3_to_fp16)[name = tensor("op_840_cast_fp16")]; + tensor angles_13_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_840_cast_fp16)[name = tensor("angles_13_cast_fp16")]; + tensor rotr_13_cast_fp16 = cos(x = angles_13_cast_fp16)[name = tensor("rotr_13_cast_fp16")]; + tensor roti_13_cast_fp16 = sin(x = angles_13_cast_fp16)[name = tensor("roti_13_cast_fp16")]; + tensor var_848 = const()[name = tensor("op_848"), val = tensor([1, 16, 32, 2])]; + tensor pairs_13_cast_fp16 = reshape(shape = var_848, x = x_13_cast_fp16)[name = tensor("pairs_13_cast_fp16")]; + tensor xr_13_begin_0 = const()[name = tensor("xr_13_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_13_end_0 = const()[name = tensor("xr_13_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_13_end_mask_0 = const()[name = tensor("xr_13_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_13_squeeze_mask_0 = const()[name = tensor("xr_13_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_13_cast_fp16 = slice_by_index(begin = xr_13_begin_0, end = xr_13_end_0, end_mask = xr_13_end_mask_0, squeeze_mask = xr_13_squeeze_mask_0, x = pairs_13_cast_fp16)[name = tensor("xr_13_cast_fp16")]; + tensor xi_13_begin_0 = const()[name = tensor("xi_13_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_13_end_0 = const()[name = tensor("xi_13_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_13_end_mask_0 = const()[name = tensor("xi_13_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_13_squeeze_mask_0 = const()[name = tensor("xi_13_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_13_cast_fp16 = slice_by_index(begin = xi_13_begin_0, end = xi_13_end_0, end_mask = xi_13_end_mask_0, squeeze_mask = xi_13_squeeze_mask_0, x = pairs_13_cast_fp16)[name = tensor("xi_13_cast_fp16")]; + tensor var_856_cast_fp16 = mul(x = xr_13_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_856_cast_fp16")]; + tensor var_857_cast_fp16 = mul(x = xi_13_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_857_cast_fp16")]; + tensor xor__13_cast_fp16 = sub(x = var_856_cast_fp16, y = var_857_cast_fp16)[name = tensor("xor__13_cast_fp16")]; + tensor var_860_cast_fp16 = mul(x = xr_13_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_860_cast_fp16")]; + tensor var_861_cast_fp16 = mul(x = xi_13_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_861_cast_fp16")]; + tensor xoi_13_cast_fp16 = add(x = var_860_cast_fp16, y = var_861_cast_fp16)[name = tensor("xoi_13_cast_fp16")]; + tensor out_13_axis_0 = const()[name = tensor("out_13_axis_0"), val = tensor(-1)]; + tensor out_13_cast_fp16 = stack(axis = out_13_axis_0, values = (xor__13_cast_fp16, xoi_13_cast_fp16))[name = tensor("out_13_cast_fp16")]; + tensor var_884 = const()[name = tensor("op_884"), val = tensor([1, 16, 32, 2])]; + tensor pairs_15_cast_fp16 = reshape(shape = var_884, x = x_15_cast_fp16)[name = tensor("pairs_15_cast_fp16")]; + tensor xr_15_begin_0 = const()[name = tensor("xr_15_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_15_end_0 = const()[name = tensor("xr_15_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_15_end_mask_0 = const()[name = tensor("xr_15_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_15_squeeze_mask_0 = const()[name = tensor("xr_15_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_15_cast_fp16 = slice_by_index(begin = xr_15_begin_0, end = xr_15_end_0, end_mask = xr_15_end_mask_0, squeeze_mask = xr_15_squeeze_mask_0, x = pairs_15_cast_fp16)[name = tensor("xr_15_cast_fp16")]; + tensor xi_15_begin_0 = const()[name = tensor("xi_15_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_15_end_0 = const()[name = tensor("xi_15_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_15_end_mask_0 = const()[name = tensor("xi_15_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_15_squeeze_mask_0 = const()[name = tensor("xi_15_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_15_cast_fp16 = slice_by_index(begin = xi_15_begin_0, end = xi_15_end_0, end_mask = xi_15_end_mask_0, squeeze_mask = xi_15_squeeze_mask_0, x = pairs_15_cast_fp16)[name = tensor("xi_15_cast_fp16")]; + tensor var_892_cast_fp16 = mul(x = xr_15_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_892_cast_fp16")]; + tensor var_893_cast_fp16 = mul(x = xi_15_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_893_cast_fp16")]; + tensor xor__15_cast_fp16 = sub(x = var_892_cast_fp16, y = var_893_cast_fp16)[name = tensor("xor__15_cast_fp16")]; + tensor var_896_cast_fp16 = mul(x = xr_15_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_896_cast_fp16")]; + tensor var_897_cast_fp16 = mul(x = xi_15_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_897_cast_fp16")]; + tensor xoi_15_cast_fp16 = add(x = var_896_cast_fp16, y = var_897_cast_fp16)[name = tensor("xoi_15_cast_fp16")]; + tensor out_15_axis_0 = const()[name = tensor("out_15_axis_0"), val = tensor(-1)]; + tensor out_15_cast_fp16 = stack(axis = out_15_axis_0, values = (xor__15_cast_fp16, xoi_15_cast_fp16))[name = tensor("out_15_cast_fp16")]; + tensor _inversed_909_y_0_to_fp16 = const()[name = tensor("_inversed_909_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_909_cast_fp16 = mul(x = position3_to_fp16, y = _inversed_909_y_0_to_fp16)[name = tensor("_inversed_909_cast_fp16")]; + tensor var_910_cast_fp16 = floor(x = _inversed_909_cast_fp16)[name = tensor("op_910_cast_fp16")]; + tensor var_911_to_fp16 = const()[name = tensor("op_911_to_fp16"), val = tensor(0x1p+9)]; + tensor var_912_cast_fp16 = mul(x = var_910_cast_fp16, y = var_911_to_fp16)[name = tensor("op_912_cast_fp16")]; + tensor write_idx_7_cast_fp16 = sub(x = position3_to_fp16, y = var_912_cast_fp16)[name = tensor("write_idx_7_cast_fp16")]; + tensor var_915_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_7_cast_fp16)[name = tensor("op_915_cast_fp16")]; + tensor var_925 = const()[name = tensor("op_925"), val = tensor([1, 512, 1, 1])]; + tensor onehot_13_to_fp16_dtype_0 = const()[name = tensor("onehot_13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_915_cast_fp16_to_fp16 = cast(dtype = onehot_13_to_fp16_dtype_0, x = var_915_cast_fp16)[name = tensor("cast_120")]; + tensor onehot_15_cast_fp16 = reshape(shape = var_925, x = var_915_cast_fp16_to_fp16)[name = tensor("onehot_15_cast_fp16")]; + tensor var_927_to_fp16 = const()[name = tensor("op_927_to_fp16"), val = tensor(0x1p+0)]; + tensor var_929_cast_fp16 = sub(x = var_927_to_fp16, y = onehot_15_cast_fp16)[name = tensor("op_929_cast_fp16")]; + tensor k_cache3_to_fp16_dtype_0 = const()[name = tensor("k_cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache3_to_fp16 = cast(dtype = k_cache3_to_fp16_dtype_0, x = k_cache3)[name = tensor("cast_119")]; + tensor var_930_cast_fp16 = mul(x = k_cache3_to_fp16, y = var_929_cast_fp16)[name = tensor("op_930_cast_fp16")]; + tensor var_935 = const()[name = tensor("op_935"), val = tensor([1, 1, 16, 64])]; + tensor var_936_cast_fp16 = reshape(shape = var_935, x = out_15_cast_fp16)[name = tensor("op_936_cast_fp16")]; + tensor var_937_cast_fp16 = mul(x = var_936_cast_fp16, y = onehot_15_cast_fp16)[name = tensor("op_937_cast_fp16")]; + tensor new_k_cache_7_cast_fp16 = add(x = var_930_cast_fp16, y = var_937_cast_fp16)[name = tensor("new_k_cache_7_cast_fp16")]; + tensor new_k_cache_7_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_7_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache3_to_fp16_dtype_0 = const()[name = tensor("v_cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache3_to_fp16 = cast(dtype = v_cache3_to_fp16_dtype_0, x = v_cache3)[name = tensor("cast_117")]; + tensor var_943_cast_fp16 = mul(x = v_cache3_to_fp16, y = var_929_cast_fp16)[name = tensor("op_943_cast_fp16")]; + tensor var_948 = const()[name = tensor("op_948"), val = tensor([1, 1, 16, 64])]; + tensor var_949_cast_fp16 = reshape(shape = var_948, x = v_7_cast_fp16)[name = tensor("op_949_cast_fp16")]; + tensor var_950_cast_fp16 = mul(x = var_949_cast_fp16, y = onehot_15_cast_fp16)[name = tensor("op_950_cast_fp16")]; + tensor new_v_cache_7_cast_fp16 = add(x = var_943_cast_fp16, y = var_950_cast_fp16)[name = tensor("new_v_cache_7_cast_fp16")]; + tensor new_v_cache_7_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_7_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_3_perm_0 = const()[name = tensor("transpose_3_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_963 = const()[name = tensor("op_963"), val = tensor([0, 2, 1, 3])]; + tensor var_969 = const()[name = tensor("op_969"), val = tensor([1, 16, 1, 64])]; + tensor q4_7_cast_fp16 = reshape(shape = var_969, x = out_13_cast_fp16)[name = tensor("q4_7_cast_fp16")]; + tensor var_974_transpose_x_0 = const()[name = tensor("op_974_transpose_x_0"), val = tensor(false)]; + tensor var_974_transpose_y_0 = const()[name = tensor("op_974_transpose_y_0"), val = tensor(false)]; + tensor transpose_3_cast_fp16 = transpose(perm = transpose_3_perm_0, x = new_k_cache_7_cast_fp16)[name = tensor("transpose_14")]; + tensor var_974_cast_fp16 = matmul(transpose_x = var_974_transpose_x_0, transpose_y = var_974_transpose_y_0, x = q4_7_cast_fp16, y = transpose_3_cast_fp16)[name = tensor("op_974_cast_fp16")]; + tensor var_975_to_fp16 = const()[name = tensor("op_975_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_13_cast_fp16 = mul(x = var_974_cast_fp16, y = var_975_to_fp16)[name = tensor("scores_13_cast_fp16")]; + tensor var_979 = const()[name = tensor("op_979"), val = tensor([1, 1])]; + tensor var_980_cast_fp16 = reshape(shape = var_979, x = position3_to_fp16)[name = tensor("op_980_cast_fp16")]; + tensor var_981_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_980_cast_fp16)[name = tensor("op_981_cast_fp16")]; + tensor mask_7_to_fp16_dtype_0 = const()[name = tensor("mask_7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_988_to_fp16 = const()[name = tensor("op_988_to_fp16"), val = tensor(0x1p+0)]; + tensor var_981_cast_fp16_to_fp16 = cast(dtype = mask_7_to_fp16_dtype_0, x = var_981_cast_fp16)[name = tensor("cast_115")]; + tensor var_989_cast_fp16 = sub(x = var_981_cast_fp16_to_fp16, y = var_988_to_fp16)[name = tensor("op_989_cast_fp16")]; + tensor var_994 = const()[name = tensor("op_994"), val = tensor([1, 1, 1, 512])]; + tensor var_995_cast_fp16 = reshape(shape = var_994, x = var_989_cast_fp16)[name = tensor("op_995_cast_fp16")]; + tensor var_996_to_fp16 = const()[name = tensor("op_996_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_997_cast_fp16 = mul(x = var_995_cast_fp16, y = var_996_to_fp16)[name = tensor("op_997_cast_fp16")]; + tensor scores_15_cast_fp16 = add(x = scores_13_cast_fp16, y = var_997_cast_fp16)[name = tensor("scores_15_cast_fp16")]; + tensor var_1000 = const()[name = tensor("op_1000"), val = tensor(-1)]; + tensor weights_7_cast_fp16 = softmax(axis = var_1000, x = scores_15_cast_fp16)[name = tensor("weights_7_cast_fp16")]; + tensor attn_7_transpose_x_0 = const()[name = tensor("attn_7_transpose_x_0"), val = tensor(false)]; + tensor attn_7_transpose_y_0 = const()[name = tensor("attn_7_transpose_y_0"), val = tensor(false)]; + tensor values_7_cast_fp16 = transpose(perm = var_963, x = new_v_cache_7_cast_fp16)[name = tensor("transpose_13")]; + tensor attn_7_cast_fp16 = matmul(transpose_x = attn_7_transpose_x_0, transpose_y = attn_7_transpose_y_0, x = weights_7_cast_fp16, y = values_7_cast_fp16)[name = tensor("attn_7_cast_fp16")]; + tensor var_1008 = const()[name = tensor("op_1008"), val = tensor([0, 2, 1, 3])]; + tensor var_1013 = const()[name = tensor("op_1013"), val = tensor([1, 1, 1024])]; + tensor var_1009_cast_fp16 = transpose(perm = var_1008, x = attn_7_cast_fp16)[name = tensor("transpose_12")]; + tensor input_43_cast_fp16 = reshape(shape = var_1013, x = var_1009_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor attn3_out_proj_weight_to_fp16 = const()[name = tensor("attn3_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(81902784)))]; + tensor linear_14_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn3_out_proj_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor var_1019_to_fp16 = const()[name = tensor("op_1019_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1020_cast_fp16 = add(x = position3_to_fp16, y = var_1019_to_fp16)[name = tensor("op_1020_cast_fp16")]; + tensor var_1020_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1020_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_45_cast_fp16 = add(x = input_39_cast_fp16, y = linear_14_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor input_47_axes_0 = const()[name = tensor("input_47_axes_0"), val = tensor([-1])]; + tensor norm3_2_weight_to_fp16 = const()[name = tensor("norm3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84000000)))]; + tensor norm3_2_bias_to_fp16 = const()[name = tensor("norm3_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84002112)))]; + tensor var_1024_to_fp16 = const()[name = tensor("op_1024_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_47_cast_fp16 = layer_norm(axes = input_47_axes_0, beta = norm3_2_bias_to_fp16, epsilon = var_1024_to_fp16, gamma = norm3_2_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor linear3_1_weight_to_fp16 = const()[name = tensor("linear3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84004224)))]; + tensor linear_15_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear3_1_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor input_49_mode_0 = const()[name = tensor("input_49_mode_0"), val = tensor("EXACT")]; + tensor input_49_cast_fp16 = gelu(mode = input_49_mode_0, x = linear_15_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor linear3_2_weight_to_fp16 = const()[name = tensor("linear3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(92392896)))]; + tensor linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear3_2_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor input_51_cast_fp16 = add(x = input_45_cast_fp16, y = linear_16_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor input_53_axes_0 = const()[name = tensor("input_53_axes_0"), val = tensor([-1])]; + tensor norm4_1_weight_to_fp16 = const()[name = tensor("norm4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100781568)))]; + tensor norm4_1_bias_to_fp16 = const()[name = tensor("norm4_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100783680)))]; + tensor var_1041_to_fp16 = const()[name = tensor("op_1041_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_53_cast_fp16 = layer_norm(axes = input_53_axes_0, beta = norm4_1_bias_to_fp16, epsilon = var_1041_to_fp16, gamma = norm4_1_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor attn4_in_proj_weight_to_fp16 = const()[name = tensor("attn4_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100785792)))]; + tensor linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn4_in_proj_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor var_1054 = const()[name = tensor("op_1054"), val = tensor([1, 3, 16, 64])]; + tensor qkv_9_cast_fp16 = reshape(shape = var_1054, x = linear_17_cast_fp16)[name = tensor("qkv_9_cast_fp16")]; + tensor x_17_begin_0 = const()[name = tensor("x_17_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_17_end_0 = const()[name = tensor("x_17_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_17_end_mask_0 = const()[name = tensor("x_17_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_17_squeeze_mask_0 = const()[name = tensor("x_17_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_17_cast_fp16 = slice_by_index(begin = x_17_begin_0, end = x_17_end_0, end_mask = x_17_end_mask_0, squeeze_mask = x_17_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("x_17_cast_fp16")]; + tensor x_19_begin_0 = const()[name = tensor("x_19_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_19_end_0 = const()[name = tensor("x_19_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_19_end_mask_0 = const()[name = tensor("x_19_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_19_squeeze_mask_0 = const()[name = tensor("x_19_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_19_cast_fp16 = slice_by_index(begin = x_19_begin_0, end = x_19_end_0, end_mask = x_19_end_mask_0, squeeze_mask = x_19_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("x_19_cast_fp16")]; + tensor v_9_begin_0 = const()[name = tensor("v_9_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_9_end_0 = const()[name = tensor("v_9_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_9_end_mask_0 = const()[name = tensor("v_9_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_9_squeeze_mask_0 = const()[name = tensor("v_9_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_9_cast_fp16 = slice_by_index(begin = v_9_begin_0, end = v_9_end_0, end_mask = v_9_end_mask_0, squeeze_mask = v_9_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("v_9_cast_fp16")]; + tensor var_1083 = const()[name = tensor("op_1083"), val = tensor([1, 1, 1])]; + tensor position4_to_fp16_dtype_0 = const()[name = tensor("position4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position4_to_fp16 = cast(dtype = position4_to_fp16_dtype_0, x = position4)[name = tensor("cast_113")]; + tensor var_1084_cast_fp16 = reshape(shape = var_1083, x = position4_to_fp16)[name = tensor("op_1084_cast_fp16")]; + tensor angles_17_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_1084_cast_fp16)[name = tensor("angles_17_cast_fp16")]; + tensor rotr_17_cast_fp16 = cos(x = angles_17_cast_fp16)[name = tensor("rotr_17_cast_fp16")]; + tensor roti_17_cast_fp16 = sin(x = angles_17_cast_fp16)[name = tensor("roti_17_cast_fp16")]; + tensor var_1092 = const()[name = tensor("op_1092"), val = tensor([1, 16, 32, 2])]; + tensor pairs_17_cast_fp16 = reshape(shape = var_1092, x = x_17_cast_fp16)[name = tensor("pairs_17_cast_fp16")]; + tensor xr_17_begin_0 = const()[name = tensor("xr_17_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_17_end_0 = const()[name = tensor("xr_17_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_17_end_mask_0 = const()[name = tensor("xr_17_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_17_squeeze_mask_0 = const()[name = tensor("xr_17_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_17_cast_fp16 = slice_by_index(begin = xr_17_begin_0, end = xr_17_end_0, end_mask = xr_17_end_mask_0, squeeze_mask = xr_17_squeeze_mask_0, x = pairs_17_cast_fp16)[name = tensor("xr_17_cast_fp16")]; + tensor xi_17_begin_0 = const()[name = tensor("xi_17_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_17_end_0 = const()[name = tensor("xi_17_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_17_end_mask_0 = const()[name = tensor("xi_17_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_17_squeeze_mask_0 = const()[name = tensor("xi_17_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_17_cast_fp16 = slice_by_index(begin = xi_17_begin_0, end = xi_17_end_0, end_mask = xi_17_end_mask_0, squeeze_mask = xi_17_squeeze_mask_0, x = pairs_17_cast_fp16)[name = tensor("xi_17_cast_fp16")]; + tensor var_1100_cast_fp16 = mul(x = xr_17_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_1100_cast_fp16")]; + tensor var_1101_cast_fp16 = mul(x = xi_17_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_1101_cast_fp16")]; + tensor xor__17_cast_fp16 = sub(x = var_1100_cast_fp16, y = var_1101_cast_fp16)[name = tensor("xor__17_cast_fp16")]; + tensor var_1104_cast_fp16 = mul(x = xr_17_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_1104_cast_fp16")]; + tensor var_1105_cast_fp16 = mul(x = xi_17_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_1105_cast_fp16")]; + tensor xoi_17_cast_fp16 = add(x = var_1104_cast_fp16, y = var_1105_cast_fp16)[name = tensor("xoi_17_cast_fp16")]; + tensor out_17_axis_0 = const()[name = tensor("out_17_axis_0"), val = tensor(-1)]; + tensor out_17_cast_fp16 = stack(axis = out_17_axis_0, values = (xor__17_cast_fp16, xoi_17_cast_fp16))[name = tensor("out_17_cast_fp16")]; + tensor var_1128 = const()[name = tensor("op_1128"), val = tensor([1, 16, 32, 2])]; + tensor pairs_19_cast_fp16 = reshape(shape = var_1128, x = x_19_cast_fp16)[name = tensor("pairs_19_cast_fp16")]; + tensor xr_19_begin_0 = const()[name = tensor("xr_19_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_19_end_0 = const()[name = tensor("xr_19_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_19_end_mask_0 = const()[name = tensor("xr_19_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_19_squeeze_mask_0 = const()[name = tensor("xr_19_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_19_cast_fp16 = slice_by_index(begin = xr_19_begin_0, end = xr_19_end_0, end_mask = xr_19_end_mask_0, squeeze_mask = xr_19_squeeze_mask_0, x = pairs_19_cast_fp16)[name = tensor("xr_19_cast_fp16")]; + tensor xi_19_begin_0 = const()[name = tensor("xi_19_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_19_end_0 = const()[name = tensor("xi_19_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_19_end_mask_0 = const()[name = tensor("xi_19_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_19_squeeze_mask_0 = const()[name = tensor("xi_19_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_19_cast_fp16 = slice_by_index(begin = xi_19_begin_0, end = xi_19_end_0, end_mask = xi_19_end_mask_0, squeeze_mask = xi_19_squeeze_mask_0, x = pairs_19_cast_fp16)[name = tensor("xi_19_cast_fp16")]; + tensor var_1136_cast_fp16 = mul(x = xr_19_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_1136_cast_fp16")]; + tensor var_1137_cast_fp16 = mul(x = xi_19_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_1137_cast_fp16")]; + tensor xor__19_cast_fp16 = sub(x = var_1136_cast_fp16, y = var_1137_cast_fp16)[name = tensor("xor__19_cast_fp16")]; + tensor var_1140_cast_fp16 = mul(x = xr_19_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_1140_cast_fp16")]; + tensor var_1141_cast_fp16 = mul(x = xi_19_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_1141_cast_fp16")]; + tensor xoi_19_cast_fp16 = add(x = var_1140_cast_fp16, y = var_1141_cast_fp16)[name = tensor("xoi_19_cast_fp16")]; + tensor out_19_axis_0 = const()[name = tensor("out_19_axis_0"), val = tensor(-1)]; + tensor out_19_cast_fp16 = stack(axis = out_19_axis_0, values = (xor__19_cast_fp16, xoi_19_cast_fp16))[name = tensor("out_19_cast_fp16")]; + tensor _inversed_1153_y_0_to_fp16 = const()[name = tensor("_inversed_1153_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1153_cast_fp16 = mul(x = position4_to_fp16, y = _inversed_1153_y_0_to_fp16)[name = tensor("_inversed_1153_cast_fp16")]; + tensor var_1154_cast_fp16 = floor(x = _inversed_1153_cast_fp16)[name = tensor("op_1154_cast_fp16")]; + tensor var_1155_to_fp16 = const()[name = tensor("op_1155_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1156_cast_fp16 = mul(x = var_1154_cast_fp16, y = var_1155_to_fp16)[name = tensor("op_1156_cast_fp16")]; + tensor write_idx_9_cast_fp16 = sub(x = position4_to_fp16, y = var_1156_cast_fp16)[name = tensor("write_idx_9_cast_fp16")]; + tensor var_1159_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_9_cast_fp16)[name = tensor("op_1159_cast_fp16")]; + tensor var_1169 = const()[name = tensor("op_1169"), val = tensor([1, 512, 1, 1])]; + tensor onehot_17_to_fp16_dtype_0 = const()[name = tensor("onehot_17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_1159_cast_fp16_to_fp16 = cast(dtype = onehot_17_to_fp16_dtype_0, x = var_1159_cast_fp16)[name = tensor("cast_112")]; + tensor onehot_19_cast_fp16 = reshape(shape = var_1169, x = var_1159_cast_fp16_to_fp16)[name = tensor("onehot_19_cast_fp16")]; + tensor var_1171_to_fp16 = const()[name = tensor("op_1171_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1173_cast_fp16 = sub(x = var_1171_to_fp16, y = onehot_19_cast_fp16)[name = tensor("op_1173_cast_fp16")]; + tensor k_cache4_to_fp16_dtype_0 = const()[name = tensor("k_cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache4_to_fp16 = cast(dtype = k_cache4_to_fp16_dtype_0, x = k_cache4)[name = tensor("cast_111")]; + tensor var_1174_cast_fp16 = mul(x = k_cache4_to_fp16, y = var_1173_cast_fp16)[name = tensor("op_1174_cast_fp16")]; + tensor var_1179 = const()[name = tensor("op_1179"), val = tensor([1, 1, 16, 64])]; + tensor var_1180_cast_fp16 = reshape(shape = var_1179, x = out_19_cast_fp16)[name = tensor("op_1180_cast_fp16")]; + tensor var_1181_cast_fp16 = mul(x = var_1180_cast_fp16, y = onehot_19_cast_fp16)[name = tensor("op_1181_cast_fp16")]; + tensor new_k_cache_9_cast_fp16 = add(x = var_1174_cast_fp16, y = var_1181_cast_fp16)[name = tensor("new_k_cache_9_cast_fp16")]; + tensor new_k_cache_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache4_to_fp16_dtype_0 = const()[name = tensor("v_cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache4_to_fp16 = cast(dtype = v_cache4_to_fp16_dtype_0, x = v_cache4)[name = tensor("cast_109")]; + tensor var_1187_cast_fp16 = mul(x = v_cache4_to_fp16, y = var_1173_cast_fp16)[name = tensor("op_1187_cast_fp16")]; + tensor var_1192 = const()[name = tensor("op_1192"), val = tensor([1, 1, 16, 64])]; + tensor var_1193_cast_fp16 = reshape(shape = var_1192, x = v_9_cast_fp16)[name = tensor("op_1193_cast_fp16")]; + tensor var_1194_cast_fp16 = mul(x = var_1193_cast_fp16, y = onehot_19_cast_fp16)[name = tensor("op_1194_cast_fp16")]; + tensor new_v_cache_9_cast_fp16 = add(x = var_1187_cast_fp16, y = var_1194_cast_fp16)[name = tensor("new_v_cache_9_cast_fp16")]; + tensor new_v_cache_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_4_perm_0 = const()[name = tensor("transpose_4_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_1207 = const()[name = tensor("op_1207"), val = tensor([0, 2, 1, 3])]; + tensor var_1213 = const()[name = tensor("op_1213"), val = tensor([1, 16, 1, 64])]; + tensor q4_9_cast_fp16 = reshape(shape = var_1213, x = out_17_cast_fp16)[name = tensor("q4_9_cast_fp16")]; + tensor var_1218_transpose_x_0 = const()[name = tensor("op_1218_transpose_x_0"), val = tensor(false)]; + tensor var_1218_transpose_y_0 = const()[name = tensor("op_1218_transpose_y_0"), val = tensor(false)]; + tensor transpose_4_cast_fp16 = transpose(perm = transpose_4_perm_0, x = new_k_cache_9_cast_fp16)[name = tensor("transpose_11")]; + tensor var_1218_cast_fp16 = matmul(transpose_x = var_1218_transpose_x_0, transpose_y = var_1218_transpose_y_0, x = q4_9_cast_fp16, y = transpose_4_cast_fp16)[name = tensor("op_1218_cast_fp16")]; + tensor var_1219_to_fp16 = const()[name = tensor("op_1219_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_17_cast_fp16 = mul(x = var_1218_cast_fp16, y = var_1219_to_fp16)[name = tensor("scores_17_cast_fp16")]; + tensor var_1223 = const()[name = tensor("op_1223"), val = tensor([1, 1])]; + tensor var_1224_cast_fp16 = reshape(shape = var_1223, x = position4_to_fp16)[name = tensor("op_1224_cast_fp16")]; + tensor var_1225_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_1224_cast_fp16)[name = tensor("op_1225_cast_fp16")]; + tensor mask_9_to_fp16_dtype_0 = const()[name = tensor("mask_9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_1232_to_fp16 = const()[name = tensor("op_1232_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1225_cast_fp16_to_fp16 = cast(dtype = mask_9_to_fp16_dtype_0, x = var_1225_cast_fp16)[name = tensor("cast_107")]; + tensor var_1233_cast_fp16 = sub(x = var_1225_cast_fp16_to_fp16, y = var_1232_to_fp16)[name = tensor("op_1233_cast_fp16")]; + tensor var_1238 = const()[name = tensor("op_1238"), val = tensor([1, 1, 1, 512])]; + tensor var_1239_cast_fp16 = reshape(shape = var_1238, x = var_1233_cast_fp16)[name = tensor("op_1239_cast_fp16")]; + tensor var_1240_to_fp16 = const()[name = tensor("op_1240_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_1241_cast_fp16 = mul(x = var_1239_cast_fp16, y = var_1240_to_fp16)[name = tensor("op_1241_cast_fp16")]; + tensor scores_19_cast_fp16 = add(x = scores_17_cast_fp16, y = var_1241_cast_fp16)[name = tensor("scores_19_cast_fp16")]; + tensor var_1244 = const()[name = tensor("op_1244"), val = tensor(-1)]; + tensor weights_9_cast_fp16 = softmax(axis = var_1244, x = scores_19_cast_fp16)[name = tensor("weights_9_cast_fp16")]; + tensor attn_9_transpose_x_0 = const()[name = tensor("attn_9_transpose_x_0"), val = tensor(false)]; + tensor attn_9_transpose_y_0 = const()[name = tensor("attn_9_transpose_y_0"), val = tensor(false)]; + tensor values_9_cast_fp16 = transpose(perm = var_1207, x = new_v_cache_9_cast_fp16)[name = tensor("transpose_10")]; + tensor attn_9_cast_fp16 = matmul(transpose_x = attn_9_transpose_x_0, transpose_y = attn_9_transpose_y_0, x = weights_9_cast_fp16, y = values_9_cast_fp16)[name = tensor("attn_9_cast_fp16")]; + tensor var_1252 = const()[name = tensor("op_1252"), val = tensor([0, 2, 1, 3])]; + tensor var_1257 = const()[name = tensor("op_1257"), val = tensor([1, 1, 1024])]; + tensor var_1253_cast_fp16 = transpose(perm = var_1252, x = attn_9_cast_fp16)[name = tensor("transpose_9")]; + tensor input_55_cast_fp16 = reshape(shape = var_1257, x = var_1253_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor attn4_out_proj_weight_to_fp16 = const()[name = tensor("attn4_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(107077312)))]; + tensor linear_18_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn4_out_proj_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor var_1263_to_fp16 = const()[name = tensor("op_1263_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1264_cast_fp16 = add(x = position4_to_fp16, y = var_1263_to_fp16)[name = tensor("op_1264_cast_fp16")]; + tensor var_1264_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1264_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_57_cast_fp16 = add(x = input_51_cast_fp16, y = linear_18_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor input_59_axes_0 = const()[name = tensor("input_59_axes_0"), val = tensor([-1])]; + tensor norm4_2_weight_to_fp16 = const()[name = tensor("norm4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109174528)))]; + tensor norm4_2_bias_to_fp16 = const()[name = tensor("norm4_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109176640)))]; + tensor var_1268_to_fp16 = const()[name = tensor("op_1268_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_59_cast_fp16 = layer_norm(axes = input_59_axes_0, beta = norm4_2_bias_to_fp16, epsilon = var_1268_to_fp16, gamma = norm4_2_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor linear4_1_weight_to_fp16 = const()[name = tensor("linear4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109178752)))]; + tensor linear_19_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear4_1_weight_to_fp16, x = input_59_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_61_mode_0 = const()[name = tensor("input_61_mode_0"), val = tensor("EXACT")]; + tensor input_61_cast_fp16 = gelu(mode = input_61_mode_0, x = linear_19_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor linear4_2_weight_to_fp16 = const()[name = tensor("linear4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(117567424)))]; + tensor linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear4_2_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor input_63_cast_fp16 = add(x = input_57_cast_fp16, y = linear_20_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor input_65_axes_0 = const()[name = tensor("input_65_axes_0"), val = tensor([-1])]; + tensor norm5_1_weight_to_fp16 = const()[name = tensor("norm5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125956096)))]; + tensor norm5_1_bias_to_fp16 = const()[name = tensor("norm5_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125958208)))]; + tensor var_1285_to_fp16 = const()[name = tensor("op_1285_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_65_cast_fp16 = layer_norm(axes = input_65_axes_0, beta = norm5_1_bias_to_fp16, epsilon = var_1285_to_fp16, gamma = norm5_1_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("input_65_cast_fp16")]; + tensor attn5_in_proj_weight_to_fp16 = const()[name = tensor("attn5_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125960320)))]; + tensor linear_21_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn5_in_proj_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("linear_21_cast_fp16")]; + tensor var_1298 = const()[name = tensor("op_1298"), val = tensor([1, 3, 16, 64])]; + tensor qkv_cast_fp16 = reshape(shape = var_1298, x = linear_21_cast_fp16)[name = tensor("qkv_cast_fp16")]; + tensor x_21_begin_0 = const()[name = tensor("x_21_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_21_end_0 = const()[name = tensor("x_21_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_21_end_mask_0 = const()[name = tensor("x_21_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_21_squeeze_mask_0 = const()[name = tensor("x_21_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_21_cast_fp16 = slice_by_index(begin = x_21_begin_0, end = x_21_end_0, end_mask = x_21_end_mask_0, squeeze_mask = x_21_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("x_21_cast_fp16")]; + tensor x_begin_0 = const()[name = tensor("x_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_end_0 = const()[name = tensor("x_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_end_mask_0 = const()[name = tensor("x_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_squeeze_mask_0 = const()[name = tensor("x_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_cast_fp16 = slice_by_index(begin = x_begin_0, end = x_end_0, end_mask = x_end_mask_0, squeeze_mask = x_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("x_cast_fp16")]; + tensor v_begin_0 = const()[name = tensor("v_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_end_0 = const()[name = tensor("v_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_end_mask_0 = const()[name = tensor("v_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_squeeze_mask_0 = const()[name = tensor("v_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_cast_fp16 = slice_by_index(begin = v_begin_0, end = v_end_0, end_mask = v_end_mask_0, squeeze_mask = v_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("v_cast_fp16")]; + tensor var_1327 = const()[name = tensor("op_1327"), val = tensor([1, 1, 1])]; + tensor position5_to_fp16_dtype_0 = const()[name = tensor("position5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position5_to_fp16 = cast(dtype = position5_to_fp16_dtype_0, x = position5)[name = tensor("cast_105")]; + tensor var_1328_cast_fp16 = reshape(shape = var_1327, x = position5_to_fp16)[name = tensor("op_1328_cast_fp16")]; + tensor angles_21_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_1328_cast_fp16)[name = tensor("angles_21_cast_fp16")]; + tensor rotr_21_cast_fp16 = cos(x = angles_21_cast_fp16)[name = tensor("rotr_21_cast_fp16")]; + tensor roti_21_cast_fp16 = sin(x = angles_21_cast_fp16)[name = tensor("roti_21_cast_fp16")]; + tensor var_1336 = const()[name = tensor("op_1336"), val = tensor([1, 16, 32, 2])]; + tensor pairs_21_cast_fp16 = reshape(shape = var_1336, x = x_21_cast_fp16)[name = tensor("pairs_21_cast_fp16")]; + tensor xr_21_begin_0 = const()[name = tensor("xr_21_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_21_end_0 = const()[name = tensor("xr_21_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_21_end_mask_0 = const()[name = tensor("xr_21_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_21_squeeze_mask_0 = const()[name = tensor("xr_21_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_21_cast_fp16 = slice_by_index(begin = xr_21_begin_0, end = xr_21_end_0, end_mask = xr_21_end_mask_0, squeeze_mask = xr_21_squeeze_mask_0, x = pairs_21_cast_fp16)[name = tensor("xr_21_cast_fp16")]; + tensor xi_21_begin_0 = const()[name = tensor("xi_21_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_21_end_0 = const()[name = tensor("xi_21_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_21_end_mask_0 = const()[name = tensor("xi_21_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_21_squeeze_mask_0 = const()[name = tensor("xi_21_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_21_cast_fp16 = slice_by_index(begin = xi_21_begin_0, end = xi_21_end_0, end_mask = xi_21_end_mask_0, squeeze_mask = xi_21_squeeze_mask_0, x = pairs_21_cast_fp16)[name = tensor("xi_21_cast_fp16")]; + tensor var_1344_cast_fp16 = mul(x = xr_21_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_1344_cast_fp16")]; + tensor var_1345_cast_fp16 = mul(x = xi_21_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_1345_cast_fp16")]; + tensor xor__21_cast_fp16 = sub(x = var_1344_cast_fp16, y = var_1345_cast_fp16)[name = tensor("xor__21_cast_fp16")]; + tensor var_1348_cast_fp16 = mul(x = xr_21_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_1348_cast_fp16")]; + tensor var_1349_cast_fp16 = mul(x = xi_21_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_1349_cast_fp16")]; + tensor xoi_21_cast_fp16 = add(x = var_1348_cast_fp16, y = var_1349_cast_fp16)[name = tensor("xoi_21_cast_fp16")]; + tensor out_21_axis_0 = const()[name = tensor("out_21_axis_0"), val = tensor(-1)]; + tensor out_21_cast_fp16 = stack(axis = out_21_axis_0, values = (xor__21_cast_fp16, xoi_21_cast_fp16))[name = tensor("out_21_cast_fp16")]; + tensor var_1372 = const()[name = tensor("op_1372"), val = tensor([1, 16, 32, 2])]; + tensor pairs_cast_fp16 = reshape(shape = var_1372, x = x_cast_fp16)[name = tensor("pairs_cast_fp16")]; + tensor xr_begin_0 = const()[name = tensor("xr_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_end_0 = const()[name = tensor("xr_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_end_mask_0 = const()[name = tensor("xr_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_squeeze_mask_0 = const()[name = tensor("xr_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_cast_fp16 = slice_by_index(begin = xr_begin_0, end = xr_end_0, end_mask = xr_end_mask_0, squeeze_mask = xr_squeeze_mask_0, x = pairs_cast_fp16)[name = tensor("xr_cast_fp16")]; + tensor xi_begin_0 = const()[name = tensor("xi_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_end_0 = const()[name = tensor("xi_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_end_mask_0 = const()[name = tensor("xi_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_squeeze_mask_0 = const()[name = tensor("xi_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_cast_fp16 = slice_by_index(begin = xi_begin_0, end = xi_end_0, end_mask = xi_end_mask_0, squeeze_mask = xi_squeeze_mask_0, x = pairs_cast_fp16)[name = tensor("xi_cast_fp16")]; + tensor var_1380_cast_fp16 = mul(x = xr_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_1380_cast_fp16")]; + tensor var_1381_cast_fp16 = mul(x = xi_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_1381_cast_fp16")]; + tensor xor__cast_fp16 = sub(x = var_1380_cast_fp16, y = var_1381_cast_fp16)[name = tensor("xor__cast_fp16")]; + tensor var_1384_cast_fp16 = mul(x = xr_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_1384_cast_fp16")]; + tensor var_1385_cast_fp16 = mul(x = xi_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_1385_cast_fp16")]; + tensor xoi_cast_fp16 = add(x = var_1384_cast_fp16, y = var_1385_cast_fp16)[name = tensor("xoi_cast_fp16")]; + tensor out_axis_0 = const()[name = tensor("out_axis_0"), val = tensor(-1)]; + tensor out_cast_fp16 = stack(axis = out_axis_0, values = (xor__cast_fp16, xoi_cast_fp16))[name = tensor("out_cast_fp16")]; + tensor _inversed_1397_y_0_to_fp16 = const()[name = tensor("_inversed_1397_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1397_cast_fp16 = mul(x = position5_to_fp16, y = _inversed_1397_y_0_to_fp16)[name = tensor("_inversed_1397_cast_fp16")]; + tensor var_1398_cast_fp16 = floor(x = _inversed_1397_cast_fp16)[name = tensor("op_1398_cast_fp16")]; + tensor var_1399_to_fp16 = const()[name = tensor("op_1399_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1400_cast_fp16 = mul(x = var_1398_cast_fp16, y = var_1399_to_fp16)[name = tensor("op_1400_cast_fp16")]; + tensor write_idx_cast_fp16 = sub(x = position5_to_fp16, y = var_1400_cast_fp16)[name = tensor("write_idx_cast_fp16")]; + tensor var_1403_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_cast_fp16)[name = tensor("op_1403_cast_fp16")]; + tensor var_1413 = const()[name = tensor("op_1413"), val = tensor([1, 512, 1, 1])]; + tensor onehot_21_to_fp16_dtype_0 = const()[name = tensor("onehot_21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_1403_cast_fp16_to_fp16 = cast(dtype = onehot_21_to_fp16_dtype_0, x = var_1403_cast_fp16)[name = tensor("cast_104")]; + tensor onehot_cast_fp16 = reshape(shape = var_1413, x = var_1403_cast_fp16_to_fp16)[name = tensor("onehot_cast_fp16")]; + tensor var_1415_to_fp16 = const()[name = tensor("op_1415_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1417_cast_fp16 = sub(x = var_1415_to_fp16, y = onehot_cast_fp16)[name = tensor("op_1417_cast_fp16")]; + tensor k_cache5_to_fp16_dtype_0 = const()[name = tensor("k_cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache5_to_fp16 = cast(dtype = k_cache5_to_fp16_dtype_0, x = k_cache5)[name = tensor("cast_103")]; + tensor var_1418_cast_fp16 = mul(x = k_cache5_to_fp16, y = var_1417_cast_fp16)[name = tensor("op_1418_cast_fp16")]; + tensor var_1423 = const()[name = tensor("op_1423"), val = tensor([1, 1, 16, 64])]; + tensor var_1424_cast_fp16 = reshape(shape = var_1423, x = out_cast_fp16)[name = tensor("op_1424_cast_fp16")]; + tensor var_1425_cast_fp16 = mul(x = var_1424_cast_fp16, y = onehot_cast_fp16)[name = tensor("op_1425_cast_fp16")]; + tensor new_k_cache_cast_fp16 = add(x = var_1418_cast_fp16, y = var_1425_cast_fp16)[name = tensor("new_k_cache_cast_fp16")]; + tensor new_k_cache_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache5_to_fp16_dtype_0 = const()[name = tensor("v_cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache5_to_fp16 = cast(dtype = v_cache5_to_fp16_dtype_0, x = v_cache5)[name = tensor("cast_101")]; + tensor var_1431_cast_fp16 = mul(x = v_cache5_to_fp16, y = var_1417_cast_fp16)[name = tensor("op_1431_cast_fp16")]; + tensor var_1436 = const()[name = tensor("op_1436"), val = tensor([1, 1, 16, 64])]; + tensor var_1437_cast_fp16 = reshape(shape = var_1436, x = v_cast_fp16)[name = tensor("op_1437_cast_fp16")]; + tensor var_1438_cast_fp16 = mul(x = var_1437_cast_fp16, y = onehot_cast_fp16)[name = tensor("op_1438_cast_fp16")]; + tensor new_v_cache_cast_fp16 = add(x = var_1431_cast_fp16, y = var_1438_cast_fp16)[name = tensor("new_v_cache_cast_fp16")]; + tensor new_v_cache_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_5_perm_0 = const()[name = tensor("transpose_5_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_1451 = const()[name = tensor("op_1451"), val = tensor([0, 2, 1, 3])]; + tensor var_1457 = const()[name = tensor("op_1457"), val = tensor([1, 16, 1, 64])]; + tensor q4_cast_fp16 = reshape(shape = var_1457, x = out_21_cast_fp16)[name = tensor("q4_cast_fp16")]; + tensor var_1462_transpose_x_0 = const()[name = tensor("op_1462_transpose_x_0"), val = tensor(false)]; + tensor var_1462_transpose_y_0 = const()[name = tensor("op_1462_transpose_y_0"), val = tensor(false)]; + tensor transpose_5_cast_fp16 = transpose(perm = transpose_5_perm_0, x = new_k_cache_cast_fp16)[name = tensor("transpose_8")]; + tensor var_1462_cast_fp16 = matmul(transpose_x = var_1462_transpose_x_0, transpose_y = var_1462_transpose_y_0, x = q4_cast_fp16, y = transpose_5_cast_fp16)[name = tensor("op_1462_cast_fp16")]; + tensor var_1463_to_fp16 = const()[name = tensor("op_1463_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_21_cast_fp16 = mul(x = var_1462_cast_fp16, y = var_1463_to_fp16)[name = tensor("scores_21_cast_fp16")]; + tensor var_1467 = const()[name = tensor("op_1467"), val = tensor([1, 1])]; + tensor var_1468_cast_fp16 = reshape(shape = var_1467, x = position5_to_fp16)[name = tensor("op_1468_cast_fp16")]; + tensor var_1469_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_1468_cast_fp16)[name = tensor("op_1469_cast_fp16")]; + tensor mask_to_fp16_dtype_0 = const()[name = tensor("mask_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_1476_to_fp16 = const()[name = tensor("op_1476_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1469_cast_fp16_to_fp16 = cast(dtype = mask_to_fp16_dtype_0, x = var_1469_cast_fp16)[name = tensor("cast_99")]; + tensor var_1477_cast_fp16 = sub(x = var_1469_cast_fp16_to_fp16, y = var_1476_to_fp16)[name = tensor("op_1477_cast_fp16")]; + tensor var_1482 = const()[name = tensor("op_1482"), val = tensor([1, 1, 1, 512])]; + tensor var_1483_cast_fp16 = reshape(shape = var_1482, x = var_1477_cast_fp16)[name = tensor("op_1483_cast_fp16")]; + tensor var_1484_to_fp16 = const()[name = tensor("op_1484_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_1485_cast_fp16 = mul(x = var_1483_cast_fp16, y = var_1484_to_fp16)[name = tensor("op_1485_cast_fp16")]; + tensor scores_cast_fp16 = add(x = scores_21_cast_fp16, y = var_1485_cast_fp16)[name = tensor("scores_cast_fp16")]; + tensor var_1488 = const()[name = tensor("op_1488"), val = tensor(-1)]; + tensor weights_cast_fp16 = softmax(axis = var_1488, x = scores_cast_fp16)[name = tensor("weights_cast_fp16")]; + tensor attn_transpose_x_0 = const()[name = tensor("attn_transpose_x_0"), val = tensor(false)]; + tensor attn_transpose_y_0 = const()[name = tensor("attn_transpose_y_0"), val = tensor(false)]; + tensor values_cast_fp16 = transpose(perm = var_1451, x = new_v_cache_cast_fp16)[name = tensor("transpose_7")]; + tensor attn_cast_fp16 = matmul(transpose_x = attn_transpose_x_0, transpose_y = attn_transpose_y_0, x = weights_cast_fp16, y = values_cast_fp16)[name = tensor("attn_cast_fp16")]; + tensor var_1496 = const()[name = tensor("op_1496"), val = tensor([0, 2, 1, 3])]; + tensor var_1501 = const()[name = tensor("op_1501"), val = tensor([1, 1, 1024])]; + tensor var_1497_cast_fp16 = transpose(perm = var_1496, x = attn_cast_fp16)[name = tensor("transpose_6")]; + tensor input_67_cast_fp16 = reshape(shape = var_1501, x = var_1497_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor attn5_out_proj_weight_to_fp16 = const()[name = tensor("attn5_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132251840)))]; + tensor linear_22_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn5_out_proj_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("linear_22_cast_fp16")]; + tensor var_1507_to_fp16 = const()[name = tensor("op_1507_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1508_cast_fp16 = add(x = position5_to_fp16, y = var_1507_to_fp16)[name = tensor("op_1508_cast_fp16")]; + tensor var_1508_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1508_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_69_cast_fp16 = add(x = input_63_cast_fp16, y = linear_22_cast_fp16)[name = tensor("input_69_cast_fp16")]; + tensor input_71_axes_0 = const()[name = tensor("input_71_axes_0"), val = tensor([-1])]; + tensor norm5_2_weight_to_fp16 = const()[name = tensor("norm5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134349056)))]; + tensor norm5_2_bias_to_fp16 = const()[name = tensor("norm5_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134351168)))]; + tensor var_1512_to_fp16 = const()[name = tensor("op_1512_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_71_cast_fp16 = layer_norm(axes = input_71_axes_0, beta = norm5_2_bias_to_fp16, epsilon = var_1512_to_fp16, gamma = norm5_2_weight_to_fp16, x = input_69_cast_fp16)[name = tensor("input_71_cast_fp16")]; + tensor linear5_1_weight_to_fp16 = const()[name = tensor("linear5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134353280)))]; + tensor linear_23_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear5_1_weight_to_fp16, x = input_71_cast_fp16)[name = tensor("linear_23_cast_fp16")]; + tensor input_73_mode_0 = const()[name = tensor("input_73_mode_0"), val = tensor("EXACT")]; + tensor input_73_cast_fp16 = gelu(mode = input_73_mode_0, x = linear_23_cast_fp16)[name = tensor("input_73_cast_fp16")]; + tensor linear5_2_weight_to_fp16 = const()[name = tensor("linear5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142741952)))]; + tensor linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear5_2_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("linear_24_cast_fp16")]; + tensor input_75_cast_fp16 = add(x = input_69_cast_fp16, y = linear_24_cast_fp16)[name = tensor("input_75_cast_fp16")]; + tensor input_axes_0 = const()[name = tensor("input_axes_0"), val = tensor([-1])]; + tensor out_norm_weight_to_fp16 = const()[name = tensor("out_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151130624)))]; + tensor out_norm_bias_to_fp16 = const()[name = tensor("out_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151132736)))]; + tensor var_1529_to_fp16 = const()[name = tensor("op_1529_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_cast_fp16 = layer_norm(axes = input_axes_0, beta = out_norm_bias_to_fp16, epsilon = var_1529_to_fp16, gamma = out_norm_weight_to_fp16, x = input_75_cast_fp16)[name = tensor("input_cast_fp16")]; + tensor input_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("input_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor out_eos_weight_to_fp16 = const()[name = tensor("out_eos_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151134848)))]; + tensor out_eos_bias_to_fp16 = const()[name = tensor("out_eos_bias_to_fp16"), val = tensor([-0x1.5ap-3])]; + tensor linear_25_cast_fp16 = linear(bias = out_eos_bias_to_fp16, weight = out_eos_weight_to_fp16, x = input_cast_fp16)[name = tensor("linear_25_cast_fp16")]; + tensor linear_25_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("linear_25_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor is_eos = cast(dtype = linear_25_cast_fp16_to_fp32_dtype_0, x = linear_25_cast_fp16)[name = tensor("cast_96")]; + tensor transformer_out = cast(dtype = input_cast_fp16_to_fp32_dtype_0, x = input_cast_fp16)[name = tensor("cast_97")]; + tensor new_position5 = cast(dtype = var_1508_cast_fp16_to_fp32_dtype_0, x = var_1508_cast_fp16)[name = tensor("cast_98")]; + tensor new_v_cache5 = cast(dtype = new_v_cache_cast_fp16_to_fp32_dtype_0, x = new_v_cache_cast_fp16)[name = tensor("cast_100")]; + tensor new_k_cache5 = cast(dtype = new_k_cache_cast_fp16_to_fp32_dtype_0, x = new_k_cache_cast_fp16)[name = tensor("cast_102")]; + tensor new_position4 = cast(dtype = var_1264_cast_fp16_to_fp32_dtype_0, x = var_1264_cast_fp16)[name = tensor("cast_106")]; + tensor new_v_cache4 = cast(dtype = new_v_cache_9_cast_fp16_to_fp32_dtype_0, x = new_v_cache_9_cast_fp16)[name = tensor("cast_108")]; + tensor new_k_cache4 = cast(dtype = new_k_cache_9_cast_fp16_to_fp32_dtype_0, x = new_k_cache_9_cast_fp16)[name = tensor("cast_110")]; + tensor new_position3 = cast(dtype = var_1020_cast_fp16_to_fp32_dtype_0, x = var_1020_cast_fp16)[name = tensor("cast_114")]; + tensor new_v_cache3 = cast(dtype = new_v_cache_7_cast_fp16_to_fp32_dtype_0, x = new_v_cache_7_cast_fp16)[name = tensor("cast_116")]; + tensor new_k_cache3 = cast(dtype = new_k_cache_7_cast_fp16_to_fp32_dtype_0, x = new_k_cache_7_cast_fp16)[name = tensor("cast_118")]; + tensor new_position2 = cast(dtype = var_776_cast_fp16_to_fp32_dtype_0, x = var_776_cast_fp16)[name = tensor("cast_122")]; + tensor new_v_cache2 = cast(dtype = new_v_cache_5_cast_fp16_to_fp32_dtype_0, x = new_v_cache_5_cast_fp16)[name = tensor("cast_124")]; + tensor new_k_cache2 = cast(dtype = new_k_cache_5_cast_fp16_to_fp32_dtype_0, x = new_k_cache_5_cast_fp16)[name = tensor("cast_126")]; + tensor new_position1 = cast(dtype = var_532_cast_fp16_to_fp32_dtype_0, x = var_532_cast_fp16)[name = tensor("cast_130")]; + tensor new_v_cache1 = cast(dtype = new_v_cache_3_cast_fp16_to_fp32_dtype_0, x = new_v_cache_3_cast_fp16)[name = tensor("cast_132")]; + tensor new_k_cache1 = cast(dtype = new_k_cache_3_cast_fp16_to_fp32_dtype_0, x = new_k_cache_3_cast_fp16)[name = tensor("cast_134")]; + tensor new_position0 = cast(dtype = var_288_cast_fp16_to_fp32_dtype_0, x = var_288_cast_fp16)[name = tensor("cast_138")]; + tensor new_v_cache0 = cast(dtype = new_v_cache_1_cast_fp16_to_fp32_dtype_0, x = new_v_cache_1_cast_fp16)[name = tensor("cast_140")]; + tensor new_k_cache0 = cast(dtype = new_k_cache_1_cast_fp16_to_fp32_dtype_0, x = new_k_cache_1_cast_fp16)[name = tensor("cast_142")]; + } -> (transformer_out, is_eos, new_k_cache0, new_v_cache0, new_position0, new_k_cache1, new_v_cache1, new_position1, new_k_cache2, new_v_cache2, new_position2, new_k_cache3, new_v_cache3, new_position3, new_k_cache4, new_v_cache4, new_position4, new_k_cache5, new_v_cache5, new_position5); +} \ No newline at end of file diff --git a/v2.1/portuguese/flowlm_step_ane.mlmodelc/weights/weight.bin b/v2.1/portuguese/flowlm_step_ane.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..c452438198e804015c5819cdd5d4af1eb8ced383 --- /dev/null +++ b/v2.1/portuguese/flowlm_step_ane.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f6b5338be4a5ec5afa9b8ba61a23690eb7545a4b5b89d3f0c913f9f1dd1deb41 +size 151136960 diff --git a/v2.1/portuguese/flowlm_step_ane.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/portuguese/flowlm_step_ane.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..d970ee3a6dee9947a157bea74f5c1dd5f84835a3 --- /dev/null +++ b/v2.1/portuguese/flowlm_step_ane.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:47861ae271841bec1b9c561118a1e817d8ff0ae7779a825a27f136f4c5c22c4b +size 137731 diff --git a/v2.1/portuguese/flowlm_step_ane.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/portuguese/flowlm_step_ane.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..c452438198e804015c5819cdd5d4af1eb8ced383 --- /dev/null +++ b/v2.1/portuguese/flowlm_step_ane.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f6b5338be4a5ec5afa9b8ba61a23690eb7545a4b5b89d3f0c913f9f1dd1deb41 +size 151136960 diff --git a/v2.1/portuguese/flowlm_step_ane.mlpackage/Manifest.json b/v2.1/portuguese/flowlm_step_ane.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..19b72b8d526aecd2b8145310448087a301d13151 --- /dev/null +++ b/v2.1/portuguese/flowlm_step_ane.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "41A2837B-5334-4CD1-8D4B-A7E94DE0E450": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "F90B0A09-BA2B-4EF9-9ABB-709B1736F1CE": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "41A2837B-5334-4CD1-8D4B-A7E94DE0E450" +} diff --git a/v2.1/portuguese_24l/cond_prefill_ane.mlmodelc/analytics/coremldata.bin b/v2.1/portuguese_24l/cond_prefill_ane.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..b0abb91a7ae8ade90270ad13a18f3994fa23eb9e --- /dev/null +++ b/v2.1/portuguese_24l/cond_prefill_ane.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d96fc4b215e0d736b42451b160f81d200275871afb1164d74b84323e891d2ab1 +size 243 diff --git a/v2.1/portuguese_24l/cond_prefill_ane.mlmodelc/coremldata.bin b/v2.1/portuguese_24l/cond_prefill_ane.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..41a21d4a0797dfa60e9f931d3c2d8b8cc7197150 --- /dev/null +++ b/v2.1/portuguese_24l/cond_prefill_ane.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9fc9861887894f291cee0056ad3b7911ce9a5fb9fef2f201820d698dbb573669 +size 4496 diff --git a/v2.1/portuguese_24l/cond_prefill_ane.mlmodelc/model.mil b/v2.1/portuguese_24l/cond_prefill_ane.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..f102b6e147b28d002e028bb659c273f9e9c7f4c9 --- /dev/null +++ b/v2.1/portuguese_24l/cond_prefill_ane.mlmodelc/model.mil @@ -0,0 +1,3198 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.9.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor conditioning, tensor k_cache0, tensor k_cache1, tensor k_cache10, tensor k_cache11, tensor k_cache12, tensor k_cache13, tensor k_cache14, tensor k_cache15, tensor k_cache16, tensor k_cache17, tensor k_cache18, tensor k_cache19, tensor k_cache2, tensor k_cache20, tensor k_cache21, tensor k_cache22, tensor k_cache23, tensor k_cache3, tensor k_cache4, tensor k_cache5, tensor k_cache6, tensor k_cache7, tensor k_cache8, tensor k_cache9, tensor position0, tensor position1, tensor position10, tensor position11, tensor position12, tensor position13, tensor position14, tensor position15, tensor position16, tensor position17, tensor position18, tensor position19, tensor position2, tensor position20, tensor position21, tensor position22, tensor position23, tensor position3, tensor position4, tensor position5, tensor position6, tensor position7, tensor position8, tensor position9, tensor v_cache0, tensor v_cache1, tensor v_cache10, tensor v_cache11, tensor v_cache12, tensor v_cache13, tensor v_cache14, tensor v_cache15, tensor v_cache16, tensor v_cache17, tensor v_cache18, tensor v_cache19, tensor v_cache2, tensor v_cache20, tensor v_cache21, tensor v_cache22, tensor v_cache23, tensor v_cache3, tensor v_cache4, tensor v_cache5, tensor v_cache6, tensor v_cache7, tensor v_cache8, tensor v_cache9, tensor valid_len) { + tensor token_indices_to_fp16 = const()[name = tensor("token_indices_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor position0_to_fp16_dtype_0 = const()[name = tensor("position0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position0_to_fp16 = cast(dtype = position0_to_fp16_dtype_0, x = position0)[name = tensor("cast_470")]; + tensor ts_cast_fp16 = add(x = token_indices_to_fp16, y = position0_to_fp16)[name = tensor("ts_cast_fp16")]; + tensor var_231 = const()[name = tensor("op_231"), val = tensor([256, 1, 1])]; + tensor var_232_cast_fp16 = reshape(shape = var_231, x = ts_cast_fp16)[name = tensor("op_232_cast_fp16")]; + tensor var_227_to_fp16 = const()[name = tensor("op_227_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(640)))]; + tensor angles_cast_fp16 = mul(x = var_227_to_fp16, y = var_232_cast_fp16)[name = tensor("angles_cast_fp16")]; + tensor rotr_cast_fp16 = cos(x = angles_cast_fp16)[name = tensor("rotr_cast_fp16")]; + tensor roti_cast_fp16 = sin(x = angles_cast_fp16)[name = tensor("roti_cast_fp16")]; + tensor _inversed_237_y_0_to_fp16 = const()[name = tensor("_inversed_237_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_237_cast_fp16 = mul(x = ts_cast_fp16, y = _inversed_237_y_0_to_fp16)[name = tensor("_inversed_237_cast_fp16")]; + tensor var_238_cast_fp16 = floor(x = _inversed_237_cast_fp16)[name = tensor("op_238_cast_fp16")]; + tensor var_239_to_fp16 = const()[name = tensor("op_239_to_fp16"), val = tensor(0x1p+9)]; + tensor var_240_cast_fp16 = mul(x = var_238_cast_fp16, y = var_239_to_fp16)[name = tensor("op_240_cast_fp16")]; + tensor write_idx_1_cast_fp16 = sub(x = ts_cast_fp16, y = var_240_cast_fp16)[name = tensor("write_idx_1_cast_fp16")]; + tensor valid_len_to_fp16_dtype_0 = const()[name = tensor("valid_len_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor valid_len_to_fp16 = cast(dtype = valid_len_to_fp16_dtype_0, x = valid_len)[name = tensor("cast_469")]; + tensor var_246_cast_fp16 = less(x = token_indices_to_fp16, y = valid_len_to_fp16)[name = tensor("op_246_cast_fp16")]; + tensor token_valid_to_fp16_dtype_0 = const()[name = tensor("token_valid_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_246_cast_fp16_to_fp16 = cast(dtype = token_valid_to_fp16_dtype_0, x = var_246_cast_fp16)[name = tensor("cast_468")]; + tensor var_252_cast_fp16 = mul(x = write_idx_1_cast_fp16, y = var_246_cast_fp16_to_fp16)[name = tensor("op_252_cast_fp16")]; + tensor var_253_to_fp16 = const()[name = tensor("op_253_to_fp16"), val = tensor(0x1p+0)]; + tensor var_255_cast_fp16 = sub(x = var_253_to_fp16, y = var_246_cast_fp16_to_fp16)[name = tensor("op_255_cast_fp16")]; + tensor var_256_to_fp16 = const()[name = tensor("op_256_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_257_cast_fp16 = mul(x = var_255_cast_fp16, y = var_256_to_fp16)[name = tensor("op_257_cast_fp16")]; + tensor write_idx_cast_fp16 = add(x = var_252_cast_fp16, y = var_257_cast_fp16)[name = tensor("write_idx_cast_fp16")]; + tensor var_262 = const()[name = tensor("op_262"), val = tensor([256, 1])]; + tensor var_263_cast_fp16 = reshape(shape = var_262, x = write_idx_cast_fp16)[name = tensor("op_263_cast_fp16")]; + tensor var_267_to_fp16 = const()[name = tensor("op_267_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(768)))]; + tensor var_268_cast_fp16 = equal(x = var_263_cast_fp16, y = var_267_to_fp16)[name = tensor("op_268_cast_fp16")]; + tensor var_278_axes_0 = const()[name = tensor("op_278_axes_0"), val = tensor([0])]; + tensor var_278_keep_dims_0 = const()[name = tensor("op_278_keep_dims_0"), val = tensor(false)]; + tensor assign_to_fp16_dtype_0 = const()[name = tensor("assign_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_268_cast_fp16_to_fp16 = cast(dtype = assign_to_fp16_dtype_0, x = var_268_cast_fp16)[name = tensor("cast_467")]; + tensor var_278_cast_fp16 = reduce_sum(axes = var_278_axes_0, keep_dims = var_278_keep_dims_0, x = var_268_cast_fp16_to_fp16)[name = tensor("op_278_cast_fp16")]; + tensor const_0_to_fp16 = const()[name = tensor("const_0_to_fp16"), val = tensor(-inf)]; + tensor var_280_to_fp16 = const()[name = tensor("op_280_to_fp16"), val = tensor(0x1p+0)]; + tensor clip_0_cast_fp16 = clip(alpha = const_0_to_fp16, beta = var_280_to_fp16, x = var_278_cast_fp16)[name = tensor("clip_0_cast_fp16")]; + tensor var_284 = const()[name = tensor("op_284"), val = tensor([1, 1])]; + tensor var_285_cast_fp16 = reshape(shape = var_284, x = position0_to_fp16)[name = tensor("op_285_cast_fp16")]; + tensor var_289_to_fp16 = const()[name = tensor("op_289_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1856)))]; + tensor q_positions_cast_fp16 = add(x = var_285_cast_fp16, y = var_289_to_fp16)[name = tensor("q_positions_cast_fp16")]; + tensor var_297_cast_fp16 = add(x = position0_to_fp16, y = valid_len_to_fp16)[name = tensor("op_297_cast_fp16")]; + tensor var_300 = const()[name = tensor("op_300"), val = tensor([1, 1])]; + tensor valid_end_cast_fp16 = reshape(shape = var_300, x = var_297_cast_fp16)[name = tensor("valid_end_cast_fp16")]; + tensor var_302_cast_fp16 = less(x = var_267_to_fp16, y = valid_end_cast_fp16)[name = tensor("op_302_cast_fp16")]; + tensor var_303_cast_fp16 = less_equal(x = var_267_to_fp16, y = q_positions_cast_fp16)[name = tensor("op_303_cast_fp16")]; + tensor var_304 = logical_and(x = var_302_cast_fp16, y = var_303_cast_fp16)[name = tensor("op_304")]; + tensor mask_to_fp16_dtype_0 = const()[name = tensor("mask_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_311_to_fp16 = const()[name = tensor("op_311_to_fp16"), val = tensor(0x1p+0)]; + tensor var_304_to_fp16 = cast(dtype = mask_to_fp16_dtype_0, x = var_304)[name = tensor("cast_466")]; + tensor var_312_cast_fp16 = sub(x = var_304_to_fp16, y = var_311_to_fp16)[name = tensor("op_312_cast_fp16")]; + tensor var_313_to_fp16 = const()[name = tensor("op_313_to_fp16"), val = tensor(0x1.388p+13)]; + tensor mask_bias_cast_fp16 = mul(x = var_312_cast_fp16, y = var_313_to_fp16)[name = tensor("mask_bias_cast_fp16")]; + tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; + tensor conditioning_to_fp16_dtype_0 = const()[name = tensor("conditioning_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor norm0_1_weight_to_fp16 = const()[name = tensor("norm0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2432)))]; + tensor norm0_1_bias_to_fp16 = const()[name = tensor("norm0_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4544)))]; + tensor var_316_to_fp16 = const()[name = tensor("op_316_to_fp16"), val = tensor(0x1.5p-17)]; + tensor conditioning_to_fp16 = cast(dtype = conditioning_to_fp16_dtype_0, x = conditioning)[name = tensor("cast_465")]; + tensor input_3_cast_fp16 = layer_norm(axes = input_3_axes_0, beta = norm0_1_bias_to_fp16, epsilon = var_316_to_fp16, gamma = norm0_1_weight_to_fp16, x = conditioning_to_fp16)[name = tensor("input_3_cast_fp16")]; + tensor attn0_in_proj_weight_to_fp16 = const()[name = tensor("attn0_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6656)))]; + tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6298176)))]; + tensor linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn0_in_proj_weight_to_fp16, x = input_3_cast_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor var_330 = const()[name = tensor("op_330"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_1_cast_fp16 = reshape(shape = var_330, x = linear_0_cast_fp16)[name = tensor("qkv_1_cast_fp16")]; + tensor x_1_begin_0 = const()[name = tensor("x_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_1_end_0 = const()[name = tensor("x_1_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_1_end_mask_0 = const()[name = tensor("x_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_1_squeeze_mask_0 = const()[name = tensor("x_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_1_cast_fp16 = slice_by_index(begin = x_1_begin_0, end = x_1_end_0, end_mask = x_1_end_mask_0, squeeze_mask = x_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("x_1_cast_fp16")]; + tensor x_3_begin_0 = const()[name = tensor("x_3_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_3_end_0 = const()[name = tensor("x_3_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_3_end_mask_0 = const()[name = tensor("x_3_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_3_squeeze_mask_0 = const()[name = tensor("x_3_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_3_cast_fp16 = slice_by_index(begin = x_3_begin_0, end = x_3_end_0, end_mask = x_3_end_mask_0, squeeze_mask = x_3_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("x_3_cast_fp16")]; + tensor v_1_begin_0 = const()[name = tensor("v_1_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_1_end_0 = const()[name = tensor("v_1_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_1_end_mask_0 = const()[name = tensor("v_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_1_squeeze_mask_0 = const()[name = tensor("v_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_1_cast_fp16 = slice_by_index(begin = v_1_begin_0, end = v_1_end_0, end_mask = v_1_end_mask_0, squeeze_mask = v_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("v_1_cast_fp16")]; + tensor var_375 = const()[name = tensor("op_375"), val = tensor([256, 16, 32, 2])]; + tensor pairs_1_cast_fp16 = reshape(shape = var_375, x = x_1_cast_fp16)[name = tensor("pairs_1_cast_fp16")]; + tensor xr_1_begin_0 = const()[name = tensor("xr_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_1_end_0 = const()[name = tensor("xr_1_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_1_end_mask_0 = const()[name = tensor("xr_1_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_1_squeeze_mask_0 = const()[name = tensor("xr_1_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_1_cast_fp16 = slice_by_index(begin = xr_1_begin_0, end = xr_1_end_0, end_mask = xr_1_end_mask_0, squeeze_mask = xr_1_squeeze_mask_0, x = pairs_1_cast_fp16)[name = tensor("xr_1_cast_fp16")]; + tensor xi_1_begin_0 = const()[name = tensor("xi_1_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_1_end_0 = const()[name = tensor("xi_1_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_1_end_mask_0 = const()[name = tensor("xi_1_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_1_squeeze_mask_0 = const()[name = tensor("xi_1_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_1_cast_fp16 = slice_by_index(begin = xi_1_begin_0, end = xi_1_end_0, end_mask = xi_1_end_mask_0, squeeze_mask = xi_1_squeeze_mask_0, x = pairs_1_cast_fp16)[name = tensor("xi_1_cast_fp16")]; + tensor var_383_cast_fp16 = mul(x = xr_1_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_383_cast_fp16")]; + tensor var_384_cast_fp16 = mul(x = xi_1_cast_fp16, y = roti_cast_fp16)[name = tensor("op_384_cast_fp16")]; + tensor xor__1_cast_fp16 = sub(x = var_383_cast_fp16, y = var_384_cast_fp16)[name = tensor("xor__1_cast_fp16")]; + tensor var_387_cast_fp16 = mul(x = xr_1_cast_fp16, y = roti_cast_fp16)[name = tensor("op_387_cast_fp16")]; + tensor var_388_cast_fp16 = mul(x = xi_1_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_388_cast_fp16")]; + tensor xoi_1_cast_fp16 = add(x = var_387_cast_fp16, y = var_388_cast_fp16)[name = tensor("xoi_1_cast_fp16")]; + tensor out_1_axis_0 = const()[name = tensor("out_1_axis_0"), val = tensor(-1)]; + tensor out_1_cast_fp16 = stack(axis = out_1_axis_0, values = (xor__1_cast_fp16, xoi_1_cast_fp16))[name = tensor("out_1_cast_fp16")]; + tensor var_398 = const()[name = tensor("op_398"), val = tensor([1, 256, 16, 64])]; + tensor q_1_cast_fp16 = reshape(shape = var_398, x = out_1_cast_fp16)[name = tensor("q_1_cast_fp16")]; + tensor var_404 = const()[name = tensor("op_404"), val = tensor([256, 16, 32, 2])]; + tensor pairs_3_cast_fp16 = reshape(shape = var_404, x = x_3_cast_fp16)[name = tensor("pairs_3_cast_fp16")]; + tensor xr_3_begin_0 = const()[name = tensor("xr_3_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_3_end_0 = const()[name = tensor("xr_3_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_3_end_mask_0 = const()[name = tensor("xr_3_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_3_squeeze_mask_0 = const()[name = tensor("xr_3_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_3_cast_fp16 = slice_by_index(begin = xr_3_begin_0, end = xr_3_end_0, end_mask = xr_3_end_mask_0, squeeze_mask = xr_3_squeeze_mask_0, x = pairs_3_cast_fp16)[name = tensor("xr_3_cast_fp16")]; + tensor xi_3_begin_0 = const()[name = tensor("xi_3_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_3_end_0 = const()[name = tensor("xi_3_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_3_end_mask_0 = const()[name = tensor("xi_3_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_3_squeeze_mask_0 = const()[name = tensor("xi_3_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_3_cast_fp16 = slice_by_index(begin = xi_3_begin_0, end = xi_3_end_0, end_mask = xi_3_end_mask_0, squeeze_mask = xi_3_squeeze_mask_0, x = pairs_3_cast_fp16)[name = tensor("xi_3_cast_fp16")]; + tensor var_412_cast_fp16 = mul(x = xr_3_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_412_cast_fp16")]; + tensor var_413_cast_fp16 = mul(x = xi_3_cast_fp16, y = roti_cast_fp16)[name = tensor("op_413_cast_fp16")]; + tensor xor__3_cast_fp16 = sub(x = var_412_cast_fp16, y = var_413_cast_fp16)[name = tensor("xor__3_cast_fp16")]; + tensor var_416_cast_fp16 = mul(x = xr_3_cast_fp16, y = roti_cast_fp16)[name = tensor("op_416_cast_fp16")]; + tensor var_417_cast_fp16 = mul(x = xi_3_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_417_cast_fp16")]; + tensor xoi_3_cast_fp16 = add(x = var_416_cast_fp16, y = var_417_cast_fp16)[name = tensor("xoi_3_cast_fp16")]; + tensor out_3_axis_0 = const()[name = tensor("out_3_axis_0"), val = tensor(-1)]; + tensor out_3_cast_fp16 = stack(axis = out_3_axis_0, values = (xor__3_cast_fp16, xoi_3_cast_fp16))[name = tensor("out_3_cast_fp16")]; + tensor var_434 = const()[name = tensor("op_434"), val = tensor([256, 1024])]; + tensor var_435_cast_fp16 = reshape(shape = var_434, x = out_3_cast_fp16)[name = tensor("op_435_cast_fp16")]; + tensor written_k_1_transpose_x_1 = const()[name = tensor("written_k_1_transpose_x_1"), val = tensor(true)]; + tensor written_k_1_transpose_y_1 = const()[name = tensor("written_k_1_transpose_y_1"), val = tensor(false)]; + tensor written_k_1_cast_fp16 = matmul(transpose_x = written_k_1_transpose_x_1, transpose_y = written_k_1_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_435_cast_fp16)[name = tensor("written_k_1_cast_fp16")]; + tensor var_442 = const()[name = tensor("op_442"), val = tensor([256, 1024])]; + tensor var_443_cast_fp16 = reshape(shape = var_442, x = v_1_cast_fp16)[name = tensor("op_443_cast_fp16")]; + tensor written_v_1_transpose_x_1 = const()[name = tensor("written_v_1_transpose_x_1"), val = tensor(true)]; + tensor written_v_1_transpose_y_1 = const()[name = tensor("written_v_1_transpose_y_1"), val = tensor(false)]; + tensor written_v_1_cast_fp16 = matmul(transpose_x = written_v_1_transpose_x_1, transpose_y = written_v_1_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_443_cast_fp16)[name = tensor("written_v_1_cast_fp16")]; + tensor var_445_to_fp16 = const()[name = tensor("op_445_to_fp16"), val = tensor(0x1p+0)]; + tensor var_447_cast_fp16 = sub(x = var_445_to_fp16, y = clip_0_cast_fp16)[name = tensor("op_447_cast_fp16")]; + tensor var_452 = const()[name = tensor("op_452"), val = tensor([1, 512, 1, 1])]; + tensor keep_1_cast_fp16 = reshape(shape = var_452, x = var_447_cast_fp16)[name = tensor("keep_1_cast_fp16")]; + tensor k_cache0_to_fp16_dtype_0 = const()[name = tensor("k_cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache0_to_fp16 = cast(dtype = k_cache0_to_fp16_dtype_0, x = k_cache0)[name = tensor("cast_464")]; + tensor var_454_cast_fp16 = mul(x = k_cache0_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_454_cast_fp16")]; + tensor var_459 = const()[name = tensor("op_459"), val = tensor([1, 512, 16, 64])]; + tensor var_460_cast_fp16 = reshape(shape = var_459, x = written_k_1_cast_fp16)[name = tensor("op_460_cast_fp16")]; + tensor new_k_cache_1_cast_fp16 = add(x = var_454_cast_fp16, y = var_460_cast_fp16)[name = tensor("new_k_cache_1_cast_fp16")]; + tensor new_k_cache_1_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_1_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache0_to_fp16_dtype_0 = const()[name = tensor("v_cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache0_to_fp16 = cast(dtype = v_cache0_to_fp16_dtype_0, x = v_cache0)[name = tensor("cast_462")]; + tensor var_463_cast_fp16 = mul(x = v_cache0_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_463_cast_fp16")]; + tensor var_468 = const()[name = tensor("op_468"), val = tensor([1, 512, 16, 64])]; + tensor var_469_cast_fp16 = reshape(shape = var_468, x = written_v_1_cast_fp16)[name = tensor("op_469_cast_fp16")]; + tensor new_v_cache_1_cast_fp16 = add(x = var_463_cast_fp16, y = var_469_cast_fp16)[name = tensor("new_v_cache_1_cast_fp16")]; + tensor new_v_cache_1_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_1_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_482 = const()[name = tensor("op_482"), val = tensor([0, 2, 1, 3])]; + tensor var_493_transpose_x_0 = const()[name = tensor("op_493_transpose_x_0"), val = tensor(false)]; + tensor var_493_transpose_y_0 = const()[name = tensor("op_493_transpose_y_0"), val = tensor(false)]; + tensor transpose_69_perm_0 = const()[name = tensor("transpose_69_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_70_perm_0 = const()[name = tensor("transpose_70_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_70 = transpose(perm = transpose_70_perm_0, x = new_k_cache_1_cast_fp16)[name = tensor("transpose_204")]; + tensor transpose_69 = transpose(perm = transpose_69_perm_0, x = q_1_cast_fp16)[name = tensor("transpose_205")]; + tensor var_493_cast_fp16 = matmul(transpose_x = var_493_transpose_x_0, transpose_y = var_493_transpose_y_0, x = transpose_69, y = transpose_70)[name = tensor("op_493_cast_fp16")]; + tensor var_494_to_fp16 = const()[name = tensor("op_494_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_1_cast_fp16 = mul(x = var_493_cast_fp16, y = var_494_to_fp16)[name = tensor("scores_1_cast_fp16")]; + tensor var_500 = const()[name = tensor("op_500"), val = tensor([1, 1, 256, 512])]; + tensor var_501_cast_fp16 = reshape(shape = var_500, x = mask_bias_cast_fp16)[name = tensor("op_501_cast_fp16")]; + tensor scores_3_cast_fp16 = add(x = scores_1_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_3_cast_fp16")]; + tensor var_504 = const()[name = tensor("op_504"), val = tensor(-1)]; + tensor weights_1_cast_fp16 = softmax(axis = var_504, x = scores_3_cast_fp16)[name = tensor("weights_1_cast_fp16")]; + tensor attn_1_transpose_x_0 = const()[name = tensor("attn_1_transpose_x_0"), val = tensor(false)]; + tensor attn_1_transpose_y_0 = const()[name = tensor("attn_1_transpose_y_0"), val = tensor(false)]; + tensor values_1_cast_fp16 = transpose(perm = var_482, x = new_v_cache_1_cast_fp16)[name = tensor("transpose_206")]; + tensor attn_1_cast_fp16 = matmul(transpose_x = attn_1_transpose_x_0, transpose_y = attn_1_transpose_y_0, x = weights_1_cast_fp16, y = values_1_cast_fp16)[name = tensor("attn_1_cast_fp16")]; + tensor var_512 = const()[name = tensor("op_512"), val = tensor([0, 2, 1, 3])]; + tensor var_517 = const()[name = tensor("op_517"), val = tensor([1, 256, 1024])]; + tensor var_513_cast_fp16 = transpose(perm = var_512, x = attn_1_cast_fp16)[name = tensor("transpose_203")]; + tensor input_5_cast_fp16 = reshape(shape = var_517, x = var_513_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor attn0_out_proj_weight_to_fp16 = const()[name = tensor("attn0_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6304384)))]; + tensor linear_1_bias_0_to_fp16 = const()[name = tensor("linear_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401600)))]; + tensor linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn0_out_proj_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("linear_1_cast_fp16")]; + tensor var_526_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_526_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_7_cast_fp16 = add(x = conditioning_to_fp16, y = linear_1_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor input_9_axes_0 = const()[name = tensor("input_9_axes_0"), val = tensor([-1])]; + tensor norm0_2_weight_to_fp16 = const()[name = tensor("norm0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8403712)))]; + tensor norm0_2_bias_to_fp16 = const()[name = tensor("norm0_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8405824)))]; + tensor var_530_to_fp16 = const()[name = tensor("op_530_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_9_cast_fp16 = layer_norm(axes = input_9_axes_0, beta = norm0_2_bias_to_fp16, epsilon = var_530_to_fp16, gamma = norm0_2_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor linear0_1_weight_to_fp16 = const()[name = tensor("linear0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8407936)))]; + tensor linear_2_bias_0_to_fp16 = const()[name = tensor("linear_2_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16796608)))]; + tensor linear_2_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear0_1_weight_to_fp16, x = input_9_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor input_11_mode_0 = const()[name = tensor("input_11_mode_0"), val = tensor("EXACT")]; + tensor input_11_cast_fp16 = gelu(mode = input_11_mode_0, x = linear_2_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor linear0_2_weight_to_fp16 = const()[name = tensor("linear0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16804864)))]; + tensor linear_3_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear0_2_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("linear_3_cast_fp16")]; + tensor input_13_cast_fp16 = add(x = input_7_cast_fp16, y = linear_3_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor input_15_axes_0 = const()[name = tensor("input_15_axes_0"), val = tensor([-1])]; + tensor norm1_1_weight_to_fp16 = const()[name = tensor("norm1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25193536)))]; + tensor norm1_1_bias_to_fp16 = const()[name = tensor("norm1_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25195648)))]; + tensor var_547_to_fp16 = const()[name = tensor("op_547_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_15_cast_fp16 = layer_norm(axes = input_15_axes_0, beta = norm1_1_bias_to_fp16, epsilon = var_547_to_fp16, gamma = norm1_1_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor attn1_in_proj_weight_to_fp16 = const()[name = tensor("attn1_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25197760)))]; + tensor linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn1_in_proj_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor var_561 = const()[name = tensor("op_561"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_3_cast_fp16 = reshape(shape = var_561, x = linear_4_cast_fp16)[name = tensor("qkv_3_cast_fp16")]; + tensor x_5_begin_0 = const()[name = tensor("x_5_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_5_end_0 = const()[name = tensor("x_5_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_5_end_mask_0 = const()[name = tensor("x_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_5_squeeze_mask_0 = const()[name = tensor("x_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_5_cast_fp16 = slice_by_index(begin = x_5_begin_0, end = x_5_end_0, end_mask = x_5_end_mask_0, squeeze_mask = x_5_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("x_5_cast_fp16")]; + tensor x_7_begin_0 = const()[name = tensor("x_7_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_7_end_0 = const()[name = tensor("x_7_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_7_end_mask_0 = const()[name = tensor("x_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_7_squeeze_mask_0 = const()[name = tensor("x_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_7_cast_fp16 = slice_by_index(begin = x_7_begin_0, end = x_7_end_0, end_mask = x_7_end_mask_0, squeeze_mask = x_7_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("x_7_cast_fp16")]; + tensor v_3_begin_0 = const()[name = tensor("v_3_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_3_end_0 = const()[name = tensor("v_3_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_3_end_mask_0 = const()[name = tensor("v_3_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_3_squeeze_mask_0 = const()[name = tensor("v_3_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_3_cast_fp16 = slice_by_index(begin = v_3_begin_0, end = v_3_end_0, end_mask = v_3_end_mask_0, squeeze_mask = v_3_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("v_3_cast_fp16")]; + tensor var_606 = const()[name = tensor("op_606"), val = tensor([256, 16, 32, 2])]; + tensor pairs_5_cast_fp16 = reshape(shape = var_606, x = x_5_cast_fp16)[name = tensor("pairs_5_cast_fp16")]; + tensor xr_5_begin_0 = const()[name = tensor("xr_5_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_5_end_0 = const()[name = tensor("xr_5_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_5_end_mask_0 = const()[name = tensor("xr_5_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_5_squeeze_mask_0 = const()[name = tensor("xr_5_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_5_cast_fp16 = slice_by_index(begin = xr_5_begin_0, end = xr_5_end_0, end_mask = xr_5_end_mask_0, squeeze_mask = xr_5_squeeze_mask_0, x = pairs_5_cast_fp16)[name = tensor("xr_5_cast_fp16")]; + tensor xi_5_begin_0 = const()[name = tensor("xi_5_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_5_end_0 = const()[name = tensor("xi_5_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_5_end_mask_0 = const()[name = tensor("xi_5_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_5_squeeze_mask_0 = const()[name = tensor("xi_5_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_5_cast_fp16 = slice_by_index(begin = xi_5_begin_0, end = xi_5_end_0, end_mask = xi_5_end_mask_0, squeeze_mask = xi_5_squeeze_mask_0, x = pairs_5_cast_fp16)[name = tensor("xi_5_cast_fp16")]; + tensor var_614_cast_fp16 = mul(x = xr_5_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_614_cast_fp16")]; + tensor var_615_cast_fp16 = mul(x = xi_5_cast_fp16, y = roti_cast_fp16)[name = tensor("op_615_cast_fp16")]; + tensor xor__5_cast_fp16 = sub(x = var_614_cast_fp16, y = var_615_cast_fp16)[name = tensor("xor__5_cast_fp16")]; + tensor var_618_cast_fp16 = mul(x = xr_5_cast_fp16, y = roti_cast_fp16)[name = tensor("op_618_cast_fp16")]; + tensor var_619_cast_fp16 = mul(x = xi_5_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_619_cast_fp16")]; + tensor xoi_5_cast_fp16 = add(x = var_618_cast_fp16, y = var_619_cast_fp16)[name = tensor("xoi_5_cast_fp16")]; + tensor out_5_axis_0 = const()[name = tensor("out_5_axis_0"), val = tensor(-1)]; + tensor out_5_cast_fp16 = stack(axis = out_5_axis_0, values = (xor__5_cast_fp16, xoi_5_cast_fp16))[name = tensor("out_5_cast_fp16")]; + tensor var_629 = const()[name = tensor("op_629"), val = tensor([1, 256, 16, 64])]; + tensor q_3_cast_fp16 = reshape(shape = var_629, x = out_5_cast_fp16)[name = tensor("q_3_cast_fp16")]; + tensor var_635 = const()[name = tensor("op_635"), val = tensor([256, 16, 32, 2])]; + tensor pairs_7_cast_fp16 = reshape(shape = var_635, x = x_7_cast_fp16)[name = tensor("pairs_7_cast_fp16")]; + tensor xr_7_begin_0 = const()[name = tensor("xr_7_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_7_end_0 = const()[name = tensor("xr_7_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_7_end_mask_0 = const()[name = tensor("xr_7_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_7_squeeze_mask_0 = const()[name = tensor("xr_7_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_7_cast_fp16 = slice_by_index(begin = xr_7_begin_0, end = xr_7_end_0, end_mask = xr_7_end_mask_0, squeeze_mask = xr_7_squeeze_mask_0, x = pairs_7_cast_fp16)[name = tensor("xr_7_cast_fp16")]; + tensor xi_7_begin_0 = const()[name = tensor("xi_7_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_7_end_0 = const()[name = tensor("xi_7_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_7_end_mask_0 = const()[name = tensor("xi_7_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_7_squeeze_mask_0 = const()[name = tensor("xi_7_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_7_cast_fp16 = slice_by_index(begin = xi_7_begin_0, end = xi_7_end_0, end_mask = xi_7_end_mask_0, squeeze_mask = xi_7_squeeze_mask_0, x = pairs_7_cast_fp16)[name = tensor("xi_7_cast_fp16")]; + tensor var_643_cast_fp16 = mul(x = xr_7_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_643_cast_fp16")]; + tensor var_644_cast_fp16 = mul(x = xi_7_cast_fp16, y = roti_cast_fp16)[name = tensor("op_644_cast_fp16")]; + tensor xor__7_cast_fp16 = sub(x = var_643_cast_fp16, y = var_644_cast_fp16)[name = tensor("xor__7_cast_fp16")]; + tensor var_647_cast_fp16 = mul(x = xr_7_cast_fp16, y = roti_cast_fp16)[name = tensor("op_647_cast_fp16")]; + tensor var_648_cast_fp16 = mul(x = xi_7_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_648_cast_fp16")]; + tensor xoi_7_cast_fp16 = add(x = var_647_cast_fp16, y = var_648_cast_fp16)[name = tensor("xoi_7_cast_fp16")]; + tensor out_7_axis_0 = const()[name = tensor("out_7_axis_0"), val = tensor(-1)]; + tensor out_7_cast_fp16 = stack(axis = out_7_axis_0, values = (xor__7_cast_fp16, xoi_7_cast_fp16))[name = tensor("out_7_cast_fp16")]; + tensor var_665 = const()[name = tensor("op_665"), val = tensor([256, 1024])]; + tensor var_666_cast_fp16 = reshape(shape = var_665, x = out_7_cast_fp16)[name = tensor("op_666_cast_fp16")]; + tensor written_k_3_transpose_x_1 = const()[name = tensor("written_k_3_transpose_x_1"), val = tensor(true)]; + tensor written_k_3_transpose_y_1 = const()[name = tensor("written_k_3_transpose_y_1"), val = tensor(false)]; + tensor written_k_3_cast_fp16 = matmul(transpose_x = written_k_3_transpose_x_1, transpose_y = written_k_3_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_666_cast_fp16)[name = tensor("written_k_3_cast_fp16")]; + tensor var_673 = const()[name = tensor("op_673"), val = tensor([256, 1024])]; + tensor var_674_cast_fp16 = reshape(shape = var_673, x = v_3_cast_fp16)[name = tensor("op_674_cast_fp16")]; + tensor written_v_3_transpose_x_1 = const()[name = tensor("written_v_3_transpose_x_1"), val = tensor(true)]; + tensor written_v_3_transpose_y_1 = const()[name = tensor("written_v_3_transpose_y_1"), val = tensor(false)]; + tensor written_v_3_cast_fp16 = matmul(transpose_x = written_v_3_transpose_x_1, transpose_y = written_v_3_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_674_cast_fp16)[name = tensor("written_v_3_cast_fp16")]; + tensor k_cache1_to_fp16_dtype_0 = const()[name = tensor("k_cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache1_to_fp16 = cast(dtype = k_cache1_to_fp16_dtype_0, x = k_cache1)[name = tensor("cast_459")]; + tensor var_685_cast_fp16 = mul(x = k_cache1_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_685_cast_fp16")]; + tensor var_690 = const()[name = tensor("op_690"), val = tensor([1, 512, 16, 64])]; + tensor var_691_cast_fp16 = reshape(shape = var_690, x = written_k_3_cast_fp16)[name = tensor("op_691_cast_fp16")]; + tensor new_k_cache_3_cast_fp16 = add(x = var_685_cast_fp16, y = var_691_cast_fp16)[name = tensor("new_k_cache_3_cast_fp16")]; + tensor new_k_cache_3_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_3_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache1_to_fp16_dtype_0 = const()[name = tensor("v_cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache1_to_fp16 = cast(dtype = v_cache1_to_fp16_dtype_0, x = v_cache1)[name = tensor("cast_457")]; + tensor var_694_cast_fp16 = mul(x = v_cache1_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_694_cast_fp16")]; + tensor var_699 = const()[name = tensor("op_699"), val = tensor([1, 512, 16, 64])]; + tensor var_700_cast_fp16 = reshape(shape = var_699, x = written_v_3_cast_fp16)[name = tensor("op_700_cast_fp16")]; + tensor new_v_cache_3_cast_fp16 = add(x = var_694_cast_fp16, y = var_700_cast_fp16)[name = tensor("new_v_cache_3_cast_fp16")]; + tensor new_v_cache_3_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_3_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_713 = const()[name = tensor("op_713"), val = tensor([0, 2, 1, 3])]; + tensor var_724_transpose_x_0 = const()[name = tensor("op_724_transpose_x_0"), val = tensor(false)]; + tensor var_724_transpose_y_0 = const()[name = tensor("op_724_transpose_y_0"), val = tensor(false)]; + tensor transpose_71_perm_0 = const()[name = tensor("transpose_71_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_72_perm_0 = const()[name = tensor("transpose_72_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_72 = transpose(perm = transpose_72_perm_0, x = new_k_cache_3_cast_fp16)[name = tensor("transpose_200")]; + tensor transpose_71 = transpose(perm = transpose_71_perm_0, x = q_3_cast_fp16)[name = tensor("transpose_201")]; + tensor var_724_cast_fp16 = matmul(transpose_x = var_724_transpose_x_0, transpose_y = var_724_transpose_y_0, x = transpose_71, y = transpose_72)[name = tensor("op_724_cast_fp16")]; + tensor var_725_to_fp16 = const()[name = tensor("op_725_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_5_cast_fp16 = mul(x = var_724_cast_fp16, y = var_725_to_fp16)[name = tensor("scores_5_cast_fp16")]; + tensor scores_7_cast_fp16 = add(x = scores_5_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_7_cast_fp16")]; + tensor var_735 = const()[name = tensor("op_735"), val = tensor(-1)]; + tensor weights_3_cast_fp16 = softmax(axis = var_735, x = scores_7_cast_fp16)[name = tensor("weights_3_cast_fp16")]; + tensor attn_3_transpose_x_0 = const()[name = tensor("attn_3_transpose_x_0"), val = tensor(false)]; + tensor attn_3_transpose_y_0 = const()[name = tensor("attn_3_transpose_y_0"), val = tensor(false)]; + tensor values_3_cast_fp16 = transpose(perm = var_713, x = new_v_cache_3_cast_fp16)[name = tensor("transpose_202")]; + tensor attn_3_cast_fp16 = matmul(transpose_x = attn_3_transpose_x_0, transpose_y = attn_3_transpose_y_0, x = weights_3_cast_fp16, y = values_3_cast_fp16)[name = tensor("attn_3_cast_fp16")]; + tensor var_743 = const()[name = tensor("op_743"), val = tensor([0, 2, 1, 3])]; + tensor var_748 = const()[name = tensor("op_748"), val = tensor([1, 256, 1024])]; + tensor var_744_cast_fp16 = transpose(perm = var_743, x = attn_3_cast_fp16)[name = tensor("transpose_199")]; + tensor input_17_cast_fp16 = reshape(shape = var_748, x = var_744_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor attn1_out_proj_weight_to_fp16 = const()[name = tensor("attn1_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31489280)))]; + tensor linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn1_out_proj_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor position1_to_fp16_dtype_0 = const()[name = tensor("position1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position1_to_fp16 = cast(dtype = position1_to_fp16_dtype_0, x = position1)[name = tensor("cast_455")]; + tensor var_757_cast_fp16 = add(x = position1_to_fp16, y = valid_len_to_fp16)[name = tensor("op_757_cast_fp16")]; + tensor var_757_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_757_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_19_cast_fp16 = add(x = input_13_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor input_21_axes_0 = const()[name = tensor("input_21_axes_0"), val = tensor([-1])]; + tensor norm1_2_weight_to_fp16 = const()[name = tensor("norm1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33586496)))]; + tensor norm1_2_bias_to_fp16 = const()[name = tensor("norm1_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33588608)))]; + tensor var_761_to_fp16 = const()[name = tensor("op_761_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_21_cast_fp16 = layer_norm(axes = input_21_axes_0, beta = norm1_2_bias_to_fp16, epsilon = var_761_to_fp16, gamma = norm1_2_weight_to_fp16, x = input_19_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor linear1_1_weight_to_fp16 = const()[name = tensor("linear1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33590720)))]; + tensor linear_6_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear1_1_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor input_23_mode_0 = const()[name = tensor("input_23_mode_0"), val = tensor("EXACT")]; + tensor input_23_cast_fp16 = gelu(mode = input_23_mode_0, x = linear_6_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor linear1_2_weight_to_fp16 = const()[name = tensor("linear1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41979392)))]; + tensor linear_7_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear1_2_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_25_cast_fp16 = add(x = input_19_cast_fp16, y = linear_7_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor input_27_axes_0 = const()[name = tensor("input_27_axes_0"), val = tensor([-1])]; + tensor norm2_1_weight_to_fp16 = const()[name = tensor("norm2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50368064)))]; + tensor norm2_1_bias_to_fp16 = const()[name = tensor("norm2_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50370176)))]; + tensor var_778_to_fp16 = const()[name = tensor("op_778_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = norm2_1_bias_to_fp16, epsilon = var_778_to_fp16, gamma = norm2_1_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor attn2_in_proj_weight_to_fp16 = const()[name = tensor("attn2_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50372288)))]; + tensor linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn2_in_proj_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor var_792 = const()[name = tensor("op_792"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_5_cast_fp16 = reshape(shape = var_792, x = linear_8_cast_fp16)[name = tensor("qkv_5_cast_fp16")]; + tensor x_9_begin_0 = const()[name = tensor("x_9_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_9_end_0 = const()[name = tensor("x_9_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_9_end_mask_0 = const()[name = tensor("x_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_9_squeeze_mask_0 = const()[name = tensor("x_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_9_cast_fp16 = slice_by_index(begin = x_9_begin_0, end = x_9_end_0, end_mask = x_9_end_mask_0, squeeze_mask = x_9_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("x_9_cast_fp16")]; + tensor x_11_begin_0 = const()[name = tensor("x_11_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_11_end_0 = const()[name = tensor("x_11_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_11_end_mask_0 = const()[name = tensor("x_11_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_11_squeeze_mask_0 = const()[name = tensor("x_11_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_11_cast_fp16 = slice_by_index(begin = x_11_begin_0, end = x_11_end_0, end_mask = x_11_end_mask_0, squeeze_mask = x_11_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("x_11_cast_fp16")]; + tensor v_5_begin_0 = const()[name = tensor("v_5_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_5_end_0 = const()[name = tensor("v_5_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_5_end_mask_0 = const()[name = tensor("v_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_5_squeeze_mask_0 = const()[name = tensor("v_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_5_cast_fp16 = slice_by_index(begin = v_5_begin_0, end = v_5_end_0, end_mask = v_5_end_mask_0, squeeze_mask = v_5_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("v_5_cast_fp16")]; + tensor var_837 = const()[name = tensor("op_837"), val = tensor([256, 16, 32, 2])]; + tensor pairs_9_cast_fp16 = reshape(shape = var_837, x = x_9_cast_fp16)[name = tensor("pairs_9_cast_fp16")]; + tensor xr_9_begin_0 = const()[name = tensor("xr_9_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_9_end_0 = const()[name = tensor("xr_9_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_9_end_mask_0 = const()[name = tensor("xr_9_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_9_squeeze_mask_0 = const()[name = tensor("xr_9_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_9_cast_fp16 = slice_by_index(begin = xr_9_begin_0, end = xr_9_end_0, end_mask = xr_9_end_mask_0, squeeze_mask = xr_9_squeeze_mask_0, x = pairs_9_cast_fp16)[name = tensor("xr_9_cast_fp16")]; + tensor xi_9_begin_0 = const()[name = tensor("xi_9_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_9_end_0 = const()[name = tensor("xi_9_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_9_end_mask_0 = const()[name = tensor("xi_9_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_9_squeeze_mask_0 = const()[name = tensor("xi_9_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_9_cast_fp16 = slice_by_index(begin = xi_9_begin_0, end = xi_9_end_0, end_mask = xi_9_end_mask_0, squeeze_mask = xi_9_squeeze_mask_0, x = pairs_9_cast_fp16)[name = tensor("xi_9_cast_fp16")]; + tensor var_845_cast_fp16 = mul(x = xr_9_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_845_cast_fp16")]; + tensor var_846_cast_fp16 = mul(x = xi_9_cast_fp16, y = roti_cast_fp16)[name = tensor("op_846_cast_fp16")]; + tensor xor__9_cast_fp16 = sub(x = var_845_cast_fp16, y = var_846_cast_fp16)[name = tensor("xor__9_cast_fp16")]; + tensor var_849_cast_fp16 = mul(x = xr_9_cast_fp16, y = roti_cast_fp16)[name = tensor("op_849_cast_fp16")]; + tensor var_850_cast_fp16 = mul(x = xi_9_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_850_cast_fp16")]; + tensor xoi_9_cast_fp16 = add(x = var_849_cast_fp16, y = var_850_cast_fp16)[name = tensor("xoi_9_cast_fp16")]; + tensor out_9_axis_0 = const()[name = tensor("out_9_axis_0"), val = tensor(-1)]; + tensor out_9_cast_fp16 = stack(axis = out_9_axis_0, values = (xor__9_cast_fp16, xoi_9_cast_fp16))[name = tensor("out_9_cast_fp16")]; + tensor var_860 = const()[name = tensor("op_860"), val = tensor([1, 256, 16, 64])]; + tensor q_5_cast_fp16 = reshape(shape = var_860, x = out_9_cast_fp16)[name = tensor("q_5_cast_fp16")]; + tensor var_866 = const()[name = tensor("op_866"), val = tensor([256, 16, 32, 2])]; + tensor pairs_11_cast_fp16 = reshape(shape = var_866, x = x_11_cast_fp16)[name = tensor("pairs_11_cast_fp16")]; + tensor xr_11_begin_0 = const()[name = tensor("xr_11_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_11_end_0 = const()[name = tensor("xr_11_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_11_end_mask_0 = const()[name = tensor("xr_11_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_11_squeeze_mask_0 = const()[name = tensor("xr_11_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_11_cast_fp16 = slice_by_index(begin = xr_11_begin_0, end = xr_11_end_0, end_mask = xr_11_end_mask_0, squeeze_mask = xr_11_squeeze_mask_0, x = pairs_11_cast_fp16)[name = tensor("xr_11_cast_fp16")]; + tensor xi_11_begin_0 = const()[name = tensor("xi_11_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_11_end_0 = const()[name = tensor("xi_11_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_11_end_mask_0 = const()[name = tensor("xi_11_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_11_squeeze_mask_0 = const()[name = tensor("xi_11_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_11_cast_fp16 = slice_by_index(begin = xi_11_begin_0, end = xi_11_end_0, end_mask = xi_11_end_mask_0, squeeze_mask = xi_11_squeeze_mask_0, x = pairs_11_cast_fp16)[name = tensor("xi_11_cast_fp16")]; + tensor var_874_cast_fp16 = mul(x = xr_11_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_874_cast_fp16")]; + tensor var_875_cast_fp16 = mul(x = xi_11_cast_fp16, y = roti_cast_fp16)[name = tensor("op_875_cast_fp16")]; + tensor xor__11_cast_fp16 = sub(x = var_874_cast_fp16, y = var_875_cast_fp16)[name = tensor("xor__11_cast_fp16")]; + tensor var_878_cast_fp16 = mul(x = xr_11_cast_fp16, y = roti_cast_fp16)[name = tensor("op_878_cast_fp16")]; + tensor var_879_cast_fp16 = mul(x = xi_11_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_879_cast_fp16")]; + tensor xoi_11_cast_fp16 = add(x = var_878_cast_fp16, y = var_879_cast_fp16)[name = tensor("xoi_11_cast_fp16")]; + tensor out_11_axis_0 = const()[name = tensor("out_11_axis_0"), val = tensor(-1)]; + tensor out_11_cast_fp16 = stack(axis = out_11_axis_0, values = (xor__11_cast_fp16, xoi_11_cast_fp16))[name = tensor("out_11_cast_fp16")]; + tensor var_896 = const()[name = tensor("op_896"), val = tensor([256, 1024])]; + tensor var_897_cast_fp16 = reshape(shape = var_896, x = out_11_cast_fp16)[name = tensor("op_897_cast_fp16")]; + tensor written_k_5_transpose_x_1 = const()[name = tensor("written_k_5_transpose_x_1"), val = tensor(true)]; + tensor written_k_5_transpose_y_1 = const()[name = tensor("written_k_5_transpose_y_1"), val = tensor(false)]; + tensor written_k_5_cast_fp16 = matmul(transpose_x = written_k_5_transpose_x_1, transpose_y = written_k_5_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_897_cast_fp16)[name = tensor("written_k_5_cast_fp16")]; + tensor var_904 = const()[name = tensor("op_904"), val = tensor([256, 1024])]; + tensor var_905_cast_fp16 = reshape(shape = var_904, x = v_5_cast_fp16)[name = tensor("op_905_cast_fp16")]; + tensor written_v_5_transpose_x_1 = const()[name = tensor("written_v_5_transpose_x_1"), val = tensor(true)]; + tensor written_v_5_transpose_y_1 = const()[name = tensor("written_v_5_transpose_y_1"), val = tensor(false)]; + tensor written_v_5_cast_fp16 = matmul(transpose_x = written_v_5_transpose_x_1, transpose_y = written_v_5_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_905_cast_fp16)[name = tensor("written_v_5_cast_fp16")]; + tensor k_cache2_to_fp16_dtype_0 = const()[name = tensor("k_cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache2_to_fp16 = cast(dtype = k_cache2_to_fp16_dtype_0, x = k_cache2)[name = tensor("cast_453")]; + tensor var_916_cast_fp16 = mul(x = k_cache2_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_916_cast_fp16")]; + tensor var_921 = const()[name = tensor("op_921"), val = tensor([1, 512, 16, 64])]; + tensor var_922_cast_fp16 = reshape(shape = var_921, x = written_k_5_cast_fp16)[name = tensor("op_922_cast_fp16")]; + tensor new_k_cache_5_cast_fp16 = add(x = var_916_cast_fp16, y = var_922_cast_fp16)[name = tensor("new_k_cache_5_cast_fp16")]; + tensor new_k_cache_5_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_5_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache2_to_fp16_dtype_0 = const()[name = tensor("v_cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache2_to_fp16 = cast(dtype = v_cache2_to_fp16_dtype_0, x = v_cache2)[name = tensor("cast_451")]; + tensor var_925_cast_fp16 = mul(x = v_cache2_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_925_cast_fp16")]; + tensor var_930 = const()[name = tensor("op_930"), val = tensor([1, 512, 16, 64])]; + tensor var_931_cast_fp16 = reshape(shape = var_930, x = written_v_5_cast_fp16)[name = tensor("op_931_cast_fp16")]; + tensor new_v_cache_5_cast_fp16 = add(x = var_925_cast_fp16, y = var_931_cast_fp16)[name = tensor("new_v_cache_5_cast_fp16")]; + tensor new_v_cache_5_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_5_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_944 = const()[name = tensor("op_944"), val = tensor([0, 2, 1, 3])]; + tensor var_955_transpose_x_0 = const()[name = tensor("op_955_transpose_x_0"), val = tensor(false)]; + tensor var_955_transpose_y_0 = const()[name = tensor("op_955_transpose_y_0"), val = tensor(false)]; + tensor transpose_73_perm_0 = const()[name = tensor("transpose_73_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_74_perm_0 = const()[name = tensor("transpose_74_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_74 = transpose(perm = transpose_74_perm_0, x = new_k_cache_5_cast_fp16)[name = tensor("transpose_196")]; + tensor transpose_73 = transpose(perm = transpose_73_perm_0, x = q_5_cast_fp16)[name = tensor("transpose_197")]; + tensor var_955_cast_fp16 = matmul(transpose_x = var_955_transpose_x_0, transpose_y = var_955_transpose_y_0, x = transpose_73, y = transpose_74)[name = tensor("op_955_cast_fp16")]; + tensor var_956_to_fp16 = const()[name = tensor("op_956_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_9_cast_fp16 = mul(x = var_955_cast_fp16, y = var_956_to_fp16)[name = tensor("scores_9_cast_fp16")]; + tensor scores_11_cast_fp16 = add(x = scores_9_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_11_cast_fp16")]; + tensor var_966 = const()[name = tensor("op_966"), val = tensor(-1)]; + tensor weights_5_cast_fp16 = softmax(axis = var_966, x = scores_11_cast_fp16)[name = tensor("weights_5_cast_fp16")]; + tensor attn_5_transpose_x_0 = const()[name = tensor("attn_5_transpose_x_0"), val = tensor(false)]; + tensor attn_5_transpose_y_0 = const()[name = tensor("attn_5_transpose_y_0"), val = tensor(false)]; + tensor values_5_cast_fp16 = transpose(perm = var_944, x = new_v_cache_5_cast_fp16)[name = tensor("transpose_198")]; + tensor attn_5_cast_fp16 = matmul(transpose_x = attn_5_transpose_x_0, transpose_y = attn_5_transpose_y_0, x = weights_5_cast_fp16, y = values_5_cast_fp16)[name = tensor("attn_5_cast_fp16")]; + tensor var_974 = const()[name = tensor("op_974"), val = tensor([0, 2, 1, 3])]; + tensor var_979 = const()[name = tensor("op_979"), val = tensor([1, 256, 1024])]; + tensor var_975_cast_fp16 = transpose(perm = var_974, x = attn_5_cast_fp16)[name = tensor("transpose_195")]; + tensor input_29_cast_fp16 = reshape(shape = var_979, x = var_975_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor attn2_out_proj_weight_to_fp16 = const()[name = tensor("attn2_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(56663808)))]; + tensor linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn2_out_proj_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor position2_to_fp16_dtype_0 = const()[name = tensor("position2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position2_to_fp16 = cast(dtype = position2_to_fp16_dtype_0, x = position2)[name = tensor("cast_449")]; + tensor var_988_cast_fp16 = add(x = position2_to_fp16, y = valid_len_to_fp16)[name = tensor("op_988_cast_fp16")]; + tensor var_988_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_988_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_31_cast_fp16 = add(x = input_25_cast_fp16, y = linear_9_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; + tensor norm2_2_weight_to_fp16 = const()[name = tensor("norm2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58761024)))]; + tensor norm2_2_bias_to_fp16 = const()[name = tensor("norm2_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58763136)))]; + tensor var_992_to_fp16 = const()[name = tensor("op_992_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_33_cast_fp16 = layer_norm(axes = input_33_axes_0, beta = norm2_2_bias_to_fp16, epsilon = var_992_to_fp16, gamma = norm2_2_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor linear2_1_weight_to_fp16 = const()[name = tensor("linear2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58765248)))]; + tensor linear_10_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear2_1_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor input_35_mode_0 = const()[name = tensor("input_35_mode_0"), val = tensor("EXACT")]; + tensor input_35_cast_fp16 = gelu(mode = input_35_mode_0, x = linear_10_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor linear2_2_weight_to_fp16 = const()[name = tensor("linear2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67153920)))]; + tensor linear_11_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear2_2_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor input_37_cast_fp16 = add(x = input_31_cast_fp16, y = linear_11_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor input_39_axes_0 = const()[name = tensor("input_39_axes_0"), val = tensor([-1])]; + tensor norm3_1_weight_to_fp16 = const()[name = tensor("norm3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75542592)))]; + tensor norm3_1_bias_to_fp16 = const()[name = tensor("norm3_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75544704)))]; + tensor var_1009_to_fp16 = const()[name = tensor("op_1009_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_39_cast_fp16 = layer_norm(axes = input_39_axes_0, beta = norm3_1_bias_to_fp16, epsilon = var_1009_to_fp16, gamma = norm3_1_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor attn3_in_proj_weight_to_fp16 = const()[name = tensor("attn3_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75546816)))]; + tensor linear_12_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn3_in_proj_weight_to_fp16, x = input_39_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor var_1023 = const()[name = tensor("op_1023"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_7_cast_fp16 = reshape(shape = var_1023, x = linear_12_cast_fp16)[name = tensor("qkv_7_cast_fp16")]; + tensor x_13_begin_0 = const()[name = tensor("x_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_13_end_0 = const()[name = tensor("x_13_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_13_end_mask_0 = const()[name = tensor("x_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_13_squeeze_mask_0 = const()[name = tensor("x_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_13_cast_fp16 = slice_by_index(begin = x_13_begin_0, end = x_13_end_0, end_mask = x_13_end_mask_0, squeeze_mask = x_13_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("x_13_cast_fp16")]; + tensor x_15_begin_0 = const()[name = tensor("x_15_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_15_end_0 = const()[name = tensor("x_15_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_15_end_mask_0 = const()[name = tensor("x_15_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_15_squeeze_mask_0 = const()[name = tensor("x_15_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_15_cast_fp16 = slice_by_index(begin = x_15_begin_0, end = x_15_end_0, end_mask = x_15_end_mask_0, squeeze_mask = x_15_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("x_15_cast_fp16")]; + tensor v_7_begin_0 = const()[name = tensor("v_7_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_7_end_0 = const()[name = tensor("v_7_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_7_end_mask_0 = const()[name = tensor("v_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_7_squeeze_mask_0 = const()[name = tensor("v_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_7_cast_fp16 = slice_by_index(begin = v_7_begin_0, end = v_7_end_0, end_mask = v_7_end_mask_0, squeeze_mask = v_7_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("v_7_cast_fp16")]; + tensor var_1068 = const()[name = tensor("op_1068"), val = tensor([256, 16, 32, 2])]; + tensor pairs_13_cast_fp16 = reshape(shape = var_1068, x = x_13_cast_fp16)[name = tensor("pairs_13_cast_fp16")]; + tensor xr_13_begin_0 = const()[name = tensor("xr_13_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_13_end_0 = const()[name = tensor("xr_13_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_13_end_mask_0 = const()[name = tensor("xr_13_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_13_squeeze_mask_0 = const()[name = tensor("xr_13_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_13_cast_fp16 = slice_by_index(begin = xr_13_begin_0, end = xr_13_end_0, end_mask = xr_13_end_mask_0, squeeze_mask = xr_13_squeeze_mask_0, x = pairs_13_cast_fp16)[name = tensor("xr_13_cast_fp16")]; + tensor xi_13_begin_0 = const()[name = tensor("xi_13_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_13_end_0 = const()[name = tensor("xi_13_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_13_end_mask_0 = const()[name = tensor("xi_13_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_13_squeeze_mask_0 = const()[name = tensor("xi_13_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_13_cast_fp16 = slice_by_index(begin = xi_13_begin_0, end = xi_13_end_0, end_mask = xi_13_end_mask_0, squeeze_mask = xi_13_squeeze_mask_0, x = pairs_13_cast_fp16)[name = tensor("xi_13_cast_fp16")]; + tensor var_1076_cast_fp16 = mul(x = xr_13_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1076_cast_fp16")]; + tensor var_1077_cast_fp16 = mul(x = xi_13_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1077_cast_fp16")]; + tensor xor__13_cast_fp16 = sub(x = var_1076_cast_fp16, y = var_1077_cast_fp16)[name = tensor("xor__13_cast_fp16")]; + tensor var_1080_cast_fp16 = mul(x = xr_13_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1080_cast_fp16")]; + tensor var_1081_cast_fp16 = mul(x = xi_13_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1081_cast_fp16")]; + tensor xoi_13_cast_fp16 = add(x = var_1080_cast_fp16, y = var_1081_cast_fp16)[name = tensor("xoi_13_cast_fp16")]; + tensor out_13_axis_0 = const()[name = tensor("out_13_axis_0"), val = tensor(-1)]; + tensor out_13_cast_fp16 = stack(axis = out_13_axis_0, values = (xor__13_cast_fp16, xoi_13_cast_fp16))[name = tensor("out_13_cast_fp16")]; + tensor var_1091 = const()[name = tensor("op_1091"), val = tensor([1, 256, 16, 64])]; + tensor q_7_cast_fp16 = reshape(shape = var_1091, x = out_13_cast_fp16)[name = tensor("q_7_cast_fp16")]; + tensor var_1097 = const()[name = tensor("op_1097"), val = tensor([256, 16, 32, 2])]; + tensor pairs_15_cast_fp16 = reshape(shape = var_1097, x = x_15_cast_fp16)[name = tensor("pairs_15_cast_fp16")]; + tensor xr_15_begin_0 = const()[name = tensor("xr_15_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_15_end_0 = const()[name = tensor("xr_15_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_15_end_mask_0 = const()[name = tensor("xr_15_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_15_squeeze_mask_0 = const()[name = tensor("xr_15_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_15_cast_fp16 = slice_by_index(begin = xr_15_begin_0, end = xr_15_end_0, end_mask = xr_15_end_mask_0, squeeze_mask = xr_15_squeeze_mask_0, x = pairs_15_cast_fp16)[name = tensor("xr_15_cast_fp16")]; + tensor xi_15_begin_0 = const()[name = tensor("xi_15_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_15_end_0 = const()[name = tensor("xi_15_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_15_end_mask_0 = const()[name = tensor("xi_15_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_15_squeeze_mask_0 = const()[name = tensor("xi_15_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_15_cast_fp16 = slice_by_index(begin = xi_15_begin_0, end = xi_15_end_0, end_mask = xi_15_end_mask_0, squeeze_mask = xi_15_squeeze_mask_0, x = pairs_15_cast_fp16)[name = tensor("xi_15_cast_fp16")]; + tensor var_1105_cast_fp16 = mul(x = xr_15_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1105_cast_fp16")]; + tensor var_1106_cast_fp16 = mul(x = xi_15_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1106_cast_fp16")]; + tensor xor__15_cast_fp16 = sub(x = var_1105_cast_fp16, y = var_1106_cast_fp16)[name = tensor("xor__15_cast_fp16")]; + tensor var_1109_cast_fp16 = mul(x = xr_15_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1109_cast_fp16")]; + tensor var_1110_cast_fp16 = mul(x = xi_15_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1110_cast_fp16")]; + tensor xoi_15_cast_fp16 = add(x = var_1109_cast_fp16, y = var_1110_cast_fp16)[name = tensor("xoi_15_cast_fp16")]; + tensor out_15_axis_0 = const()[name = tensor("out_15_axis_0"), val = tensor(-1)]; + tensor out_15_cast_fp16 = stack(axis = out_15_axis_0, values = (xor__15_cast_fp16, xoi_15_cast_fp16))[name = tensor("out_15_cast_fp16")]; + tensor var_1127 = const()[name = tensor("op_1127"), val = tensor([256, 1024])]; + tensor var_1128_cast_fp16 = reshape(shape = var_1127, x = out_15_cast_fp16)[name = tensor("op_1128_cast_fp16")]; + tensor written_k_7_transpose_x_1 = const()[name = tensor("written_k_7_transpose_x_1"), val = tensor(true)]; + tensor written_k_7_transpose_y_1 = const()[name = tensor("written_k_7_transpose_y_1"), val = tensor(false)]; + tensor written_k_7_cast_fp16 = matmul(transpose_x = written_k_7_transpose_x_1, transpose_y = written_k_7_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_1128_cast_fp16)[name = tensor("written_k_7_cast_fp16")]; + tensor var_1135 = const()[name = tensor("op_1135"), val = tensor([256, 1024])]; + tensor var_1136_cast_fp16 = reshape(shape = var_1135, x = v_7_cast_fp16)[name = tensor("op_1136_cast_fp16")]; + tensor written_v_7_transpose_x_1 = const()[name = tensor("written_v_7_transpose_x_1"), val = tensor(true)]; + tensor written_v_7_transpose_y_1 = const()[name = tensor("written_v_7_transpose_y_1"), val = tensor(false)]; + tensor written_v_7_cast_fp16 = matmul(transpose_x = written_v_7_transpose_x_1, transpose_y = written_v_7_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_1136_cast_fp16)[name = tensor("written_v_7_cast_fp16")]; + tensor k_cache3_to_fp16_dtype_0 = const()[name = tensor("k_cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache3_to_fp16 = cast(dtype = k_cache3_to_fp16_dtype_0, x = k_cache3)[name = tensor("cast_447")]; + tensor var_1147_cast_fp16 = mul(x = k_cache3_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_1147_cast_fp16")]; + tensor var_1152 = const()[name = tensor("op_1152"), val = tensor([1, 512, 16, 64])]; + tensor var_1153_cast_fp16 = reshape(shape = var_1152, x = written_k_7_cast_fp16)[name = tensor("op_1153_cast_fp16")]; + tensor new_k_cache_7_cast_fp16 = add(x = var_1147_cast_fp16, y = var_1153_cast_fp16)[name = tensor("new_k_cache_7_cast_fp16")]; + tensor new_k_cache_7_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_7_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache3_to_fp16_dtype_0 = const()[name = tensor("v_cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache3_to_fp16 = cast(dtype = v_cache3_to_fp16_dtype_0, x = v_cache3)[name = tensor("cast_445")]; + tensor var_1156_cast_fp16 = mul(x = v_cache3_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_1156_cast_fp16")]; + tensor var_1161 = const()[name = tensor("op_1161"), val = tensor([1, 512, 16, 64])]; + tensor var_1162_cast_fp16 = reshape(shape = var_1161, x = written_v_7_cast_fp16)[name = tensor("op_1162_cast_fp16")]; + tensor new_v_cache_7_cast_fp16 = add(x = var_1156_cast_fp16, y = var_1162_cast_fp16)[name = tensor("new_v_cache_7_cast_fp16")]; + tensor new_v_cache_7_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_7_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_1175 = const()[name = tensor("op_1175"), val = tensor([0, 2, 1, 3])]; + tensor var_1186_transpose_x_0 = const()[name = tensor("op_1186_transpose_x_0"), val = tensor(false)]; + tensor var_1186_transpose_y_0 = const()[name = tensor("op_1186_transpose_y_0"), val = tensor(false)]; + tensor transpose_75_perm_0 = const()[name = tensor("transpose_75_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_76_perm_0 = const()[name = tensor("transpose_76_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_76 = transpose(perm = transpose_76_perm_0, x = new_k_cache_7_cast_fp16)[name = tensor("transpose_192")]; + tensor transpose_75 = transpose(perm = transpose_75_perm_0, x = q_7_cast_fp16)[name = tensor("transpose_193")]; + tensor var_1186_cast_fp16 = matmul(transpose_x = var_1186_transpose_x_0, transpose_y = var_1186_transpose_y_0, x = transpose_75, y = transpose_76)[name = tensor("op_1186_cast_fp16")]; + tensor var_1187_to_fp16 = const()[name = tensor("op_1187_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_13_cast_fp16 = mul(x = var_1186_cast_fp16, y = var_1187_to_fp16)[name = tensor("scores_13_cast_fp16")]; + tensor scores_15_cast_fp16 = add(x = scores_13_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_15_cast_fp16")]; + tensor var_1197 = const()[name = tensor("op_1197"), val = tensor(-1)]; + tensor weights_7_cast_fp16 = softmax(axis = var_1197, x = scores_15_cast_fp16)[name = tensor("weights_7_cast_fp16")]; + tensor attn_7_transpose_x_0 = const()[name = tensor("attn_7_transpose_x_0"), val = tensor(false)]; + tensor attn_7_transpose_y_0 = const()[name = tensor("attn_7_transpose_y_0"), val = tensor(false)]; + tensor values_7_cast_fp16 = transpose(perm = var_1175, x = new_v_cache_7_cast_fp16)[name = tensor("transpose_194")]; + tensor attn_7_cast_fp16 = matmul(transpose_x = attn_7_transpose_x_0, transpose_y = attn_7_transpose_y_0, x = weights_7_cast_fp16, y = values_7_cast_fp16)[name = tensor("attn_7_cast_fp16")]; + tensor var_1205 = const()[name = tensor("op_1205"), val = tensor([0, 2, 1, 3])]; + tensor var_1210 = const()[name = tensor("op_1210"), val = tensor([1, 256, 1024])]; + tensor var_1206_cast_fp16 = transpose(perm = var_1205, x = attn_7_cast_fp16)[name = tensor("transpose_191")]; + tensor input_41_cast_fp16 = reshape(shape = var_1210, x = var_1206_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor attn3_out_proj_weight_to_fp16 = const()[name = tensor("attn3_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(81838336)))]; + tensor linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn3_out_proj_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor position3_to_fp16_dtype_0 = const()[name = tensor("position3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position3_to_fp16 = cast(dtype = position3_to_fp16_dtype_0, x = position3)[name = tensor("cast_443")]; + tensor var_1219_cast_fp16 = add(x = position3_to_fp16, y = valid_len_to_fp16)[name = tensor("op_1219_cast_fp16")]; + tensor var_1219_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1219_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_43_cast_fp16 = add(x = input_37_cast_fp16, y = linear_13_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; + tensor norm3_2_weight_to_fp16 = const()[name = tensor("norm3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(83935552)))]; + tensor norm3_2_bias_to_fp16 = const()[name = tensor("norm3_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(83937664)))]; + tensor var_1223_to_fp16 = const()[name = tensor("op_1223_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_45_cast_fp16 = layer_norm(axes = input_45_axes_0, beta = norm3_2_bias_to_fp16, epsilon = var_1223_to_fp16, gamma = norm3_2_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor linear3_1_weight_to_fp16 = const()[name = tensor("linear3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(83939776)))]; + tensor linear_14_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear3_1_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor input_47_mode_0 = const()[name = tensor("input_47_mode_0"), val = tensor("EXACT")]; + tensor input_47_cast_fp16 = gelu(mode = input_47_mode_0, x = linear_14_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor linear3_2_weight_to_fp16 = const()[name = tensor("linear3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(92328448)))]; + tensor linear_15_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear3_2_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor input_49_cast_fp16 = add(x = input_43_cast_fp16, y = linear_15_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor input_51_axes_0 = const()[name = tensor("input_51_axes_0"), val = tensor([-1])]; + tensor norm4_1_weight_to_fp16 = const()[name = tensor("norm4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100717120)))]; + tensor norm4_1_bias_to_fp16 = const()[name = tensor("norm4_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100719232)))]; + tensor var_1240_to_fp16 = const()[name = tensor("op_1240_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_51_cast_fp16 = layer_norm(axes = input_51_axes_0, beta = norm4_1_bias_to_fp16, epsilon = var_1240_to_fp16, gamma = norm4_1_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor attn4_in_proj_weight_to_fp16 = const()[name = tensor("attn4_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100721344)))]; + tensor linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn4_in_proj_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor var_1254 = const()[name = tensor("op_1254"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_9_cast_fp16 = reshape(shape = var_1254, x = linear_16_cast_fp16)[name = tensor("qkv_9_cast_fp16")]; + tensor x_17_begin_0 = const()[name = tensor("x_17_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_17_end_0 = const()[name = tensor("x_17_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_17_end_mask_0 = const()[name = tensor("x_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_17_squeeze_mask_0 = const()[name = tensor("x_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_17_cast_fp16 = slice_by_index(begin = x_17_begin_0, end = x_17_end_0, end_mask = x_17_end_mask_0, squeeze_mask = x_17_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("x_17_cast_fp16")]; + tensor x_19_begin_0 = const()[name = tensor("x_19_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_19_end_0 = const()[name = tensor("x_19_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_19_end_mask_0 = const()[name = tensor("x_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_19_squeeze_mask_0 = const()[name = tensor("x_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_19_cast_fp16 = slice_by_index(begin = x_19_begin_0, end = x_19_end_0, end_mask = x_19_end_mask_0, squeeze_mask = x_19_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("x_19_cast_fp16")]; + tensor v_9_begin_0 = const()[name = tensor("v_9_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_9_end_0 = const()[name = tensor("v_9_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_9_end_mask_0 = const()[name = tensor("v_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_9_squeeze_mask_0 = const()[name = tensor("v_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_9_cast_fp16 = slice_by_index(begin = v_9_begin_0, end = v_9_end_0, end_mask = v_9_end_mask_0, squeeze_mask = v_9_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("v_9_cast_fp16")]; + tensor var_1299 = const()[name = tensor("op_1299"), val = tensor([256, 16, 32, 2])]; + tensor pairs_17_cast_fp16 = reshape(shape = var_1299, x = x_17_cast_fp16)[name = tensor("pairs_17_cast_fp16")]; + tensor xr_17_begin_0 = const()[name = tensor("xr_17_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_17_end_0 = const()[name = tensor("xr_17_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_17_end_mask_0 = const()[name = tensor("xr_17_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_17_squeeze_mask_0 = const()[name = tensor("xr_17_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_17_cast_fp16 = slice_by_index(begin = xr_17_begin_0, end = xr_17_end_0, end_mask = xr_17_end_mask_0, squeeze_mask = xr_17_squeeze_mask_0, x = pairs_17_cast_fp16)[name = tensor("xr_17_cast_fp16")]; + tensor xi_17_begin_0 = const()[name = tensor("xi_17_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_17_end_0 = const()[name = tensor("xi_17_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_17_end_mask_0 = const()[name = tensor("xi_17_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_17_squeeze_mask_0 = const()[name = tensor("xi_17_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_17_cast_fp16 = slice_by_index(begin = xi_17_begin_0, end = xi_17_end_0, end_mask = xi_17_end_mask_0, squeeze_mask = xi_17_squeeze_mask_0, x = pairs_17_cast_fp16)[name = tensor("xi_17_cast_fp16")]; + tensor var_1307_cast_fp16 = mul(x = xr_17_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1307_cast_fp16")]; + tensor var_1308_cast_fp16 = mul(x = xi_17_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1308_cast_fp16")]; + tensor xor__17_cast_fp16 = sub(x = var_1307_cast_fp16, y = var_1308_cast_fp16)[name = tensor("xor__17_cast_fp16")]; + tensor var_1311_cast_fp16 = mul(x = xr_17_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1311_cast_fp16")]; + tensor var_1312_cast_fp16 = mul(x = xi_17_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1312_cast_fp16")]; + tensor xoi_17_cast_fp16 = add(x = var_1311_cast_fp16, y = var_1312_cast_fp16)[name = tensor("xoi_17_cast_fp16")]; + tensor out_17_axis_0 = const()[name = tensor("out_17_axis_0"), val = tensor(-1)]; + tensor out_17_cast_fp16 = stack(axis = out_17_axis_0, values = (xor__17_cast_fp16, xoi_17_cast_fp16))[name = tensor("out_17_cast_fp16")]; + tensor var_1322 = const()[name = tensor("op_1322"), val = tensor([1, 256, 16, 64])]; + tensor q_9_cast_fp16 = reshape(shape = var_1322, x = out_17_cast_fp16)[name = tensor("q_9_cast_fp16")]; + tensor var_1328 = const()[name = tensor("op_1328"), val = tensor([256, 16, 32, 2])]; + tensor pairs_19_cast_fp16 = reshape(shape = var_1328, x = x_19_cast_fp16)[name = tensor("pairs_19_cast_fp16")]; + tensor xr_19_begin_0 = const()[name = tensor("xr_19_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_19_end_0 = const()[name = tensor("xr_19_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_19_end_mask_0 = const()[name = tensor("xr_19_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_19_squeeze_mask_0 = const()[name = tensor("xr_19_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_19_cast_fp16 = slice_by_index(begin = xr_19_begin_0, end = xr_19_end_0, end_mask = xr_19_end_mask_0, squeeze_mask = xr_19_squeeze_mask_0, x = pairs_19_cast_fp16)[name = tensor("xr_19_cast_fp16")]; + tensor xi_19_begin_0 = const()[name = tensor("xi_19_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_19_end_0 = const()[name = tensor("xi_19_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_19_end_mask_0 = const()[name = tensor("xi_19_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_19_squeeze_mask_0 = const()[name = tensor("xi_19_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_19_cast_fp16 = slice_by_index(begin = xi_19_begin_0, end = xi_19_end_0, end_mask = xi_19_end_mask_0, squeeze_mask = xi_19_squeeze_mask_0, x = pairs_19_cast_fp16)[name = tensor("xi_19_cast_fp16")]; + tensor var_1336_cast_fp16 = mul(x = xr_19_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1336_cast_fp16")]; + tensor var_1337_cast_fp16 = mul(x = xi_19_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1337_cast_fp16")]; + tensor xor__19_cast_fp16 = sub(x = var_1336_cast_fp16, y = var_1337_cast_fp16)[name = tensor("xor__19_cast_fp16")]; + tensor var_1340_cast_fp16 = mul(x = xr_19_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1340_cast_fp16")]; + tensor var_1341_cast_fp16 = mul(x = xi_19_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1341_cast_fp16")]; + tensor xoi_19_cast_fp16 = add(x = var_1340_cast_fp16, y = var_1341_cast_fp16)[name = tensor("xoi_19_cast_fp16")]; + tensor out_19_axis_0 = const()[name = tensor("out_19_axis_0"), val = tensor(-1)]; + tensor out_19_cast_fp16 = stack(axis = out_19_axis_0, values = (xor__19_cast_fp16, xoi_19_cast_fp16))[name = tensor("out_19_cast_fp16")]; + tensor var_1358 = const()[name = tensor("op_1358"), val = tensor([256, 1024])]; + tensor var_1359_cast_fp16 = reshape(shape = var_1358, x = out_19_cast_fp16)[name = tensor("op_1359_cast_fp16")]; + tensor written_k_9_transpose_x_1 = const()[name = tensor("written_k_9_transpose_x_1"), val = tensor(true)]; + tensor written_k_9_transpose_y_1 = const()[name = tensor("written_k_9_transpose_y_1"), val = tensor(false)]; + tensor written_k_9_cast_fp16 = matmul(transpose_x = written_k_9_transpose_x_1, transpose_y = written_k_9_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_1359_cast_fp16)[name = tensor("written_k_9_cast_fp16")]; + tensor var_1366 = const()[name = tensor("op_1366"), val = tensor([256, 1024])]; + tensor var_1367_cast_fp16 = reshape(shape = var_1366, x = v_9_cast_fp16)[name = tensor("op_1367_cast_fp16")]; + tensor written_v_9_transpose_x_1 = const()[name = tensor("written_v_9_transpose_x_1"), val = tensor(true)]; + tensor written_v_9_transpose_y_1 = const()[name = tensor("written_v_9_transpose_y_1"), val = tensor(false)]; + tensor written_v_9_cast_fp16 = matmul(transpose_x = written_v_9_transpose_x_1, transpose_y = written_v_9_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_1367_cast_fp16)[name = tensor("written_v_9_cast_fp16")]; + tensor k_cache4_to_fp16_dtype_0 = const()[name = tensor("k_cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache4_to_fp16 = cast(dtype = k_cache4_to_fp16_dtype_0, x = k_cache4)[name = tensor("cast_441")]; + tensor var_1378_cast_fp16 = mul(x = k_cache4_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_1378_cast_fp16")]; + tensor var_1383 = const()[name = tensor("op_1383"), val = tensor([1, 512, 16, 64])]; + tensor var_1384_cast_fp16 = reshape(shape = var_1383, x = written_k_9_cast_fp16)[name = tensor("op_1384_cast_fp16")]; + tensor new_k_cache_9_cast_fp16 = add(x = var_1378_cast_fp16, y = var_1384_cast_fp16)[name = tensor("new_k_cache_9_cast_fp16")]; + tensor new_k_cache_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache4_to_fp16_dtype_0 = const()[name = tensor("v_cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache4_to_fp16 = cast(dtype = v_cache4_to_fp16_dtype_0, x = v_cache4)[name = tensor("cast_439")]; + tensor var_1387_cast_fp16 = mul(x = v_cache4_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_1387_cast_fp16")]; + tensor var_1392 = const()[name = tensor("op_1392"), val = tensor([1, 512, 16, 64])]; + tensor var_1393_cast_fp16 = reshape(shape = var_1392, x = written_v_9_cast_fp16)[name = tensor("op_1393_cast_fp16")]; + tensor new_v_cache_9_cast_fp16 = add(x = var_1387_cast_fp16, y = var_1393_cast_fp16)[name = tensor("new_v_cache_9_cast_fp16")]; + tensor new_v_cache_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_1406 = const()[name = tensor("op_1406"), val = tensor([0, 2, 1, 3])]; + tensor var_1417_transpose_x_0 = const()[name = tensor("op_1417_transpose_x_0"), val = tensor(false)]; + tensor var_1417_transpose_y_0 = const()[name = tensor("op_1417_transpose_y_0"), val = tensor(false)]; + tensor transpose_77_perm_0 = const()[name = tensor("transpose_77_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_78_perm_0 = const()[name = tensor("transpose_78_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_78 = transpose(perm = transpose_78_perm_0, x = new_k_cache_9_cast_fp16)[name = tensor("transpose_188")]; + tensor transpose_77 = transpose(perm = transpose_77_perm_0, x = q_9_cast_fp16)[name = tensor("transpose_189")]; + tensor var_1417_cast_fp16 = matmul(transpose_x = var_1417_transpose_x_0, transpose_y = var_1417_transpose_y_0, x = transpose_77, y = transpose_78)[name = tensor("op_1417_cast_fp16")]; + tensor var_1418_to_fp16 = const()[name = tensor("op_1418_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_17_cast_fp16 = mul(x = var_1417_cast_fp16, y = var_1418_to_fp16)[name = tensor("scores_17_cast_fp16")]; + tensor scores_19_cast_fp16 = add(x = scores_17_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_19_cast_fp16")]; + tensor var_1428 = const()[name = tensor("op_1428"), val = tensor(-1)]; + tensor weights_9_cast_fp16 = softmax(axis = var_1428, x = scores_19_cast_fp16)[name = tensor("weights_9_cast_fp16")]; + tensor attn_9_transpose_x_0 = const()[name = tensor("attn_9_transpose_x_0"), val = tensor(false)]; + tensor attn_9_transpose_y_0 = const()[name = tensor("attn_9_transpose_y_0"), val = tensor(false)]; + tensor values_9_cast_fp16 = transpose(perm = var_1406, x = new_v_cache_9_cast_fp16)[name = tensor("transpose_190")]; + tensor attn_9_cast_fp16 = matmul(transpose_x = attn_9_transpose_x_0, transpose_y = attn_9_transpose_y_0, x = weights_9_cast_fp16, y = values_9_cast_fp16)[name = tensor("attn_9_cast_fp16")]; + tensor var_1436 = const()[name = tensor("op_1436"), val = tensor([0, 2, 1, 3])]; + tensor var_1441 = const()[name = tensor("op_1441"), val = tensor([1, 256, 1024])]; + tensor var_1437_cast_fp16 = transpose(perm = var_1436, x = attn_9_cast_fp16)[name = tensor("transpose_187")]; + tensor input_53_cast_fp16 = reshape(shape = var_1441, x = var_1437_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor attn4_out_proj_weight_to_fp16 = const()[name = tensor("attn4_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(107012864)))]; + tensor linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn4_out_proj_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor position4_to_fp16_dtype_0 = const()[name = tensor("position4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position4_to_fp16 = cast(dtype = position4_to_fp16_dtype_0, x = position4)[name = tensor("cast_437")]; + tensor var_1450_cast_fp16 = add(x = position4_to_fp16, y = valid_len_to_fp16)[name = tensor("op_1450_cast_fp16")]; + tensor var_1450_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1450_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_55_cast_fp16 = add(x = input_49_cast_fp16, y = linear_17_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor input_57_axes_0 = const()[name = tensor("input_57_axes_0"), val = tensor([-1])]; + tensor norm4_2_weight_to_fp16 = const()[name = tensor("norm4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109110080)))]; + tensor norm4_2_bias_to_fp16 = const()[name = tensor("norm4_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109112192)))]; + tensor var_1454_to_fp16 = const()[name = tensor("op_1454_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_57_cast_fp16 = layer_norm(axes = input_57_axes_0, beta = norm4_2_bias_to_fp16, epsilon = var_1454_to_fp16, gamma = norm4_2_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor linear4_1_weight_to_fp16 = const()[name = tensor("linear4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109114304)))]; + tensor linear_18_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear4_1_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor input_59_mode_0 = const()[name = tensor("input_59_mode_0"), val = tensor("EXACT")]; + tensor input_59_cast_fp16 = gelu(mode = input_59_mode_0, x = linear_18_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor linear4_2_weight_to_fp16 = const()[name = tensor("linear4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(117502976)))]; + tensor linear_19_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear4_2_weight_to_fp16, x = input_59_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_61_cast_fp16 = add(x = input_55_cast_fp16, y = linear_19_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor input_63_axes_0 = const()[name = tensor("input_63_axes_0"), val = tensor([-1])]; + tensor norm5_1_weight_to_fp16 = const()[name = tensor("norm5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125891648)))]; + tensor norm5_1_bias_to_fp16 = const()[name = tensor("norm5_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125893760)))]; + tensor var_1471_to_fp16 = const()[name = tensor("op_1471_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_63_cast_fp16 = layer_norm(axes = input_63_axes_0, beta = norm5_1_bias_to_fp16, epsilon = var_1471_to_fp16, gamma = norm5_1_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor attn5_in_proj_weight_to_fp16 = const()[name = tensor("attn5_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125895872)))]; + tensor linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn5_in_proj_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor var_1485 = const()[name = tensor("op_1485"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_11_cast_fp16 = reshape(shape = var_1485, x = linear_20_cast_fp16)[name = tensor("qkv_11_cast_fp16")]; + tensor x_21_begin_0 = const()[name = tensor("x_21_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_21_end_0 = const()[name = tensor("x_21_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_21_end_mask_0 = const()[name = tensor("x_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_21_squeeze_mask_0 = const()[name = tensor("x_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_21_cast_fp16 = slice_by_index(begin = x_21_begin_0, end = x_21_end_0, end_mask = x_21_end_mask_0, squeeze_mask = x_21_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("x_21_cast_fp16")]; + tensor x_23_begin_0 = const()[name = tensor("x_23_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_23_end_0 = const()[name = tensor("x_23_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_23_end_mask_0 = const()[name = tensor("x_23_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_23_squeeze_mask_0 = const()[name = tensor("x_23_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_23_cast_fp16 = slice_by_index(begin = x_23_begin_0, end = x_23_end_0, end_mask = x_23_end_mask_0, squeeze_mask = x_23_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("x_23_cast_fp16")]; + tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_11_cast_fp16 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("v_11_cast_fp16")]; + tensor var_1530 = const()[name = tensor("op_1530"), val = tensor([256, 16, 32, 2])]; + tensor pairs_21_cast_fp16 = reshape(shape = var_1530, x = x_21_cast_fp16)[name = tensor("pairs_21_cast_fp16")]; + tensor xr_21_begin_0 = const()[name = tensor("xr_21_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_21_end_0 = const()[name = tensor("xr_21_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_21_end_mask_0 = const()[name = tensor("xr_21_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_21_squeeze_mask_0 = const()[name = tensor("xr_21_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_21_cast_fp16 = slice_by_index(begin = xr_21_begin_0, end = xr_21_end_0, end_mask = xr_21_end_mask_0, squeeze_mask = xr_21_squeeze_mask_0, x = pairs_21_cast_fp16)[name = tensor("xr_21_cast_fp16")]; + tensor xi_21_begin_0 = const()[name = tensor("xi_21_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_21_end_0 = const()[name = tensor("xi_21_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_21_end_mask_0 = const()[name = tensor("xi_21_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_21_squeeze_mask_0 = const()[name = tensor("xi_21_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_21_cast_fp16 = slice_by_index(begin = xi_21_begin_0, end = xi_21_end_0, end_mask = xi_21_end_mask_0, squeeze_mask = xi_21_squeeze_mask_0, x = pairs_21_cast_fp16)[name = tensor("xi_21_cast_fp16")]; + tensor var_1538_cast_fp16 = mul(x = xr_21_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1538_cast_fp16")]; + tensor var_1539_cast_fp16 = mul(x = xi_21_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1539_cast_fp16")]; + tensor xor__21_cast_fp16 = sub(x = var_1538_cast_fp16, y = var_1539_cast_fp16)[name = tensor("xor__21_cast_fp16")]; + tensor var_1542_cast_fp16 = mul(x = xr_21_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1542_cast_fp16")]; + tensor var_1543_cast_fp16 = mul(x = xi_21_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1543_cast_fp16")]; + tensor xoi_21_cast_fp16 = add(x = var_1542_cast_fp16, y = var_1543_cast_fp16)[name = tensor("xoi_21_cast_fp16")]; + tensor out_21_axis_0 = const()[name = tensor("out_21_axis_0"), val = tensor(-1)]; + tensor out_21_cast_fp16 = stack(axis = out_21_axis_0, values = (xor__21_cast_fp16, xoi_21_cast_fp16))[name = tensor("out_21_cast_fp16")]; + tensor var_1553 = const()[name = tensor("op_1553"), val = tensor([1, 256, 16, 64])]; + tensor q_11_cast_fp16 = reshape(shape = var_1553, x = out_21_cast_fp16)[name = tensor("q_11_cast_fp16")]; + tensor var_1559 = const()[name = tensor("op_1559"), val = tensor([256, 16, 32, 2])]; + tensor pairs_23_cast_fp16 = reshape(shape = var_1559, x = x_23_cast_fp16)[name = tensor("pairs_23_cast_fp16")]; + tensor xr_23_begin_0 = const()[name = tensor("xr_23_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_23_end_0 = const()[name = tensor("xr_23_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_23_end_mask_0 = const()[name = tensor("xr_23_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_23_squeeze_mask_0 = const()[name = tensor("xr_23_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_23_cast_fp16 = slice_by_index(begin = xr_23_begin_0, end = xr_23_end_0, end_mask = xr_23_end_mask_0, squeeze_mask = xr_23_squeeze_mask_0, x = pairs_23_cast_fp16)[name = tensor("xr_23_cast_fp16")]; + tensor xi_23_begin_0 = const()[name = tensor("xi_23_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_23_end_0 = const()[name = tensor("xi_23_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_23_end_mask_0 = const()[name = tensor("xi_23_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_23_squeeze_mask_0 = const()[name = tensor("xi_23_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_23_cast_fp16 = slice_by_index(begin = xi_23_begin_0, end = xi_23_end_0, end_mask = xi_23_end_mask_0, squeeze_mask = xi_23_squeeze_mask_0, x = pairs_23_cast_fp16)[name = tensor("xi_23_cast_fp16")]; + tensor var_1567_cast_fp16 = mul(x = xr_23_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1567_cast_fp16")]; + tensor var_1568_cast_fp16 = mul(x = xi_23_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1568_cast_fp16")]; + tensor xor__23_cast_fp16 = sub(x = var_1567_cast_fp16, y = var_1568_cast_fp16)[name = tensor("xor__23_cast_fp16")]; + tensor var_1571_cast_fp16 = mul(x = xr_23_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1571_cast_fp16")]; + tensor var_1572_cast_fp16 = mul(x = xi_23_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1572_cast_fp16")]; + tensor xoi_23_cast_fp16 = add(x = var_1571_cast_fp16, y = var_1572_cast_fp16)[name = tensor("xoi_23_cast_fp16")]; + tensor out_23_axis_0 = const()[name = tensor("out_23_axis_0"), val = tensor(-1)]; + tensor out_23_cast_fp16 = stack(axis = out_23_axis_0, values = (xor__23_cast_fp16, xoi_23_cast_fp16))[name = tensor("out_23_cast_fp16")]; + tensor var_1589 = const()[name = tensor("op_1589"), val = tensor([256, 1024])]; + tensor var_1590_cast_fp16 = reshape(shape = var_1589, x = out_23_cast_fp16)[name = tensor("op_1590_cast_fp16")]; + tensor written_k_11_transpose_x_1 = const()[name = tensor("written_k_11_transpose_x_1"), val = tensor(true)]; + tensor written_k_11_transpose_y_1 = const()[name = tensor("written_k_11_transpose_y_1"), val = tensor(false)]; + tensor written_k_11_cast_fp16 = matmul(transpose_x = written_k_11_transpose_x_1, transpose_y = written_k_11_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_1590_cast_fp16)[name = tensor("written_k_11_cast_fp16")]; + tensor var_1597 = const()[name = tensor("op_1597"), val = tensor([256, 1024])]; + tensor var_1598_cast_fp16 = reshape(shape = var_1597, x = v_11_cast_fp16)[name = tensor("op_1598_cast_fp16")]; + tensor written_v_11_transpose_x_1 = const()[name = tensor("written_v_11_transpose_x_1"), val = tensor(true)]; + tensor written_v_11_transpose_y_1 = const()[name = tensor("written_v_11_transpose_y_1"), val = tensor(false)]; + tensor written_v_11_cast_fp16 = matmul(transpose_x = written_v_11_transpose_x_1, transpose_y = written_v_11_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_1598_cast_fp16)[name = tensor("written_v_11_cast_fp16")]; + tensor k_cache5_to_fp16_dtype_0 = const()[name = tensor("k_cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache5_to_fp16 = cast(dtype = k_cache5_to_fp16_dtype_0, x = k_cache5)[name = tensor("cast_435")]; + tensor var_1609_cast_fp16 = mul(x = k_cache5_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_1609_cast_fp16")]; + tensor var_1614 = const()[name = tensor("op_1614"), val = tensor([1, 512, 16, 64])]; + tensor var_1615_cast_fp16 = reshape(shape = var_1614, x = written_k_11_cast_fp16)[name = tensor("op_1615_cast_fp16")]; + tensor new_k_cache_11_cast_fp16 = add(x = var_1609_cast_fp16, y = var_1615_cast_fp16)[name = tensor("new_k_cache_11_cast_fp16")]; + tensor new_k_cache_11_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_11_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache5_to_fp16_dtype_0 = const()[name = tensor("v_cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache5_to_fp16 = cast(dtype = v_cache5_to_fp16_dtype_0, x = v_cache5)[name = tensor("cast_433")]; + tensor var_1618_cast_fp16 = mul(x = v_cache5_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_1618_cast_fp16")]; + tensor var_1623 = const()[name = tensor("op_1623"), val = tensor([1, 512, 16, 64])]; + tensor var_1624_cast_fp16 = reshape(shape = var_1623, x = written_v_11_cast_fp16)[name = tensor("op_1624_cast_fp16")]; + tensor new_v_cache_11_cast_fp16 = add(x = var_1618_cast_fp16, y = var_1624_cast_fp16)[name = tensor("new_v_cache_11_cast_fp16")]; + tensor new_v_cache_11_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_11_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_1637 = const()[name = tensor("op_1637"), val = tensor([0, 2, 1, 3])]; + tensor var_1648_transpose_x_0 = const()[name = tensor("op_1648_transpose_x_0"), val = tensor(false)]; + tensor var_1648_transpose_y_0 = const()[name = tensor("op_1648_transpose_y_0"), val = tensor(false)]; + tensor transpose_79_perm_0 = const()[name = tensor("transpose_79_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_80_perm_0 = const()[name = tensor("transpose_80_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_80 = transpose(perm = transpose_80_perm_0, x = new_k_cache_11_cast_fp16)[name = tensor("transpose_184")]; + tensor transpose_79 = transpose(perm = transpose_79_perm_0, x = q_11_cast_fp16)[name = tensor("transpose_185")]; + tensor var_1648_cast_fp16 = matmul(transpose_x = var_1648_transpose_x_0, transpose_y = var_1648_transpose_y_0, x = transpose_79, y = transpose_80)[name = tensor("op_1648_cast_fp16")]; + tensor var_1649_to_fp16 = const()[name = tensor("op_1649_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_21_cast_fp16 = mul(x = var_1648_cast_fp16, y = var_1649_to_fp16)[name = tensor("scores_21_cast_fp16")]; + tensor scores_23_cast_fp16 = add(x = scores_21_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_23_cast_fp16")]; + tensor var_1659 = const()[name = tensor("op_1659"), val = tensor(-1)]; + tensor weights_11_cast_fp16 = softmax(axis = var_1659, x = scores_23_cast_fp16)[name = tensor("weights_11_cast_fp16")]; + tensor attn_11_transpose_x_0 = const()[name = tensor("attn_11_transpose_x_0"), val = tensor(false)]; + tensor attn_11_transpose_y_0 = const()[name = tensor("attn_11_transpose_y_0"), val = tensor(false)]; + tensor values_11_cast_fp16 = transpose(perm = var_1637, x = new_v_cache_11_cast_fp16)[name = tensor("transpose_186")]; + tensor attn_11_cast_fp16 = matmul(transpose_x = attn_11_transpose_x_0, transpose_y = attn_11_transpose_y_0, x = weights_11_cast_fp16, y = values_11_cast_fp16)[name = tensor("attn_11_cast_fp16")]; + tensor var_1667 = const()[name = tensor("op_1667"), val = tensor([0, 2, 1, 3])]; + tensor var_1672 = const()[name = tensor("op_1672"), val = tensor([1, 256, 1024])]; + tensor var_1668_cast_fp16 = transpose(perm = var_1667, x = attn_11_cast_fp16)[name = tensor("transpose_183")]; + tensor input_65_cast_fp16 = reshape(shape = var_1672, x = var_1668_cast_fp16)[name = tensor("input_65_cast_fp16")]; + tensor attn5_out_proj_weight_to_fp16 = const()[name = tensor("attn5_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132187392)))]; + tensor linear_21_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn5_out_proj_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("linear_21_cast_fp16")]; + tensor position5_to_fp16_dtype_0 = const()[name = tensor("position5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position5_to_fp16 = cast(dtype = position5_to_fp16_dtype_0, x = position5)[name = tensor("cast_431")]; + tensor var_1681_cast_fp16 = add(x = position5_to_fp16, y = valid_len_to_fp16)[name = tensor("op_1681_cast_fp16")]; + tensor var_1681_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1681_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_67_cast_fp16 = add(x = input_61_cast_fp16, y = linear_21_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor input_69_axes_0 = const()[name = tensor("input_69_axes_0"), val = tensor([-1])]; + tensor norm5_2_weight_to_fp16 = const()[name = tensor("norm5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134284608)))]; + tensor norm5_2_bias_to_fp16 = const()[name = tensor("norm5_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134286720)))]; + tensor var_1685_to_fp16 = const()[name = tensor("op_1685_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_69_cast_fp16 = layer_norm(axes = input_69_axes_0, beta = norm5_2_bias_to_fp16, epsilon = var_1685_to_fp16, gamma = norm5_2_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("input_69_cast_fp16")]; + tensor linear5_1_weight_to_fp16 = const()[name = tensor("linear5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134288832)))]; + tensor linear_22_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear5_1_weight_to_fp16, x = input_69_cast_fp16)[name = tensor("linear_22_cast_fp16")]; + tensor input_71_mode_0 = const()[name = tensor("input_71_mode_0"), val = tensor("EXACT")]; + tensor input_71_cast_fp16 = gelu(mode = input_71_mode_0, x = linear_22_cast_fp16)[name = tensor("input_71_cast_fp16")]; + tensor linear5_2_weight_to_fp16 = const()[name = tensor("linear5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142677504)))]; + tensor linear_23_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear5_2_weight_to_fp16, x = input_71_cast_fp16)[name = tensor("linear_23_cast_fp16")]; + tensor input_73_cast_fp16 = add(x = input_67_cast_fp16, y = linear_23_cast_fp16)[name = tensor("input_73_cast_fp16")]; + tensor input_75_axes_0 = const()[name = tensor("input_75_axes_0"), val = tensor([-1])]; + tensor norm6_1_weight_to_fp16 = const()[name = tensor("norm6_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151066176)))]; + tensor norm6_1_bias_to_fp16 = const()[name = tensor("norm6_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151068288)))]; + tensor var_1702_to_fp16 = const()[name = tensor("op_1702_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_75_cast_fp16 = layer_norm(axes = input_75_axes_0, beta = norm6_1_bias_to_fp16, epsilon = var_1702_to_fp16, gamma = norm6_1_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("input_75_cast_fp16")]; + tensor attn6_in_proj_weight_to_fp16 = const()[name = tensor("attn6_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151070400)))]; + tensor linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn6_in_proj_weight_to_fp16, x = input_75_cast_fp16)[name = tensor("linear_24_cast_fp16")]; + tensor var_1716 = const()[name = tensor("op_1716"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_13_cast_fp16 = reshape(shape = var_1716, x = linear_24_cast_fp16)[name = tensor("qkv_13_cast_fp16")]; + tensor x_25_begin_0 = const()[name = tensor("x_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_25_end_0 = const()[name = tensor("x_25_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_25_end_mask_0 = const()[name = tensor("x_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_25_squeeze_mask_0 = const()[name = tensor("x_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_25_cast_fp16 = slice_by_index(begin = x_25_begin_0, end = x_25_end_0, end_mask = x_25_end_mask_0, squeeze_mask = x_25_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("x_25_cast_fp16")]; + tensor x_27_begin_0 = const()[name = tensor("x_27_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_27_end_0 = const()[name = tensor("x_27_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_27_end_mask_0 = const()[name = tensor("x_27_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_27_squeeze_mask_0 = const()[name = tensor("x_27_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_27_cast_fp16 = slice_by_index(begin = x_27_begin_0, end = x_27_end_0, end_mask = x_27_end_mask_0, squeeze_mask = x_27_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("x_27_cast_fp16")]; + tensor v_13_begin_0 = const()[name = tensor("v_13_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_13_end_0 = const()[name = tensor("v_13_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_13_end_mask_0 = const()[name = tensor("v_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_13_squeeze_mask_0 = const()[name = tensor("v_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_13_cast_fp16 = slice_by_index(begin = v_13_begin_0, end = v_13_end_0, end_mask = v_13_end_mask_0, squeeze_mask = v_13_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("v_13_cast_fp16")]; + tensor var_1761 = const()[name = tensor("op_1761"), val = tensor([256, 16, 32, 2])]; + tensor pairs_25_cast_fp16 = reshape(shape = var_1761, x = x_25_cast_fp16)[name = tensor("pairs_25_cast_fp16")]; + tensor xr_25_begin_0 = const()[name = tensor("xr_25_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_25_end_0 = const()[name = tensor("xr_25_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_25_end_mask_0 = const()[name = tensor("xr_25_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_25_squeeze_mask_0 = const()[name = tensor("xr_25_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_25_cast_fp16 = slice_by_index(begin = xr_25_begin_0, end = xr_25_end_0, end_mask = xr_25_end_mask_0, squeeze_mask = xr_25_squeeze_mask_0, x = pairs_25_cast_fp16)[name = tensor("xr_25_cast_fp16")]; + tensor xi_25_begin_0 = const()[name = tensor("xi_25_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_25_end_0 = const()[name = tensor("xi_25_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_25_end_mask_0 = const()[name = tensor("xi_25_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_25_squeeze_mask_0 = const()[name = tensor("xi_25_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_25_cast_fp16 = slice_by_index(begin = xi_25_begin_0, end = xi_25_end_0, end_mask = xi_25_end_mask_0, squeeze_mask = xi_25_squeeze_mask_0, x = pairs_25_cast_fp16)[name = tensor("xi_25_cast_fp16")]; + tensor var_1769_cast_fp16 = mul(x = xr_25_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1769_cast_fp16")]; + tensor var_1770_cast_fp16 = mul(x = xi_25_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1770_cast_fp16")]; + tensor xor__25_cast_fp16 = sub(x = var_1769_cast_fp16, y = var_1770_cast_fp16)[name = tensor("xor__25_cast_fp16")]; + tensor var_1773_cast_fp16 = mul(x = xr_25_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1773_cast_fp16")]; + tensor var_1774_cast_fp16 = mul(x = xi_25_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1774_cast_fp16")]; + tensor xoi_25_cast_fp16 = add(x = var_1773_cast_fp16, y = var_1774_cast_fp16)[name = tensor("xoi_25_cast_fp16")]; + tensor out_25_axis_0 = const()[name = tensor("out_25_axis_0"), val = tensor(-1)]; + tensor out_25_cast_fp16 = stack(axis = out_25_axis_0, values = (xor__25_cast_fp16, xoi_25_cast_fp16))[name = tensor("out_25_cast_fp16")]; + tensor var_1784 = const()[name = tensor("op_1784"), val = tensor([1, 256, 16, 64])]; + tensor q_13_cast_fp16 = reshape(shape = var_1784, x = out_25_cast_fp16)[name = tensor("q_13_cast_fp16")]; + tensor var_1790 = const()[name = tensor("op_1790"), val = tensor([256, 16, 32, 2])]; + tensor pairs_27_cast_fp16 = reshape(shape = var_1790, x = x_27_cast_fp16)[name = tensor("pairs_27_cast_fp16")]; + tensor xr_27_begin_0 = const()[name = tensor("xr_27_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_27_end_0 = const()[name = tensor("xr_27_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_27_end_mask_0 = const()[name = tensor("xr_27_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_27_squeeze_mask_0 = const()[name = tensor("xr_27_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_27_cast_fp16 = slice_by_index(begin = xr_27_begin_0, end = xr_27_end_0, end_mask = xr_27_end_mask_0, squeeze_mask = xr_27_squeeze_mask_0, x = pairs_27_cast_fp16)[name = tensor("xr_27_cast_fp16")]; + tensor xi_27_begin_0 = const()[name = tensor("xi_27_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_27_end_0 = const()[name = tensor("xi_27_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_27_end_mask_0 = const()[name = tensor("xi_27_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_27_squeeze_mask_0 = const()[name = tensor("xi_27_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_27_cast_fp16 = slice_by_index(begin = xi_27_begin_0, end = xi_27_end_0, end_mask = xi_27_end_mask_0, squeeze_mask = xi_27_squeeze_mask_0, x = pairs_27_cast_fp16)[name = tensor("xi_27_cast_fp16")]; + tensor var_1798_cast_fp16 = mul(x = xr_27_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1798_cast_fp16")]; + tensor var_1799_cast_fp16 = mul(x = xi_27_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1799_cast_fp16")]; + tensor xor__27_cast_fp16 = sub(x = var_1798_cast_fp16, y = var_1799_cast_fp16)[name = tensor("xor__27_cast_fp16")]; + tensor var_1802_cast_fp16 = mul(x = xr_27_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1802_cast_fp16")]; + tensor var_1803_cast_fp16 = mul(x = xi_27_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1803_cast_fp16")]; + tensor xoi_27_cast_fp16 = add(x = var_1802_cast_fp16, y = var_1803_cast_fp16)[name = tensor("xoi_27_cast_fp16")]; + tensor out_27_axis_0 = const()[name = tensor("out_27_axis_0"), val = tensor(-1)]; + tensor out_27_cast_fp16 = stack(axis = out_27_axis_0, values = (xor__27_cast_fp16, xoi_27_cast_fp16))[name = tensor("out_27_cast_fp16")]; + tensor var_1820 = const()[name = tensor("op_1820"), val = tensor([256, 1024])]; + tensor var_1821_cast_fp16 = reshape(shape = var_1820, x = out_27_cast_fp16)[name = tensor("op_1821_cast_fp16")]; + tensor written_k_13_transpose_x_1 = const()[name = tensor("written_k_13_transpose_x_1"), val = tensor(true)]; + tensor written_k_13_transpose_y_1 = const()[name = tensor("written_k_13_transpose_y_1"), val = tensor(false)]; + tensor written_k_13_cast_fp16 = matmul(transpose_x = written_k_13_transpose_x_1, transpose_y = written_k_13_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_1821_cast_fp16)[name = tensor("written_k_13_cast_fp16")]; + tensor var_1828 = const()[name = tensor("op_1828"), val = tensor([256, 1024])]; + tensor var_1829_cast_fp16 = reshape(shape = var_1828, x = v_13_cast_fp16)[name = tensor("op_1829_cast_fp16")]; + tensor written_v_13_transpose_x_1 = const()[name = tensor("written_v_13_transpose_x_1"), val = tensor(true)]; + tensor written_v_13_transpose_y_1 = const()[name = tensor("written_v_13_transpose_y_1"), val = tensor(false)]; + tensor written_v_13_cast_fp16 = matmul(transpose_x = written_v_13_transpose_x_1, transpose_y = written_v_13_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_1829_cast_fp16)[name = tensor("written_v_13_cast_fp16")]; + tensor k_cache6_to_fp16_dtype_0 = const()[name = tensor("k_cache6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache6_to_fp16 = cast(dtype = k_cache6_to_fp16_dtype_0, x = k_cache6)[name = tensor("cast_429")]; + tensor var_1840_cast_fp16 = mul(x = k_cache6_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_1840_cast_fp16")]; + tensor var_1845 = const()[name = tensor("op_1845"), val = tensor([1, 512, 16, 64])]; + tensor var_1846_cast_fp16 = reshape(shape = var_1845, x = written_k_13_cast_fp16)[name = tensor("op_1846_cast_fp16")]; + tensor new_k_cache_13_cast_fp16 = add(x = var_1840_cast_fp16, y = var_1846_cast_fp16)[name = tensor("new_k_cache_13_cast_fp16")]; + tensor new_k_cache_13_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_13_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache6_to_fp16_dtype_0 = const()[name = tensor("v_cache6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache6_to_fp16 = cast(dtype = v_cache6_to_fp16_dtype_0, x = v_cache6)[name = tensor("cast_427")]; + tensor var_1849_cast_fp16 = mul(x = v_cache6_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_1849_cast_fp16")]; + tensor var_1854 = const()[name = tensor("op_1854"), val = tensor([1, 512, 16, 64])]; + tensor var_1855_cast_fp16 = reshape(shape = var_1854, x = written_v_13_cast_fp16)[name = tensor("op_1855_cast_fp16")]; + tensor new_v_cache_13_cast_fp16 = add(x = var_1849_cast_fp16, y = var_1855_cast_fp16)[name = tensor("new_v_cache_13_cast_fp16")]; + tensor new_v_cache_13_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_13_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_1868 = const()[name = tensor("op_1868"), val = tensor([0, 2, 1, 3])]; + tensor var_1879_transpose_x_0 = const()[name = tensor("op_1879_transpose_x_0"), val = tensor(false)]; + tensor var_1879_transpose_y_0 = const()[name = tensor("op_1879_transpose_y_0"), val = tensor(false)]; + tensor transpose_81_perm_0 = const()[name = tensor("transpose_81_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_82_perm_0 = const()[name = tensor("transpose_82_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_82 = transpose(perm = transpose_82_perm_0, x = new_k_cache_13_cast_fp16)[name = tensor("transpose_180")]; + tensor transpose_81 = transpose(perm = transpose_81_perm_0, x = q_13_cast_fp16)[name = tensor("transpose_181")]; + tensor var_1879_cast_fp16 = matmul(transpose_x = var_1879_transpose_x_0, transpose_y = var_1879_transpose_y_0, x = transpose_81, y = transpose_82)[name = tensor("op_1879_cast_fp16")]; + tensor var_1880_to_fp16 = const()[name = tensor("op_1880_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_25_cast_fp16 = mul(x = var_1879_cast_fp16, y = var_1880_to_fp16)[name = tensor("scores_25_cast_fp16")]; + tensor scores_27_cast_fp16 = add(x = scores_25_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_27_cast_fp16")]; + tensor var_1890 = const()[name = tensor("op_1890"), val = tensor(-1)]; + tensor weights_13_cast_fp16 = softmax(axis = var_1890, x = scores_27_cast_fp16)[name = tensor("weights_13_cast_fp16")]; + tensor attn_13_transpose_x_0 = const()[name = tensor("attn_13_transpose_x_0"), val = tensor(false)]; + tensor attn_13_transpose_y_0 = const()[name = tensor("attn_13_transpose_y_0"), val = tensor(false)]; + tensor values_13_cast_fp16 = transpose(perm = var_1868, x = new_v_cache_13_cast_fp16)[name = tensor("transpose_182")]; + tensor attn_13_cast_fp16 = matmul(transpose_x = attn_13_transpose_x_0, transpose_y = attn_13_transpose_y_0, x = weights_13_cast_fp16, y = values_13_cast_fp16)[name = tensor("attn_13_cast_fp16")]; + tensor var_1898 = const()[name = tensor("op_1898"), val = tensor([0, 2, 1, 3])]; + tensor var_1903 = const()[name = tensor("op_1903"), val = tensor([1, 256, 1024])]; + tensor var_1899_cast_fp16 = transpose(perm = var_1898, x = attn_13_cast_fp16)[name = tensor("transpose_179")]; + tensor input_77_cast_fp16 = reshape(shape = var_1903, x = var_1899_cast_fp16)[name = tensor("input_77_cast_fp16")]; + tensor attn6_out_proj_weight_to_fp16 = const()[name = tensor("attn6_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(157361920)))]; + tensor linear_25_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn6_out_proj_weight_to_fp16, x = input_77_cast_fp16)[name = tensor("linear_25_cast_fp16")]; + tensor position6_to_fp16_dtype_0 = const()[name = tensor("position6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position6_to_fp16 = cast(dtype = position6_to_fp16_dtype_0, x = position6)[name = tensor("cast_425")]; + tensor var_1912_cast_fp16 = add(x = position6_to_fp16, y = valid_len_to_fp16)[name = tensor("op_1912_cast_fp16")]; + tensor var_1912_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1912_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_79_cast_fp16 = add(x = input_73_cast_fp16, y = linear_25_cast_fp16)[name = tensor("input_79_cast_fp16")]; + tensor input_81_axes_0 = const()[name = tensor("input_81_axes_0"), val = tensor([-1])]; + tensor norm6_2_weight_to_fp16 = const()[name = tensor("norm6_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159459136)))]; + tensor norm6_2_bias_to_fp16 = const()[name = tensor("norm6_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159461248)))]; + tensor var_1916_to_fp16 = const()[name = tensor("op_1916_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_81_cast_fp16 = layer_norm(axes = input_81_axes_0, beta = norm6_2_bias_to_fp16, epsilon = var_1916_to_fp16, gamma = norm6_2_weight_to_fp16, x = input_79_cast_fp16)[name = tensor("input_81_cast_fp16")]; + tensor linear6_1_weight_to_fp16 = const()[name = tensor("linear6_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159463360)))]; + tensor linear_26_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear6_1_weight_to_fp16, x = input_81_cast_fp16)[name = tensor("linear_26_cast_fp16")]; + tensor input_83_mode_0 = const()[name = tensor("input_83_mode_0"), val = tensor("EXACT")]; + tensor input_83_cast_fp16 = gelu(mode = input_83_mode_0, x = linear_26_cast_fp16)[name = tensor("input_83_cast_fp16")]; + tensor linear6_2_weight_to_fp16 = const()[name = tensor("linear6_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(167852032)))]; + tensor linear_27_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear6_2_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_27_cast_fp16")]; + tensor input_85_cast_fp16 = add(x = input_79_cast_fp16, y = linear_27_cast_fp16)[name = tensor("input_85_cast_fp16")]; + tensor input_87_axes_0 = const()[name = tensor("input_87_axes_0"), val = tensor([-1])]; + tensor norm7_1_weight_to_fp16 = const()[name = tensor("norm7_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176240704)))]; + tensor norm7_1_bias_to_fp16 = const()[name = tensor("norm7_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176242816)))]; + tensor var_1933_to_fp16 = const()[name = tensor("op_1933_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_87_cast_fp16 = layer_norm(axes = input_87_axes_0, beta = norm7_1_bias_to_fp16, epsilon = var_1933_to_fp16, gamma = norm7_1_weight_to_fp16, x = input_85_cast_fp16)[name = tensor("input_87_cast_fp16")]; + tensor attn7_in_proj_weight_to_fp16 = const()[name = tensor("attn7_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176244928)))]; + tensor linear_28_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn7_in_proj_weight_to_fp16, x = input_87_cast_fp16)[name = tensor("linear_28_cast_fp16")]; + tensor var_1947 = const()[name = tensor("op_1947"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_15_cast_fp16 = reshape(shape = var_1947, x = linear_28_cast_fp16)[name = tensor("qkv_15_cast_fp16")]; + tensor x_29_begin_0 = const()[name = tensor("x_29_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_29_end_0 = const()[name = tensor("x_29_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_29_end_mask_0 = const()[name = tensor("x_29_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_29_squeeze_mask_0 = const()[name = tensor("x_29_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_29_cast_fp16 = slice_by_index(begin = x_29_begin_0, end = x_29_end_0, end_mask = x_29_end_mask_0, squeeze_mask = x_29_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("x_29_cast_fp16")]; + tensor x_31_begin_0 = const()[name = tensor("x_31_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_31_end_0 = const()[name = tensor("x_31_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_31_end_mask_0 = const()[name = tensor("x_31_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_31_squeeze_mask_0 = const()[name = tensor("x_31_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_31_cast_fp16 = slice_by_index(begin = x_31_begin_0, end = x_31_end_0, end_mask = x_31_end_mask_0, squeeze_mask = x_31_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("x_31_cast_fp16")]; + tensor v_15_begin_0 = const()[name = tensor("v_15_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_15_end_0 = const()[name = tensor("v_15_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_15_end_mask_0 = const()[name = tensor("v_15_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_15_squeeze_mask_0 = const()[name = tensor("v_15_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_15_cast_fp16 = slice_by_index(begin = v_15_begin_0, end = v_15_end_0, end_mask = v_15_end_mask_0, squeeze_mask = v_15_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("v_15_cast_fp16")]; + tensor var_1992 = const()[name = tensor("op_1992"), val = tensor([256, 16, 32, 2])]; + tensor pairs_29_cast_fp16 = reshape(shape = var_1992, x = x_29_cast_fp16)[name = tensor("pairs_29_cast_fp16")]; + tensor xr_29_begin_0 = const()[name = tensor("xr_29_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_29_end_0 = const()[name = tensor("xr_29_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_29_end_mask_0 = const()[name = tensor("xr_29_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_29_squeeze_mask_0 = const()[name = tensor("xr_29_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_29_cast_fp16 = slice_by_index(begin = xr_29_begin_0, end = xr_29_end_0, end_mask = xr_29_end_mask_0, squeeze_mask = xr_29_squeeze_mask_0, x = pairs_29_cast_fp16)[name = tensor("xr_29_cast_fp16")]; + tensor xi_29_begin_0 = const()[name = tensor("xi_29_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_29_end_0 = const()[name = tensor("xi_29_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_29_end_mask_0 = const()[name = tensor("xi_29_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_29_squeeze_mask_0 = const()[name = tensor("xi_29_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_29_cast_fp16 = slice_by_index(begin = xi_29_begin_0, end = xi_29_end_0, end_mask = xi_29_end_mask_0, squeeze_mask = xi_29_squeeze_mask_0, x = pairs_29_cast_fp16)[name = tensor("xi_29_cast_fp16")]; + tensor var_2000_cast_fp16 = mul(x = xr_29_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2000_cast_fp16")]; + tensor var_2001_cast_fp16 = mul(x = xi_29_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2001_cast_fp16")]; + tensor xor__29_cast_fp16 = sub(x = var_2000_cast_fp16, y = var_2001_cast_fp16)[name = tensor("xor__29_cast_fp16")]; + tensor var_2004_cast_fp16 = mul(x = xr_29_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2004_cast_fp16")]; + tensor var_2005_cast_fp16 = mul(x = xi_29_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2005_cast_fp16")]; + tensor xoi_29_cast_fp16 = add(x = var_2004_cast_fp16, y = var_2005_cast_fp16)[name = tensor("xoi_29_cast_fp16")]; + tensor out_29_axis_0 = const()[name = tensor("out_29_axis_0"), val = tensor(-1)]; + tensor out_29_cast_fp16 = stack(axis = out_29_axis_0, values = (xor__29_cast_fp16, xoi_29_cast_fp16))[name = tensor("out_29_cast_fp16")]; + tensor var_2015 = const()[name = tensor("op_2015"), val = tensor([1, 256, 16, 64])]; + tensor q_15_cast_fp16 = reshape(shape = var_2015, x = out_29_cast_fp16)[name = tensor("q_15_cast_fp16")]; + tensor var_2021 = const()[name = tensor("op_2021"), val = tensor([256, 16, 32, 2])]; + tensor pairs_31_cast_fp16 = reshape(shape = var_2021, x = x_31_cast_fp16)[name = tensor("pairs_31_cast_fp16")]; + tensor xr_31_begin_0 = const()[name = tensor("xr_31_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_31_end_0 = const()[name = tensor("xr_31_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_31_end_mask_0 = const()[name = tensor("xr_31_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_31_squeeze_mask_0 = const()[name = tensor("xr_31_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_31_cast_fp16 = slice_by_index(begin = xr_31_begin_0, end = xr_31_end_0, end_mask = xr_31_end_mask_0, squeeze_mask = xr_31_squeeze_mask_0, x = pairs_31_cast_fp16)[name = tensor("xr_31_cast_fp16")]; + tensor xi_31_begin_0 = const()[name = tensor("xi_31_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_31_end_0 = const()[name = tensor("xi_31_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_31_end_mask_0 = const()[name = tensor("xi_31_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_31_squeeze_mask_0 = const()[name = tensor("xi_31_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_31_cast_fp16 = slice_by_index(begin = xi_31_begin_0, end = xi_31_end_0, end_mask = xi_31_end_mask_0, squeeze_mask = xi_31_squeeze_mask_0, x = pairs_31_cast_fp16)[name = tensor("xi_31_cast_fp16")]; + tensor var_2029_cast_fp16 = mul(x = xr_31_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2029_cast_fp16")]; + tensor var_2030_cast_fp16 = mul(x = xi_31_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2030_cast_fp16")]; + tensor xor__31_cast_fp16 = sub(x = var_2029_cast_fp16, y = var_2030_cast_fp16)[name = tensor("xor__31_cast_fp16")]; + tensor var_2033_cast_fp16 = mul(x = xr_31_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2033_cast_fp16")]; + tensor var_2034_cast_fp16 = mul(x = xi_31_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2034_cast_fp16")]; + tensor xoi_31_cast_fp16 = add(x = var_2033_cast_fp16, y = var_2034_cast_fp16)[name = tensor("xoi_31_cast_fp16")]; + tensor out_31_axis_0 = const()[name = tensor("out_31_axis_0"), val = tensor(-1)]; + tensor out_31_cast_fp16 = stack(axis = out_31_axis_0, values = (xor__31_cast_fp16, xoi_31_cast_fp16))[name = tensor("out_31_cast_fp16")]; + tensor var_2051 = const()[name = tensor("op_2051"), val = tensor([256, 1024])]; + tensor var_2052_cast_fp16 = reshape(shape = var_2051, x = out_31_cast_fp16)[name = tensor("op_2052_cast_fp16")]; + tensor written_k_15_transpose_x_1 = const()[name = tensor("written_k_15_transpose_x_1"), val = tensor(true)]; + tensor written_k_15_transpose_y_1 = const()[name = tensor("written_k_15_transpose_y_1"), val = tensor(false)]; + tensor written_k_15_cast_fp16 = matmul(transpose_x = written_k_15_transpose_x_1, transpose_y = written_k_15_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_2052_cast_fp16)[name = tensor("written_k_15_cast_fp16")]; + tensor var_2059 = const()[name = tensor("op_2059"), val = tensor([256, 1024])]; + tensor var_2060_cast_fp16 = reshape(shape = var_2059, x = v_15_cast_fp16)[name = tensor("op_2060_cast_fp16")]; + tensor written_v_15_transpose_x_1 = const()[name = tensor("written_v_15_transpose_x_1"), val = tensor(true)]; + tensor written_v_15_transpose_y_1 = const()[name = tensor("written_v_15_transpose_y_1"), val = tensor(false)]; + tensor written_v_15_cast_fp16 = matmul(transpose_x = written_v_15_transpose_x_1, transpose_y = written_v_15_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_2060_cast_fp16)[name = tensor("written_v_15_cast_fp16")]; + tensor k_cache7_to_fp16_dtype_0 = const()[name = tensor("k_cache7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache7_to_fp16 = cast(dtype = k_cache7_to_fp16_dtype_0, x = k_cache7)[name = tensor("cast_423")]; + tensor var_2071_cast_fp16 = mul(x = k_cache7_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_2071_cast_fp16")]; + tensor var_2076 = const()[name = tensor("op_2076"), val = tensor([1, 512, 16, 64])]; + tensor var_2077_cast_fp16 = reshape(shape = var_2076, x = written_k_15_cast_fp16)[name = tensor("op_2077_cast_fp16")]; + tensor new_k_cache_15_cast_fp16 = add(x = var_2071_cast_fp16, y = var_2077_cast_fp16)[name = tensor("new_k_cache_15_cast_fp16")]; + tensor new_k_cache_15_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_15_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache7_to_fp16_dtype_0 = const()[name = tensor("v_cache7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache7_to_fp16 = cast(dtype = v_cache7_to_fp16_dtype_0, x = v_cache7)[name = tensor("cast_421")]; + tensor var_2080_cast_fp16 = mul(x = v_cache7_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_2080_cast_fp16")]; + tensor var_2085 = const()[name = tensor("op_2085"), val = tensor([1, 512, 16, 64])]; + tensor var_2086_cast_fp16 = reshape(shape = var_2085, x = written_v_15_cast_fp16)[name = tensor("op_2086_cast_fp16")]; + tensor new_v_cache_15_cast_fp16 = add(x = var_2080_cast_fp16, y = var_2086_cast_fp16)[name = tensor("new_v_cache_15_cast_fp16")]; + tensor new_v_cache_15_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_15_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_2099 = const()[name = tensor("op_2099"), val = tensor([0, 2, 1, 3])]; + tensor var_2110_transpose_x_0 = const()[name = tensor("op_2110_transpose_x_0"), val = tensor(false)]; + tensor var_2110_transpose_y_0 = const()[name = tensor("op_2110_transpose_y_0"), val = tensor(false)]; + tensor transpose_83_perm_0 = const()[name = tensor("transpose_83_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_84_perm_0 = const()[name = tensor("transpose_84_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_84 = transpose(perm = transpose_84_perm_0, x = new_k_cache_15_cast_fp16)[name = tensor("transpose_176")]; + tensor transpose_83 = transpose(perm = transpose_83_perm_0, x = q_15_cast_fp16)[name = tensor("transpose_177")]; + tensor var_2110_cast_fp16 = matmul(transpose_x = var_2110_transpose_x_0, transpose_y = var_2110_transpose_y_0, x = transpose_83, y = transpose_84)[name = tensor("op_2110_cast_fp16")]; + tensor var_2111_to_fp16 = const()[name = tensor("op_2111_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_29_cast_fp16 = mul(x = var_2110_cast_fp16, y = var_2111_to_fp16)[name = tensor("scores_29_cast_fp16")]; + tensor scores_31_cast_fp16 = add(x = scores_29_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_31_cast_fp16")]; + tensor var_2121 = const()[name = tensor("op_2121"), val = tensor(-1)]; + tensor weights_15_cast_fp16 = softmax(axis = var_2121, x = scores_31_cast_fp16)[name = tensor("weights_15_cast_fp16")]; + tensor attn_15_transpose_x_0 = const()[name = tensor("attn_15_transpose_x_0"), val = tensor(false)]; + tensor attn_15_transpose_y_0 = const()[name = tensor("attn_15_transpose_y_0"), val = tensor(false)]; + tensor values_15_cast_fp16 = transpose(perm = var_2099, x = new_v_cache_15_cast_fp16)[name = tensor("transpose_178")]; + tensor attn_15_cast_fp16 = matmul(transpose_x = attn_15_transpose_x_0, transpose_y = attn_15_transpose_y_0, x = weights_15_cast_fp16, y = values_15_cast_fp16)[name = tensor("attn_15_cast_fp16")]; + tensor var_2129 = const()[name = tensor("op_2129"), val = tensor([0, 2, 1, 3])]; + tensor var_2134 = const()[name = tensor("op_2134"), val = tensor([1, 256, 1024])]; + tensor var_2130_cast_fp16 = transpose(perm = var_2129, x = attn_15_cast_fp16)[name = tensor("transpose_175")]; + tensor input_89_cast_fp16 = reshape(shape = var_2134, x = var_2130_cast_fp16)[name = tensor("input_89_cast_fp16")]; + tensor attn7_out_proj_weight_to_fp16 = const()[name = tensor("attn7_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(182536448)))]; + tensor linear_29_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn7_out_proj_weight_to_fp16, x = input_89_cast_fp16)[name = tensor("linear_29_cast_fp16")]; + tensor position7_to_fp16_dtype_0 = const()[name = tensor("position7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position7_to_fp16 = cast(dtype = position7_to_fp16_dtype_0, x = position7)[name = tensor("cast_419")]; + tensor var_2143_cast_fp16 = add(x = position7_to_fp16, y = valid_len_to_fp16)[name = tensor("op_2143_cast_fp16")]; + tensor var_2143_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2143_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_91_cast_fp16 = add(x = input_85_cast_fp16, y = linear_29_cast_fp16)[name = tensor("input_91_cast_fp16")]; + tensor input_93_axes_0 = const()[name = tensor("input_93_axes_0"), val = tensor([-1])]; + tensor norm7_2_weight_to_fp16 = const()[name = tensor("norm7_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184633664)))]; + tensor norm7_2_bias_to_fp16 = const()[name = tensor("norm7_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184635776)))]; + tensor var_2147_to_fp16 = const()[name = tensor("op_2147_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_93_cast_fp16 = layer_norm(axes = input_93_axes_0, beta = norm7_2_bias_to_fp16, epsilon = var_2147_to_fp16, gamma = norm7_2_weight_to_fp16, x = input_91_cast_fp16)[name = tensor("input_93_cast_fp16")]; + tensor linear7_1_weight_to_fp16 = const()[name = tensor("linear7_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184637888)))]; + tensor linear_30_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear7_1_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("linear_30_cast_fp16")]; + tensor input_95_mode_0 = const()[name = tensor("input_95_mode_0"), val = tensor("EXACT")]; + tensor input_95_cast_fp16 = gelu(mode = input_95_mode_0, x = linear_30_cast_fp16)[name = tensor("input_95_cast_fp16")]; + tensor linear7_2_weight_to_fp16 = const()[name = tensor("linear7_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(193026560)))]; + tensor linear_31_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear7_2_weight_to_fp16, x = input_95_cast_fp16)[name = tensor("linear_31_cast_fp16")]; + tensor input_97_cast_fp16 = add(x = input_91_cast_fp16, y = linear_31_cast_fp16)[name = tensor("input_97_cast_fp16")]; + tensor input_99_axes_0 = const()[name = tensor("input_99_axes_0"), val = tensor([-1])]; + tensor norm8_1_weight_to_fp16 = const()[name = tensor("norm8_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201415232)))]; + tensor norm8_1_bias_to_fp16 = const()[name = tensor("norm8_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201417344)))]; + tensor var_2164_to_fp16 = const()[name = tensor("op_2164_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_99_cast_fp16 = layer_norm(axes = input_99_axes_0, beta = norm8_1_bias_to_fp16, epsilon = var_2164_to_fp16, gamma = norm8_1_weight_to_fp16, x = input_97_cast_fp16)[name = tensor("input_99_cast_fp16")]; + tensor attn8_in_proj_weight_to_fp16 = const()[name = tensor("attn8_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201419456)))]; + tensor linear_32_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn8_in_proj_weight_to_fp16, x = input_99_cast_fp16)[name = tensor("linear_32_cast_fp16")]; + tensor var_2178 = const()[name = tensor("op_2178"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_17_cast_fp16 = reshape(shape = var_2178, x = linear_32_cast_fp16)[name = tensor("qkv_17_cast_fp16")]; + tensor x_33_begin_0 = const()[name = tensor("x_33_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_33_end_0 = const()[name = tensor("x_33_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_33_end_mask_0 = const()[name = tensor("x_33_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_33_squeeze_mask_0 = const()[name = tensor("x_33_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_33_cast_fp16 = slice_by_index(begin = x_33_begin_0, end = x_33_end_0, end_mask = x_33_end_mask_0, squeeze_mask = x_33_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("x_33_cast_fp16")]; + tensor x_35_begin_0 = const()[name = tensor("x_35_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_35_end_0 = const()[name = tensor("x_35_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_35_end_mask_0 = const()[name = tensor("x_35_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_35_squeeze_mask_0 = const()[name = tensor("x_35_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_35_cast_fp16 = slice_by_index(begin = x_35_begin_0, end = x_35_end_0, end_mask = x_35_end_mask_0, squeeze_mask = x_35_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("x_35_cast_fp16")]; + tensor v_17_begin_0 = const()[name = tensor("v_17_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_17_end_0 = const()[name = tensor("v_17_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_17_end_mask_0 = const()[name = tensor("v_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_17_squeeze_mask_0 = const()[name = tensor("v_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_17_cast_fp16 = slice_by_index(begin = v_17_begin_0, end = v_17_end_0, end_mask = v_17_end_mask_0, squeeze_mask = v_17_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("v_17_cast_fp16")]; + tensor var_2223 = const()[name = tensor("op_2223"), val = tensor([256, 16, 32, 2])]; + tensor pairs_33_cast_fp16 = reshape(shape = var_2223, x = x_33_cast_fp16)[name = tensor("pairs_33_cast_fp16")]; + tensor xr_33_begin_0 = const()[name = tensor("xr_33_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_33_end_0 = const()[name = tensor("xr_33_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_33_end_mask_0 = const()[name = tensor("xr_33_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_33_squeeze_mask_0 = const()[name = tensor("xr_33_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_33_cast_fp16 = slice_by_index(begin = xr_33_begin_0, end = xr_33_end_0, end_mask = xr_33_end_mask_0, squeeze_mask = xr_33_squeeze_mask_0, x = pairs_33_cast_fp16)[name = tensor("xr_33_cast_fp16")]; + tensor xi_33_begin_0 = const()[name = tensor("xi_33_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_33_end_0 = const()[name = tensor("xi_33_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_33_end_mask_0 = const()[name = tensor("xi_33_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_33_squeeze_mask_0 = const()[name = tensor("xi_33_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_33_cast_fp16 = slice_by_index(begin = xi_33_begin_0, end = xi_33_end_0, end_mask = xi_33_end_mask_0, squeeze_mask = xi_33_squeeze_mask_0, x = pairs_33_cast_fp16)[name = tensor("xi_33_cast_fp16")]; + tensor var_2231_cast_fp16 = mul(x = xr_33_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2231_cast_fp16")]; + tensor var_2232_cast_fp16 = mul(x = xi_33_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2232_cast_fp16")]; + tensor xor__33_cast_fp16 = sub(x = var_2231_cast_fp16, y = var_2232_cast_fp16)[name = tensor("xor__33_cast_fp16")]; + tensor var_2235_cast_fp16 = mul(x = xr_33_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2235_cast_fp16")]; + tensor var_2236_cast_fp16 = mul(x = xi_33_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2236_cast_fp16")]; + tensor xoi_33_cast_fp16 = add(x = var_2235_cast_fp16, y = var_2236_cast_fp16)[name = tensor("xoi_33_cast_fp16")]; + tensor out_33_axis_0 = const()[name = tensor("out_33_axis_0"), val = tensor(-1)]; + tensor out_33_cast_fp16 = stack(axis = out_33_axis_0, values = (xor__33_cast_fp16, xoi_33_cast_fp16))[name = tensor("out_33_cast_fp16")]; + tensor var_2246 = const()[name = tensor("op_2246"), val = tensor([1, 256, 16, 64])]; + tensor q_17_cast_fp16 = reshape(shape = var_2246, x = out_33_cast_fp16)[name = tensor("q_17_cast_fp16")]; + tensor var_2252 = const()[name = tensor("op_2252"), val = tensor([256, 16, 32, 2])]; + tensor pairs_35_cast_fp16 = reshape(shape = var_2252, x = x_35_cast_fp16)[name = tensor("pairs_35_cast_fp16")]; + tensor xr_35_begin_0 = const()[name = tensor("xr_35_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_35_end_0 = const()[name = tensor("xr_35_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_35_end_mask_0 = const()[name = tensor("xr_35_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_35_squeeze_mask_0 = const()[name = tensor("xr_35_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_35_cast_fp16 = slice_by_index(begin = xr_35_begin_0, end = xr_35_end_0, end_mask = xr_35_end_mask_0, squeeze_mask = xr_35_squeeze_mask_0, x = pairs_35_cast_fp16)[name = tensor("xr_35_cast_fp16")]; + tensor xi_35_begin_0 = const()[name = tensor("xi_35_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_35_end_0 = const()[name = tensor("xi_35_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_35_end_mask_0 = const()[name = tensor("xi_35_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_35_squeeze_mask_0 = const()[name = tensor("xi_35_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_35_cast_fp16 = slice_by_index(begin = xi_35_begin_0, end = xi_35_end_0, end_mask = xi_35_end_mask_0, squeeze_mask = xi_35_squeeze_mask_0, x = pairs_35_cast_fp16)[name = tensor("xi_35_cast_fp16")]; + tensor var_2260_cast_fp16 = mul(x = xr_35_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2260_cast_fp16")]; + tensor var_2261_cast_fp16 = mul(x = xi_35_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2261_cast_fp16")]; + tensor xor__35_cast_fp16 = sub(x = var_2260_cast_fp16, y = var_2261_cast_fp16)[name = tensor("xor__35_cast_fp16")]; + tensor var_2264_cast_fp16 = mul(x = xr_35_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2264_cast_fp16")]; + tensor var_2265_cast_fp16 = mul(x = xi_35_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2265_cast_fp16")]; + tensor xoi_35_cast_fp16 = add(x = var_2264_cast_fp16, y = var_2265_cast_fp16)[name = tensor("xoi_35_cast_fp16")]; + tensor out_35_axis_0 = const()[name = tensor("out_35_axis_0"), val = tensor(-1)]; + tensor out_35_cast_fp16 = stack(axis = out_35_axis_0, values = (xor__35_cast_fp16, xoi_35_cast_fp16))[name = tensor("out_35_cast_fp16")]; + tensor var_2282 = const()[name = tensor("op_2282"), val = tensor([256, 1024])]; + tensor var_2283_cast_fp16 = reshape(shape = var_2282, x = out_35_cast_fp16)[name = tensor("op_2283_cast_fp16")]; + tensor written_k_17_transpose_x_1 = const()[name = tensor("written_k_17_transpose_x_1"), val = tensor(true)]; + tensor written_k_17_transpose_y_1 = const()[name = tensor("written_k_17_transpose_y_1"), val = tensor(false)]; + tensor written_k_17_cast_fp16 = matmul(transpose_x = written_k_17_transpose_x_1, transpose_y = written_k_17_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_2283_cast_fp16)[name = tensor("written_k_17_cast_fp16")]; + tensor var_2290 = const()[name = tensor("op_2290"), val = tensor([256, 1024])]; + tensor var_2291_cast_fp16 = reshape(shape = var_2290, x = v_17_cast_fp16)[name = tensor("op_2291_cast_fp16")]; + tensor written_v_17_transpose_x_1 = const()[name = tensor("written_v_17_transpose_x_1"), val = tensor(true)]; + tensor written_v_17_transpose_y_1 = const()[name = tensor("written_v_17_transpose_y_1"), val = tensor(false)]; + tensor written_v_17_cast_fp16 = matmul(transpose_x = written_v_17_transpose_x_1, transpose_y = written_v_17_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_2291_cast_fp16)[name = tensor("written_v_17_cast_fp16")]; + tensor k_cache8_to_fp16_dtype_0 = const()[name = tensor("k_cache8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache8_to_fp16 = cast(dtype = k_cache8_to_fp16_dtype_0, x = k_cache8)[name = tensor("cast_417")]; + tensor var_2302_cast_fp16 = mul(x = k_cache8_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_2302_cast_fp16")]; + tensor var_2307 = const()[name = tensor("op_2307"), val = tensor([1, 512, 16, 64])]; + tensor var_2308_cast_fp16 = reshape(shape = var_2307, x = written_k_17_cast_fp16)[name = tensor("op_2308_cast_fp16")]; + tensor new_k_cache_17_cast_fp16 = add(x = var_2302_cast_fp16, y = var_2308_cast_fp16)[name = tensor("new_k_cache_17_cast_fp16")]; + tensor new_k_cache_17_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_17_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache8_to_fp16_dtype_0 = const()[name = tensor("v_cache8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache8_to_fp16 = cast(dtype = v_cache8_to_fp16_dtype_0, x = v_cache8)[name = tensor("cast_415")]; + tensor var_2311_cast_fp16 = mul(x = v_cache8_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_2311_cast_fp16")]; + tensor var_2316 = const()[name = tensor("op_2316"), val = tensor([1, 512, 16, 64])]; + tensor var_2317_cast_fp16 = reshape(shape = var_2316, x = written_v_17_cast_fp16)[name = tensor("op_2317_cast_fp16")]; + tensor new_v_cache_17_cast_fp16 = add(x = var_2311_cast_fp16, y = var_2317_cast_fp16)[name = tensor("new_v_cache_17_cast_fp16")]; + tensor new_v_cache_17_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_17_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_2330 = const()[name = tensor("op_2330"), val = tensor([0, 2, 1, 3])]; + tensor var_2341_transpose_x_0 = const()[name = tensor("op_2341_transpose_x_0"), val = tensor(false)]; + tensor var_2341_transpose_y_0 = const()[name = tensor("op_2341_transpose_y_0"), val = tensor(false)]; + tensor transpose_85_perm_0 = const()[name = tensor("transpose_85_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_86_perm_0 = const()[name = tensor("transpose_86_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_86 = transpose(perm = transpose_86_perm_0, x = new_k_cache_17_cast_fp16)[name = tensor("transpose_172")]; + tensor transpose_85 = transpose(perm = transpose_85_perm_0, x = q_17_cast_fp16)[name = tensor("transpose_173")]; + tensor var_2341_cast_fp16 = matmul(transpose_x = var_2341_transpose_x_0, transpose_y = var_2341_transpose_y_0, x = transpose_85, y = transpose_86)[name = tensor("op_2341_cast_fp16")]; + tensor var_2342_to_fp16 = const()[name = tensor("op_2342_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_33_cast_fp16 = mul(x = var_2341_cast_fp16, y = var_2342_to_fp16)[name = tensor("scores_33_cast_fp16")]; + tensor scores_35_cast_fp16 = add(x = scores_33_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_35_cast_fp16")]; + tensor var_2352 = const()[name = tensor("op_2352"), val = tensor(-1)]; + tensor weights_17_cast_fp16 = softmax(axis = var_2352, x = scores_35_cast_fp16)[name = tensor("weights_17_cast_fp16")]; + tensor attn_17_transpose_x_0 = const()[name = tensor("attn_17_transpose_x_0"), val = tensor(false)]; + tensor attn_17_transpose_y_0 = const()[name = tensor("attn_17_transpose_y_0"), val = tensor(false)]; + tensor values_17_cast_fp16 = transpose(perm = var_2330, x = new_v_cache_17_cast_fp16)[name = tensor("transpose_174")]; + tensor attn_17_cast_fp16 = matmul(transpose_x = attn_17_transpose_x_0, transpose_y = attn_17_transpose_y_0, x = weights_17_cast_fp16, y = values_17_cast_fp16)[name = tensor("attn_17_cast_fp16")]; + tensor var_2360 = const()[name = tensor("op_2360"), val = tensor([0, 2, 1, 3])]; + tensor var_2365 = const()[name = tensor("op_2365"), val = tensor([1, 256, 1024])]; + tensor var_2361_cast_fp16 = transpose(perm = var_2360, x = attn_17_cast_fp16)[name = tensor("transpose_171")]; + tensor input_101_cast_fp16 = reshape(shape = var_2365, x = var_2361_cast_fp16)[name = tensor("input_101_cast_fp16")]; + tensor attn8_out_proj_weight_to_fp16 = const()[name = tensor("attn8_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(207710976)))]; + tensor linear_33_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn8_out_proj_weight_to_fp16, x = input_101_cast_fp16)[name = tensor("linear_33_cast_fp16")]; + tensor position8_to_fp16_dtype_0 = const()[name = tensor("position8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position8_to_fp16 = cast(dtype = position8_to_fp16_dtype_0, x = position8)[name = tensor("cast_413")]; + tensor var_2374_cast_fp16 = add(x = position8_to_fp16, y = valid_len_to_fp16)[name = tensor("op_2374_cast_fp16")]; + tensor var_2374_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2374_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_103_cast_fp16 = add(x = input_97_cast_fp16, y = linear_33_cast_fp16)[name = tensor("input_103_cast_fp16")]; + tensor input_105_axes_0 = const()[name = tensor("input_105_axes_0"), val = tensor([-1])]; + tensor norm8_2_weight_to_fp16 = const()[name = tensor("norm8_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209808192)))]; + tensor norm8_2_bias_to_fp16 = const()[name = tensor("norm8_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209810304)))]; + tensor var_2378_to_fp16 = const()[name = tensor("op_2378_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_105_cast_fp16 = layer_norm(axes = input_105_axes_0, beta = norm8_2_bias_to_fp16, epsilon = var_2378_to_fp16, gamma = norm8_2_weight_to_fp16, x = input_103_cast_fp16)[name = tensor("input_105_cast_fp16")]; + tensor linear8_1_weight_to_fp16 = const()[name = tensor("linear8_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209812416)))]; + tensor linear_34_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear8_1_weight_to_fp16, x = input_105_cast_fp16)[name = tensor("linear_34_cast_fp16")]; + tensor input_107_mode_0 = const()[name = tensor("input_107_mode_0"), val = tensor("EXACT")]; + tensor input_107_cast_fp16 = gelu(mode = input_107_mode_0, x = linear_34_cast_fp16)[name = tensor("input_107_cast_fp16")]; + tensor linear8_2_weight_to_fp16 = const()[name = tensor("linear8_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(218201088)))]; + tensor linear_35_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear8_2_weight_to_fp16, x = input_107_cast_fp16)[name = tensor("linear_35_cast_fp16")]; + tensor input_109_cast_fp16 = add(x = input_103_cast_fp16, y = linear_35_cast_fp16)[name = tensor("input_109_cast_fp16")]; + tensor input_111_axes_0 = const()[name = tensor("input_111_axes_0"), val = tensor([-1])]; + tensor norm9_1_weight_to_fp16 = const()[name = tensor("norm9_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226589760)))]; + tensor norm9_1_bias_to_fp16 = const()[name = tensor("norm9_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226591872)))]; + tensor var_2395_to_fp16 = const()[name = tensor("op_2395_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_111_cast_fp16 = layer_norm(axes = input_111_axes_0, beta = norm9_1_bias_to_fp16, epsilon = var_2395_to_fp16, gamma = norm9_1_weight_to_fp16, x = input_109_cast_fp16)[name = tensor("input_111_cast_fp16")]; + tensor attn9_in_proj_weight_to_fp16 = const()[name = tensor("attn9_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226593984)))]; + tensor linear_36_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn9_in_proj_weight_to_fp16, x = input_111_cast_fp16)[name = tensor("linear_36_cast_fp16")]; + tensor var_2409 = const()[name = tensor("op_2409"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_19_cast_fp16 = reshape(shape = var_2409, x = linear_36_cast_fp16)[name = tensor("qkv_19_cast_fp16")]; + tensor x_37_begin_0 = const()[name = tensor("x_37_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_37_end_0 = const()[name = tensor("x_37_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_37_end_mask_0 = const()[name = tensor("x_37_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_37_squeeze_mask_0 = const()[name = tensor("x_37_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_37_cast_fp16 = slice_by_index(begin = x_37_begin_0, end = x_37_end_0, end_mask = x_37_end_mask_0, squeeze_mask = x_37_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("x_37_cast_fp16")]; + tensor x_39_begin_0 = const()[name = tensor("x_39_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_39_end_0 = const()[name = tensor("x_39_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_39_end_mask_0 = const()[name = tensor("x_39_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_39_squeeze_mask_0 = const()[name = tensor("x_39_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_39_cast_fp16 = slice_by_index(begin = x_39_begin_0, end = x_39_end_0, end_mask = x_39_end_mask_0, squeeze_mask = x_39_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("x_39_cast_fp16")]; + tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_19_cast_fp16 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("v_19_cast_fp16")]; + tensor var_2454 = const()[name = tensor("op_2454"), val = tensor([256, 16, 32, 2])]; + tensor pairs_37_cast_fp16 = reshape(shape = var_2454, x = x_37_cast_fp16)[name = tensor("pairs_37_cast_fp16")]; + tensor xr_37_begin_0 = const()[name = tensor("xr_37_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_37_end_0 = const()[name = tensor("xr_37_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_37_end_mask_0 = const()[name = tensor("xr_37_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_37_squeeze_mask_0 = const()[name = tensor("xr_37_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_37_cast_fp16 = slice_by_index(begin = xr_37_begin_0, end = xr_37_end_0, end_mask = xr_37_end_mask_0, squeeze_mask = xr_37_squeeze_mask_0, x = pairs_37_cast_fp16)[name = tensor("xr_37_cast_fp16")]; + tensor xi_37_begin_0 = const()[name = tensor("xi_37_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_37_end_0 = const()[name = tensor("xi_37_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_37_end_mask_0 = const()[name = tensor("xi_37_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_37_squeeze_mask_0 = const()[name = tensor("xi_37_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_37_cast_fp16 = slice_by_index(begin = xi_37_begin_0, end = xi_37_end_0, end_mask = xi_37_end_mask_0, squeeze_mask = xi_37_squeeze_mask_0, x = pairs_37_cast_fp16)[name = tensor("xi_37_cast_fp16")]; + tensor var_2462_cast_fp16 = mul(x = xr_37_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2462_cast_fp16")]; + tensor var_2463_cast_fp16 = mul(x = xi_37_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2463_cast_fp16")]; + tensor xor__37_cast_fp16 = sub(x = var_2462_cast_fp16, y = var_2463_cast_fp16)[name = tensor("xor__37_cast_fp16")]; + tensor var_2466_cast_fp16 = mul(x = xr_37_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2466_cast_fp16")]; + tensor var_2467_cast_fp16 = mul(x = xi_37_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2467_cast_fp16")]; + tensor xoi_37_cast_fp16 = add(x = var_2466_cast_fp16, y = var_2467_cast_fp16)[name = tensor("xoi_37_cast_fp16")]; + tensor out_37_axis_0 = const()[name = tensor("out_37_axis_0"), val = tensor(-1)]; + tensor out_37_cast_fp16 = stack(axis = out_37_axis_0, values = (xor__37_cast_fp16, xoi_37_cast_fp16))[name = tensor("out_37_cast_fp16")]; + tensor var_2477 = const()[name = tensor("op_2477"), val = tensor([1, 256, 16, 64])]; + tensor q_19_cast_fp16 = reshape(shape = var_2477, x = out_37_cast_fp16)[name = tensor("q_19_cast_fp16")]; + tensor var_2483 = const()[name = tensor("op_2483"), val = tensor([256, 16, 32, 2])]; + tensor pairs_39_cast_fp16 = reshape(shape = var_2483, x = x_39_cast_fp16)[name = tensor("pairs_39_cast_fp16")]; + tensor xr_39_begin_0 = const()[name = tensor("xr_39_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_39_end_0 = const()[name = tensor("xr_39_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_39_end_mask_0 = const()[name = tensor("xr_39_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_39_squeeze_mask_0 = const()[name = tensor("xr_39_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_39_cast_fp16 = slice_by_index(begin = xr_39_begin_0, end = xr_39_end_0, end_mask = xr_39_end_mask_0, squeeze_mask = xr_39_squeeze_mask_0, x = pairs_39_cast_fp16)[name = tensor("xr_39_cast_fp16")]; + tensor xi_39_begin_0 = const()[name = tensor("xi_39_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_39_end_0 = const()[name = tensor("xi_39_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_39_end_mask_0 = const()[name = tensor("xi_39_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_39_squeeze_mask_0 = const()[name = tensor("xi_39_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_39_cast_fp16 = slice_by_index(begin = xi_39_begin_0, end = xi_39_end_0, end_mask = xi_39_end_mask_0, squeeze_mask = xi_39_squeeze_mask_0, x = pairs_39_cast_fp16)[name = tensor("xi_39_cast_fp16")]; + tensor var_2491_cast_fp16 = mul(x = xr_39_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2491_cast_fp16")]; + tensor var_2492_cast_fp16 = mul(x = xi_39_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2492_cast_fp16")]; + tensor xor__39_cast_fp16 = sub(x = var_2491_cast_fp16, y = var_2492_cast_fp16)[name = tensor("xor__39_cast_fp16")]; + tensor var_2495_cast_fp16 = mul(x = xr_39_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2495_cast_fp16")]; + tensor var_2496_cast_fp16 = mul(x = xi_39_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2496_cast_fp16")]; + tensor xoi_39_cast_fp16 = add(x = var_2495_cast_fp16, y = var_2496_cast_fp16)[name = tensor("xoi_39_cast_fp16")]; + tensor out_39_axis_0 = const()[name = tensor("out_39_axis_0"), val = tensor(-1)]; + tensor out_39_cast_fp16 = stack(axis = out_39_axis_0, values = (xor__39_cast_fp16, xoi_39_cast_fp16))[name = tensor("out_39_cast_fp16")]; + tensor var_2513 = const()[name = tensor("op_2513"), val = tensor([256, 1024])]; + tensor var_2514_cast_fp16 = reshape(shape = var_2513, x = out_39_cast_fp16)[name = tensor("op_2514_cast_fp16")]; + tensor written_k_19_transpose_x_1 = const()[name = tensor("written_k_19_transpose_x_1"), val = tensor(true)]; + tensor written_k_19_transpose_y_1 = const()[name = tensor("written_k_19_transpose_y_1"), val = tensor(false)]; + tensor written_k_19_cast_fp16 = matmul(transpose_x = written_k_19_transpose_x_1, transpose_y = written_k_19_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_2514_cast_fp16)[name = tensor("written_k_19_cast_fp16")]; + tensor var_2521 = const()[name = tensor("op_2521"), val = tensor([256, 1024])]; + tensor var_2522_cast_fp16 = reshape(shape = var_2521, x = v_19_cast_fp16)[name = tensor("op_2522_cast_fp16")]; + tensor written_v_19_transpose_x_1 = const()[name = tensor("written_v_19_transpose_x_1"), val = tensor(true)]; + tensor written_v_19_transpose_y_1 = const()[name = tensor("written_v_19_transpose_y_1"), val = tensor(false)]; + tensor written_v_19_cast_fp16 = matmul(transpose_x = written_v_19_transpose_x_1, transpose_y = written_v_19_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_2522_cast_fp16)[name = tensor("written_v_19_cast_fp16")]; + tensor k_cache9_to_fp16_dtype_0 = const()[name = tensor("k_cache9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache9_to_fp16 = cast(dtype = k_cache9_to_fp16_dtype_0, x = k_cache9)[name = tensor("cast_411")]; + tensor var_2533_cast_fp16 = mul(x = k_cache9_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_2533_cast_fp16")]; + tensor var_2538 = const()[name = tensor("op_2538"), val = tensor([1, 512, 16, 64])]; + tensor var_2539_cast_fp16 = reshape(shape = var_2538, x = written_k_19_cast_fp16)[name = tensor("op_2539_cast_fp16")]; + tensor new_k_cache_19_cast_fp16 = add(x = var_2533_cast_fp16, y = var_2539_cast_fp16)[name = tensor("new_k_cache_19_cast_fp16")]; + tensor new_k_cache_19_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_19_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache9_to_fp16_dtype_0 = const()[name = tensor("v_cache9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache9_to_fp16 = cast(dtype = v_cache9_to_fp16_dtype_0, x = v_cache9)[name = tensor("cast_409")]; + tensor var_2542_cast_fp16 = mul(x = v_cache9_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_2542_cast_fp16")]; + tensor var_2547 = const()[name = tensor("op_2547"), val = tensor([1, 512, 16, 64])]; + tensor var_2548_cast_fp16 = reshape(shape = var_2547, x = written_v_19_cast_fp16)[name = tensor("op_2548_cast_fp16")]; + tensor new_v_cache_19_cast_fp16 = add(x = var_2542_cast_fp16, y = var_2548_cast_fp16)[name = tensor("new_v_cache_19_cast_fp16")]; + tensor new_v_cache_19_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_19_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_2561 = const()[name = tensor("op_2561"), val = tensor([0, 2, 1, 3])]; + tensor var_2572_transpose_x_0 = const()[name = tensor("op_2572_transpose_x_0"), val = tensor(false)]; + tensor var_2572_transpose_y_0 = const()[name = tensor("op_2572_transpose_y_0"), val = tensor(false)]; + tensor transpose_87_perm_0 = const()[name = tensor("transpose_87_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_88_perm_0 = const()[name = tensor("transpose_88_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_88 = transpose(perm = transpose_88_perm_0, x = new_k_cache_19_cast_fp16)[name = tensor("transpose_168")]; + tensor transpose_87 = transpose(perm = transpose_87_perm_0, x = q_19_cast_fp16)[name = tensor("transpose_169")]; + tensor var_2572_cast_fp16 = matmul(transpose_x = var_2572_transpose_x_0, transpose_y = var_2572_transpose_y_0, x = transpose_87, y = transpose_88)[name = tensor("op_2572_cast_fp16")]; + tensor var_2573_to_fp16 = const()[name = tensor("op_2573_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_37_cast_fp16 = mul(x = var_2572_cast_fp16, y = var_2573_to_fp16)[name = tensor("scores_37_cast_fp16")]; + tensor scores_39_cast_fp16 = add(x = scores_37_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_39_cast_fp16")]; + tensor var_2583 = const()[name = tensor("op_2583"), val = tensor(-1)]; + tensor weights_19_cast_fp16 = softmax(axis = var_2583, x = scores_39_cast_fp16)[name = tensor("weights_19_cast_fp16")]; + tensor attn_19_transpose_x_0 = const()[name = tensor("attn_19_transpose_x_0"), val = tensor(false)]; + tensor attn_19_transpose_y_0 = const()[name = tensor("attn_19_transpose_y_0"), val = tensor(false)]; + tensor values_19_cast_fp16 = transpose(perm = var_2561, x = new_v_cache_19_cast_fp16)[name = tensor("transpose_170")]; + tensor attn_19_cast_fp16 = matmul(transpose_x = attn_19_transpose_x_0, transpose_y = attn_19_transpose_y_0, x = weights_19_cast_fp16, y = values_19_cast_fp16)[name = tensor("attn_19_cast_fp16")]; + tensor var_2591 = const()[name = tensor("op_2591"), val = tensor([0, 2, 1, 3])]; + tensor var_2596 = const()[name = tensor("op_2596"), val = tensor([1, 256, 1024])]; + tensor var_2592_cast_fp16 = transpose(perm = var_2591, x = attn_19_cast_fp16)[name = tensor("transpose_167")]; + tensor input_113_cast_fp16 = reshape(shape = var_2596, x = var_2592_cast_fp16)[name = tensor("input_113_cast_fp16")]; + tensor attn9_out_proj_weight_to_fp16 = const()[name = tensor("attn9_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(232885504)))]; + tensor linear_37_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn9_out_proj_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("linear_37_cast_fp16")]; + tensor position9_to_fp16_dtype_0 = const()[name = tensor("position9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position9_to_fp16 = cast(dtype = position9_to_fp16_dtype_0, x = position9)[name = tensor("cast_407")]; + tensor var_2605_cast_fp16 = add(x = position9_to_fp16, y = valid_len_to_fp16)[name = tensor("op_2605_cast_fp16")]; + tensor var_2605_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2605_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_115_cast_fp16 = add(x = input_109_cast_fp16, y = linear_37_cast_fp16)[name = tensor("input_115_cast_fp16")]; + tensor input_117_axes_0 = const()[name = tensor("input_117_axes_0"), val = tensor([-1])]; + tensor norm9_2_weight_to_fp16 = const()[name = tensor("norm9_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(234982720)))]; + tensor norm9_2_bias_to_fp16 = const()[name = tensor("norm9_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(234984832)))]; + tensor var_2609_to_fp16 = const()[name = tensor("op_2609_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_117_cast_fp16 = layer_norm(axes = input_117_axes_0, beta = norm9_2_bias_to_fp16, epsilon = var_2609_to_fp16, gamma = norm9_2_weight_to_fp16, x = input_115_cast_fp16)[name = tensor("input_117_cast_fp16")]; + tensor linear9_1_weight_to_fp16 = const()[name = tensor("linear9_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(234986944)))]; + tensor linear_38_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear9_1_weight_to_fp16, x = input_117_cast_fp16)[name = tensor("linear_38_cast_fp16")]; + tensor input_119_mode_0 = const()[name = tensor("input_119_mode_0"), val = tensor("EXACT")]; + tensor input_119_cast_fp16 = gelu(mode = input_119_mode_0, x = linear_38_cast_fp16)[name = tensor("input_119_cast_fp16")]; + tensor linear9_2_weight_to_fp16 = const()[name = tensor("linear9_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(243375616)))]; + tensor linear_39_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear9_2_weight_to_fp16, x = input_119_cast_fp16)[name = tensor("linear_39_cast_fp16")]; + tensor input_121_cast_fp16 = add(x = input_115_cast_fp16, y = linear_39_cast_fp16)[name = tensor("input_121_cast_fp16")]; + tensor input_123_axes_0 = const()[name = tensor("input_123_axes_0"), val = tensor([-1])]; + tensor norm10_1_weight_to_fp16 = const()[name = tensor("norm10_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(251764288)))]; + tensor norm10_1_bias_to_fp16 = const()[name = tensor("norm10_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(251766400)))]; + tensor var_2626_to_fp16 = const()[name = tensor("op_2626_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_123_cast_fp16 = layer_norm(axes = input_123_axes_0, beta = norm10_1_bias_to_fp16, epsilon = var_2626_to_fp16, gamma = norm10_1_weight_to_fp16, x = input_121_cast_fp16)[name = tensor("input_123_cast_fp16")]; + tensor attn10_in_proj_weight_to_fp16 = const()[name = tensor("attn10_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(251768512)))]; + tensor linear_40_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn10_in_proj_weight_to_fp16, x = input_123_cast_fp16)[name = tensor("linear_40_cast_fp16")]; + tensor var_2640 = const()[name = tensor("op_2640"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_21_cast_fp16 = reshape(shape = var_2640, x = linear_40_cast_fp16)[name = tensor("qkv_21_cast_fp16")]; + tensor x_41_begin_0 = const()[name = tensor("x_41_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_41_end_0 = const()[name = tensor("x_41_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_41_end_mask_0 = const()[name = tensor("x_41_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_41_squeeze_mask_0 = const()[name = tensor("x_41_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_41_cast_fp16 = slice_by_index(begin = x_41_begin_0, end = x_41_end_0, end_mask = x_41_end_mask_0, squeeze_mask = x_41_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("x_41_cast_fp16")]; + tensor x_43_begin_0 = const()[name = tensor("x_43_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_43_end_0 = const()[name = tensor("x_43_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_43_end_mask_0 = const()[name = tensor("x_43_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_43_squeeze_mask_0 = const()[name = tensor("x_43_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_43_cast_fp16 = slice_by_index(begin = x_43_begin_0, end = x_43_end_0, end_mask = x_43_end_mask_0, squeeze_mask = x_43_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("x_43_cast_fp16")]; + tensor v_21_begin_0 = const()[name = tensor("v_21_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_21_end_0 = const()[name = tensor("v_21_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_21_end_mask_0 = const()[name = tensor("v_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_21_squeeze_mask_0 = const()[name = tensor("v_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_21_cast_fp16 = slice_by_index(begin = v_21_begin_0, end = v_21_end_0, end_mask = v_21_end_mask_0, squeeze_mask = v_21_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("v_21_cast_fp16")]; + tensor var_2685 = const()[name = tensor("op_2685"), val = tensor([256, 16, 32, 2])]; + tensor pairs_41_cast_fp16 = reshape(shape = var_2685, x = x_41_cast_fp16)[name = tensor("pairs_41_cast_fp16")]; + tensor xr_41_begin_0 = const()[name = tensor("xr_41_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_41_end_0 = const()[name = tensor("xr_41_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_41_end_mask_0 = const()[name = tensor("xr_41_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_41_squeeze_mask_0 = const()[name = tensor("xr_41_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_41_cast_fp16 = slice_by_index(begin = xr_41_begin_0, end = xr_41_end_0, end_mask = xr_41_end_mask_0, squeeze_mask = xr_41_squeeze_mask_0, x = pairs_41_cast_fp16)[name = tensor("xr_41_cast_fp16")]; + tensor xi_41_begin_0 = const()[name = tensor("xi_41_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_41_end_0 = const()[name = tensor("xi_41_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_41_end_mask_0 = const()[name = tensor("xi_41_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_41_squeeze_mask_0 = const()[name = tensor("xi_41_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_41_cast_fp16 = slice_by_index(begin = xi_41_begin_0, end = xi_41_end_0, end_mask = xi_41_end_mask_0, squeeze_mask = xi_41_squeeze_mask_0, x = pairs_41_cast_fp16)[name = tensor("xi_41_cast_fp16")]; + tensor var_2693_cast_fp16 = mul(x = xr_41_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2693_cast_fp16")]; + tensor var_2694_cast_fp16 = mul(x = xi_41_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2694_cast_fp16")]; + tensor xor__41_cast_fp16 = sub(x = var_2693_cast_fp16, y = var_2694_cast_fp16)[name = tensor("xor__41_cast_fp16")]; + tensor var_2697_cast_fp16 = mul(x = xr_41_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2697_cast_fp16")]; + tensor var_2698_cast_fp16 = mul(x = xi_41_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2698_cast_fp16")]; + tensor xoi_41_cast_fp16 = add(x = var_2697_cast_fp16, y = var_2698_cast_fp16)[name = tensor("xoi_41_cast_fp16")]; + tensor out_41_axis_0 = const()[name = tensor("out_41_axis_0"), val = tensor(-1)]; + tensor out_41_cast_fp16 = stack(axis = out_41_axis_0, values = (xor__41_cast_fp16, xoi_41_cast_fp16))[name = tensor("out_41_cast_fp16")]; + tensor var_2708 = const()[name = tensor("op_2708"), val = tensor([1, 256, 16, 64])]; + tensor q_21_cast_fp16 = reshape(shape = var_2708, x = out_41_cast_fp16)[name = tensor("q_21_cast_fp16")]; + tensor var_2714 = const()[name = tensor("op_2714"), val = tensor([256, 16, 32, 2])]; + tensor pairs_43_cast_fp16 = reshape(shape = var_2714, x = x_43_cast_fp16)[name = tensor("pairs_43_cast_fp16")]; + tensor xr_43_begin_0 = const()[name = tensor("xr_43_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_43_end_0 = const()[name = tensor("xr_43_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_43_end_mask_0 = const()[name = tensor("xr_43_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_43_squeeze_mask_0 = const()[name = tensor("xr_43_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_43_cast_fp16 = slice_by_index(begin = xr_43_begin_0, end = xr_43_end_0, end_mask = xr_43_end_mask_0, squeeze_mask = xr_43_squeeze_mask_0, x = pairs_43_cast_fp16)[name = tensor("xr_43_cast_fp16")]; + tensor xi_43_begin_0 = const()[name = tensor("xi_43_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_43_end_0 = const()[name = tensor("xi_43_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_43_end_mask_0 = const()[name = tensor("xi_43_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_43_squeeze_mask_0 = const()[name = tensor("xi_43_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_43_cast_fp16 = slice_by_index(begin = xi_43_begin_0, end = xi_43_end_0, end_mask = xi_43_end_mask_0, squeeze_mask = xi_43_squeeze_mask_0, x = pairs_43_cast_fp16)[name = tensor("xi_43_cast_fp16")]; + tensor var_2722_cast_fp16 = mul(x = xr_43_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2722_cast_fp16")]; + tensor var_2723_cast_fp16 = mul(x = xi_43_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2723_cast_fp16")]; + tensor xor__43_cast_fp16 = sub(x = var_2722_cast_fp16, y = var_2723_cast_fp16)[name = tensor("xor__43_cast_fp16")]; + tensor var_2726_cast_fp16 = mul(x = xr_43_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2726_cast_fp16")]; + tensor var_2727_cast_fp16 = mul(x = xi_43_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2727_cast_fp16")]; + tensor xoi_43_cast_fp16 = add(x = var_2726_cast_fp16, y = var_2727_cast_fp16)[name = tensor("xoi_43_cast_fp16")]; + tensor out_43_axis_0 = const()[name = tensor("out_43_axis_0"), val = tensor(-1)]; + tensor out_43_cast_fp16 = stack(axis = out_43_axis_0, values = (xor__43_cast_fp16, xoi_43_cast_fp16))[name = tensor("out_43_cast_fp16")]; + tensor var_2744 = const()[name = tensor("op_2744"), val = tensor([256, 1024])]; + tensor var_2745_cast_fp16 = reshape(shape = var_2744, x = out_43_cast_fp16)[name = tensor("op_2745_cast_fp16")]; + tensor written_k_21_transpose_x_1 = const()[name = tensor("written_k_21_transpose_x_1"), val = tensor(true)]; + tensor written_k_21_transpose_y_1 = const()[name = tensor("written_k_21_transpose_y_1"), val = tensor(false)]; + tensor written_k_21_cast_fp16 = matmul(transpose_x = written_k_21_transpose_x_1, transpose_y = written_k_21_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_2745_cast_fp16)[name = tensor("written_k_21_cast_fp16")]; + tensor var_2752 = const()[name = tensor("op_2752"), val = tensor([256, 1024])]; + tensor var_2753_cast_fp16 = reshape(shape = var_2752, x = v_21_cast_fp16)[name = tensor("op_2753_cast_fp16")]; + tensor written_v_21_transpose_x_1 = const()[name = tensor("written_v_21_transpose_x_1"), val = tensor(true)]; + tensor written_v_21_transpose_y_1 = const()[name = tensor("written_v_21_transpose_y_1"), val = tensor(false)]; + tensor written_v_21_cast_fp16 = matmul(transpose_x = written_v_21_transpose_x_1, transpose_y = written_v_21_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_2753_cast_fp16)[name = tensor("written_v_21_cast_fp16")]; + tensor k_cache10_to_fp16_dtype_0 = const()[name = tensor("k_cache10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache10_to_fp16 = cast(dtype = k_cache10_to_fp16_dtype_0, x = k_cache10)[name = tensor("cast_405")]; + tensor var_2764_cast_fp16 = mul(x = k_cache10_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_2764_cast_fp16")]; + tensor var_2769 = const()[name = tensor("op_2769"), val = tensor([1, 512, 16, 64])]; + tensor var_2770_cast_fp16 = reshape(shape = var_2769, x = written_k_21_cast_fp16)[name = tensor("op_2770_cast_fp16")]; + tensor new_k_cache_21_cast_fp16 = add(x = var_2764_cast_fp16, y = var_2770_cast_fp16)[name = tensor("new_k_cache_21_cast_fp16")]; + tensor new_k_cache_21_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_21_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache10_to_fp16_dtype_0 = const()[name = tensor("v_cache10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache10_to_fp16 = cast(dtype = v_cache10_to_fp16_dtype_0, x = v_cache10)[name = tensor("cast_403")]; + tensor var_2773_cast_fp16 = mul(x = v_cache10_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_2773_cast_fp16")]; + tensor var_2778 = const()[name = tensor("op_2778"), val = tensor([1, 512, 16, 64])]; + tensor var_2779_cast_fp16 = reshape(shape = var_2778, x = written_v_21_cast_fp16)[name = tensor("op_2779_cast_fp16")]; + tensor new_v_cache_21_cast_fp16 = add(x = var_2773_cast_fp16, y = var_2779_cast_fp16)[name = tensor("new_v_cache_21_cast_fp16")]; + tensor new_v_cache_21_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_21_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_2792 = const()[name = tensor("op_2792"), val = tensor([0, 2, 1, 3])]; + tensor var_2803_transpose_x_0 = const()[name = tensor("op_2803_transpose_x_0"), val = tensor(false)]; + tensor var_2803_transpose_y_0 = const()[name = tensor("op_2803_transpose_y_0"), val = tensor(false)]; + tensor transpose_89_perm_0 = const()[name = tensor("transpose_89_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_90_perm_0 = const()[name = tensor("transpose_90_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_90 = transpose(perm = transpose_90_perm_0, x = new_k_cache_21_cast_fp16)[name = tensor("transpose_164")]; + tensor transpose_89 = transpose(perm = transpose_89_perm_0, x = q_21_cast_fp16)[name = tensor("transpose_165")]; + tensor var_2803_cast_fp16 = matmul(transpose_x = var_2803_transpose_x_0, transpose_y = var_2803_transpose_y_0, x = transpose_89, y = transpose_90)[name = tensor("op_2803_cast_fp16")]; + tensor var_2804_to_fp16 = const()[name = tensor("op_2804_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_41_cast_fp16 = mul(x = var_2803_cast_fp16, y = var_2804_to_fp16)[name = tensor("scores_41_cast_fp16")]; + tensor scores_43_cast_fp16 = add(x = scores_41_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_43_cast_fp16")]; + tensor var_2814 = const()[name = tensor("op_2814"), val = tensor(-1)]; + tensor weights_21_cast_fp16 = softmax(axis = var_2814, x = scores_43_cast_fp16)[name = tensor("weights_21_cast_fp16")]; + tensor attn_21_transpose_x_0 = const()[name = tensor("attn_21_transpose_x_0"), val = tensor(false)]; + tensor attn_21_transpose_y_0 = const()[name = tensor("attn_21_transpose_y_0"), val = tensor(false)]; + tensor values_21_cast_fp16 = transpose(perm = var_2792, x = new_v_cache_21_cast_fp16)[name = tensor("transpose_166")]; + tensor attn_21_cast_fp16 = matmul(transpose_x = attn_21_transpose_x_0, transpose_y = attn_21_transpose_y_0, x = weights_21_cast_fp16, y = values_21_cast_fp16)[name = tensor("attn_21_cast_fp16")]; + tensor var_2822 = const()[name = tensor("op_2822"), val = tensor([0, 2, 1, 3])]; + tensor var_2827 = const()[name = tensor("op_2827"), val = tensor([1, 256, 1024])]; + tensor var_2823_cast_fp16 = transpose(perm = var_2822, x = attn_21_cast_fp16)[name = tensor("transpose_163")]; + tensor input_125_cast_fp16 = reshape(shape = var_2827, x = var_2823_cast_fp16)[name = tensor("input_125_cast_fp16")]; + tensor attn10_out_proj_weight_to_fp16 = const()[name = tensor("attn10_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(258060032)))]; + tensor linear_41_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn10_out_proj_weight_to_fp16, x = input_125_cast_fp16)[name = tensor("linear_41_cast_fp16")]; + tensor position10_to_fp16_dtype_0 = const()[name = tensor("position10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position10_to_fp16 = cast(dtype = position10_to_fp16_dtype_0, x = position10)[name = tensor("cast_401")]; + tensor var_2836_cast_fp16 = add(x = position10_to_fp16, y = valid_len_to_fp16)[name = tensor("op_2836_cast_fp16")]; + tensor var_2836_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2836_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_127_cast_fp16 = add(x = input_121_cast_fp16, y = linear_41_cast_fp16)[name = tensor("input_127_cast_fp16")]; + tensor input_129_axes_0 = const()[name = tensor("input_129_axes_0"), val = tensor([-1])]; + tensor norm10_2_weight_to_fp16 = const()[name = tensor("norm10_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260157248)))]; + tensor norm10_2_bias_to_fp16 = const()[name = tensor("norm10_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260159360)))]; + tensor var_2840_to_fp16 = const()[name = tensor("op_2840_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_129_cast_fp16 = layer_norm(axes = input_129_axes_0, beta = norm10_2_bias_to_fp16, epsilon = var_2840_to_fp16, gamma = norm10_2_weight_to_fp16, x = input_127_cast_fp16)[name = tensor("input_129_cast_fp16")]; + tensor linear10_1_weight_to_fp16 = const()[name = tensor("linear10_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260161472)))]; + tensor linear_42_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear10_1_weight_to_fp16, x = input_129_cast_fp16)[name = tensor("linear_42_cast_fp16")]; + tensor input_131_mode_0 = const()[name = tensor("input_131_mode_0"), val = tensor("EXACT")]; + tensor input_131_cast_fp16 = gelu(mode = input_131_mode_0, x = linear_42_cast_fp16)[name = tensor("input_131_cast_fp16")]; + tensor linear10_2_weight_to_fp16 = const()[name = tensor("linear10_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(268550144)))]; + tensor linear_43_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear10_2_weight_to_fp16, x = input_131_cast_fp16)[name = tensor("linear_43_cast_fp16")]; + tensor input_133_cast_fp16 = add(x = input_127_cast_fp16, y = linear_43_cast_fp16)[name = tensor("input_133_cast_fp16")]; + tensor input_135_axes_0 = const()[name = tensor("input_135_axes_0"), val = tensor([-1])]; + tensor norm11_1_weight_to_fp16 = const()[name = tensor("norm11_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(276938816)))]; + tensor norm11_1_bias_to_fp16 = const()[name = tensor("norm11_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(276940928)))]; + tensor var_2857_to_fp16 = const()[name = tensor("op_2857_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_135_cast_fp16 = layer_norm(axes = input_135_axes_0, beta = norm11_1_bias_to_fp16, epsilon = var_2857_to_fp16, gamma = norm11_1_weight_to_fp16, x = input_133_cast_fp16)[name = tensor("input_135_cast_fp16")]; + tensor attn11_in_proj_weight_to_fp16 = const()[name = tensor("attn11_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(276943040)))]; + tensor linear_44_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn11_in_proj_weight_to_fp16, x = input_135_cast_fp16)[name = tensor("linear_44_cast_fp16")]; + tensor var_2871 = const()[name = tensor("op_2871"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_23_cast_fp16 = reshape(shape = var_2871, x = linear_44_cast_fp16)[name = tensor("qkv_23_cast_fp16")]; + tensor x_45_begin_0 = const()[name = tensor("x_45_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_45_end_0 = const()[name = tensor("x_45_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_45_end_mask_0 = const()[name = tensor("x_45_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_45_squeeze_mask_0 = const()[name = tensor("x_45_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_45_cast_fp16 = slice_by_index(begin = x_45_begin_0, end = x_45_end_0, end_mask = x_45_end_mask_0, squeeze_mask = x_45_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("x_45_cast_fp16")]; + tensor x_47_begin_0 = const()[name = tensor("x_47_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_47_end_0 = const()[name = tensor("x_47_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_47_end_mask_0 = const()[name = tensor("x_47_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_47_squeeze_mask_0 = const()[name = tensor("x_47_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_47_cast_fp16 = slice_by_index(begin = x_47_begin_0, end = x_47_end_0, end_mask = x_47_end_mask_0, squeeze_mask = x_47_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("x_47_cast_fp16")]; + tensor v_23_begin_0 = const()[name = tensor("v_23_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_23_end_0 = const()[name = tensor("v_23_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_23_end_mask_0 = const()[name = tensor("v_23_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_23_squeeze_mask_0 = const()[name = tensor("v_23_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_23_cast_fp16 = slice_by_index(begin = v_23_begin_0, end = v_23_end_0, end_mask = v_23_end_mask_0, squeeze_mask = v_23_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("v_23_cast_fp16")]; + tensor var_2916 = const()[name = tensor("op_2916"), val = tensor([256, 16, 32, 2])]; + tensor pairs_45_cast_fp16 = reshape(shape = var_2916, x = x_45_cast_fp16)[name = tensor("pairs_45_cast_fp16")]; + tensor xr_45_begin_0 = const()[name = tensor("xr_45_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_45_end_0 = const()[name = tensor("xr_45_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_45_end_mask_0 = const()[name = tensor("xr_45_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_45_squeeze_mask_0 = const()[name = tensor("xr_45_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_45_cast_fp16 = slice_by_index(begin = xr_45_begin_0, end = xr_45_end_0, end_mask = xr_45_end_mask_0, squeeze_mask = xr_45_squeeze_mask_0, x = pairs_45_cast_fp16)[name = tensor("xr_45_cast_fp16")]; + tensor xi_45_begin_0 = const()[name = tensor("xi_45_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_45_end_0 = const()[name = tensor("xi_45_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_45_end_mask_0 = const()[name = tensor("xi_45_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_45_squeeze_mask_0 = const()[name = tensor("xi_45_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_45_cast_fp16 = slice_by_index(begin = xi_45_begin_0, end = xi_45_end_0, end_mask = xi_45_end_mask_0, squeeze_mask = xi_45_squeeze_mask_0, x = pairs_45_cast_fp16)[name = tensor("xi_45_cast_fp16")]; + tensor var_2924_cast_fp16 = mul(x = xr_45_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2924_cast_fp16")]; + tensor var_2925_cast_fp16 = mul(x = xi_45_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2925_cast_fp16")]; + tensor xor__45_cast_fp16 = sub(x = var_2924_cast_fp16, y = var_2925_cast_fp16)[name = tensor("xor__45_cast_fp16")]; + tensor var_2928_cast_fp16 = mul(x = xr_45_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2928_cast_fp16")]; + tensor var_2929_cast_fp16 = mul(x = xi_45_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2929_cast_fp16")]; + tensor xoi_45_cast_fp16 = add(x = var_2928_cast_fp16, y = var_2929_cast_fp16)[name = tensor("xoi_45_cast_fp16")]; + tensor out_45_axis_0 = const()[name = tensor("out_45_axis_0"), val = tensor(-1)]; + tensor out_45_cast_fp16 = stack(axis = out_45_axis_0, values = (xor__45_cast_fp16, xoi_45_cast_fp16))[name = tensor("out_45_cast_fp16")]; + tensor var_2939 = const()[name = tensor("op_2939"), val = tensor([1, 256, 16, 64])]; + tensor q_23_cast_fp16 = reshape(shape = var_2939, x = out_45_cast_fp16)[name = tensor("q_23_cast_fp16")]; + tensor var_2945 = const()[name = tensor("op_2945"), val = tensor([256, 16, 32, 2])]; + tensor pairs_47_cast_fp16 = reshape(shape = var_2945, x = x_47_cast_fp16)[name = tensor("pairs_47_cast_fp16")]; + tensor xr_47_begin_0 = const()[name = tensor("xr_47_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_47_end_0 = const()[name = tensor("xr_47_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_47_end_mask_0 = const()[name = tensor("xr_47_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_47_squeeze_mask_0 = const()[name = tensor("xr_47_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_47_cast_fp16 = slice_by_index(begin = xr_47_begin_0, end = xr_47_end_0, end_mask = xr_47_end_mask_0, squeeze_mask = xr_47_squeeze_mask_0, x = pairs_47_cast_fp16)[name = tensor("xr_47_cast_fp16")]; + tensor xi_47_begin_0 = const()[name = tensor("xi_47_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_47_end_0 = const()[name = tensor("xi_47_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_47_end_mask_0 = const()[name = tensor("xi_47_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_47_squeeze_mask_0 = const()[name = tensor("xi_47_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_47_cast_fp16 = slice_by_index(begin = xi_47_begin_0, end = xi_47_end_0, end_mask = xi_47_end_mask_0, squeeze_mask = xi_47_squeeze_mask_0, x = pairs_47_cast_fp16)[name = tensor("xi_47_cast_fp16")]; + tensor var_2953_cast_fp16 = mul(x = xr_47_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2953_cast_fp16")]; + tensor var_2954_cast_fp16 = mul(x = xi_47_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2954_cast_fp16")]; + tensor xor__47_cast_fp16 = sub(x = var_2953_cast_fp16, y = var_2954_cast_fp16)[name = tensor("xor__47_cast_fp16")]; + tensor var_2957_cast_fp16 = mul(x = xr_47_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2957_cast_fp16")]; + tensor var_2958_cast_fp16 = mul(x = xi_47_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2958_cast_fp16")]; + tensor xoi_47_cast_fp16 = add(x = var_2957_cast_fp16, y = var_2958_cast_fp16)[name = tensor("xoi_47_cast_fp16")]; + tensor out_47_axis_0 = const()[name = tensor("out_47_axis_0"), val = tensor(-1)]; + tensor out_47_cast_fp16 = stack(axis = out_47_axis_0, values = (xor__47_cast_fp16, xoi_47_cast_fp16))[name = tensor("out_47_cast_fp16")]; + tensor var_2975 = const()[name = tensor("op_2975"), val = tensor([256, 1024])]; + tensor var_2976_cast_fp16 = reshape(shape = var_2975, x = out_47_cast_fp16)[name = tensor("op_2976_cast_fp16")]; + tensor written_k_23_transpose_x_1 = const()[name = tensor("written_k_23_transpose_x_1"), val = tensor(true)]; + tensor written_k_23_transpose_y_1 = const()[name = tensor("written_k_23_transpose_y_1"), val = tensor(false)]; + tensor written_k_23_cast_fp16 = matmul(transpose_x = written_k_23_transpose_x_1, transpose_y = written_k_23_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_2976_cast_fp16)[name = tensor("written_k_23_cast_fp16")]; + tensor var_2983 = const()[name = tensor("op_2983"), val = tensor([256, 1024])]; + tensor var_2984_cast_fp16 = reshape(shape = var_2983, x = v_23_cast_fp16)[name = tensor("op_2984_cast_fp16")]; + tensor written_v_23_transpose_x_1 = const()[name = tensor("written_v_23_transpose_x_1"), val = tensor(true)]; + tensor written_v_23_transpose_y_1 = const()[name = tensor("written_v_23_transpose_y_1"), val = tensor(false)]; + tensor written_v_23_cast_fp16 = matmul(transpose_x = written_v_23_transpose_x_1, transpose_y = written_v_23_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_2984_cast_fp16)[name = tensor("written_v_23_cast_fp16")]; + tensor k_cache11_to_fp16_dtype_0 = const()[name = tensor("k_cache11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache11_to_fp16 = cast(dtype = k_cache11_to_fp16_dtype_0, x = k_cache11)[name = tensor("cast_399")]; + tensor var_2995_cast_fp16 = mul(x = k_cache11_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_2995_cast_fp16")]; + tensor var_3000 = const()[name = tensor("op_3000"), val = tensor([1, 512, 16, 64])]; + tensor var_3001_cast_fp16 = reshape(shape = var_3000, x = written_k_23_cast_fp16)[name = tensor("op_3001_cast_fp16")]; + tensor new_k_cache_23_cast_fp16 = add(x = var_2995_cast_fp16, y = var_3001_cast_fp16)[name = tensor("new_k_cache_23_cast_fp16")]; + tensor new_k_cache_23_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_23_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache11_to_fp16_dtype_0 = const()[name = tensor("v_cache11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache11_to_fp16 = cast(dtype = v_cache11_to_fp16_dtype_0, x = v_cache11)[name = tensor("cast_397")]; + tensor var_3004_cast_fp16 = mul(x = v_cache11_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_3004_cast_fp16")]; + tensor var_3009 = const()[name = tensor("op_3009"), val = tensor([1, 512, 16, 64])]; + tensor var_3010_cast_fp16 = reshape(shape = var_3009, x = written_v_23_cast_fp16)[name = tensor("op_3010_cast_fp16")]; + tensor new_v_cache_23_cast_fp16 = add(x = var_3004_cast_fp16, y = var_3010_cast_fp16)[name = tensor("new_v_cache_23_cast_fp16")]; + tensor new_v_cache_23_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_23_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_3023 = const()[name = tensor("op_3023"), val = tensor([0, 2, 1, 3])]; + tensor var_3034_transpose_x_0 = const()[name = tensor("op_3034_transpose_x_0"), val = tensor(false)]; + tensor var_3034_transpose_y_0 = const()[name = tensor("op_3034_transpose_y_0"), val = tensor(false)]; + tensor transpose_91_perm_0 = const()[name = tensor("transpose_91_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_92_perm_0 = const()[name = tensor("transpose_92_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_92 = transpose(perm = transpose_92_perm_0, x = new_k_cache_23_cast_fp16)[name = tensor("transpose_160")]; + tensor transpose_91 = transpose(perm = transpose_91_perm_0, x = q_23_cast_fp16)[name = tensor("transpose_161")]; + tensor var_3034_cast_fp16 = matmul(transpose_x = var_3034_transpose_x_0, transpose_y = var_3034_transpose_y_0, x = transpose_91, y = transpose_92)[name = tensor("op_3034_cast_fp16")]; + tensor var_3035_to_fp16 = const()[name = tensor("op_3035_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_45_cast_fp16 = mul(x = var_3034_cast_fp16, y = var_3035_to_fp16)[name = tensor("scores_45_cast_fp16")]; + tensor scores_47_cast_fp16 = add(x = scores_45_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_47_cast_fp16")]; + tensor var_3045 = const()[name = tensor("op_3045"), val = tensor(-1)]; + tensor weights_23_cast_fp16 = softmax(axis = var_3045, x = scores_47_cast_fp16)[name = tensor("weights_23_cast_fp16")]; + tensor attn_23_transpose_x_0 = const()[name = tensor("attn_23_transpose_x_0"), val = tensor(false)]; + tensor attn_23_transpose_y_0 = const()[name = tensor("attn_23_transpose_y_0"), val = tensor(false)]; + tensor values_23_cast_fp16 = transpose(perm = var_3023, x = new_v_cache_23_cast_fp16)[name = tensor("transpose_162")]; + tensor attn_23_cast_fp16 = matmul(transpose_x = attn_23_transpose_x_0, transpose_y = attn_23_transpose_y_0, x = weights_23_cast_fp16, y = values_23_cast_fp16)[name = tensor("attn_23_cast_fp16")]; + tensor var_3053 = const()[name = tensor("op_3053"), val = tensor([0, 2, 1, 3])]; + tensor var_3058 = const()[name = tensor("op_3058"), val = tensor([1, 256, 1024])]; + tensor var_3054_cast_fp16 = transpose(perm = var_3053, x = attn_23_cast_fp16)[name = tensor("transpose_159")]; + tensor input_137_cast_fp16 = reshape(shape = var_3058, x = var_3054_cast_fp16)[name = tensor("input_137_cast_fp16")]; + tensor attn11_out_proj_weight_to_fp16 = const()[name = tensor("attn11_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(283234560)))]; + tensor linear_45_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn11_out_proj_weight_to_fp16, x = input_137_cast_fp16)[name = tensor("linear_45_cast_fp16")]; + tensor position11_to_fp16_dtype_0 = const()[name = tensor("position11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position11_to_fp16 = cast(dtype = position11_to_fp16_dtype_0, x = position11)[name = tensor("cast_395")]; + tensor var_3067_cast_fp16 = add(x = position11_to_fp16, y = valid_len_to_fp16)[name = tensor("op_3067_cast_fp16")]; + tensor var_3067_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3067_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_139_cast_fp16 = add(x = input_133_cast_fp16, y = linear_45_cast_fp16)[name = tensor("input_139_cast_fp16")]; + tensor input_141_axes_0 = const()[name = tensor("input_141_axes_0"), val = tensor([-1])]; + tensor norm11_2_weight_to_fp16 = const()[name = tensor("norm11_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285331776)))]; + tensor norm11_2_bias_to_fp16 = const()[name = tensor("norm11_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285333888)))]; + tensor var_3071_to_fp16 = const()[name = tensor("op_3071_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_141_cast_fp16 = layer_norm(axes = input_141_axes_0, beta = norm11_2_bias_to_fp16, epsilon = var_3071_to_fp16, gamma = norm11_2_weight_to_fp16, x = input_139_cast_fp16)[name = tensor("input_141_cast_fp16")]; + tensor linear11_1_weight_to_fp16 = const()[name = tensor("linear11_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285336000)))]; + tensor linear_46_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear11_1_weight_to_fp16, x = input_141_cast_fp16)[name = tensor("linear_46_cast_fp16")]; + tensor input_143_mode_0 = const()[name = tensor("input_143_mode_0"), val = tensor("EXACT")]; + tensor input_143_cast_fp16 = gelu(mode = input_143_mode_0, x = linear_46_cast_fp16)[name = tensor("input_143_cast_fp16")]; + tensor linear11_2_weight_to_fp16 = const()[name = tensor("linear11_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293724672)))]; + tensor linear_47_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear11_2_weight_to_fp16, x = input_143_cast_fp16)[name = tensor("linear_47_cast_fp16")]; + tensor input_145_cast_fp16 = add(x = input_139_cast_fp16, y = linear_47_cast_fp16)[name = tensor("input_145_cast_fp16")]; + tensor input_147_axes_0 = const()[name = tensor("input_147_axes_0"), val = tensor([-1])]; + tensor norm12_1_weight_to_fp16 = const()[name = tensor("norm12_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302113344)))]; + tensor norm12_1_bias_to_fp16 = const()[name = tensor("norm12_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302115456)))]; + tensor var_3088_to_fp16 = const()[name = tensor("op_3088_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_147_cast_fp16 = layer_norm(axes = input_147_axes_0, beta = norm12_1_bias_to_fp16, epsilon = var_3088_to_fp16, gamma = norm12_1_weight_to_fp16, x = input_145_cast_fp16)[name = tensor("input_147_cast_fp16")]; + tensor attn12_in_proj_weight_to_fp16 = const()[name = tensor("attn12_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302117568)))]; + tensor linear_48_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn12_in_proj_weight_to_fp16, x = input_147_cast_fp16)[name = tensor("linear_48_cast_fp16")]; + tensor var_3102 = const()[name = tensor("op_3102"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_25_cast_fp16 = reshape(shape = var_3102, x = linear_48_cast_fp16)[name = tensor("qkv_25_cast_fp16")]; + tensor x_49_begin_0 = const()[name = tensor("x_49_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_49_end_0 = const()[name = tensor("x_49_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_49_end_mask_0 = const()[name = tensor("x_49_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_49_squeeze_mask_0 = const()[name = tensor("x_49_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_49_cast_fp16 = slice_by_index(begin = x_49_begin_0, end = x_49_end_0, end_mask = x_49_end_mask_0, squeeze_mask = x_49_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("x_49_cast_fp16")]; + tensor x_51_begin_0 = const()[name = tensor("x_51_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_51_end_0 = const()[name = tensor("x_51_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_51_end_mask_0 = const()[name = tensor("x_51_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_51_squeeze_mask_0 = const()[name = tensor("x_51_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_51_cast_fp16 = slice_by_index(begin = x_51_begin_0, end = x_51_end_0, end_mask = x_51_end_mask_0, squeeze_mask = x_51_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("x_51_cast_fp16")]; + tensor v_25_begin_0 = const()[name = tensor("v_25_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_25_end_0 = const()[name = tensor("v_25_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_25_end_mask_0 = const()[name = tensor("v_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_25_squeeze_mask_0 = const()[name = tensor("v_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_25_cast_fp16 = slice_by_index(begin = v_25_begin_0, end = v_25_end_0, end_mask = v_25_end_mask_0, squeeze_mask = v_25_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("v_25_cast_fp16")]; + tensor var_3147 = const()[name = tensor("op_3147"), val = tensor([256, 16, 32, 2])]; + tensor pairs_49_cast_fp16 = reshape(shape = var_3147, x = x_49_cast_fp16)[name = tensor("pairs_49_cast_fp16")]; + tensor xr_49_begin_0 = const()[name = tensor("xr_49_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_49_end_0 = const()[name = tensor("xr_49_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_49_end_mask_0 = const()[name = tensor("xr_49_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_49_squeeze_mask_0 = const()[name = tensor("xr_49_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_49_cast_fp16 = slice_by_index(begin = xr_49_begin_0, end = xr_49_end_0, end_mask = xr_49_end_mask_0, squeeze_mask = xr_49_squeeze_mask_0, x = pairs_49_cast_fp16)[name = tensor("xr_49_cast_fp16")]; + tensor xi_49_begin_0 = const()[name = tensor("xi_49_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_49_end_0 = const()[name = tensor("xi_49_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_49_end_mask_0 = const()[name = tensor("xi_49_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_49_squeeze_mask_0 = const()[name = tensor("xi_49_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_49_cast_fp16 = slice_by_index(begin = xi_49_begin_0, end = xi_49_end_0, end_mask = xi_49_end_mask_0, squeeze_mask = xi_49_squeeze_mask_0, x = pairs_49_cast_fp16)[name = tensor("xi_49_cast_fp16")]; + tensor var_3155_cast_fp16 = mul(x = xr_49_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3155_cast_fp16")]; + tensor var_3156_cast_fp16 = mul(x = xi_49_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3156_cast_fp16")]; + tensor xor__49_cast_fp16 = sub(x = var_3155_cast_fp16, y = var_3156_cast_fp16)[name = tensor("xor__49_cast_fp16")]; + tensor var_3159_cast_fp16 = mul(x = xr_49_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3159_cast_fp16")]; + tensor var_3160_cast_fp16 = mul(x = xi_49_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3160_cast_fp16")]; + tensor xoi_49_cast_fp16 = add(x = var_3159_cast_fp16, y = var_3160_cast_fp16)[name = tensor("xoi_49_cast_fp16")]; + tensor out_49_axis_0 = const()[name = tensor("out_49_axis_0"), val = tensor(-1)]; + tensor out_49_cast_fp16 = stack(axis = out_49_axis_0, values = (xor__49_cast_fp16, xoi_49_cast_fp16))[name = tensor("out_49_cast_fp16")]; + tensor var_3170 = const()[name = tensor("op_3170"), val = tensor([1, 256, 16, 64])]; + tensor q_25_cast_fp16 = reshape(shape = var_3170, x = out_49_cast_fp16)[name = tensor("q_25_cast_fp16")]; + tensor var_3176 = const()[name = tensor("op_3176"), val = tensor([256, 16, 32, 2])]; + tensor pairs_51_cast_fp16 = reshape(shape = var_3176, x = x_51_cast_fp16)[name = tensor("pairs_51_cast_fp16")]; + tensor xr_51_begin_0 = const()[name = tensor("xr_51_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_51_end_0 = const()[name = tensor("xr_51_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_51_end_mask_0 = const()[name = tensor("xr_51_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_51_squeeze_mask_0 = const()[name = tensor("xr_51_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_51_cast_fp16 = slice_by_index(begin = xr_51_begin_0, end = xr_51_end_0, end_mask = xr_51_end_mask_0, squeeze_mask = xr_51_squeeze_mask_0, x = pairs_51_cast_fp16)[name = tensor("xr_51_cast_fp16")]; + tensor xi_51_begin_0 = const()[name = tensor("xi_51_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_51_end_0 = const()[name = tensor("xi_51_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_51_end_mask_0 = const()[name = tensor("xi_51_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_51_squeeze_mask_0 = const()[name = tensor("xi_51_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_51_cast_fp16 = slice_by_index(begin = xi_51_begin_0, end = xi_51_end_0, end_mask = xi_51_end_mask_0, squeeze_mask = xi_51_squeeze_mask_0, x = pairs_51_cast_fp16)[name = tensor("xi_51_cast_fp16")]; + tensor var_3184_cast_fp16 = mul(x = xr_51_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3184_cast_fp16")]; + tensor var_3185_cast_fp16 = mul(x = xi_51_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3185_cast_fp16")]; + tensor xor__51_cast_fp16 = sub(x = var_3184_cast_fp16, y = var_3185_cast_fp16)[name = tensor("xor__51_cast_fp16")]; + tensor var_3188_cast_fp16 = mul(x = xr_51_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3188_cast_fp16")]; + tensor var_3189_cast_fp16 = mul(x = xi_51_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3189_cast_fp16")]; + tensor xoi_51_cast_fp16 = add(x = var_3188_cast_fp16, y = var_3189_cast_fp16)[name = tensor("xoi_51_cast_fp16")]; + tensor out_51_axis_0 = const()[name = tensor("out_51_axis_0"), val = tensor(-1)]; + tensor out_51_cast_fp16 = stack(axis = out_51_axis_0, values = (xor__51_cast_fp16, xoi_51_cast_fp16))[name = tensor("out_51_cast_fp16")]; + tensor var_3206 = const()[name = tensor("op_3206"), val = tensor([256, 1024])]; + tensor var_3207_cast_fp16 = reshape(shape = var_3206, x = out_51_cast_fp16)[name = tensor("op_3207_cast_fp16")]; + tensor written_k_25_transpose_x_1 = const()[name = tensor("written_k_25_transpose_x_1"), val = tensor(true)]; + tensor written_k_25_transpose_y_1 = const()[name = tensor("written_k_25_transpose_y_1"), val = tensor(false)]; + tensor written_k_25_cast_fp16 = matmul(transpose_x = written_k_25_transpose_x_1, transpose_y = written_k_25_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_3207_cast_fp16)[name = tensor("written_k_25_cast_fp16")]; + tensor var_3214 = const()[name = tensor("op_3214"), val = tensor([256, 1024])]; + tensor var_3215_cast_fp16 = reshape(shape = var_3214, x = v_25_cast_fp16)[name = tensor("op_3215_cast_fp16")]; + tensor written_v_25_transpose_x_1 = const()[name = tensor("written_v_25_transpose_x_1"), val = tensor(true)]; + tensor written_v_25_transpose_y_1 = const()[name = tensor("written_v_25_transpose_y_1"), val = tensor(false)]; + tensor written_v_25_cast_fp16 = matmul(transpose_x = written_v_25_transpose_x_1, transpose_y = written_v_25_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_3215_cast_fp16)[name = tensor("written_v_25_cast_fp16")]; + tensor k_cache12_to_fp16_dtype_0 = const()[name = tensor("k_cache12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache12_to_fp16 = cast(dtype = k_cache12_to_fp16_dtype_0, x = k_cache12)[name = tensor("cast_393")]; + tensor var_3226_cast_fp16 = mul(x = k_cache12_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_3226_cast_fp16")]; + tensor var_3231 = const()[name = tensor("op_3231"), val = tensor([1, 512, 16, 64])]; + tensor var_3232_cast_fp16 = reshape(shape = var_3231, x = written_k_25_cast_fp16)[name = tensor("op_3232_cast_fp16")]; + tensor new_k_cache_25_cast_fp16 = add(x = var_3226_cast_fp16, y = var_3232_cast_fp16)[name = tensor("new_k_cache_25_cast_fp16")]; + tensor new_k_cache_25_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_25_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache12_to_fp16_dtype_0 = const()[name = tensor("v_cache12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache12_to_fp16 = cast(dtype = v_cache12_to_fp16_dtype_0, x = v_cache12)[name = tensor("cast_391")]; + tensor var_3235_cast_fp16 = mul(x = v_cache12_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_3235_cast_fp16")]; + tensor var_3240 = const()[name = tensor("op_3240"), val = tensor([1, 512, 16, 64])]; + tensor var_3241_cast_fp16 = reshape(shape = var_3240, x = written_v_25_cast_fp16)[name = tensor("op_3241_cast_fp16")]; + tensor new_v_cache_25_cast_fp16 = add(x = var_3235_cast_fp16, y = var_3241_cast_fp16)[name = tensor("new_v_cache_25_cast_fp16")]; + tensor new_v_cache_25_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_25_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_3254 = const()[name = tensor("op_3254"), val = tensor([0, 2, 1, 3])]; + tensor var_3265_transpose_x_0 = const()[name = tensor("op_3265_transpose_x_0"), val = tensor(false)]; + tensor var_3265_transpose_y_0 = const()[name = tensor("op_3265_transpose_y_0"), val = tensor(false)]; + tensor transpose_93_perm_0 = const()[name = tensor("transpose_93_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_94_perm_0 = const()[name = tensor("transpose_94_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_94 = transpose(perm = transpose_94_perm_0, x = new_k_cache_25_cast_fp16)[name = tensor("transpose_156")]; + tensor transpose_93 = transpose(perm = transpose_93_perm_0, x = q_25_cast_fp16)[name = tensor("transpose_157")]; + tensor var_3265_cast_fp16 = matmul(transpose_x = var_3265_transpose_x_0, transpose_y = var_3265_transpose_y_0, x = transpose_93, y = transpose_94)[name = tensor("op_3265_cast_fp16")]; + tensor var_3266_to_fp16 = const()[name = tensor("op_3266_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_49_cast_fp16 = mul(x = var_3265_cast_fp16, y = var_3266_to_fp16)[name = tensor("scores_49_cast_fp16")]; + tensor scores_51_cast_fp16 = add(x = scores_49_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_51_cast_fp16")]; + tensor var_3276 = const()[name = tensor("op_3276"), val = tensor(-1)]; + tensor weights_25_cast_fp16 = softmax(axis = var_3276, x = scores_51_cast_fp16)[name = tensor("weights_25_cast_fp16")]; + tensor attn_25_transpose_x_0 = const()[name = tensor("attn_25_transpose_x_0"), val = tensor(false)]; + tensor attn_25_transpose_y_0 = const()[name = tensor("attn_25_transpose_y_0"), val = tensor(false)]; + tensor values_25_cast_fp16 = transpose(perm = var_3254, x = new_v_cache_25_cast_fp16)[name = tensor("transpose_158")]; + tensor attn_25_cast_fp16 = matmul(transpose_x = attn_25_transpose_x_0, transpose_y = attn_25_transpose_y_0, x = weights_25_cast_fp16, y = values_25_cast_fp16)[name = tensor("attn_25_cast_fp16")]; + tensor var_3284 = const()[name = tensor("op_3284"), val = tensor([0, 2, 1, 3])]; + tensor var_3289 = const()[name = tensor("op_3289"), val = tensor([1, 256, 1024])]; + tensor var_3285_cast_fp16 = transpose(perm = var_3284, x = attn_25_cast_fp16)[name = tensor("transpose_155")]; + tensor input_149_cast_fp16 = reshape(shape = var_3289, x = var_3285_cast_fp16)[name = tensor("input_149_cast_fp16")]; + tensor attn12_out_proj_weight_to_fp16 = const()[name = tensor("attn12_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(308409088)))]; + tensor linear_49_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn12_out_proj_weight_to_fp16, x = input_149_cast_fp16)[name = tensor("linear_49_cast_fp16")]; + tensor position12_to_fp16_dtype_0 = const()[name = tensor("position12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position12_to_fp16 = cast(dtype = position12_to_fp16_dtype_0, x = position12)[name = tensor("cast_389")]; + tensor var_3298_cast_fp16 = add(x = position12_to_fp16, y = valid_len_to_fp16)[name = tensor("op_3298_cast_fp16")]; + tensor var_3298_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3298_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_151_cast_fp16 = add(x = input_145_cast_fp16, y = linear_49_cast_fp16)[name = tensor("input_151_cast_fp16")]; + tensor input_153_axes_0 = const()[name = tensor("input_153_axes_0"), val = tensor([-1])]; + tensor norm12_2_weight_to_fp16 = const()[name = tensor("norm12_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(310506304)))]; + tensor norm12_2_bias_to_fp16 = const()[name = tensor("norm12_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(310508416)))]; + tensor var_3302_to_fp16 = const()[name = tensor("op_3302_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_153_cast_fp16 = layer_norm(axes = input_153_axes_0, beta = norm12_2_bias_to_fp16, epsilon = var_3302_to_fp16, gamma = norm12_2_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("input_153_cast_fp16")]; + tensor linear12_1_weight_to_fp16 = const()[name = tensor("linear12_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(310510528)))]; + tensor linear_50_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear12_1_weight_to_fp16, x = input_153_cast_fp16)[name = tensor("linear_50_cast_fp16")]; + tensor input_155_mode_0 = const()[name = tensor("input_155_mode_0"), val = tensor("EXACT")]; + tensor input_155_cast_fp16 = gelu(mode = input_155_mode_0, x = linear_50_cast_fp16)[name = tensor("input_155_cast_fp16")]; + tensor linear12_2_weight_to_fp16 = const()[name = tensor("linear12_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(318899200)))]; + tensor linear_51_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear12_2_weight_to_fp16, x = input_155_cast_fp16)[name = tensor("linear_51_cast_fp16")]; + tensor input_157_cast_fp16 = add(x = input_151_cast_fp16, y = linear_51_cast_fp16)[name = tensor("input_157_cast_fp16")]; + tensor input_159_axes_0 = const()[name = tensor("input_159_axes_0"), val = tensor([-1])]; + tensor norm13_1_weight_to_fp16 = const()[name = tensor("norm13_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(327287872)))]; + tensor norm13_1_bias_to_fp16 = const()[name = tensor("norm13_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(327289984)))]; + tensor var_3319_to_fp16 = const()[name = tensor("op_3319_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_159_cast_fp16 = layer_norm(axes = input_159_axes_0, beta = norm13_1_bias_to_fp16, epsilon = var_3319_to_fp16, gamma = norm13_1_weight_to_fp16, x = input_157_cast_fp16)[name = tensor("input_159_cast_fp16")]; + tensor attn13_in_proj_weight_to_fp16 = const()[name = tensor("attn13_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(327292096)))]; + tensor linear_52_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn13_in_proj_weight_to_fp16, x = input_159_cast_fp16)[name = tensor("linear_52_cast_fp16")]; + tensor var_3333 = const()[name = tensor("op_3333"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_27_cast_fp16 = reshape(shape = var_3333, x = linear_52_cast_fp16)[name = tensor("qkv_27_cast_fp16")]; + tensor x_53_begin_0 = const()[name = tensor("x_53_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_53_end_0 = const()[name = tensor("x_53_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_53_end_mask_0 = const()[name = tensor("x_53_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_53_squeeze_mask_0 = const()[name = tensor("x_53_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_53_cast_fp16 = slice_by_index(begin = x_53_begin_0, end = x_53_end_0, end_mask = x_53_end_mask_0, squeeze_mask = x_53_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("x_53_cast_fp16")]; + tensor x_55_begin_0 = const()[name = tensor("x_55_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_55_end_0 = const()[name = tensor("x_55_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_55_end_mask_0 = const()[name = tensor("x_55_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_55_squeeze_mask_0 = const()[name = tensor("x_55_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_55_cast_fp16 = slice_by_index(begin = x_55_begin_0, end = x_55_end_0, end_mask = x_55_end_mask_0, squeeze_mask = x_55_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("x_55_cast_fp16")]; + tensor v_27_begin_0 = const()[name = tensor("v_27_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_27_end_0 = const()[name = tensor("v_27_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_27_end_mask_0 = const()[name = tensor("v_27_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_27_squeeze_mask_0 = const()[name = tensor("v_27_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_27_cast_fp16 = slice_by_index(begin = v_27_begin_0, end = v_27_end_0, end_mask = v_27_end_mask_0, squeeze_mask = v_27_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("v_27_cast_fp16")]; + tensor var_3378 = const()[name = tensor("op_3378"), val = tensor([256, 16, 32, 2])]; + tensor pairs_53_cast_fp16 = reshape(shape = var_3378, x = x_53_cast_fp16)[name = tensor("pairs_53_cast_fp16")]; + tensor xr_53_begin_0 = const()[name = tensor("xr_53_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_53_end_0 = const()[name = tensor("xr_53_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_53_end_mask_0 = const()[name = tensor("xr_53_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_53_squeeze_mask_0 = const()[name = tensor("xr_53_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_53_cast_fp16 = slice_by_index(begin = xr_53_begin_0, end = xr_53_end_0, end_mask = xr_53_end_mask_0, squeeze_mask = xr_53_squeeze_mask_0, x = pairs_53_cast_fp16)[name = tensor("xr_53_cast_fp16")]; + tensor xi_53_begin_0 = const()[name = tensor("xi_53_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_53_end_0 = const()[name = tensor("xi_53_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_53_end_mask_0 = const()[name = tensor("xi_53_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_53_squeeze_mask_0 = const()[name = tensor("xi_53_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_53_cast_fp16 = slice_by_index(begin = xi_53_begin_0, end = xi_53_end_0, end_mask = xi_53_end_mask_0, squeeze_mask = xi_53_squeeze_mask_0, x = pairs_53_cast_fp16)[name = tensor("xi_53_cast_fp16")]; + tensor var_3386_cast_fp16 = mul(x = xr_53_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3386_cast_fp16")]; + tensor var_3387_cast_fp16 = mul(x = xi_53_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3387_cast_fp16")]; + tensor xor__53_cast_fp16 = sub(x = var_3386_cast_fp16, y = var_3387_cast_fp16)[name = tensor("xor__53_cast_fp16")]; + tensor var_3390_cast_fp16 = mul(x = xr_53_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3390_cast_fp16")]; + tensor var_3391_cast_fp16 = mul(x = xi_53_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3391_cast_fp16")]; + tensor xoi_53_cast_fp16 = add(x = var_3390_cast_fp16, y = var_3391_cast_fp16)[name = tensor("xoi_53_cast_fp16")]; + tensor out_53_axis_0 = const()[name = tensor("out_53_axis_0"), val = tensor(-1)]; + tensor out_53_cast_fp16 = stack(axis = out_53_axis_0, values = (xor__53_cast_fp16, xoi_53_cast_fp16))[name = tensor("out_53_cast_fp16")]; + tensor var_3401 = const()[name = tensor("op_3401"), val = tensor([1, 256, 16, 64])]; + tensor q_27_cast_fp16 = reshape(shape = var_3401, x = out_53_cast_fp16)[name = tensor("q_27_cast_fp16")]; + tensor var_3407 = const()[name = tensor("op_3407"), val = tensor([256, 16, 32, 2])]; + tensor pairs_55_cast_fp16 = reshape(shape = var_3407, x = x_55_cast_fp16)[name = tensor("pairs_55_cast_fp16")]; + tensor xr_55_begin_0 = const()[name = tensor("xr_55_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_55_end_0 = const()[name = tensor("xr_55_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_55_end_mask_0 = const()[name = tensor("xr_55_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_55_squeeze_mask_0 = const()[name = tensor("xr_55_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_55_cast_fp16 = slice_by_index(begin = xr_55_begin_0, end = xr_55_end_0, end_mask = xr_55_end_mask_0, squeeze_mask = xr_55_squeeze_mask_0, x = pairs_55_cast_fp16)[name = tensor("xr_55_cast_fp16")]; + tensor xi_55_begin_0 = const()[name = tensor("xi_55_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_55_end_0 = const()[name = tensor("xi_55_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_55_end_mask_0 = const()[name = tensor("xi_55_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_55_squeeze_mask_0 = const()[name = tensor("xi_55_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_55_cast_fp16 = slice_by_index(begin = xi_55_begin_0, end = xi_55_end_0, end_mask = xi_55_end_mask_0, squeeze_mask = xi_55_squeeze_mask_0, x = pairs_55_cast_fp16)[name = tensor("xi_55_cast_fp16")]; + tensor var_3415_cast_fp16 = mul(x = xr_55_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3415_cast_fp16")]; + tensor var_3416_cast_fp16 = mul(x = xi_55_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3416_cast_fp16")]; + tensor xor__55_cast_fp16 = sub(x = var_3415_cast_fp16, y = var_3416_cast_fp16)[name = tensor("xor__55_cast_fp16")]; + tensor var_3419_cast_fp16 = mul(x = xr_55_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3419_cast_fp16")]; + tensor var_3420_cast_fp16 = mul(x = xi_55_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3420_cast_fp16")]; + tensor xoi_55_cast_fp16 = add(x = var_3419_cast_fp16, y = var_3420_cast_fp16)[name = tensor("xoi_55_cast_fp16")]; + tensor out_55_axis_0 = const()[name = tensor("out_55_axis_0"), val = tensor(-1)]; + tensor out_55_cast_fp16 = stack(axis = out_55_axis_0, values = (xor__55_cast_fp16, xoi_55_cast_fp16))[name = tensor("out_55_cast_fp16")]; + tensor var_3437 = const()[name = tensor("op_3437"), val = tensor([256, 1024])]; + tensor var_3438_cast_fp16 = reshape(shape = var_3437, x = out_55_cast_fp16)[name = tensor("op_3438_cast_fp16")]; + tensor written_k_27_transpose_x_1 = const()[name = tensor("written_k_27_transpose_x_1"), val = tensor(true)]; + tensor written_k_27_transpose_y_1 = const()[name = tensor("written_k_27_transpose_y_1"), val = tensor(false)]; + tensor written_k_27_cast_fp16 = matmul(transpose_x = written_k_27_transpose_x_1, transpose_y = written_k_27_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_3438_cast_fp16)[name = tensor("written_k_27_cast_fp16")]; + tensor var_3445 = const()[name = tensor("op_3445"), val = tensor([256, 1024])]; + tensor var_3446_cast_fp16 = reshape(shape = var_3445, x = v_27_cast_fp16)[name = tensor("op_3446_cast_fp16")]; + tensor written_v_27_transpose_x_1 = const()[name = tensor("written_v_27_transpose_x_1"), val = tensor(true)]; + tensor written_v_27_transpose_y_1 = const()[name = tensor("written_v_27_transpose_y_1"), val = tensor(false)]; + tensor written_v_27_cast_fp16 = matmul(transpose_x = written_v_27_transpose_x_1, transpose_y = written_v_27_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_3446_cast_fp16)[name = tensor("written_v_27_cast_fp16")]; + tensor k_cache13_to_fp16_dtype_0 = const()[name = tensor("k_cache13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache13_to_fp16 = cast(dtype = k_cache13_to_fp16_dtype_0, x = k_cache13)[name = tensor("cast_387")]; + tensor var_3457_cast_fp16 = mul(x = k_cache13_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_3457_cast_fp16")]; + tensor var_3462 = const()[name = tensor("op_3462"), val = tensor([1, 512, 16, 64])]; + tensor var_3463_cast_fp16 = reshape(shape = var_3462, x = written_k_27_cast_fp16)[name = tensor("op_3463_cast_fp16")]; + tensor new_k_cache_27_cast_fp16 = add(x = var_3457_cast_fp16, y = var_3463_cast_fp16)[name = tensor("new_k_cache_27_cast_fp16")]; + tensor new_k_cache_27_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_27_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache13_to_fp16_dtype_0 = const()[name = tensor("v_cache13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache13_to_fp16 = cast(dtype = v_cache13_to_fp16_dtype_0, x = v_cache13)[name = tensor("cast_385")]; + tensor var_3466_cast_fp16 = mul(x = v_cache13_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_3466_cast_fp16")]; + tensor var_3471 = const()[name = tensor("op_3471"), val = tensor([1, 512, 16, 64])]; + tensor var_3472_cast_fp16 = reshape(shape = var_3471, x = written_v_27_cast_fp16)[name = tensor("op_3472_cast_fp16")]; + tensor new_v_cache_27_cast_fp16 = add(x = var_3466_cast_fp16, y = var_3472_cast_fp16)[name = tensor("new_v_cache_27_cast_fp16")]; + tensor new_v_cache_27_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_27_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_3485 = const()[name = tensor("op_3485"), val = tensor([0, 2, 1, 3])]; + tensor var_3496_transpose_x_0 = const()[name = tensor("op_3496_transpose_x_0"), val = tensor(false)]; + tensor var_3496_transpose_y_0 = const()[name = tensor("op_3496_transpose_y_0"), val = tensor(false)]; + tensor transpose_95_perm_0 = const()[name = tensor("transpose_95_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_96_perm_0 = const()[name = tensor("transpose_96_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_96 = transpose(perm = transpose_96_perm_0, x = new_k_cache_27_cast_fp16)[name = tensor("transpose_152")]; + tensor transpose_95 = transpose(perm = transpose_95_perm_0, x = q_27_cast_fp16)[name = tensor("transpose_153")]; + tensor var_3496_cast_fp16 = matmul(transpose_x = var_3496_transpose_x_0, transpose_y = var_3496_transpose_y_0, x = transpose_95, y = transpose_96)[name = tensor("op_3496_cast_fp16")]; + tensor var_3497_to_fp16 = const()[name = tensor("op_3497_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_53_cast_fp16 = mul(x = var_3496_cast_fp16, y = var_3497_to_fp16)[name = tensor("scores_53_cast_fp16")]; + tensor scores_55_cast_fp16 = add(x = scores_53_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_55_cast_fp16")]; + tensor var_3507 = const()[name = tensor("op_3507"), val = tensor(-1)]; + tensor weights_27_cast_fp16 = softmax(axis = var_3507, x = scores_55_cast_fp16)[name = tensor("weights_27_cast_fp16")]; + tensor attn_27_transpose_x_0 = const()[name = tensor("attn_27_transpose_x_0"), val = tensor(false)]; + tensor attn_27_transpose_y_0 = const()[name = tensor("attn_27_transpose_y_0"), val = tensor(false)]; + tensor values_27_cast_fp16 = transpose(perm = var_3485, x = new_v_cache_27_cast_fp16)[name = tensor("transpose_154")]; + tensor attn_27_cast_fp16 = matmul(transpose_x = attn_27_transpose_x_0, transpose_y = attn_27_transpose_y_0, x = weights_27_cast_fp16, y = values_27_cast_fp16)[name = tensor("attn_27_cast_fp16")]; + tensor var_3515 = const()[name = tensor("op_3515"), val = tensor([0, 2, 1, 3])]; + tensor var_3520 = const()[name = tensor("op_3520"), val = tensor([1, 256, 1024])]; + tensor var_3516_cast_fp16 = transpose(perm = var_3515, x = attn_27_cast_fp16)[name = tensor("transpose_151")]; + tensor input_161_cast_fp16 = reshape(shape = var_3520, x = var_3516_cast_fp16)[name = tensor("input_161_cast_fp16")]; + tensor attn13_out_proj_weight_to_fp16 = const()[name = tensor("attn13_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(333583616)))]; + tensor linear_53_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn13_out_proj_weight_to_fp16, x = input_161_cast_fp16)[name = tensor("linear_53_cast_fp16")]; + tensor position13_to_fp16_dtype_0 = const()[name = tensor("position13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position13_to_fp16 = cast(dtype = position13_to_fp16_dtype_0, x = position13)[name = tensor("cast_383")]; + tensor var_3529_cast_fp16 = add(x = position13_to_fp16, y = valid_len_to_fp16)[name = tensor("op_3529_cast_fp16")]; + tensor var_3529_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3529_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_163_cast_fp16 = add(x = input_157_cast_fp16, y = linear_53_cast_fp16)[name = tensor("input_163_cast_fp16")]; + tensor input_165_axes_0 = const()[name = tensor("input_165_axes_0"), val = tensor([-1])]; + tensor norm13_2_weight_to_fp16 = const()[name = tensor("norm13_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335680832)))]; + tensor norm13_2_bias_to_fp16 = const()[name = tensor("norm13_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335682944)))]; + tensor var_3533_to_fp16 = const()[name = tensor("op_3533_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_165_cast_fp16 = layer_norm(axes = input_165_axes_0, beta = norm13_2_bias_to_fp16, epsilon = var_3533_to_fp16, gamma = norm13_2_weight_to_fp16, x = input_163_cast_fp16)[name = tensor("input_165_cast_fp16")]; + tensor linear13_1_weight_to_fp16 = const()[name = tensor("linear13_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335685056)))]; + tensor linear_54_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear13_1_weight_to_fp16, x = input_165_cast_fp16)[name = tensor("linear_54_cast_fp16")]; + tensor input_167_mode_0 = const()[name = tensor("input_167_mode_0"), val = tensor("EXACT")]; + tensor input_167_cast_fp16 = gelu(mode = input_167_mode_0, x = linear_54_cast_fp16)[name = tensor("input_167_cast_fp16")]; + tensor linear13_2_weight_to_fp16 = const()[name = tensor("linear13_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(344073728)))]; + tensor linear_55_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear13_2_weight_to_fp16, x = input_167_cast_fp16)[name = tensor("linear_55_cast_fp16")]; + tensor input_169_cast_fp16 = add(x = input_163_cast_fp16, y = linear_55_cast_fp16)[name = tensor("input_169_cast_fp16")]; + tensor input_171_axes_0 = const()[name = tensor("input_171_axes_0"), val = tensor([-1])]; + tensor norm14_1_weight_to_fp16 = const()[name = tensor("norm14_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(352462400)))]; + tensor norm14_1_bias_to_fp16 = const()[name = tensor("norm14_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(352464512)))]; + tensor var_3550_to_fp16 = const()[name = tensor("op_3550_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_171_cast_fp16 = layer_norm(axes = input_171_axes_0, beta = norm14_1_bias_to_fp16, epsilon = var_3550_to_fp16, gamma = norm14_1_weight_to_fp16, x = input_169_cast_fp16)[name = tensor("input_171_cast_fp16")]; + tensor attn14_in_proj_weight_to_fp16 = const()[name = tensor("attn14_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(352466624)))]; + tensor linear_56_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn14_in_proj_weight_to_fp16, x = input_171_cast_fp16)[name = tensor("linear_56_cast_fp16")]; + tensor var_3564 = const()[name = tensor("op_3564"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_29_cast_fp16 = reshape(shape = var_3564, x = linear_56_cast_fp16)[name = tensor("qkv_29_cast_fp16")]; + tensor x_57_begin_0 = const()[name = tensor("x_57_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_57_end_0 = const()[name = tensor("x_57_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_57_end_mask_0 = const()[name = tensor("x_57_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_57_squeeze_mask_0 = const()[name = tensor("x_57_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_57_cast_fp16 = slice_by_index(begin = x_57_begin_0, end = x_57_end_0, end_mask = x_57_end_mask_0, squeeze_mask = x_57_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("x_57_cast_fp16")]; + tensor x_59_begin_0 = const()[name = tensor("x_59_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_59_end_0 = const()[name = tensor("x_59_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_59_end_mask_0 = const()[name = tensor("x_59_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_59_squeeze_mask_0 = const()[name = tensor("x_59_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_59_cast_fp16 = slice_by_index(begin = x_59_begin_0, end = x_59_end_0, end_mask = x_59_end_mask_0, squeeze_mask = x_59_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("x_59_cast_fp16")]; + tensor v_29_begin_0 = const()[name = tensor("v_29_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_29_end_0 = const()[name = tensor("v_29_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_29_end_mask_0 = const()[name = tensor("v_29_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_29_squeeze_mask_0 = const()[name = tensor("v_29_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_29_cast_fp16 = slice_by_index(begin = v_29_begin_0, end = v_29_end_0, end_mask = v_29_end_mask_0, squeeze_mask = v_29_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("v_29_cast_fp16")]; + tensor var_3609 = const()[name = tensor("op_3609"), val = tensor([256, 16, 32, 2])]; + tensor pairs_57_cast_fp16 = reshape(shape = var_3609, x = x_57_cast_fp16)[name = tensor("pairs_57_cast_fp16")]; + tensor xr_57_begin_0 = const()[name = tensor("xr_57_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_57_end_0 = const()[name = tensor("xr_57_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_57_end_mask_0 = const()[name = tensor("xr_57_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_57_squeeze_mask_0 = const()[name = tensor("xr_57_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_57_cast_fp16 = slice_by_index(begin = xr_57_begin_0, end = xr_57_end_0, end_mask = xr_57_end_mask_0, squeeze_mask = xr_57_squeeze_mask_0, x = pairs_57_cast_fp16)[name = tensor("xr_57_cast_fp16")]; + tensor xi_57_begin_0 = const()[name = tensor("xi_57_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_57_end_0 = const()[name = tensor("xi_57_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_57_end_mask_0 = const()[name = tensor("xi_57_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_57_squeeze_mask_0 = const()[name = tensor("xi_57_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_57_cast_fp16 = slice_by_index(begin = xi_57_begin_0, end = xi_57_end_0, end_mask = xi_57_end_mask_0, squeeze_mask = xi_57_squeeze_mask_0, x = pairs_57_cast_fp16)[name = tensor("xi_57_cast_fp16")]; + tensor var_3617_cast_fp16 = mul(x = xr_57_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3617_cast_fp16")]; + tensor var_3618_cast_fp16 = mul(x = xi_57_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3618_cast_fp16")]; + tensor xor__57_cast_fp16 = sub(x = var_3617_cast_fp16, y = var_3618_cast_fp16)[name = tensor("xor__57_cast_fp16")]; + tensor var_3621_cast_fp16 = mul(x = xr_57_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3621_cast_fp16")]; + tensor var_3622_cast_fp16 = mul(x = xi_57_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3622_cast_fp16")]; + tensor xoi_57_cast_fp16 = add(x = var_3621_cast_fp16, y = var_3622_cast_fp16)[name = tensor("xoi_57_cast_fp16")]; + tensor out_57_axis_0 = const()[name = tensor("out_57_axis_0"), val = tensor(-1)]; + tensor out_57_cast_fp16 = stack(axis = out_57_axis_0, values = (xor__57_cast_fp16, xoi_57_cast_fp16))[name = tensor("out_57_cast_fp16")]; + tensor var_3632 = const()[name = tensor("op_3632"), val = tensor([1, 256, 16, 64])]; + tensor q_29_cast_fp16 = reshape(shape = var_3632, x = out_57_cast_fp16)[name = tensor("q_29_cast_fp16")]; + tensor var_3638 = const()[name = tensor("op_3638"), val = tensor([256, 16, 32, 2])]; + tensor pairs_59_cast_fp16 = reshape(shape = var_3638, x = x_59_cast_fp16)[name = tensor("pairs_59_cast_fp16")]; + tensor xr_59_begin_0 = const()[name = tensor("xr_59_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_59_end_0 = const()[name = tensor("xr_59_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_59_end_mask_0 = const()[name = tensor("xr_59_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_59_squeeze_mask_0 = const()[name = tensor("xr_59_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_59_cast_fp16 = slice_by_index(begin = xr_59_begin_0, end = xr_59_end_0, end_mask = xr_59_end_mask_0, squeeze_mask = xr_59_squeeze_mask_0, x = pairs_59_cast_fp16)[name = tensor("xr_59_cast_fp16")]; + tensor xi_59_begin_0 = const()[name = tensor("xi_59_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_59_end_0 = const()[name = tensor("xi_59_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_59_end_mask_0 = const()[name = tensor("xi_59_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_59_squeeze_mask_0 = const()[name = tensor("xi_59_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_59_cast_fp16 = slice_by_index(begin = xi_59_begin_0, end = xi_59_end_0, end_mask = xi_59_end_mask_0, squeeze_mask = xi_59_squeeze_mask_0, x = pairs_59_cast_fp16)[name = tensor("xi_59_cast_fp16")]; + tensor var_3646_cast_fp16 = mul(x = xr_59_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3646_cast_fp16")]; + tensor var_3647_cast_fp16 = mul(x = xi_59_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3647_cast_fp16")]; + tensor xor__59_cast_fp16 = sub(x = var_3646_cast_fp16, y = var_3647_cast_fp16)[name = tensor("xor__59_cast_fp16")]; + tensor var_3650_cast_fp16 = mul(x = xr_59_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3650_cast_fp16")]; + tensor var_3651_cast_fp16 = mul(x = xi_59_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3651_cast_fp16")]; + tensor xoi_59_cast_fp16 = add(x = var_3650_cast_fp16, y = var_3651_cast_fp16)[name = tensor("xoi_59_cast_fp16")]; + tensor out_59_axis_0 = const()[name = tensor("out_59_axis_0"), val = tensor(-1)]; + tensor out_59_cast_fp16 = stack(axis = out_59_axis_0, values = (xor__59_cast_fp16, xoi_59_cast_fp16))[name = tensor("out_59_cast_fp16")]; + tensor var_3668 = const()[name = tensor("op_3668"), val = tensor([256, 1024])]; + tensor var_3669_cast_fp16 = reshape(shape = var_3668, x = out_59_cast_fp16)[name = tensor("op_3669_cast_fp16")]; + tensor written_k_29_transpose_x_1 = const()[name = tensor("written_k_29_transpose_x_1"), val = tensor(true)]; + tensor written_k_29_transpose_y_1 = const()[name = tensor("written_k_29_transpose_y_1"), val = tensor(false)]; + tensor written_k_29_cast_fp16 = matmul(transpose_x = written_k_29_transpose_x_1, transpose_y = written_k_29_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_3669_cast_fp16)[name = tensor("written_k_29_cast_fp16")]; + tensor var_3676 = const()[name = tensor("op_3676"), val = tensor([256, 1024])]; + tensor var_3677_cast_fp16 = reshape(shape = var_3676, x = v_29_cast_fp16)[name = tensor("op_3677_cast_fp16")]; + tensor written_v_29_transpose_x_1 = const()[name = tensor("written_v_29_transpose_x_1"), val = tensor(true)]; + tensor written_v_29_transpose_y_1 = const()[name = tensor("written_v_29_transpose_y_1"), val = tensor(false)]; + tensor written_v_29_cast_fp16 = matmul(transpose_x = written_v_29_transpose_x_1, transpose_y = written_v_29_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_3677_cast_fp16)[name = tensor("written_v_29_cast_fp16")]; + tensor k_cache14_to_fp16_dtype_0 = const()[name = tensor("k_cache14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache14_to_fp16 = cast(dtype = k_cache14_to_fp16_dtype_0, x = k_cache14)[name = tensor("cast_381")]; + tensor var_3688_cast_fp16 = mul(x = k_cache14_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_3688_cast_fp16")]; + tensor var_3693 = const()[name = tensor("op_3693"), val = tensor([1, 512, 16, 64])]; + tensor var_3694_cast_fp16 = reshape(shape = var_3693, x = written_k_29_cast_fp16)[name = tensor("op_3694_cast_fp16")]; + tensor new_k_cache_29_cast_fp16 = add(x = var_3688_cast_fp16, y = var_3694_cast_fp16)[name = tensor("new_k_cache_29_cast_fp16")]; + tensor new_k_cache_29_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_29_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache14_to_fp16_dtype_0 = const()[name = tensor("v_cache14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache14_to_fp16 = cast(dtype = v_cache14_to_fp16_dtype_0, x = v_cache14)[name = tensor("cast_379")]; + tensor var_3697_cast_fp16 = mul(x = v_cache14_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_3697_cast_fp16")]; + tensor var_3702 = const()[name = tensor("op_3702"), val = tensor([1, 512, 16, 64])]; + tensor var_3703_cast_fp16 = reshape(shape = var_3702, x = written_v_29_cast_fp16)[name = tensor("op_3703_cast_fp16")]; + tensor new_v_cache_29_cast_fp16 = add(x = var_3697_cast_fp16, y = var_3703_cast_fp16)[name = tensor("new_v_cache_29_cast_fp16")]; + tensor new_v_cache_29_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_29_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_3716 = const()[name = tensor("op_3716"), val = tensor([0, 2, 1, 3])]; + tensor var_3727_transpose_x_0 = const()[name = tensor("op_3727_transpose_x_0"), val = tensor(false)]; + tensor var_3727_transpose_y_0 = const()[name = tensor("op_3727_transpose_y_0"), val = tensor(false)]; + tensor transpose_97_perm_0 = const()[name = tensor("transpose_97_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_98_perm_0 = const()[name = tensor("transpose_98_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_98 = transpose(perm = transpose_98_perm_0, x = new_k_cache_29_cast_fp16)[name = tensor("transpose_148")]; + tensor transpose_97 = transpose(perm = transpose_97_perm_0, x = q_29_cast_fp16)[name = tensor("transpose_149")]; + tensor var_3727_cast_fp16 = matmul(transpose_x = var_3727_transpose_x_0, transpose_y = var_3727_transpose_y_0, x = transpose_97, y = transpose_98)[name = tensor("op_3727_cast_fp16")]; + tensor var_3728_to_fp16 = const()[name = tensor("op_3728_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_57_cast_fp16 = mul(x = var_3727_cast_fp16, y = var_3728_to_fp16)[name = tensor("scores_57_cast_fp16")]; + tensor scores_59_cast_fp16 = add(x = scores_57_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_59_cast_fp16")]; + tensor var_3738 = const()[name = tensor("op_3738"), val = tensor(-1)]; + tensor weights_29_cast_fp16 = softmax(axis = var_3738, x = scores_59_cast_fp16)[name = tensor("weights_29_cast_fp16")]; + tensor attn_29_transpose_x_0 = const()[name = tensor("attn_29_transpose_x_0"), val = tensor(false)]; + tensor attn_29_transpose_y_0 = const()[name = tensor("attn_29_transpose_y_0"), val = tensor(false)]; + tensor values_29_cast_fp16 = transpose(perm = var_3716, x = new_v_cache_29_cast_fp16)[name = tensor("transpose_150")]; + tensor attn_29_cast_fp16 = matmul(transpose_x = attn_29_transpose_x_0, transpose_y = attn_29_transpose_y_0, x = weights_29_cast_fp16, y = values_29_cast_fp16)[name = tensor("attn_29_cast_fp16")]; + tensor var_3746 = const()[name = tensor("op_3746"), val = tensor([0, 2, 1, 3])]; + tensor var_3751 = const()[name = tensor("op_3751"), val = tensor([1, 256, 1024])]; + tensor var_3747_cast_fp16 = transpose(perm = var_3746, x = attn_29_cast_fp16)[name = tensor("transpose_147")]; + tensor input_173_cast_fp16 = reshape(shape = var_3751, x = var_3747_cast_fp16)[name = tensor("input_173_cast_fp16")]; + tensor attn14_out_proj_weight_to_fp16 = const()[name = tensor("attn14_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(358758144)))]; + tensor linear_57_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn14_out_proj_weight_to_fp16, x = input_173_cast_fp16)[name = tensor("linear_57_cast_fp16")]; + tensor position14_to_fp16_dtype_0 = const()[name = tensor("position14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position14_to_fp16 = cast(dtype = position14_to_fp16_dtype_0, x = position14)[name = tensor("cast_377")]; + tensor var_3760_cast_fp16 = add(x = position14_to_fp16, y = valid_len_to_fp16)[name = tensor("op_3760_cast_fp16")]; + tensor var_3760_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3760_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_175_cast_fp16 = add(x = input_169_cast_fp16, y = linear_57_cast_fp16)[name = tensor("input_175_cast_fp16")]; + tensor input_177_axes_0 = const()[name = tensor("input_177_axes_0"), val = tensor([-1])]; + tensor norm14_2_weight_to_fp16 = const()[name = tensor("norm14_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(360855360)))]; + tensor norm14_2_bias_to_fp16 = const()[name = tensor("norm14_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(360857472)))]; + tensor var_3764_to_fp16 = const()[name = tensor("op_3764_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_177_cast_fp16 = layer_norm(axes = input_177_axes_0, beta = norm14_2_bias_to_fp16, epsilon = var_3764_to_fp16, gamma = norm14_2_weight_to_fp16, x = input_175_cast_fp16)[name = tensor("input_177_cast_fp16")]; + tensor linear14_1_weight_to_fp16 = const()[name = tensor("linear14_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(360859584)))]; + tensor linear_58_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear14_1_weight_to_fp16, x = input_177_cast_fp16)[name = tensor("linear_58_cast_fp16")]; + tensor input_179_mode_0 = const()[name = tensor("input_179_mode_0"), val = tensor("EXACT")]; + tensor input_179_cast_fp16 = gelu(mode = input_179_mode_0, x = linear_58_cast_fp16)[name = tensor("input_179_cast_fp16")]; + tensor linear14_2_weight_to_fp16 = const()[name = tensor("linear14_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(369248256)))]; + tensor linear_59_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear14_2_weight_to_fp16, x = input_179_cast_fp16)[name = tensor("linear_59_cast_fp16")]; + tensor input_181_cast_fp16 = add(x = input_175_cast_fp16, y = linear_59_cast_fp16)[name = tensor("input_181_cast_fp16")]; + tensor input_183_axes_0 = const()[name = tensor("input_183_axes_0"), val = tensor([-1])]; + tensor norm15_1_weight_to_fp16 = const()[name = tensor("norm15_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(377636928)))]; + tensor norm15_1_bias_to_fp16 = const()[name = tensor("norm15_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(377639040)))]; + tensor var_3781_to_fp16 = const()[name = tensor("op_3781_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_183_cast_fp16 = layer_norm(axes = input_183_axes_0, beta = norm15_1_bias_to_fp16, epsilon = var_3781_to_fp16, gamma = norm15_1_weight_to_fp16, x = input_181_cast_fp16)[name = tensor("input_183_cast_fp16")]; + tensor attn15_in_proj_weight_to_fp16 = const()[name = tensor("attn15_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(377641152)))]; + tensor linear_60_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn15_in_proj_weight_to_fp16, x = input_183_cast_fp16)[name = tensor("linear_60_cast_fp16")]; + tensor var_3795 = const()[name = tensor("op_3795"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_31_cast_fp16 = reshape(shape = var_3795, x = linear_60_cast_fp16)[name = tensor("qkv_31_cast_fp16")]; + tensor x_61_begin_0 = const()[name = tensor("x_61_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_61_end_0 = const()[name = tensor("x_61_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_61_end_mask_0 = const()[name = tensor("x_61_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_61_squeeze_mask_0 = const()[name = tensor("x_61_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_61_cast_fp16 = slice_by_index(begin = x_61_begin_0, end = x_61_end_0, end_mask = x_61_end_mask_0, squeeze_mask = x_61_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("x_61_cast_fp16")]; + tensor x_63_begin_0 = const()[name = tensor("x_63_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_63_end_0 = const()[name = tensor("x_63_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_63_end_mask_0 = const()[name = tensor("x_63_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_63_squeeze_mask_0 = const()[name = tensor("x_63_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_63_cast_fp16 = slice_by_index(begin = x_63_begin_0, end = x_63_end_0, end_mask = x_63_end_mask_0, squeeze_mask = x_63_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("x_63_cast_fp16")]; + tensor v_31_begin_0 = const()[name = tensor("v_31_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_31_end_0 = const()[name = tensor("v_31_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_31_end_mask_0 = const()[name = tensor("v_31_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_31_squeeze_mask_0 = const()[name = tensor("v_31_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_31_cast_fp16 = slice_by_index(begin = v_31_begin_0, end = v_31_end_0, end_mask = v_31_end_mask_0, squeeze_mask = v_31_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("v_31_cast_fp16")]; + tensor var_3840 = const()[name = tensor("op_3840"), val = tensor([256, 16, 32, 2])]; + tensor pairs_61_cast_fp16 = reshape(shape = var_3840, x = x_61_cast_fp16)[name = tensor("pairs_61_cast_fp16")]; + tensor xr_61_begin_0 = const()[name = tensor("xr_61_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_61_end_0 = const()[name = tensor("xr_61_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_61_end_mask_0 = const()[name = tensor("xr_61_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_61_squeeze_mask_0 = const()[name = tensor("xr_61_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_61_cast_fp16 = slice_by_index(begin = xr_61_begin_0, end = xr_61_end_0, end_mask = xr_61_end_mask_0, squeeze_mask = xr_61_squeeze_mask_0, x = pairs_61_cast_fp16)[name = tensor("xr_61_cast_fp16")]; + tensor xi_61_begin_0 = const()[name = tensor("xi_61_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_61_end_0 = const()[name = tensor("xi_61_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_61_end_mask_0 = const()[name = tensor("xi_61_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_61_squeeze_mask_0 = const()[name = tensor("xi_61_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_61_cast_fp16 = slice_by_index(begin = xi_61_begin_0, end = xi_61_end_0, end_mask = xi_61_end_mask_0, squeeze_mask = xi_61_squeeze_mask_0, x = pairs_61_cast_fp16)[name = tensor("xi_61_cast_fp16")]; + tensor var_3848_cast_fp16 = mul(x = xr_61_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3848_cast_fp16")]; + tensor var_3849_cast_fp16 = mul(x = xi_61_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3849_cast_fp16")]; + tensor xor__61_cast_fp16 = sub(x = var_3848_cast_fp16, y = var_3849_cast_fp16)[name = tensor("xor__61_cast_fp16")]; + tensor var_3852_cast_fp16 = mul(x = xr_61_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3852_cast_fp16")]; + tensor var_3853_cast_fp16 = mul(x = xi_61_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3853_cast_fp16")]; + tensor xoi_61_cast_fp16 = add(x = var_3852_cast_fp16, y = var_3853_cast_fp16)[name = tensor("xoi_61_cast_fp16")]; + tensor out_61_axis_0 = const()[name = tensor("out_61_axis_0"), val = tensor(-1)]; + tensor out_61_cast_fp16 = stack(axis = out_61_axis_0, values = (xor__61_cast_fp16, xoi_61_cast_fp16))[name = tensor("out_61_cast_fp16")]; + tensor var_3863 = const()[name = tensor("op_3863"), val = tensor([1, 256, 16, 64])]; + tensor q_31_cast_fp16 = reshape(shape = var_3863, x = out_61_cast_fp16)[name = tensor("q_31_cast_fp16")]; + tensor var_3869 = const()[name = tensor("op_3869"), val = tensor([256, 16, 32, 2])]; + tensor pairs_63_cast_fp16 = reshape(shape = var_3869, x = x_63_cast_fp16)[name = tensor("pairs_63_cast_fp16")]; + tensor xr_63_begin_0 = const()[name = tensor("xr_63_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_63_end_0 = const()[name = tensor("xr_63_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_63_end_mask_0 = const()[name = tensor("xr_63_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_63_squeeze_mask_0 = const()[name = tensor("xr_63_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_63_cast_fp16 = slice_by_index(begin = xr_63_begin_0, end = xr_63_end_0, end_mask = xr_63_end_mask_0, squeeze_mask = xr_63_squeeze_mask_0, x = pairs_63_cast_fp16)[name = tensor("xr_63_cast_fp16")]; + tensor xi_63_begin_0 = const()[name = tensor("xi_63_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_63_end_0 = const()[name = tensor("xi_63_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_63_end_mask_0 = const()[name = tensor("xi_63_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_63_squeeze_mask_0 = const()[name = tensor("xi_63_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_63_cast_fp16 = slice_by_index(begin = xi_63_begin_0, end = xi_63_end_0, end_mask = xi_63_end_mask_0, squeeze_mask = xi_63_squeeze_mask_0, x = pairs_63_cast_fp16)[name = tensor("xi_63_cast_fp16")]; + tensor var_3877_cast_fp16 = mul(x = xr_63_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3877_cast_fp16")]; + tensor var_3878_cast_fp16 = mul(x = xi_63_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3878_cast_fp16")]; + tensor xor__63_cast_fp16 = sub(x = var_3877_cast_fp16, y = var_3878_cast_fp16)[name = tensor("xor__63_cast_fp16")]; + tensor var_3881_cast_fp16 = mul(x = xr_63_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3881_cast_fp16")]; + tensor var_3882_cast_fp16 = mul(x = xi_63_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3882_cast_fp16")]; + tensor xoi_63_cast_fp16 = add(x = var_3881_cast_fp16, y = var_3882_cast_fp16)[name = tensor("xoi_63_cast_fp16")]; + tensor out_63_axis_0 = const()[name = tensor("out_63_axis_0"), val = tensor(-1)]; + tensor out_63_cast_fp16 = stack(axis = out_63_axis_0, values = (xor__63_cast_fp16, xoi_63_cast_fp16))[name = tensor("out_63_cast_fp16")]; + tensor var_3899 = const()[name = tensor("op_3899"), val = tensor([256, 1024])]; + tensor var_3900_cast_fp16 = reshape(shape = var_3899, x = out_63_cast_fp16)[name = tensor("op_3900_cast_fp16")]; + tensor written_k_31_transpose_x_1 = const()[name = tensor("written_k_31_transpose_x_1"), val = tensor(true)]; + tensor written_k_31_transpose_y_1 = const()[name = tensor("written_k_31_transpose_y_1"), val = tensor(false)]; + tensor written_k_31_cast_fp16 = matmul(transpose_x = written_k_31_transpose_x_1, transpose_y = written_k_31_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_3900_cast_fp16)[name = tensor("written_k_31_cast_fp16")]; + tensor var_3907 = const()[name = tensor("op_3907"), val = tensor([256, 1024])]; + tensor var_3908_cast_fp16 = reshape(shape = var_3907, x = v_31_cast_fp16)[name = tensor("op_3908_cast_fp16")]; + tensor written_v_31_transpose_x_1 = const()[name = tensor("written_v_31_transpose_x_1"), val = tensor(true)]; + tensor written_v_31_transpose_y_1 = const()[name = tensor("written_v_31_transpose_y_1"), val = tensor(false)]; + tensor written_v_31_cast_fp16 = matmul(transpose_x = written_v_31_transpose_x_1, transpose_y = written_v_31_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_3908_cast_fp16)[name = tensor("written_v_31_cast_fp16")]; + tensor k_cache15_to_fp16_dtype_0 = const()[name = tensor("k_cache15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache15_to_fp16 = cast(dtype = k_cache15_to_fp16_dtype_0, x = k_cache15)[name = tensor("cast_375")]; + tensor var_3919_cast_fp16 = mul(x = k_cache15_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_3919_cast_fp16")]; + tensor var_3924 = const()[name = tensor("op_3924"), val = tensor([1, 512, 16, 64])]; + tensor var_3925_cast_fp16 = reshape(shape = var_3924, x = written_k_31_cast_fp16)[name = tensor("op_3925_cast_fp16")]; + tensor new_k_cache_31_cast_fp16 = add(x = var_3919_cast_fp16, y = var_3925_cast_fp16)[name = tensor("new_k_cache_31_cast_fp16")]; + tensor new_k_cache_31_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_31_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache15_to_fp16_dtype_0 = const()[name = tensor("v_cache15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache15_to_fp16 = cast(dtype = v_cache15_to_fp16_dtype_0, x = v_cache15)[name = tensor("cast_373")]; + tensor var_3928_cast_fp16 = mul(x = v_cache15_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_3928_cast_fp16")]; + tensor var_3933 = const()[name = tensor("op_3933"), val = tensor([1, 512, 16, 64])]; + tensor var_3934_cast_fp16 = reshape(shape = var_3933, x = written_v_31_cast_fp16)[name = tensor("op_3934_cast_fp16")]; + tensor new_v_cache_31_cast_fp16 = add(x = var_3928_cast_fp16, y = var_3934_cast_fp16)[name = tensor("new_v_cache_31_cast_fp16")]; + tensor new_v_cache_31_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_31_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_3947 = const()[name = tensor("op_3947"), val = tensor([0, 2, 1, 3])]; + tensor var_3958_transpose_x_0 = const()[name = tensor("op_3958_transpose_x_0"), val = tensor(false)]; + tensor var_3958_transpose_y_0 = const()[name = tensor("op_3958_transpose_y_0"), val = tensor(false)]; + tensor transpose_99_perm_0 = const()[name = tensor("transpose_99_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_100_perm_0 = const()[name = tensor("transpose_100_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_100 = transpose(perm = transpose_100_perm_0, x = new_k_cache_31_cast_fp16)[name = tensor("transpose_144")]; + tensor transpose_99 = transpose(perm = transpose_99_perm_0, x = q_31_cast_fp16)[name = tensor("transpose_145")]; + tensor var_3958_cast_fp16 = matmul(transpose_x = var_3958_transpose_x_0, transpose_y = var_3958_transpose_y_0, x = transpose_99, y = transpose_100)[name = tensor("op_3958_cast_fp16")]; + tensor var_3959_to_fp16 = const()[name = tensor("op_3959_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_61_cast_fp16 = mul(x = var_3958_cast_fp16, y = var_3959_to_fp16)[name = tensor("scores_61_cast_fp16")]; + tensor scores_63_cast_fp16 = add(x = scores_61_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_63_cast_fp16")]; + tensor var_3969 = const()[name = tensor("op_3969"), val = tensor(-1)]; + tensor weights_31_cast_fp16 = softmax(axis = var_3969, x = scores_63_cast_fp16)[name = tensor("weights_31_cast_fp16")]; + tensor attn_31_transpose_x_0 = const()[name = tensor("attn_31_transpose_x_0"), val = tensor(false)]; + tensor attn_31_transpose_y_0 = const()[name = tensor("attn_31_transpose_y_0"), val = tensor(false)]; + tensor values_31_cast_fp16 = transpose(perm = var_3947, x = new_v_cache_31_cast_fp16)[name = tensor("transpose_146")]; + tensor attn_31_cast_fp16 = matmul(transpose_x = attn_31_transpose_x_0, transpose_y = attn_31_transpose_y_0, x = weights_31_cast_fp16, y = values_31_cast_fp16)[name = tensor("attn_31_cast_fp16")]; + tensor var_3977 = const()[name = tensor("op_3977"), val = tensor([0, 2, 1, 3])]; + tensor var_3982 = const()[name = tensor("op_3982"), val = tensor([1, 256, 1024])]; + tensor var_3978_cast_fp16 = transpose(perm = var_3977, x = attn_31_cast_fp16)[name = tensor("transpose_143")]; + tensor input_185_cast_fp16 = reshape(shape = var_3982, x = var_3978_cast_fp16)[name = tensor("input_185_cast_fp16")]; + tensor attn15_out_proj_weight_to_fp16 = const()[name = tensor("attn15_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(383932672)))]; + tensor linear_61_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn15_out_proj_weight_to_fp16, x = input_185_cast_fp16)[name = tensor("linear_61_cast_fp16")]; + tensor position15_to_fp16_dtype_0 = const()[name = tensor("position15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position15_to_fp16 = cast(dtype = position15_to_fp16_dtype_0, x = position15)[name = tensor("cast_371")]; + tensor var_3991_cast_fp16 = add(x = position15_to_fp16, y = valid_len_to_fp16)[name = tensor("op_3991_cast_fp16")]; + tensor var_3991_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3991_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_187_cast_fp16 = add(x = input_181_cast_fp16, y = linear_61_cast_fp16)[name = tensor("input_187_cast_fp16")]; + tensor input_189_axes_0 = const()[name = tensor("input_189_axes_0"), val = tensor([-1])]; + tensor norm15_2_weight_to_fp16 = const()[name = tensor("norm15_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(386029888)))]; + tensor norm15_2_bias_to_fp16 = const()[name = tensor("norm15_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(386032000)))]; + tensor var_3995_to_fp16 = const()[name = tensor("op_3995_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_189_cast_fp16 = layer_norm(axes = input_189_axes_0, beta = norm15_2_bias_to_fp16, epsilon = var_3995_to_fp16, gamma = norm15_2_weight_to_fp16, x = input_187_cast_fp16)[name = tensor("input_189_cast_fp16")]; + tensor linear15_1_weight_to_fp16 = const()[name = tensor("linear15_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(386034112)))]; + tensor linear_62_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear15_1_weight_to_fp16, x = input_189_cast_fp16)[name = tensor("linear_62_cast_fp16")]; + tensor input_191_mode_0 = const()[name = tensor("input_191_mode_0"), val = tensor("EXACT")]; + tensor input_191_cast_fp16 = gelu(mode = input_191_mode_0, x = linear_62_cast_fp16)[name = tensor("input_191_cast_fp16")]; + tensor linear15_2_weight_to_fp16 = const()[name = tensor("linear15_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(394422784)))]; + tensor linear_63_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear15_2_weight_to_fp16, x = input_191_cast_fp16)[name = tensor("linear_63_cast_fp16")]; + tensor input_193_cast_fp16 = add(x = input_187_cast_fp16, y = linear_63_cast_fp16)[name = tensor("input_193_cast_fp16")]; + tensor input_195_axes_0 = const()[name = tensor("input_195_axes_0"), val = tensor([-1])]; + tensor norm16_1_weight_to_fp16 = const()[name = tensor("norm16_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(402811456)))]; + tensor norm16_1_bias_to_fp16 = const()[name = tensor("norm16_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(402813568)))]; + tensor var_4012_to_fp16 = const()[name = tensor("op_4012_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_195_cast_fp16 = layer_norm(axes = input_195_axes_0, beta = norm16_1_bias_to_fp16, epsilon = var_4012_to_fp16, gamma = norm16_1_weight_to_fp16, x = input_193_cast_fp16)[name = tensor("input_195_cast_fp16")]; + tensor attn16_in_proj_weight_to_fp16 = const()[name = tensor("attn16_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(402815680)))]; + tensor linear_64_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn16_in_proj_weight_to_fp16, x = input_195_cast_fp16)[name = tensor("linear_64_cast_fp16")]; + tensor var_4026 = const()[name = tensor("op_4026"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_33_cast_fp16 = reshape(shape = var_4026, x = linear_64_cast_fp16)[name = tensor("qkv_33_cast_fp16")]; + tensor x_65_begin_0 = const()[name = tensor("x_65_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_65_end_0 = const()[name = tensor("x_65_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_65_end_mask_0 = const()[name = tensor("x_65_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_65_squeeze_mask_0 = const()[name = tensor("x_65_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_65_cast_fp16 = slice_by_index(begin = x_65_begin_0, end = x_65_end_0, end_mask = x_65_end_mask_0, squeeze_mask = x_65_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("x_65_cast_fp16")]; + tensor x_67_begin_0 = const()[name = tensor("x_67_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_67_end_0 = const()[name = tensor("x_67_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_67_end_mask_0 = const()[name = tensor("x_67_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_67_squeeze_mask_0 = const()[name = tensor("x_67_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_67_cast_fp16 = slice_by_index(begin = x_67_begin_0, end = x_67_end_0, end_mask = x_67_end_mask_0, squeeze_mask = x_67_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("x_67_cast_fp16")]; + tensor v_33_begin_0 = const()[name = tensor("v_33_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_33_end_0 = const()[name = tensor("v_33_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_33_end_mask_0 = const()[name = tensor("v_33_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_33_squeeze_mask_0 = const()[name = tensor("v_33_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_33_cast_fp16 = slice_by_index(begin = v_33_begin_0, end = v_33_end_0, end_mask = v_33_end_mask_0, squeeze_mask = v_33_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("v_33_cast_fp16")]; + tensor var_4071 = const()[name = tensor("op_4071"), val = tensor([256, 16, 32, 2])]; + tensor pairs_65_cast_fp16 = reshape(shape = var_4071, x = x_65_cast_fp16)[name = tensor("pairs_65_cast_fp16")]; + tensor xr_65_begin_0 = const()[name = tensor("xr_65_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_65_end_0 = const()[name = tensor("xr_65_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_65_end_mask_0 = const()[name = tensor("xr_65_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_65_squeeze_mask_0 = const()[name = tensor("xr_65_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_65_cast_fp16 = slice_by_index(begin = xr_65_begin_0, end = xr_65_end_0, end_mask = xr_65_end_mask_0, squeeze_mask = xr_65_squeeze_mask_0, x = pairs_65_cast_fp16)[name = tensor("xr_65_cast_fp16")]; + tensor xi_65_begin_0 = const()[name = tensor("xi_65_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_65_end_0 = const()[name = tensor("xi_65_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_65_end_mask_0 = const()[name = tensor("xi_65_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_65_squeeze_mask_0 = const()[name = tensor("xi_65_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_65_cast_fp16 = slice_by_index(begin = xi_65_begin_0, end = xi_65_end_0, end_mask = xi_65_end_mask_0, squeeze_mask = xi_65_squeeze_mask_0, x = pairs_65_cast_fp16)[name = tensor("xi_65_cast_fp16")]; + tensor var_4079_cast_fp16 = mul(x = xr_65_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4079_cast_fp16")]; + tensor var_4080_cast_fp16 = mul(x = xi_65_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4080_cast_fp16")]; + tensor xor__65_cast_fp16 = sub(x = var_4079_cast_fp16, y = var_4080_cast_fp16)[name = tensor("xor__65_cast_fp16")]; + tensor var_4083_cast_fp16 = mul(x = xr_65_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4083_cast_fp16")]; + tensor var_4084_cast_fp16 = mul(x = xi_65_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4084_cast_fp16")]; + tensor xoi_65_cast_fp16 = add(x = var_4083_cast_fp16, y = var_4084_cast_fp16)[name = tensor("xoi_65_cast_fp16")]; + tensor out_65_axis_0 = const()[name = tensor("out_65_axis_0"), val = tensor(-1)]; + tensor out_65_cast_fp16 = stack(axis = out_65_axis_0, values = (xor__65_cast_fp16, xoi_65_cast_fp16))[name = tensor("out_65_cast_fp16")]; + tensor var_4094 = const()[name = tensor("op_4094"), val = tensor([1, 256, 16, 64])]; + tensor q_33_cast_fp16 = reshape(shape = var_4094, x = out_65_cast_fp16)[name = tensor("q_33_cast_fp16")]; + tensor var_4100 = const()[name = tensor("op_4100"), val = tensor([256, 16, 32, 2])]; + tensor pairs_67_cast_fp16 = reshape(shape = var_4100, x = x_67_cast_fp16)[name = tensor("pairs_67_cast_fp16")]; + tensor xr_67_begin_0 = const()[name = tensor("xr_67_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_67_end_0 = const()[name = tensor("xr_67_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_67_end_mask_0 = const()[name = tensor("xr_67_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_67_squeeze_mask_0 = const()[name = tensor("xr_67_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_67_cast_fp16 = slice_by_index(begin = xr_67_begin_0, end = xr_67_end_0, end_mask = xr_67_end_mask_0, squeeze_mask = xr_67_squeeze_mask_0, x = pairs_67_cast_fp16)[name = tensor("xr_67_cast_fp16")]; + tensor xi_67_begin_0 = const()[name = tensor("xi_67_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_67_end_0 = const()[name = tensor("xi_67_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_67_end_mask_0 = const()[name = tensor("xi_67_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_67_squeeze_mask_0 = const()[name = tensor("xi_67_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_67_cast_fp16 = slice_by_index(begin = xi_67_begin_0, end = xi_67_end_0, end_mask = xi_67_end_mask_0, squeeze_mask = xi_67_squeeze_mask_0, x = pairs_67_cast_fp16)[name = tensor("xi_67_cast_fp16")]; + tensor var_4108_cast_fp16 = mul(x = xr_67_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4108_cast_fp16")]; + tensor var_4109_cast_fp16 = mul(x = xi_67_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4109_cast_fp16")]; + tensor xor__67_cast_fp16 = sub(x = var_4108_cast_fp16, y = var_4109_cast_fp16)[name = tensor("xor__67_cast_fp16")]; + tensor var_4112_cast_fp16 = mul(x = xr_67_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4112_cast_fp16")]; + tensor var_4113_cast_fp16 = mul(x = xi_67_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4113_cast_fp16")]; + tensor xoi_67_cast_fp16 = add(x = var_4112_cast_fp16, y = var_4113_cast_fp16)[name = tensor("xoi_67_cast_fp16")]; + tensor out_67_axis_0 = const()[name = tensor("out_67_axis_0"), val = tensor(-1)]; + tensor out_67_cast_fp16 = stack(axis = out_67_axis_0, values = (xor__67_cast_fp16, xoi_67_cast_fp16))[name = tensor("out_67_cast_fp16")]; + tensor var_4130 = const()[name = tensor("op_4130"), val = tensor([256, 1024])]; + tensor var_4131_cast_fp16 = reshape(shape = var_4130, x = out_67_cast_fp16)[name = tensor("op_4131_cast_fp16")]; + tensor written_k_33_transpose_x_1 = const()[name = tensor("written_k_33_transpose_x_1"), val = tensor(true)]; + tensor written_k_33_transpose_y_1 = const()[name = tensor("written_k_33_transpose_y_1"), val = tensor(false)]; + tensor written_k_33_cast_fp16 = matmul(transpose_x = written_k_33_transpose_x_1, transpose_y = written_k_33_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_4131_cast_fp16)[name = tensor("written_k_33_cast_fp16")]; + tensor var_4138 = const()[name = tensor("op_4138"), val = tensor([256, 1024])]; + tensor var_4139_cast_fp16 = reshape(shape = var_4138, x = v_33_cast_fp16)[name = tensor("op_4139_cast_fp16")]; + tensor written_v_33_transpose_x_1 = const()[name = tensor("written_v_33_transpose_x_1"), val = tensor(true)]; + tensor written_v_33_transpose_y_1 = const()[name = tensor("written_v_33_transpose_y_1"), val = tensor(false)]; + tensor written_v_33_cast_fp16 = matmul(transpose_x = written_v_33_transpose_x_1, transpose_y = written_v_33_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_4139_cast_fp16)[name = tensor("written_v_33_cast_fp16")]; + tensor k_cache16_to_fp16_dtype_0 = const()[name = tensor("k_cache16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache16_to_fp16 = cast(dtype = k_cache16_to_fp16_dtype_0, x = k_cache16)[name = tensor("cast_369")]; + tensor var_4150_cast_fp16 = mul(x = k_cache16_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_4150_cast_fp16")]; + tensor var_4155 = const()[name = tensor("op_4155"), val = tensor([1, 512, 16, 64])]; + tensor var_4156_cast_fp16 = reshape(shape = var_4155, x = written_k_33_cast_fp16)[name = tensor("op_4156_cast_fp16")]; + tensor new_k_cache_33_cast_fp16 = add(x = var_4150_cast_fp16, y = var_4156_cast_fp16)[name = tensor("new_k_cache_33_cast_fp16")]; + tensor new_k_cache_33_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_33_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache16_to_fp16_dtype_0 = const()[name = tensor("v_cache16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache16_to_fp16 = cast(dtype = v_cache16_to_fp16_dtype_0, x = v_cache16)[name = tensor("cast_367")]; + tensor var_4159_cast_fp16 = mul(x = v_cache16_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_4159_cast_fp16")]; + tensor var_4164 = const()[name = tensor("op_4164"), val = tensor([1, 512, 16, 64])]; + tensor var_4165_cast_fp16 = reshape(shape = var_4164, x = written_v_33_cast_fp16)[name = tensor("op_4165_cast_fp16")]; + tensor new_v_cache_33_cast_fp16 = add(x = var_4159_cast_fp16, y = var_4165_cast_fp16)[name = tensor("new_v_cache_33_cast_fp16")]; + tensor new_v_cache_33_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_33_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_4178 = const()[name = tensor("op_4178"), val = tensor([0, 2, 1, 3])]; + tensor var_4189_transpose_x_0 = const()[name = tensor("op_4189_transpose_x_0"), val = tensor(false)]; + tensor var_4189_transpose_y_0 = const()[name = tensor("op_4189_transpose_y_0"), val = tensor(false)]; + tensor transpose_101_perm_0 = const()[name = tensor("transpose_101_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_102_perm_0 = const()[name = tensor("transpose_102_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_102 = transpose(perm = transpose_102_perm_0, x = new_k_cache_33_cast_fp16)[name = tensor("transpose_140")]; + tensor transpose_101 = transpose(perm = transpose_101_perm_0, x = q_33_cast_fp16)[name = tensor("transpose_141")]; + tensor var_4189_cast_fp16 = matmul(transpose_x = var_4189_transpose_x_0, transpose_y = var_4189_transpose_y_0, x = transpose_101, y = transpose_102)[name = tensor("op_4189_cast_fp16")]; + tensor var_4190_to_fp16 = const()[name = tensor("op_4190_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_65_cast_fp16 = mul(x = var_4189_cast_fp16, y = var_4190_to_fp16)[name = tensor("scores_65_cast_fp16")]; + tensor scores_67_cast_fp16 = add(x = scores_65_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_67_cast_fp16")]; + tensor var_4200 = const()[name = tensor("op_4200"), val = tensor(-1)]; + tensor weights_33_cast_fp16 = softmax(axis = var_4200, x = scores_67_cast_fp16)[name = tensor("weights_33_cast_fp16")]; + tensor attn_33_transpose_x_0 = const()[name = tensor("attn_33_transpose_x_0"), val = tensor(false)]; + tensor attn_33_transpose_y_0 = const()[name = tensor("attn_33_transpose_y_0"), val = tensor(false)]; + tensor values_33_cast_fp16 = transpose(perm = var_4178, x = new_v_cache_33_cast_fp16)[name = tensor("transpose_142")]; + tensor attn_33_cast_fp16 = matmul(transpose_x = attn_33_transpose_x_0, transpose_y = attn_33_transpose_y_0, x = weights_33_cast_fp16, y = values_33_cast_fp16)[name = tensor("attn_33_cast_fp16")]; + tensor var_4208 = const()[name = tensor("op_4208"), val = tensor([0, 2, 1, 3])]; + tensor var_4213 = const()[name = tensor("op_4213"), val = tensor([1, 256, 1024])]; + tensor var_4209_cast_fp16 = transpose(perm = var_4208, x = attn_33_cast_fp16)[name = tensor("transpose_139")]; + tensor input_197_cast_fp16 = reshape(shape = var_4213, x = var_4209_cast_fp16)[name = tensor("input_197_cast_fp16")]; + tensor attn16_out_proj_weight_to_fp16 = const()[name = tensor("attn16_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(409107200)))]; + tensor linear_65_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn16_out_proj_weight_to_fp16, x = input_197_cast_fp16)[name = tensor("linear_65_cast_fp16")]; + tensor position16_to_fp16_dtype_0 = const()[name = tensor("position16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position16_to_fp16 = cast(dtype = position16_to_fp16_dtype_0, x = position16)[name = tensor("cast_365")]; + tensor var_4222_cast_fp16 = add(x = position16_to_fp16, y = valid_len_to_fp16)[name = tensor("op_4222_cast_fp16")]; + tensor var_4222_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4222_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_199_cast_fp16 = add(x = input_193_cast_fp16, y = linear_65_cast_fp16)[name = tensor("input_199_cast_fp16")]; + tensor input_201_axes_0 = const()[name = tensor("input_201_axes_0"), val = tensor([-1])]; + tensor norm16_2_weight_to_fp16 = const()[name = tensor("norm16_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411204416)))]; + tensor norm16_2_bias_to_fp16 = const()[name = tensor("norm16_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411206528)))]; + tensor var_4226_to_fp16 = const()[name = tensor("op_4226_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_201_cast_fp16 = layer_norm(axes = input_201_axes_0, beta = norm16_2_bias_to_fp16, epsilon = var_4226_to_fp16, gamma = norm16_2_weight_to_fp16, x = input_199_cast_fp16)[name = tensor("input_201_cast_fp16")]; + tensor linear16_1_weight_to_fp16 = const()[name = tensor("linear16_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411208640)))]; + tensor linear_66_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear16_1_weight_to_fp16, x = input_201_cast_fp16)[name = tensor("linear_66_cast_fp16")]; + tensor input_203_mode_0 = const()[name = tensor("input_203_mode_0"), val = tensor("EXACT")]; + tensor input_203_cast_fp16 = gelu(mode = input_203_mode_0, x = linear_66_cast_fp16)[name = tensor("input_203_cast_fp16")]; + tensor linear16_2_weight_to_fp16 = const()[name = tensor("linear16_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(419597312)))]; + tensor linear_67_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear16_2_weight_to_fp16, x = input_203_cast_fp16)[name = tensor("linear_67_cast_fp16")]; + tensor input_205_cast_fp16 = add(x = input_199_cast_fp16, y = linear_67_cast_fp16)[name = tensor("input_205_cast_fp16")]; + tensor input_207_axes_0 = const()[name = tensor("input_207_axes_0"), val = tensor([-1])]; + tensor norm17_1_weight_to_fp16 = const()[name = tensor("norm17_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(427985984)))]; + tensor norm17_1_bias_to_fp16 = const()[name = tensor("norm17_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(427988096)))]; + tensor var_4243_to_fp16 = const()[name = tensor("op_4243_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_207_cast_fp16 = layer_norm(axes = input_207_axes_0, beta = norm17_1_bias_to_fp16, epsilon = var_4243_to_fp16, gamma = norm17_1_weight_to_fp16, x = input_205_cast_fp16)[name = tensor("input_207_cast_fp16")]; + tensor attn17_in_proj_weight_to_fp16 = const()[name = tensor("attn17_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(427990208)))]; + tensor linear_68_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn17_in_proj_weight_to_fp16, x = input_207_cast_fp16)[name = tensor("linear_68_cast_fp16")]; + tensor var_4257 = const()[name = tensor("op_4257"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_35_cast_fp16 = reshape(shape = var_4257, x = linear_68_cast_fp16)[name = tensor("qkv_35_cast_fp16")]; + tensor x_69_begin_0 = const()[name = tensor("x_69_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_69_end_0 = const()[name = tensor("x_69_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_69_end_mask_0 = const()[name = tensor("x_69_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_69_squeeze_mask_0 = const()[name = tensor("x_69_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_69_cast_fp16 = slice_by_index(begin = x_69_begin_0, end = x_69_end_0, end_mask = x_69_end_mask_0, squeeze_mask = x_69_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("x_69_cast_fp16")]; + tensor x_71_begin_0 = const()[name = tensor("x_71_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_71_end_0 = const()[name = tensor("x_71_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_71_end_mask_0 = const()[name = tensor("x_71_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_71_squeeze_mask_0 = const()[name = tensor("x_71_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_71_cast_fp16 = slice_by_index(begin = x_71_begin_0, end = x_71_end_0, end_mask = x_71_end_mask_0, squeeze_mask = x_71_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("x_71_cast_fp16")]; + tensor v_35_begin_0 = const()[name = tensor("v_35_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_35_end_0 = const()[name = tensor("v_35_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_35_end_mask_0 = const()[name = tensor("v_35_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_35_squeeze_mask_0 = const()[name = tensor("v_35_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_35_cast_fp16 = slice_by_index(begin = v_35_begin_0, end = v_35_end_0, end_mask = v_35_end_mask_0, squeeze_mask = v_35_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("v_35_cast_fp16")]; + tensor var_4302 = const()[name = tensor("op_4302"), val = tensor([256, 16, 32, 2])]; + tensor pairs_69_cast_fp16 = reshape(shape = var_4302, x = x_69_cast_fp16)[name = tensor("pairs_69_cast_fp16")]; + tensor xr_69_begin_0 = const()[name = tensor("xr_69_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_69_end_0 = const()[name = tensor("xr_69_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_69_end_mask_0 = const()[name = tensor("xr_69_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_69_squeeze_mask_0 = const()[name = tensor("xr_69_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_69_cast_fp16 = slice_by_index(begin = xr_69_begin_0, end = xr_69_end_0, end_mask = xr_69_end_mask_0, squeeze_mask = xr_69_squeeze_mask_0, x = pairs_69_cast_fp16)[name = tensor("xr_69_cast_fp16")]; + tensor xi_69_begin_0 = const()[name = tensor("xi_69_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_69_end_0 = const()[name = tensor("xi_69_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_69_end_mask_0 = const()[name = tensor("xi_69_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_69_squeeze_mask_0 = const()[name = tensor("xi_69_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_69_cast_fp16 = slice_by_index(begin = xi_69_begin_0, end = xi_69_end_0, end_mask = xi_69_end_mask_0, squeeze_mask = xi_69_squeeze_mask_0, x = pairs_69_cast_fp16)[name = tensor("xi_69_cast_fp16")]; + tensor var_4310_cast_fp16 = mul(x = xr_69_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4310_cast_fp16")]; + tensor var_4311_cast_fp16 = mul(x = xi_69_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4311_cast_fp16")]; + tensor xor__69_cast_fp16 = sub(x = var_4310_cast_fp16, y = var_4311_cast_fp16)[name = tensor("xor__69_cast_fp16")]; + tensor var_4314_cast_fp16 = mul(x = xr_69_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4314_cast_fp16")]; + tensor var_4315_cast_fp16 = mul(x = xi_69_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4315_cast_fp16")]; + tensor xoi_69_cast_fp16 = add(x = var_4314_cast_fp16, y = var_4315_cast_fp16)[name = tensor("xoi_69_cast_fp16")]; + tensor out_69_axis_0 = const()[name = tensor("out_69_axis_0"), val = tensor(-1)]; + tensor out_69_cast_fp16 = stack(axis = out_69_axis_0, values = (xor__69_cast_fp16, xoi_69_cast_fp16))[name = tensor("out_69_cast_fp16")]; + tensor var_4325 = const()[name = tensor("op_4325"), val = tensor([1, 256, 16, 64])]; + tensor q_35_cast_fp16 = reshape(shape = var_4325, x = out_69_cast_fp16)[name = tensor("q_35_cast_fp16")]; + tensor var_4331 = const()[name = tensor("op_4331"), val = tensor([256, 16, 32, 2])]; + tensor pairs_71_cast_fp16 = reshape(shape = var_4331, x = x_71_cast_fp16)[name = tensor("pairs_71_cast_fp16")]; + tensor xr_71_begin_0 = const()[name = tensor("xr_71_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_71_end_0 = const()[name = tensor("xr_71_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_71_end_mask_0 = const()[name = tensor("xr_71_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_71_squeeze_mask_0 = const()[name = tensor("xr_71_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_71_cast_fp16 = slice_by_index(begin = xr_71_begin_0, end = xr_71_end_0, end_mask = xr_71_end_mask_0, squeeze_mask = xr_71_squeeze_mask_0, x = pairs_71_cast_fp16)[name = tensor("xr_71_cast_fp16")]; + tensor xi_71_begin_0 = const()[name = tensor("xi_71_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_71_end_0 = const()[name = tensor("xi_71_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_71_end_mask_0 = const()[name = tensor("xi_71_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_71_squeeze_mask_0 = const()[name = tensor("xi_71_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_71_cast_fp16 = slice_by_index(begin = xi_71_begin_0, end = xi_71_end_0, end_mask = xi_71_end_mask_0, squeeze_mask = xi_71_squeeze_mask_0, x = pairs_71_cast_fp16)[name = tensor("xi_71_cast_fp16")]; + tensor var_4339_cast_fp16 = mul(x = xr_71_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4339_cast_fp16")]; + tensor var_4340_cast_fp16 = mul(x = xi_71_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4340_cast_fp16")]; + tensor xor__71_cast_fp16 = sub(x = var_4339_cast_fp16, y = var_4340_cast_fp16)[name = tensor("xor__71_cast_fp16")]; + tensor var_4343_cast_fp16 = mul(x = xr_71_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4343_cast_fp16")]; + tensor var_4344_cast_fp16 = mul(x = xi_71_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4344_cast_fp16")]; + tensor xoi_71_cast_fp16 = add(x = var_4343_cast_fp16, y = var_4344_cast_fp16)[name = tensor("xoi_71_cast_fp16")]; + tensor out_71_axis_0 = const()[name = tensor("out_71_axis_0"), val = tensor(-1)]; + tensor out_71_cast_fp16 = stack(axis = out_71_axis_0, values = (xor__71_cast_fp16, xoi_71_cast_fp16))[name = tensor("out_71_cast_fp16")]; + tensor var_4361 = const()[name = tensor("op_4361"), val = tensor([256, 1024])]; + tensor var_4362_cast_fp16 = reshape(shape = var_4361, x = out_71_cast_fp16)[name = tensor("op_4362_cast_fp16")]; + tensor written_k_35_transpose_x_1 = const()[name = tensor("written_k_35_transpose_x_1"), val = tensor(true)]; + tensor written_k_35_transpose_y_1 = const()[name = tensor("written_k_35_transpose_y_1"), val = tensor(false)]; + tensor written_k_35_cast_fp16 = matmul(transpose_x = written_k_35_transpose_x_1, transpose_y = written_k_35_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_4362_cast_fp16)[name = tensor("written_k_35_cast_fp16")]; + tensor var_4369 = const()[name = tensor("op_4369"), val = tensor([256, 1024])]; + tensor var_4370_cast_fp16 = reshape(shape = var_4369, x = v_35_cast_fp16)[name = tensor("op_4370_cast_fp16")]; + tensor written_v_35_transpose_x_1 = const()[name = tensor("written_v_35_transpose_x_1"), val = tensor(true)]; + tensor written_v_35_transpose_y_1 = const()[name = tensor("written_v_35_transpose_y_1"), val = tensor(false)]; + tensor written_v_35_cast_fp16 = matmul(transpose_x = written_v_35_transpose_x_1, transpose_y = written_v_35_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_4370_cast_fp16)[name = tensor("written_v_35_cast_fp16")]; + tensor k_cache17_to_fp16_dtype_0 = const()[name = tensor("k_cache17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache17_to_fp16 = cast(dtype = k_cache17_to_fp16_dtype_0, x = k_cache17)[name = tensor("cast_363")]; + tensor var_4381_cast_fp16 = mul(x = k_cache17_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_4381_cast_fp16")]; + tensor var_4386 = const()[name = tensor("op_4386"), val = tensor([1, 512, 16, 64])]; + tensor var_4387_cast_fp16 = reshape(shape = var_4386, x = written_k_35_cast_fp16)[name = tensor("op_4387_cast_fp16")]; + tensor new_k_cache_35_cast_fp16 = add(x = var_4381_cast_fp16, y = var_4387_cast_fp16)[name = tensor("new_k_cache_35_cast_fp16")]; + tensor new_k_cache_35_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_35_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache17_to_fp16_dtype_0 = const()[name = tensor("v_cache17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache17_to_fp16 = cast(dtype = v_cache17_to_fp16_dtype_0, x = v_cache17)[name = tensor("cast_361")]; + tensor var_4390_cast_fp16 = mul(x = v_cache17_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_4390_cast_fp16")]; + tensor var_4395 = const()[name = tensor("op_4395"), val = tensor([1, 512, 16, 64])]; + tensor var_4396_cast_fp16 = reshape(shape = var_4395, x = written_v_35_cast_fp16)[name = tensor("op_4396_cast_fp16")]; + tensor new_v_cache_35_cast_fp16 = add(x = var_4390_cast_fp16, y = var_4396_cast_fp16)[name = tensor("new_v_cache_35_cast_fp16")]; + tensor new_v_cache_35_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_35_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_4409 = const()[name = tensor("op_4409"), val = tensor([0, 2, 1, 3])]; + tensor var_4420_transpose_x_0 = const()[name = tensor("op_4420_transpose_x_0"), val = tensor(false)]; + tensor var_4420_transpose_y_0 = const()[name = tensor("op_4420_transpose_y_0"), val = tensor(false)]; + tensor transpose_103_perm_0 = const()[name = tensor("transpose_103_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_104_perm_0 = const()[name = tensor("transpose_104_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_104 = transpose(perm = transpose_104_perm_0, x = new_k_cache_35_cast_fp16)[name = tensor("transpose_136")]; + tensor transpose_103 = transpose(perm = transpose_103_perm_0, x = q_35_cast_fp16)[name = tensor("transpose_137")]; + tensor var_4420_cast_fp16 = matmul(transpose_x = var_4420_transpose_x_0, transpose_y = var_4420_transpose_y_0, x = transpose_103, y = transpose_104)[name = tensor("op_4420_cast_fp16")]; + tensor var_4421_to_fp16 = const()[name = tensor("op_4421_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_69_cast_fp16 = mul(x = var_4420_cast_fp16, y = var_4421_to_fp16)[name = tensor("scores_69_cast_fp16")]; + tensor scores_71_cast_fp16 = add(x = scores_69_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_71_cast_fp16")]; + tensor var_4431 = const()[name = tensor("op_4431"), val = tensor(-1)]; + tensor weights_35_cast_fp16 = softmax(axis = var_4431, x = scores_71_cast_fp16)[name = tensor("weights_35_cast_fp16")]; + tensor attn_35_transpose_x_0 = const()[name = tensor("attn_35_transpose_x_0"), val = tensor(false)]; + tensor attn_35_transpose_y_0 = const()[name = tensor("attn_35_transpose_y_0"), val = tensor(false)]; + tensor values_35_cast_fp16 = transpose(perm = var_4409, x = new_v_cache_35_cast_fp16)[name = tensor("transpose_138")]; + tensor attn_35_cast_fp16 = matmul(transpose_x = attn_35_transpose_x_0, transpose_y = attn_35_transpose_y_0, x = weights_35_cast_fp16, y = values_35_cast_fp16)[name = tensor("attn_35_cast_fp16")]; + tensor var_4439 = const()[name = tensor("op_4439"), val = tensor([0, 2, 1, 3])]; + tensor var_4444 = const()[name = tensor("op_4444"), val = tensor([1, 256, 1024])]; + tensor var_4440_cast_fp16 = transpose(perm = var_4439, x = attn_35_cast_fp16)[name = tensor("transpose_135")]; + tensor input_209_cast_fp16 = reshape(shape = var_4444, x = var_4440_cast_fp16)[name = tensor("input_209_cast_fp16")]; + tensor attn17_out_proj_weight_to_fp16 = const()[name = tensor("attn17_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(434281728)))]; + tensor linear_69_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn17_out_proj_weight_to_fp16, x = input_209_cast_fp16)[name = tensor("linear_69_cast_fp16")]; + tensor position17_to_fp16_dtype_0 = const()[name = tensor("position17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position17_to_fp16 = cast(dtype = position17_to_fp16_dtype_0, x = position17)[name = tensor("cast_359")]; + tensor var_4453_cast_fp16 = add(x = position17_to_fp16, y = valid_len_to_fp16)[name = tensor("op_4453_cast_fp16")]; + tensor var_4453_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4453_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_211_cast_fp16 = add(x = input_205_cast_fp16, y = linear_69_cast_fp16)[name = tensor("input_211_cast_fp16")]; + tensor input_213_axes_0 = const()[name = tensor("input_213_axes_0"), val = tensor([-1])]; + tensor norm17_2_weight_to_fp16 = const()[name = tensor("norm17_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(436378944)))]; + tensor norm17_2_bias_to_fp16 = const()[name = tensor("norm17_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(436381056)))]; + tensor var_4457_to_fp16 = const()[name = tensor("op_4457_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_213_cast_fp16 = layer_norm(axes = input_213_axes_0, beta = norm17_2_bias_to_fp16, epsilon = var_4457_to_fp16, gamma = norm17_2_weight_to_fp16, x = input_211_cast_fp16)[name = tensor("input_213_cast_fp16")]; + tensor linear17_1_weight_to_fp16 = const()[name = tensor("linear17_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(436383168)))]; + tensor linear_70_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear17_1_weight_to_fp16, x = input_213_cast_fp16)[name = tensor("linear_70_cast_fp16")]; + tensor input_215_mode_0 = const()[name = tensor("input_215_mode_0"), val = tensor("EXACT")]; + tensor input_215_cast_fp16 = gelu(mode = input_215_mode_0, x = linear_70_cast_fp16)[name = tensor("input_215_cast_fp16")]; + tensor linear17_2_weight_to_fp16 = const()[name = tensor("linear17_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(444771840)))]; + tensor linear_71_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear17_2_weight_to_fp16, x = input_215_cast_fp16)[name = tensor("linear_71_cast_fp16")]; + tensor input_217_cast_fp16 = add(x = input_211_cast_fp16, y = linear_71_cast_fp16)[name = tensor("input_217_cast_fp16")]; + tensor input_219_axes_0 = const()[name = tensor("input_219_axes_0"), val = tensor([-1])]; + tensor norm18_1_weight_to_fp16 = const()[name = tensor("norm18_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(453160512)))]; + tensor norm18_1_bias_to_fp16 = const()[name = tensor("norm18_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(453162624)))]; + tensor var_4474_to_fp16 = const()[name = tensor("op_4474_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_219_cast_fp16 = layer_norm(axes = input_219_axes_0, beta = norm18_1_bias_to_fp16, epsilon = var_4474_to_fp16, gamma = norm18_1_weight_to_fp16, x = input_217_cast_fp16)[name = tensor("input_219_cast_fp16")]; + tensor attn18_in_proj_weight_to_fp16 = const()[name = tensor("attn18_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(453164736)))]; + tensor linear_72_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn18_in_proj_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_72_cast_fp16")]; + tensor var_4488 = const()[name = tensor("op_4488"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_37_cast_fp16 = reshape(shape = var_4488, x = linear_72_cast_fp16)[name = tensor("qkv_37_cast_fp16")]; + tensor x_73_begin_0 = const()[name = tensor("x_73_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_73_end_0 = const()[name = tensor("x_73_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_73_end_mask_0 = const()[name = tensor("x_73_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_73_squeeze_mask_0 = const()[name = tensor("x_73_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_73_cast_fp16 = slice_by_index(begin = x_73_begin_0, end = x_73_end_0, end_mask = x_73_end_mask_0, squeeze_mask = x_73_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("x_73_cast_fp16")]; + tensor x_75_begin_0 = const()[name = tensor("x_75_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_75_end_0 = const()[name = tensor("x_75_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_75_end_mask_0 = const()[name = tensor("x_75_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_75_squeeze_mask_0 = const()[name = tensor("x_75_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_75_cast_fp16 = slice_by_index(begin = x_75_begin_0, end = x_75_end_0, end_mask = x_75_end_mask_0, squeeze_mask = x_75_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("x_75_cast_fp16")]; + tensor v_37_begin_0 = const()[name = tensor("v_37_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_37_end_0 = const()[name = tensor("v_37_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_37_end_mask_0 = const()[name = tensor("v_37_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_37_squeeze_mask_0 = const()[name = tensor("v_37_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_37_cast_fp16 = slice_by_index(begin = v_37_begin_0, end = v_37_end_0, end_mask = v_37_end_mask_0, squeeze_mask = v_37_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("v_37_cast_fp16")]; + tensor var_4533 = const()[name = tensor("op_4533"), val = tensor([256, 16, 32, 2])]; + tensor pairs_73_cast_fp16 = reshape(shape = var_4533, x = x_73_cast_fp16)[name = tensor("pairs_73_cast_fp16")]; + tensor xr_73_begin_0 = const()[name = tensor("xr_73_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_73_end_0 = const()[name = tensor("xr_73_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_73_end_mask_0 = const()[name = tensor("xr_73_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_73_squeeze_mask_0 = const()[name = tensor("xr_73_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_73_cast_fp16 = slice_by_index(begin = xr_73_begin_0, end = xr_73_end_0, end_mask = xr_73_end_mask_0, squeeze_mask = xr_73_squeeze_mask_0, x = pairs_73_cast_fp16)[name = tensor("xr_73_cast_fp16")]; + tensor xi_73_begin_0 = const()[name = tensor("xi_73_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_73_end_0 = const()[name = tensor("xi_73_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_73_end_mask_0 = const()[name = tensor("xi_73_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_73_squeeze_mask_0 = const()[name = tensor("xi_73_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_73_cast_fp16 = slice_by_index(begin = xi_73_begin_0, end = xi_73_end_0, end_mask = xi_73_end_mask_0, squeeze_mask = xi_73_squeeze_mask_0, x = pairs_73_cast_fp16)[name = tensor("xi_73_cast_fp16")]; + tensor var_4541_cast_fp16 = mul(x = xr_73_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4541_cast_fp16")]; + tensor var_4542_cast_fp16 = mul(x = xi_73_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4542_cast_fp16")]; + tensor xor__73_cast_fp16 = sub(x = var_4541_cast_fp16, y = var_4542_cast_fp16)[name = tensor("xor__73_cast_fp16")]; + tensor var_4545_cast_fp16 = mul(x = xr_73_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4545_cast_fp16")]; + tensor var_4546_cast_fp16 = mul(x = xi_73_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4546_cast_fp16")]; + tensor xoi_73_cast_fp16 = add(x = var_4545_cast_fp16, y = var_4546_cast_fp16)[name = tensor("xoi_73_cast_fp16")]; + tensor out_73_axis_0 = const()[name = tensor("out_73_axis_0"), val = tensor(-1)]; + tensor out_73_cast_fp16 = stack(axis = out_73_axis_0, values = (xor__73_cast_fp16, xoi_73_cast_fp16))[name = tensor("out_73_cast_fp16")]; + tensor var_4556 = const()[name = tensor("op_4556"), val = tensor([1, 256, 16, 64])]; + tensor q_37_cast_fp16 = reshape(shape = var_4556, x = out_73_cast_fp16)[name = tensor("q_37_cast_fp16")]; + tensor var_4562 = const()[name = tensor("op_4562"), val = tensor([256, 16, 32, 2])]; + tensor pairs_75_cast_fp16 = reshape(shape = var_4562, x = x_75_cast_fp16)[name = tensor("pairs_75_cast_fp16")]; + tensor xr_75_begin_0 = const()[name = tensor("xr_75_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_75_end_0 = const()[name = tensor("xr_75_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_75_end_mask_0 = const()[name = tensor("xr_75_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_75_squeeze_mask_0 = const()[name = tensor("xr_75_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_75_cast_fp16 = slice_by_index(begin = xr_75_begin_0, end = xr_75_end_0, end_mask = xr_75_end_mask_0, squeeze_mask = xr_75_squeeze_mask_0, x = pairs_75_cast_fp16)[name = tensor("xr_75_cast_fp16")]; + tensor xi_75_begin_0 = const()[name = tensor("xi_75_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_75_end_0 = const()[name = tensor("xi_75_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_75_end_mask_0 = const()[name = tensor("xi_75_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_75_squeeze_mask_0 = const()[name = tensor("xi_75_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_75_cast_fp16 = slice_by_index(begin = xi_75_begin_0, end = xi_75_end_0, end_mask = xi_75_end_mask_0, squeeze_mask = xi_75_squeeze_mask_0, x = pairs_75_cast_fp16)[name = tensor("xi_75_cast_fp16")]; + tensor var_4570_cast_fp16 = mul(x = xr_75_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4570_cast_fp16")]; + tensor var_4571_cast_fp16 = mul(x = xi_75_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4571_cast_fp16")]; + tensor xor__75_cast_fp16 = sub(x = var_4570_cast_fp16, y = var_4571_cast_fp16)[name = tensor("xor__75_cast_fp16")]; + tensor var_4574_cast_fp16 = mul(x = xr_75_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4574_cast_fp16")]; + tensor var_4575_cast_fp16 = mul(x = xi_75_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4575_cast_fp16")]; + tensor xoi_75_cast_fp16 = add(x = var_4574_cast_fp16, y = var_4575_cast_fp16)[name = tensor("xoi_75_cast_fp16")]; + tensor out_75_axis_0 = const()[name = tensor("out_75_axis_0"), val = tensor(-1)]; + tensor out_75_cast_fp16 = stack(axis = out_75_axis_0, values = (xor__75_cast_fp16, xoi_75_cast_fp16))[name = tensor("out_75_cast_fp16")]; + tensor var_4592 = const()[name = tensor("op_4592"), val = tensor([256, 1024])]; + tensor var_4593_cast_fp16 = reshape(shape = var_4592, x = out_75_cast_fp16)[name = tensor("op_4593_cast_fp16")]; + tensor written_k_37_transpose_x_1 = const()[name = tensor("written_k_37_transpose_x_1"), val = tensor(true)]; + tensor written_k_37_transpose_y_1 = const()[name = tensor("written_k_37_transpose_y_1"), val = tensor(false)]; + tensor written_k_37_cast_fp16 = matmul(transpose_x = written_k_37_transpose_x_1, transpose_y = written_k_37_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_4593_cast_fp16)[name = tensor("written_k_37_cast_fp16")]; + tensor var_4600 = const()[name = tensor("op_4600"), val = tensor([256, 1024])]; + tensor var_4601_cast_fp16 = reshape(shape = var_4600, x = v_37_cast_fp16)[name = tensor("op_4601_cast_fp16")]; + tensor written_v_37_transpose_x_1 = const()[name = tensor("written_v_37_transpose_x_1"), val = tensor(true)]; + tensor written_v_37_transpose_y_1 = const()[name = tensor("written_v_37_transpose_y_1"), val = tensor(false)]; + tensor written_v_37_cast_fp16 = matmul(transpose_x = written_v_37_transpose_x_1, transpose_y = written_v_37_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_4601_cast_fp16)[name = tensor("written_v_37_cast_fp16")]; + tensor k_cache18_to_fp16_dtype_0 = const()[name = tensor("k_cache18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache18_to_fp16 = cast(dtype = k_cache18_to_fp16_dtype_0, x = k_cache18)[name = tensor("cast_357")]; + tensor var_4612_cast_fp16 = mul(x = k_cache18_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_4612_cast_fp16")]; + tensor var_4617 = const()[name = tensor("op_4617"), val = tensor([1, 512, 16, 64])]; + tensor var_4618_cast_fp16 = reshape(shape = var_4617, x = written_k_37_cast_fp16)[name = tensor("op_4618_cast_fp16")]; + tensor new_k_cache_37_cast_fp16 = add(x = var_4612_cast_fp16, y = var_4618_cast_fp16)[name = tensor("new_k_cache_37_cast_fp16")]; + tensor new_k_cache_37_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_37_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache18_to_fp16_dtype_0 = const()[name = tensor("v_cache18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache18_to_fp16 = cast(dtype = v_cache18_to_fp16_dtype_0, x = v_cache18)[name = tensor("cast_355")]; + tensor var_4621_cast_fp16 = mul(x = v_cache18_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_4621_cast_fp16")]; + tensor var_4626 = const()[name = tensor("op_4626"), val = tensor([1, 512, 16, 64])]; + tensor var_4627_cast_fp16 = reshape(shape = var_4626, x = written_v_37_cast_fp16)[name = tensor("op_4627_cast_fp16")]; + tensor new_v_cache_37_cast_fp16 = add(x = var_4621_cast_fp16, y = var_4627_cast_fp16)[name = tensor("new_v_cache_37_cast_fp16")]; + tensor new_v_cache_37_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_37_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_4640 = const()[name = tensor("op_4640"), val = tensor([0, 2, 1, 3])]; + tensor var_4651_transpose_x_0 = const()[name = tensor("op_4651_transpose_x_0"), val = tensor(false)]; + tensor var_4651_transpose_y_0 = const()[name = tensor("op_4651_transpose_y_0"), val = tensor(false)]; + tensor transpose_105_perm_0 = const()[name = tensor("transpose_105_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_106_perm_0 = const()[name = tensor("transpose_106_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_106 = transpose(perm = transpose_106_perm_0, x = new_k_cache_37_cast_fp16)[name = tensor("transpose_132")]; + tensor transpose_105 = transpose(perm = transpose_105_perm_0, x = q_37_cast_fp16)[name = tensor("transpose_133")]; + tensor var_4651_cast_fp16 = matmul(transpose_x = var_4651_transpose_x_0, transpose_y = var_4651_transpose_y_0, x = transpose_105, y = transpose_106)[name = tensor("op_4651_cast_fp16")]; + tensor var_4652_to_fp16 = const()[name = tensor("op_4652_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_73_cast_fp16 = mul(x = var_4651_cast_fp16, y = var_4652_to_fp16)[name = tensor("scores_73_cast_fp16")]; + tensor scores_75_cast_fp16 = add(x = scores_73_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_75_cast_fp16")]; + tensor var_4662 = const()[name = tensor("op_4662"), val = tensor(-1)]; + tensor weights_37_cast_fp16 = softmax(axis = var_4662, x = scores_75_cast_fp16)[name = tensor("weights_37_cast_fp16")]; + tensor attn_37_transpose_x_0 = const()[name = tensor("attn_37_transpose_x_0"), val = tensor(false)]; + tensor attn_37_transpose_y_0 = const()[name = tensor("attn_37_transpose_y_0"), val = tensor(false)]; + tensor values_37_cast_fp16 = transpose(perm = var_4640, x = new_v_cache_37_cast_fp16)[name = tensor("transpose_134")]; + tensor attn_37_cast_fp16 = matmul(transpose_x = attn_37_transpose_x_0, transpose_y = attn_37_transpose_y_0, x = weights_37_cast_fp16, y = values_37_cast_fp16)[name = tensor("attn_37_cast_fp16")]; + tensor var_4670 = const()[name = tensor("op_4670"), val = tensor([0, 2, 1, 3])]; + tensor var_4675 = const()[name = tensor("op_4675"), val = tensor([1, 256, 1024])]; + tensor var_4671_cast_fp16 = transpose(perm = var_4670, x = attn_37_cast_fp16)[name = tensor("transpose_131")]; + tensor input_221_cast_fp16 = reshape(shape = var_4675, x = var_4671_cast_fp16)[name = tensor("input_221_cast_fp16")]; + tensor attn18_out_proj_weight_to_fp16 = const()[name = tensor("attn18_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(459456256)))]; + tensor linear_73_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn18_out_proj_weight_to_fp16, x = input_221_cast_fp16)[name = tensor("linear_73_cast_fp16")]; + tensor position18_to_fp16_dtype_0 = const()[name = tensor("position18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position18_to_fp16 = cast(dtype = position18_to_fp16_dtype_0, x = position18)[name = tensor("cast_353")]; + tensor var_4684_cast_fp16 = add(x = position18_to_fp16, y = valid_len_to_fp16)[name = tensor("op_4684_cast_fp16")]; + tensor var_4684_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4684_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_223_cast_fp16 = add(x = input_217_cast_fp16, y = linear_73_cast_fp16)[name = tensor("input_223_cast_fp16")]; + tensor input_225_axes_0 = const()[name = tensor("input_225_axes_0"), val = tensor([-1])]; + tensor norm18_2_weight_to_fp16 = const()[name = tensor("norm18_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(461553472)))]; + tensor norm18_2_bias_to_fp16 = const()[name = tensor("norm18_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(461555584)))]; + tensor var_4688_to_fp16 = const()[name = tensor("op_4688_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_225_cast_fp16 = layer_norm(axes = input_225_axes_0, beta = norm18_2_bias_to_fp16, epsilon = var_4688_to_fp16, gamma = norm18_2_weight_to_fp16, x = input_223_cast_fp16)[name = tensor("input_225_cast_fp16")]; + tensor linear18_1_weight_to_fp16 = const()[name = tensor("linear18_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(461557696)))]; + tensor linear_74_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear18_1_weight_to_fp16, x = input_225_cast_fp16)[name = tensor("linear_74_cast_fp16")]; + tensor input_227_mode_0 = const()[name = tensor("input_227_mode_0"), val = tensor("EXACT")]; + tensor input_227_cast_fp16 = gelu(mode = input_227_mode_0, x = linear_74_cast_fp16)[name = tensor("input_227_cast_fp16")]; + tensor linear18_2_weight_to_fp16 = const()[name = tensor("linear18_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(469946368)))]; + tensor linear_75_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear18_2_weight_to_fp16, x = input_227_cast_fp16)[name = tensor("linear_75_cast_fp16")]; + tensor input_229_cast_fp16 = add(x = input_223_cast_fp16, y = linear_75_cast_fp16)[name = tensor("input_229_cast_fp16")]; + tensor input_231_axes_0 = const()[name = tensor("input_231_axes_0"), val = tensor([-1])]; + tensor norm19_1_weight_to_fp16 = const()[name = tensor("norm19_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(478335040)))]; + tensor norm19_1_bias_to_fp16 = const()[name = tensor("norm19_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(478337152)))]; + tensor var_4705_to_fp16 = const()[name = tensor("op_4705_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_231_cast_fp16 = layer_norm(axes = input_231_axes_0, beta = norm19_1_bias_to_fp16, epsilon = var_4705_to_fp16, gamma = norm19_1_weight_to_fp16, x = input_229_cast_fp16)[name = tensor("input_231_cast_fp16")]; + tensor attn19_in_proj_weight_to_fp16 = const()[name = tensor("attn19_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(478339264)))]; + tensor linear_76_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn19_in_proj_weight_to_fp16, x = input_231_cast_fp16)[name = tensor("linear_76_cast_fp16")]; + tensor var_4719 = const()[name = tensor("op_4719"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_39_cast_fp16 = reshape(shape = var_4719, x = linear_76_cast_fp16)[name = tensor("qkv_39_cast_fp16")]; + tensor x_77_begin_0 = const()[name = tensor("x_77_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_77_end_0 = const()[name = tensor("x_77_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_77_end_mask_0 = const()[name = tensor("x_77_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_77_squeeze_mask_0 = const()[name = tensor("x_77_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_77_cast_fp16 = slice_by_index(begin = x_77_begin_0, end = x_77_end_0, end_mask = x_77_end_mask_0, squeeze_mask = x_77_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("x_77_cast_fp16")]; + tensor x_79_begin_0 = const()[name = tensor("x_79_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_79_end_0 = const()[name = tensor("x_79_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_79_end_mask_0 = const()[name = tensor("x_79_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_79_squeeze_mask_0 = const()[name = tensor("x_79_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_79_cast_fp16 = slice_by_index(begin = x_79_begin_0, end = x_79_end_0, end_mask = x_79_end_mask_0, squeeze_mask = x_79_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("x_79_cast_fp16")]; + tensor v_39_begin_0 = const()[name = tensor("v_39_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_39_end_0 = const()[name = tensor("v_39_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_39_end_mask_0 = const()[name = tensor("v_39_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_39_squeeze_mask_0 = const()[name = tensor("v_39_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_39_cast_fp16 = slice_by_index(begin = v_39_begin_0, end = v_39_end_0, end_mask = v_39_end_mask_0, squeeze_mask = v_39_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("v_39_cast_fp16")]; + tensor var_4764 = const()[name = tensor("op_4764"), val = tensor([256, 16, 32, 2])]; + tensor pairs_77_cast_fp16 = reshape(shape = var_4764, x = x_77_cast_fp16)[name = tensor("pairs_77_cast_fp16")]; + tensor xr_77_begin_0 = const()[name = tensor("xr_77_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_77_end_0 = const()[name = tensor("xr_77_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_77_end_mask_0 = const()[name = tensor("xr_77_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_77_squeeze_mask_0 = const()[name = tensor("xr_77_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_77_cast_fp16 = slice_by_index(begin = xr_77_begin_0, end = xr_77_end_0, end_mask = xr_77_end_mask_0, squeeze_mask = xr_77_squeeze_mask_0, x = pairs_77_cast_fp16)[name = tensor("xr_77_cast_fp16")]; + tensor xi_77_begin_0 = const()[name = tensor("xi_77_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_77_end_0 = const()[name = tensor("xi_77_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_77_end_mask_0 = const()[name = tensor("xi_77_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_77_squeeze_mask_0 = const()[name = tensor("xi_77_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_77_cast_fp16 = slice_by_index(begin = xi_77_begin_0, end = xi_77_end_0, end_mask = xi_77_end_mask_0, squeeze_mask = xi_77_squeeze_mask_0, x = pairs_77_cast_fp16)[name = tensor("xi_77_cast_fp16")]; + tensor var_4772_cast_fp16 = mul(x = xr_77_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4772_cast_fp16")]; + tensor var_4773_cast_fp16 = mul(x = xi_77_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4773_cast_fp16")]; + tensor xor__77_cast_fp16 = sub(x = var_4772_cast_fp16, y = var_4773_cast_fp16)[name = tensor("xor__77_cast_fp16")]; + tensor var_4776_cast_fp16 = mul(x = xr_77_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4776_cast_fp16")]; + tensor var_4777_cast_fp16 = mul(x = xi_77_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4777_cast_fp16")]; + tensor xoi_77_cast_fp16 = add(x = var_4776_cast_fp16, y = var_4777_cast_fp16)[name = tensor("xoi_77_cast_fp16")]; + tensor out_77_axis_0 = const()[name = tensor("out_77_axis_0"), val = tensor(-1)]; + tensor out_77_cast_fp16 = stack(axis = out_77_axis_0, values = (xor__77_cast_fp16, xoi_77_cast_fp16))[name = tensor("out_77_cast_fp16")]; + tensor var_4787 = const()[name = tensor("op_4787"), val = tensor([1, 256, 16, 64])]; + tensor q_39_cast_fp16 = reshape(shape = var_4787, x = out_77_cast_fp16)[name = tensor("q_39_cast_fp16")]; + tensor var_4793 = const()[name = tensor("op_4793"), val = tensor([256, 16, 32, 2])]; + tensor pairs_79_cast_fp16 = reshape(shape = var_4793, x = x_79_cast_fp16)[name = tensor("pairs_79_cast_fp16")]; + tensor xr_79_begin_0 = const()[name = tensor("xr_79_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_79_end_0 = const()[name = tensor("xr_79_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_79_end_mask_0 = const()[name = tensor("xr_79_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_79_squeeze_mask_0 = const()[name = tensor("xr_79_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_79_cast_fp16 = slice_by_index(begin = xr_79_begin_0, end = xr_79_end_0, end_mask = xr_79_end_mask_0, squeeze_mask = xr_79_squeeze_mask_0, x = pairs_79_cast_fp16)[name = tensor("xr_79_cast_fp16")]; + tensor xi_79_begin_0 = const()[name = tensor("xi_79_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_79_end_0 = const()[name = tensor("xi_79_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_79_end_mask_0 = const()[name = tensor("xi_79_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_79_squeeze_mask_0 = const()[name = tensor("xi_79_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_79_cast_fp16 = slice_by_index(begin = xi_79_begin_0, end = xi_79_end_0, end_mask = xi_79_end_mask_0, squeeze_mask = xi_79_squeeze_mask_0, x = pairs_79_cast_fp16)[name = tensor("xi_79_cast_fp16")]; + tensor var_4801_cast_fp16 = mul(x = xr_79_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4801_cast_fp16")]; + tensor var_4802_cast_fp16 = mul(x = xi_79_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4802_cast_fp16")]; + tensor xor__79_cast_fp16 = sub(x = var_4801_cast_fp16, y = var_4802_cast_fp16)[name = tensor("xor__79_cast_fp16")]; + tensor var_4805_cast_fp16 = mul(x = xr_79_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4805_cast_fp16")]; + tensor var_4806_cast_fp16 = mul(x = xi_79_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4806_cast_fp16")]; + tensor xoi_79_cast_fp16 = add(x = var_4805_cast_fp16, y = var_4806_cast_fp16)[name = tensor("xoi_79_cast_fp16")]; + tensor out_79_axis_0 = const()[name = tensor("out_79_axis_0"), val = tensor(-1)]; + tensor out_79_cast_fp16 = stack(axis = out_79_axis_0, values = (xor__79_cast_fp16, xoi_79_cast_fp16))[name = tensor("out_79_cast_fp16")]; + tensor var_4823 = const()[name = tensor("op_4823"), val = tensor([256, 1024])]; + tensor var_4824_cast_fp16 = reshape(shape = var_4823, x = out_79_cast_fp16)[name = tensor("op_4824_cast_fp16")]; + tensor written_k_39_transpose_x_1 = const()[name = tensor("written_k_39_transpose_x_1"), val = tensor(true)]; + tensor written_k_39_transpose_y_1 = const()[name = tensor("written_k_39_transpose_y_1"), val = tensor(false)]; + tensor written_k_39_cast_fp16 = matmul(transpose_x = written_k_39_transpose_x_1, transpose_y = written_k_39_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_4824_cast_fp16)[name = tensor("written_k_39_cast_fp16")]; + tensor var_4831 = const()[name = tensor("op_4831"), val = tensor([256, 1024])]; + tensor var_4832_cast_fp16 = reshape(shape = var_4831, x = v_39_cast_fp16)[name = tensor("op_4832_cast_fp16")]; + tensor written_v_39_transpose_x_1 = const()[name = tensor("written_v_39_transpose_x_1"), val = tensor(true)]; + tensor written_v_39_transpose_y_1 = const()[name = tensor("written_v_39_transpose_y_1"), val = tensor(false)]; + tensor written_v_39_cast_fp16 = matmul(transpose_x = written_v_39_transpose_x_1, transpose_y = written_v_39_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_4832_cast_fp16)[name = tensor("written_v_39_cast_fp16")]; + tensor k_cache19_to_fp16_dtype_0 = const()[name = tensor("k_cache19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache19_to_fp16 = cast(dtype = k_cache19_to_fp16_dtype_0, x = k_cache19)[name = tensor("cast_351")]; + tensor var_4843_cast_fp16 = mul(x = k_cache19_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_4843_cast_fp16")]; + tensor var_4848 = const()[name = tensor("op_4848"), val = tensor([1, 512, 16, 64])]; + tensor var_4849_cast_fp16 = reshape(shape = var_4848, x = written_k_39_cast_fp16)[name = tensor("op_4849_cast_fp16")]; + tensor new_k_cache_39_cast_fp16 = add(x = var_4843_cast_fp16, y = var_4849_cast_fp16)[name = tensor("new_k_cache_39_cast_fp16")]; + tensor new_k_cache_39_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_39_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache19_to_fp16_dtype_0 = const()[name = tensor("v_cache19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache19_to_fp16 = cast(dtype = v_cache19_to_fp16_dtype_0, x = v_cache19)[name = tensor("cast_349")]; + tensor var_4852_cast_fp16 = mul(x = v_cache19_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_4852_cast_fp16")]; + tensor var_4857 = const()[name = tensor("op_4857"), val = tensor([1, 512, 16, 64])]; + tensor var_4858_cast_fp16 = reshape(shape = var_4857, x = written_v_39_cast_fp16)[name = tensor("op_4858_cast_fp16")]; + tensor new_v_cache_39_cast_fp16 = add(x = var_4852_cast_fp16, y = var_4858_cast_fp16)[name = tensor("new_v_cache_39_cast_fp16")]; + tensor new_v_cache_39_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_39_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_4871 = const()[name = tensor("op_4871"), val = tensor([0, 2, 1, 3])]; + tensor var_4882_transpose_x_0 = const()[name = tensor("op_4882_transpose_x_0"), val = tensor(false)]; + tensor var_4882_transpose_y_0 = const()[name = tensor("op_4882_transpose_y_0"), val = tensor(false)]; + tensor transpose_107_perm_0 = const()[name = tensor("transpose_107_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_108_perm_0 = const()[name = tensor("transpose_108_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_108 = transpose(perm = transpose_108_perm_0, x = new_k_cache_39_cast_fp16)[name = tensor("transpose_128")]; + tensor transpose_107 = transpose(perm = transpose_107_perm_0, x = q_39_cast_fp16)[name = tensor("transpose_129")]; + tensor var_4882_cast_fp16 = matmul(transpose_x = var_4882_transpose_x_0, transpose_y = var_4882_transpose_y_0, x = transpose_107, y = transpose_108)[name = tensor("op_4882_cast_fp16")]; + tensor var_4883_to_fp16 = const()[name = tensor("op_4883_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_77_cast_fp16 = mul(x = var_4882_cast_fp16, y = var_4883_to_fp16)[name = tensor("scores_77_cast_fp16")]; + tensor scores_79_cast_fp16 = add(x = scores_77_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_79_cast_fp16")]; + tensor var_4893 = const()[name = tensor("op_4893"), val = tensor(-1)]; + tensor weights_39_cast_fp16 = softmax(axis = var_4893, x = scores_79_cast_fp16)[name = tensor("weights_39_cast_fp16")]; + tensor attn_39_transpose_x_0 = const()[name = tensor("attn_39_transpose_x_0"), val = tensor(false)]; + tensor attn_39_transpose_y_0 = const()[name = tensor("attn_39_transpose_y_0"), val = tensor(false)]; + tensor values_39_cast_fp16 = transpose(perm = var_4871, x = new_v_cache_39_cast_fp16)[name = tensor("transpose_130")]; + tensor attn_39_cast_fp16 = matmul(transpose_x = attn_39_transpose_x_0, transpose_y = attn_39_transpose_y_0, x = weights_39_cast_fp16, y = values_39_cast_fp16)[name = tensor("attn_39_cast_fp16")]; + tensor var_4901 = const()[name = tensor("op_4901"), val = tensor([0, 2, 1, 3])]; + tensor var_4906 = const()[name = tensor("op_4906"), val = tensor([1, 256, 1024])]; + tensor var_4902_cast_fp16 = transpose(perm = var_4901, x = attn_39_cast_fp16)[name = tensor("transpose_127")]; + tensor input_233_cast_fp16 = reshape(shape = var_4906, x = var_4902_cast_fp16)[name = tensor("input_233_cast_fp16")]; + tensor attn19_out_proj_weight_to_fp16 = const()[name = tensor("attn19_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(484630784)))]; + tensor linear_77_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn19_out_proj_weight_to_fp16, x = input_233_cast_fp16)[name = tensor("linear_77_cast_fp16")]; + tensor position19_to_fp16_dtype_0 = const()[name = tensor("position19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position19_to_fp16 = cast(dtype = position19_to_fp16_dtype_0, x = position19)[name = tensor("cast_347")]; + tensor var_4915_cast_fp16 = add(x = position19_to_fp16, y = valid_len_to_fp16)[name = tensor("op_4915_cast_fp16")]; + tensor var_4915_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4915_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_235_cast_fp16 = add(x = input_229_cast_fp16, y = linear_77_cast_fp16)[name = tensor("input_235_cast_fp16")]; + tensor input_237_axes_0 = const()[name = tensor("input_237_axes_0"), val = tensor([-1])]; + tensor norm19_2_weight_to_fp16 = const()[name = tensor("norm19_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486728000)))]; + tensor norm19_2_bias_to_fp16 = const()[name = tensor("norm19_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486730112)))]; + tensor var_4919_to_fp16 = const()[name = tensor("op_4919_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_237_cast_fp16 = layer_norm(axes = input_237_axes_0, beta = norm19_2_bias_to_fp16, epsilon = var_4919_to_fp16, gamma = norm19_2_weight_to_fp16, x = input_235_cast_fp16)[name = tensor("input_237_cast_fp16")]; + tensor linear19_1_weight_to_fp16 = const()[name = tensor("linear19_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486732224)))]; + tensor linear_78_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear19_1_weight_to_fp16, x = input_237_cast_fp16)[name = tensor("linear_78_cast_fp16")]; + tensor input_239_mode_0 = const()[name = tensor("input_239_mode_0"), val = tensor("EXACT")]; + tensor input_239_cast_fp16 = gelu(mode = input_239_mode_0, x = linear_78_cast_fp16)[name = tensor("input_239_cast_fp16")]; + tensor linear19_2_weight_to_fp16 = const()[name = tensor("linear19_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(495120896)))]; + tensor linear_79_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear19_2_weight_to_fp16, x = input_239_cast_fp16)[name = tensor("linear_79_cast_fp16")]; + tensor input_241_cast_fp16 = add(x = input_235_cast_fp16, y = linear_79_cast_fp16)[name = tensor("input_241_cast_fp16")]; + tensor input_243_axes_0 = const()[name = tensor("input_243_axes_0"), val = tensor([-1])]; + tensor norm20_1_weight_to_fp16 = const()[name = tensor("norm20_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(503509568)))]; + tensor norm20_1_bias_to_fp16 = const()[name = tensor("norm20_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(503511680)))]; + tensor var_4936_to_fp16 = const()[name = tensor("op_4936_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_243_cast_fp16 = layer_norm(axes = input_243_axes_0, beta = norm20_1_bias_to_fp16, epsilon = var_4936_to_fp16, gamma = norm20_1_weight_to_fp16, x = input_241_cast_fp16)[name = tensor("input_243_cast_fp16")]; + tensor attn20_in_proj_weight_to_fp16 = const()[name = tensor("attn20_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(503513792)))]; + tensor linear_80_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn20_in_proj_weight_to_fp16, x = input_243_cast_fp16)[name = tensor("linear_80_cast_fp16")]; + tensor var_4950 = const()[name = tensor("op_4950"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_41_cast_fp16 = reshape(shape = var_4950, x = linear_80_cast_fp16)[name = tensor("qkv_41_cast_fp16")]; + tensor x_81_begin_0 = const()[name = tensor("x_81_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_81_end_0 = const()[name = tensor("x_81_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_81_end_mask_0 = const()[name = tensor("x_81_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_81_squeeze_mask_0 = const()[name = tensor("x_81_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_81_cast_fp16 = slice_by_index(begin = x_81_begin_0, end = x_81_end_0, end_mask = x_81_end_mask_0, squeeze_mask = x_81_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("x_81_cast_fp16")]; + tensor x_83_begin_0 = const()[name = tensor("x_83_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_83_end_0 = const()[name = tensor("x_83_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_83_end_mask_0 = const()[name = tensor("x_83_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_83_squeeze_mask_0 = const()[name = tensor("x_83_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_83_cast_fp16 = slice_by_index(begin = x_83_begin_0, end = x_83_end_0, end_mask = x_83_end_mask_0, squeeze_mask = x_83_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("x_83_cast_fp16")]; + tensor v_41_begin_0 = const()[name = tensor("v_41_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_41_end_0 = const()[name = tensor("v_41_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_41_end_mask_0 = const()[name = tensor("v_41_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_41_squeeze_mask_0 = const()[name = tensor("v_41_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_41_cast_fp16 = slice_by_index(begin = v_41_begin_0, end = v_41_end_0, end_mask = v_41_end_mask_0, squeeze_mask = v_41_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("v_41_cast_fp16")]; + tensor var_4995 = const()[name = tensor("op_4995"), val = tensor([256, 16, 32, 2])]; + tensor pairs_81_cast_fp16 = reshape(shape = var_4995, x = x_81_cast_fp16)[name = tensor("pairs_81_cast_fp16")]; + tensor xr_81_begin_0 = const()[name = tensor("xr_81_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_81_end_0 = const()[name = tensor("xr_81_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_81_end_mask_0 = const()[name = tensor("xr_81_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_81_squeeze_mask_0 = const()[name = tensor("xr_81_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_81_cast_fp16 = slice_by_index(begin = xr_81_begin_0, end = xr_81_end_0, end_mask = xr_81_end_mask_0, squeeze_mask = xr_81_squeeze_mask_0, x = pairs_81_cast_fp16)[name = tensor("xr_81_cast_fp16")]; + tensor xi_81_begin_0 = const()[name = tensor("xi_81_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_81_end_0 = const()[name = tensor("xi_81_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_81_end_mask_0 = const()[name = tensor("xi_81_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_81_squeeze_mask_0 = const()[name = tensor("xi_81_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_81_cast_fp16 = slice_by_index(begin = xi_81_begin_0, end = xi_81_end_0, end_mask = xi_81_end_mask_0, squeeze_mask = xi_81_squeeze_mask_0, x = pairs_81_cast_fp16)[name = tensor("xi_81_cast_fp16")]; + tensor var_5003_cast_fp16 = mul(x = xr_81_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5003_cast_fp16")]; + tensor var_5004_cast_fp16 = mul(x = xi_81_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5004_cast_fp16")]; + tensor xor__81_cast_fp16 = sub(x = var_5003_cast_fp16, y = var_5004_cast_fp16)[name = tensor("xor__81_cast_fp16")]; + tensor var_5007_cast_fp16 = mul(x = xr_81_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5007_cast_fp16")]; + tensor var_5008_cast_fp16 = mul(x = xi_81_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5008_cast_fp16")]; + tensor xoi_81_cast_fp16 = add(x = var_5007_cast_fp16, y = var_5008_cast_fp16)[name = tensor("xoi_81_cast_fp16")]; + tensor out_81_axis_0 = const()[name = tensor("out_81_axis_0"), val = tensor(-1)]; + tensor out_81_cast_fp16 = stack(axis = out_81_axis_0, values = (xor__81_cast_fp16, xoi_81_cast_fp16))[name = tensor("out_81_cast_fp16")]; + tensor var_5018 = const()[name = tensor("op_5018"), val = tensor([1, 256, 16, 64])]; + tensor q_41_cast_fp16 = reshape(shape = var_5018, x = out_81_cast_fp16)[name = tensor("q_41_cast_fp16")]; + tensor var_5024 = const()[name = tensor("op_5024"), val = tensor([256, 16, 32, 2])]; + tensor pairs_83_cast_fp16 = reshape(shape = var_5024, x = x_83_cast_fp16)[name = tensor("pairs_83_cast_fp16")]; + tensor xr_83_begin_0 = const()[name = tensor("xr_83_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_83_end_0 = const()[name = tensor("xr_83_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_83_end_mask_0 = const()[name = tensor("xr_83_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_83_squeeze_mask_0 = const()[name = tensor("xr_83_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_83_cast_fp16 = slice_by_index(begin = xr_83_begin_0, end = xr_83_end_0, end_mask = xr_83_end_mask_0, squeeze_mask = xr_83_squeeze_mask_0, x = pairs_83_cast_fp16)[name = tensor("xr_83_cast_fp16")]; + tensor xi_83_begin_0 = const()[name = tensor("xi_83_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_83_end_0 = const()[name = tensor("xi_83_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_83_end_mask_0 = const()[name = tensor("xi_83_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_83_squeeze_mask_0 = const()[name = tensor("xi_83_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_83_cast_fp16 = slice_by_index(begin = xi_83_begin_0, end = xi_83_end_0, end_mask = xi_83_end_mask_0, squeeze_mask = xi_83_squeeze_mask_0, x = pairs_83_cast_fp16)[name = tensor("xi_83_cast_fp16")]; + tensor var_5032_cast_fp16 = mul(x = xr_83_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5032_cast_fp16")]; + tensor var_5033_cast_fp16 = mul(x = xi_83_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5033_cast_fp16")]; + tensor xor__83_cast_fp16 = sub(x = var_5032_cast_fp16, y = var_5033_cast_fp16)[name = tensor("xor__83_cast_fp16")]; + tensor var_5036_cast_fp16 = mul(x = xr_83_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5036_cast_fp16")]; + tensor var_5037_cast_fp16 = mul(x = xi_83_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5037_cast_fp16")]; + tensor xoi_83_cast_fp16 = add(x = var_5036_cast_fp16, y = var_5037_cast_fp16)[name = tensor("xoi_83_cast_fp16")]; + tensor out_83_axis_0 = const()[name = tensor("out_83_axis_0"), val = tensor(-1)]; + tensor out_83_cast_fp16 = stack(axis = out_83_axis_0, values = (xor__83_cast_fp16, xoi_83_cast_fp16))[name = tensor("out_83_cast_fp16")]; + tensor var_5054 = const()[name = tensor("op_5054"), val = tensor([256, 1024])]; + tensor var_5055_cast_fp16 = reshape(shape = var_5054, x = out_83_cast_fp16)[name = tensor("op_5055_cast_fp16")]; + tensor written_k_41_transpose_x_1 = const()[name = tensor("written_k_41_transpose_x_1"), val = tensor(true)]; + tensor written_k_41_transpose_y_1 = const()[name = tensor("written_k_41_transpose_y_1"), val = tensor(false)]; + tensor written_k_41_cast_fp16 = matmul(transpose_x = written_k_41_transpose_x_1, transpose_y = written_k_41_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_5055_cast_fp16)[name = tensor("written_k_41_cast_fp16")]; + tensor var_5062 = const()[name = tensor("op_5062"), val = tensor([256, 1024])]; + tensor var_5063_cast_fp16 = reshape(shape = var_5062, x = v_41_cast_fp16)[name = tensor("op_5063_cast_fp16")]; + tensor written_v_41_transpose_x_1 = const()[name = tensor("written_v_41_transpose_x_1"), val = tensor(true)]; + tensor written_v_41_transpose_y_1 = const()[name = tensor("written_v_41_transpose_y_1"), val = tensor(false)]; + tensor written_v_41_cast_fp16 = matmul(transpose_x = written_v_41_transpose_x_1, transpose_y = written_v_41_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_5063_cast_fp16)[name = tensor("written_v_41_cast_fp16")]; + tensor k_cache20_to_fp16_dtype_0 = const()[name = tensor("k_cache20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache20_to_fp16 = cast(dtype = k_cache20_to_fp16_dtype_0, x = k_cache20)[name = tensor("cast_345")]; + tensor var_5074_cast_fp16 = mul(x = k_cache20_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_5074_cast_fp16")]; + tensor var_5079 = const()[name = tensor("op_5079"), val = tensor([1, 512, 16, 64])]; + tensor var_5080_cast_fp16 = reshape(shape = var_5079, x = written_k_41_cast_fp16)[name = tensor("op_5080_cast_fp16")]; + tensor new_k_cache_41_cast_fp16 = add(x = var_5074_cast_fp16, y = var_5080_cast_fp16)[name = tensor("new_k_cache_41_cast_fp16")]; + tensor new_k_cache_41_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_41_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache20_to_fp16_dtype_0 = const()[name = tensor("v_cache20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache20_to_fp16 = cast(dtype = v_cache20_to_fp16_dtype_0, x = v_cache20)[name = tensor("cast_343")]; + tensor var_5083_cast_fp16 = mul(x = v_cache20_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_5083_cast_fp16")]; + tensor var_5088 = const()[name = tensor("op_5088"), val = tensor([1, 512, 16, 64])]; + tensor var_5089_cast_fp16 = reshape(shape = var_5088, x = written_v_41_cast_fp16)[name = tensor("op_5089_cast_fp16")]; + tensor new_v_cache_41_cast_fp16 = add(x = var_5083_cast_fp16, y = var_5089_cast_fp16)[name = tensor("new_v_cache_41_cast_fp16")]; + tensor new_v_cache_41_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_41_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_5102 = const()[name = tensor("op_5102"), val = tensor([0, 2, 1, 3])]; + tensor var_5113_transpose_x_0 = const()[name = tensor("op_5113_transpose_x_0"), val = tensor(false)]; + tensor var_5113_transpose_y_0 = const()[name = tensor("op_5113_transpose_y_0"), val = tensor(false)]; + tensor transpose_109_perm_0 = const()[name = tensor("transpose_109_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_110_perm_0 = const()[name = tensor("transpose_110_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_110 = transpose(perm = transpose_110_perm_0, x = new_k_cache_41_cast_fp16)[name = tensor("transpose_124")]; + tensor transpose_109 = transpose(perm = transpose_109_perm_0, x = q_41_cast_fp16)[name = tensor("transpose_125")]; + tensor var_5113_cast_fp16 = matmul(transpose_x = var_5113_transpose_x_0, transpose_y = var_5113_transpose_y_0, x = transpose_109, y = transpose_110)[name = tensor("op_5113_cast_fp16")]; + tensor var_5114_to_fp16 = const()[name = tensor("op_5114_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_81_cast_fp16 = mul(x = var_5113_cast_fp16, y = var_5114_to_fp16)[name = tensor("scores_81_cast_fp16")]; + tensor scores_83_cast_fp16 = add(x = scores_81_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_83_cast_fp16")]; + tensor var_5124 = const()[name = tensor("op_5124"), val = tensor(-1)]; + tensor weights_41_cast_fp16 = softmax(axis = var_5124, x = scores_83_cast_fp16)[name = tensor("weights_41_cast_fp16")]; + tensor attn_41_transpose_x_0 = const()[name = tensor("attn_41_transpose_x_0"), val = tensor(false)]; + tensor attn_41_transpose_y_0 = const()[name = tensor("attn_41_transpose_y_0"), val = tensor(false)]; + tensor values_41_cast_fp16 = transpose(perm = var_5102, x = new_v_cache_41_cast_fp16)[name = tensor("transpose_126")]; + tensor attn_41_cast_fp16 = matmul(transpose_x = attn_41_transpose_x_0, transpose_y = attn_41_transpose_y_0, x = weights_41_cast_fp16, y = values_41_cast_fp16)[name = tensor("attn_41_cast_fp16")]; + tensor var_5132 = const()[name = tensor("op_5132"), val = tensor([0, 2, 1, 3])]; + tensor var_5137 = const()[name = tensor("op_5137"), val = tensor([1, 256, 1024])]; + tensor var_5133_cast_fp16 = transpose(perm = var_5132, x = attn_41_cast_fp16)[name = tensor("transpose_123")]; + tensor input_245_cast_fp16 = reshape(shape = var_5137, x = var_5133_cast_fp16)[name = tensor("input_245_cast_fp16")]; + tensor attn20_out_proj_weight_to_fp16 = const()[name = tensor("attn20_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(509805312)))]; + tensor linear_81_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn20_out_proj_weight_to_fp16, x = input_245_cast_fp16)[name = tensor("linear_81_cast_fp16")]; + tensor position20_to_fp16_dtype_0 = const()[name = tensor("position20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position20_to_fp16 = cast(dtype = position20_to_fp16_dtype_0, x = position20)[name = tensor("cast_341")]; + tensor var_5146_cast_fp16 = add(x = position20_to_fp16, y = valid_len_to_fp16)[name = tensor("op_5146_cast_fp16")]; + tensor var_5146_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5146_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_247_cast_fp16 = add(x = input_241_cast_fp16, y = linear_81_cast_fp16)[name = tensor("input_247_cast_fp16")]; + tensor input_249_axes_0 = const()[name = tensor("input_249_axes_0"), val = tensor([-1])]; + tensor norm20_2_weight_to_fp16 = const()[name = tensor("norm20_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(511902528)))]; + tensor norm20_2_bias_to_fp16 = const()[name = tensor("norm20_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(511904640)))]; + tensor var_5150_to_fp16 = const()[name = tensor("op_5150_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_249_cast_fp16 = layer_norm(axes = input_249_axes_0, beta = norm20_2_bias_to_fp16, epsilon = var_5150_to_fp16, gamma = norm20_2_weight_to_fp16, x = input_247_cast_fp16)[name = tensor("input_249_cast_fp16")]; + tensor linear20_1_weight_to_fp16 = const()[name = tensor("linear20_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(511906752)))]; + tensor linear_82_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear20_1_weight_to_fp16, x = input_249_cast_fp16)[name = tensor("linear_82_cast_fp16")]; + tensor input_251_mode_0 = const()[name = tensor("input_251_mode_0"), val = tensor("EXACT")]; + tensor input_251_cast_fp16 = gelu(mode = input_251_mode_0, x = linear_82_cast_fp16)[name = tensor("input_251_cast_fp16")]; + tensor linear20_2_weight_to_fp16 = const()[name = tensor("linear20_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(520295424)))]; + tensor linear_83_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear20_2_weight_to_fp16, x = input_251_cast_fp16)[name = tensor("linear_83_cast_fp16")]; + tensor input_253_cast_fp16 = add(x = input_247_cast_fp16, y = linear_83_cast_fp16)[name = tensor("input_253_cast_fp16")]; + tensor input_255_axes_0 = const()[name = tensor("input_255_axes_0"), val = tensor([-1])]; + tensor norm21_1_weight_to_fp16 = const()[name = tensor("norm21_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(528684096)))]; + tensor norm21_1_bias_to_fp16 = const()[name = tensor("norm21_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(528686208)))]; + tensor var_5167_to_fp16 = const()[name = tensor("op_5167_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_255_cast_fp16 = layer_norm(axes = input_255_axes_0, beta = norm21_1_bias_to_fp16, epsilon = var_5167_to_fp16, gamma = norm21_1_weight_to_fp16, x = input_253_cast_fp16)[name = tensor("input_255_cast_fp16")]; + tensor attn21_in_proj_weight_to_fp16 = const()[name = tensor("attn21_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(528688320)))]; + tensor linear_84_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn21_in_proj_weight_to_fp16, x = input_255_cast_fp16)[name = tensor("linear_84_cast_fp16")]; + tensor var_5181 = const()[name = tensor("op_5181"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_43_cast_fp16 = reshape(shape = var_5181, x = linear_84_cast_fp16)[name = tensor("qkv_43_cast_fp16")]; + tensor x_85_begin_0 = const()[name = tensor("x_85_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_85_end_0 = const()[name = tensor("x_85_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_85_end_mask_0 = const()[name = tensor("x_85_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_85_squeeze_mask_0 = const()[name = tensor("x_85_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_85_cast_fp16 = slice_by_index(begin = x_85_begin_0, end = x_85_end_0, end_mask = x_85_end_mask_0, squeeze_mask = x_85_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("x_85_cast_fp16")]; + tensor x_87_begin_0 = const()[name = tensor("x_87_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_87_end_0 = const()[name = tensor("x_87_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_87_end_mask_0 = const()[name = tensor("x_87_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_87_squeeze_mask_0 = const()[name = tensor("x_87_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_87_cast_fp16 = slice_by_index(begin = x_87_begin_0, end = x_87_end_0, end_mask = x_87_end_mask_0, squeeze_mask = x_87_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("x_87_cast_fp16")]; + tensor v_43_begin_0 = const()[name = tensor("v_43_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_43_end_0 = const()[name = tensor("v_43_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_43_end_mask_0 = const()[name = tensor("v_43_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_43_squeeze_mask_0 = const()[name = tensor("v_43_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_43_cast_fp16 = slice_by_index(begin = v_43_begin_0, end = v_43_end_0, end_mask = v_43_end_mask_0, squeeze_mask = v_43_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("v_43_cast_fp16")]; + tensor var_5226 = const()[name = tensor("op_5226"), val = tensor([256, 16, 32, 2])]; + tensor pairs_85_cast_fp16 = reshape(shape = var_5226, x = x_85_cast_fp16)[name = tensor("pairs_85_cast_fp16")]; + tensor xr_85_begin_0 = const()[name = tensor("xr_85_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_85_end_0 = const()[name = tensor("xr_85_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_85_end_mask_0 = const()[name = tensor("xr_85_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_85_squeeze_mask_0 = const()[name = tensor("xr_85_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_85_cast_fp16 = slice_by_index(begin = xr_85_begin_0, end = xr_85_end_0, end_mask = xr_85_end_mask_0, squeeze_mask = xr_85_squeeze_mask_0, x = pairs_85_cast_fp16)[name = tensor("xr_85_cast_fp16")]; + tensor xi_85_begin_0 = const()[name = tensor("xi_85_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_85_end_0 = const()[name = tensor("xi_85_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_85_end_mask_0 = const()[name = tensor("xi_85_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_85_squeeze_mask_0 = const()[name = tensor("xi_85_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_85_cast_fp16 = slice_by_index(begin = xi_85_begin_0, end = xi_85_end_0, end_mask = xi_85_end_mask_0, squeeze_mask = xi_85_squeeze_mask_0, x = pairs_85_cast_fp16)[name = tensor("xi_85_cast_fp16")]; + tensor var_5234_cast_fp16 = mul(x = xr_85_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5234_cast_fp16")]; + tensor var_5235_cast_fp16 = mul(x = xi_85_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5235_cast_fp16")]; + tensor xor__85_cast_fp16 = sub(x = var_5234_cast_fp16, y = var_5235_cast_fp16)[name = tensor("xor__85_cast_fp16")]; + tensor var_5238_cast_fp16 = mul(x = xr_85_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5238_cast_fp16")]; + tensor var_5239_cast_fp16 = mul(x = xi_85_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5239_cast_fp16")]; + tensor xoi_85_cast_fp16 = add(x = var_5238_cast_fp16, y = var_5239_cast_fp16)[name = tensor("xoi_85_cast_fp16")]; + tensor out_85_axis_0 = const()[name = tensor("out_85_axis_0"), val = tensor(-1)]; + tensor out_85_cast_fp16 = stack(axis = out_85_axis_0, values = (xor__85_cast_fp16, xoi_85_cast_fp16))[name = tensor("out_85_cast_fp16")]; + tensor var_5249 = const()[name = tensor("op_5249"), val = tensor([1, 256, 16, 64])]; + tensor q_43_cast_fp16 = reshape(shape = var_5249, x = out_85_cast_fp16)[name = tensor("q_43_cast_fp16")]; + tensor var_5255 = const()[name = tensor("op_5255"), val = tensor([256, 16, 32, 2])]; + tensor pairs_87_cast_fp16 = reshape(shape = var_5255, x = x_87_cast_fp16)[name = tensor("pairs_87_cast_fp16")]; + tensor xr_87_begin_0 = const()[name = tensor("xr_87_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_87_end_0 = const()[name = tensor("xr_87_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_87_end_mask_0 = const()[name = tensor("xr_87_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_87_squeeze_mask_0 = const()[name = tensor("xr_87_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_87_cast_fp16 = slice_by_index(begin = xr_87_begin_0, end = xr_87_end_0, end_mask = xr_87_end_mask_0, squeeze_mask = xr_87_squeeze_mask_0, x = pairs_87_cast_fp16)[name = tensor("xr_87_cast_fp16")]; + tensor xi_87_begin_0 = const()[name = tensor("xi_87_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_87_end_0 = const()[name = tensor("xi_87_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_87_end_mask_0 = const()[name = tensor("xi_87_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_87_squeeze_mask_0 = const()[name = tensor("xi_87_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_87_cast_fp16 = slice_by_index(begin = xi_87_begin_0, end = xi_87_end_0, end_mask = xi_87_end_mask_0, squeeze_mask = xi_87_squeeze_mask_0, x = pairs_87_cast_fp16)[name = tensor("xi_87_cast_fp16")]; + tensor var_5263_cast_fp16 = mul(x = xr_87_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5263_cast_fp16")]; + tensor var_5264_cast_fp16 = mul(x = xi_87_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5264_cast_fp16")]; + tensor xor__87_cast_fp16 = sub(x = var_5263_cast_fp16, y = var_5264_cast_fp16)[name = tensor("xor__87_cast_fp16")]; + tensor var_5267_cast_fp16 = mul(x = xr_87_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5267_cast_fp16")]; + tensor var_5268_cast_fp16 = mul(x = xi_87_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5268_cast_fp16")]; + tensor xoi_87_cast_fp16 = add(x = var_5267_cast_fp16, y = var_5268_cast_fp16)[name = tensor("xoi_87_cast_fp16")]; + tensor out_87_axis_0 = const()[name = tensor("out_87_axis_0"), val = tensor(-1)]; + tensor out_87_cast_fp16 = stack(axis = out_87_axis_0, values = (xor__87_cast_fp16, xoi_87_cast_fp16))[name = tensor("out_87_cast_fp16")]; + tensor var_5285 = const()[name = tensor("op_5285"), val = tensor([256, 1024])]; + tensor var_5286_cast_fp16 = reshape(shape = var_5285, x = out_87_cast_fp16)[name = tensor("op_5286_cast_fp16")]; + tensor written_k_43_transpose_x_1 = const()[name = tensor("written_k_43_transpose_x_1"), val = tensor(true)]; + tensor written_k_43_transpose_y_1 = const()[name = tensor("written_k_43_transpose_y_1"), val = tensor(false)]; + tensor written_k_43_cast_fp16 = matmul(transpose_x = written_k_43_transpose_x_1, transpose_y = written_k_43_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_5286_cast_fp16)[name = tensor("written_k_43_cast_fp16")]; + tensor var_5293 = const()[name = tensor("op_5293"), val = tensor([256, 1024])]; + tensor var_5294_cast_fp16 = reshape(shape = var_5293, x = v_43_cast_fp16)[name = tensor("op_5294_cast_fp16")]; + tensor written_v_43_transpose_x_1 = const()[name = tensor("written_v_43_transpose_x_1"), val = tensor(true)]; + tensor written_v_43_transpose_y_1 = const()[name = tensor("written_v_43_transpose_y_1"), val = tensor(false)]; + tensor written_v_43_cast_fp16 = matmul(transpose_x = written_v_43_transpose_x_1, transpose_y = written_v_43_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_5294_cast_fp16)[name = tensor("written_v_43_cast_fp16")]; + tensor k_cache21_to_fp16_dtype_0 = const()[name = tensor("k_cache21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache21_to_fp16 = cast(dtype = k_cache21_to_fp16_dtype_0, x = k_cache21)[name = tensor("cast_339")]; + tensor var_5305_cast_fp16 = mul(x = k_cache21_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_5305_cast_fp16")]; + tensor var_5310 = const()[name = tensor("op_5310"), val = tensor([1, 512, 16, 64])]; + tensor var_5311_cast_fp16 = reshape(shape = var_5310, x = written_k_43_cast_fp16)[name = tensor("op_5311_cast_fp16")]; + tensor new_k_cache_43_cast_fp16 = add(x = var_5305_cast_fp16, y = var_5311_cast_fp16)[name = tensor("new_k_cache_43_cast_fp16")]; + tensor new_k_cache_43_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_43_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache21_to_fp16_dtype_0 = const()[name = tensor("v_cache21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache21_to_fp16 = cast(dtype = v_cache21_to_fp16_dtype_0, x = v_cache21)[name = tensor("cast_337")]; + tensor var_5314_cast_fp16 = mul(x = v_cache21_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_5314_cast_fp16")]; + tensor var_5319 = const()[name = tensor("op_5319"), val = tensor([1, 512, 16, 64])]; + tensor var_5320_cast_fp16 = reshape(shape = var_5319, x = written_v_43_cast_fp16)[name = tensor("op_5320_cast_fp16")]; + tensor new_v_cache_43_cast_fp16 = add(x = var_5314_cast_fp16, y = var_5320_cast_fp16)[name = tensor("new_v_cache_43_cast_fp16")]; + tensor new_v_cache_43_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_43_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_5333 = const()[name = tensor("op_5333"), val = tensor([0, 2, 1, 3])]; + tensor var_5344_transpose_x_0 = const()[name = tensor("op_5344_transpose_x_0"), val = tensor(false)]; + tensor var_5344_transpose_y_0 = const()[name = tensor("op_5344_transpose_y_0"), val = tensor(false)]; + tensor transpose_111_perm_0 = const()[name = tensor("transpose_111_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_112_perm_0 = const()[name = tensor("transpose_112_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_112 = transpose(perm = transpose_112_perm_0, x = new_k_cache_43_cast_fp16)[name = tensor("transpose_120")]; + tensor transpose_111 = transpose(perm = transpose_111_perm_0, x = q_43_cast_fp16)[name = tensor("transpose_121")]; + tensor var_5344_cast_fp16 = matmul(transpose_x = var_5344_transpose_x_0, transpose_y = var_5344_transpose_y_0, x = transpose_111, y = transpose_112)[name = tensor("op_5344_cast_fp16")]; + tensor var_5345_to_fp16 = const()[name = tensor("op_5345_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_85_cast_fp16 = mul(x = var_5344_cast_fp16, y = var_5345_to_fp16)[name = tensor("scores_85_cast_fp16")]; + tensor scores_87_cast_fp16 = add(x = scores_85_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_87_cast_fp16")]; + tensor var_5355 = const()[name = tensor("op_5355"), val = tensor(-1)]; + tensor weights_43_cast_fp16 = softmax(axis = var_5355, x = scores_87_cast_fp16)[name = tensor("weights_43_cast_fp16")]; + tensor attn_43_transpose_x_0 = const()[name = tensor("attn_43_transpose_x_0"), val = tensor(false)]; + tensor attn_43_transpose_y_0 = const()[name = tensor("attn_43_transpose_y_0"), val = tensor(false)]; + tensor values_43_cast_fp16 = transpose(perm = var_5333, x = new_v_cache_43_cast_fp16)[name = tensor("transpose_122")]; + tensor attn_43_cast_fp16 = matmul(transpose_x = attn_43_transpose_x_0, transpose_y = attn_43_transpose_y_0, x = weights_43_cast_fp16, y = values_43_cast_fp16)[name = tensor("attn_43_cast_fp16")]; + tensor var_5363 = const()[name = tensor("op_5363"), val = tensor([0, 2, 1, 3])]; + tensor var_5368 = const()[name = tensor("op_5368"), val = tensor([1, 256, 1024])]; + tensor var_5364_cast_fp16 = transpose(perm = var_5363, x = attn_43_cast_fp16)[name = tensor("transpose_119")]; + tensor input_257_cast_fp16 = reshape(shape = var_5368, x = var_5364_cast_fp16)[name = tensor("input_257_cast_fp16")]; + tensor attn21_out_proj_weight_to_fp16 = const()[name = tensor("attn21_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(534979840)))]; + tensor linear_85_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn21_out_proj_weight_to_fp16, x = input_257_cast_fp16)[name = tensor("linear_85_cast_fp16")]; + tensor position21_to_fp16_dtype_0 = const()[name = tensor("position21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position21_to_fp16 = cast(dtype = position21_to_fp16_dtype_0, x = position21)[name = tensor("cast_335")]; + tensor var_5377_cast_fp16 = add(x = position21_to_fp16, y = valid_len_to_fp16)[name = tensor("op_5377_cast_fp16")]; + tensor var_5377_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5377_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_259_cast_fp16 = add(x = input_253_cast_fp16, y = linear_85_cast_fp16)[name = tensor("input_259_cast_fp16")]; + tensor input_261_axes_0 = const()[name = tensor("input_261_axes_0"), val = tensor([-1])]; + tensor norm21_2_weight_to_fp16 = const()[name = tensor("norm21_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(537077056)))]; + tensor norm21_2_bias_to_fp16 = const()[name = tensor("norm21_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(537079168)))]; + tensor var_5381_to_fp16 = const()[name = tensor("op_5381_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_261_cast_fp16 = layer_norm(axes = input_261_axes_0, beta = norm21_2_bias_to_fp16, epsilon = var_5381_to_fp16, gamma = norm21_2_weight_to_fp16, x = input_259_cast_fp16)[name = tensor("input_261_cast_fp16")]; + tensor linear21_1_weight_to_fp16 = const()[name = tensor("linear21_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(537081280)))]; + tensor linear_86_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear21_1_weight_to_fp16, x = input_261_cast_fp16)[name = tensor("linear_86_cast_fp16")]; + tensor input_263_mode_0 = const()[name = tensor("input_263_mode_0"), val = tensor("EXACT")]; + tensor input_263_cast_fp16 = gelu(mode = input_263_mode_0, x = linear_86_cast_fp16)[name = tensor("input_263_cast_fp16")]; + tensor linear21_2_weight_to_fp16 = const()[name = tensor("linear21_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(545469952)))]; + tensor linear_87_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear21_2_weight_to_fp16, x = input_263_cast_fp16)[name = tensor("linear_87_cast_fp16")]; + tensor input_265_cast_fp16 = add(x = input_259_cast_fp16, y = linear_87_cast_fp16)[name = tensor("input_265_cast_fp16")]; + tensor input_267_axes_0 = const()[name = tensor("input_267_axes_0"), val = tensor([-1])]; + tensor norm22_1_weight_to_fp16 = const()[name = tensor("norm22_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(553858624)))]; + tensor norm22_1_bias_to_fp16 = const()[name = tensor("norm22_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(553860736)))]; + tensor var_5398_to_fp16 = const()[name = tensor("op_5398_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_267_cast_fp16 = layer_norm(axes = input_267_axes_0, beta = norm22_1_bias_to_fp16, epsilon = var_5398_to_fp16, gamma = norm22_1_weight_to_fp16, x = input_265_cast_fp16)[name = tensor("input_267_cast_fp16")]; + tensor attn22_in_proj_weight_to_fp16 = const()[name = tensor("attn22_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(553862848)))]; + tensor linear_88_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn22_in_proj_weight_to_fp16, x = input_267_cast_fp16)[name = tensor("linear_88_cast_fp16")]; + tensor var_5412 = const()[name = tensor("op_5412"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_45_cast_fp16 = reshape(shape = var_5412, x = linear_88_cast_fp16)[name = tensor("qkv_45_cast_fp16")]; + tensor x_89_begin_0 = const()[name = tensor("x_89_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_89_end_0 = const()[name = tensor("x_89_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_89_end_mask_0 = const()[name = tensor("x_89_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_89_squeeze_mask_0 = const()[name = tensor("x_89_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_89_cast_fp16 = slice_by_index(begin = x_89_begin_0, end = x_89_end_0, end_mask = x_89_end_mask_0, squeeze_mask = x_89_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("x_89_cast_fp16")]; + tensor x_91_begin_0 = const()[name = tensor("x_91_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_91_end_0 = const()[name = tensor("x_91_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_91_end_mask_0 = const()[name = tensor("x_91_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_91_squeeze_mask_0 = const()[name = tensor("x_91_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_91_cast_fp16 = slice_by_index(begin = x_91_begin_0, end = x_91_end_0, end_mask = x_91_end_mask_0, squeeze_mask = x_91_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("x_91_cast_fp16")]; + tensor v_45_begin_0 = const()[name = tensor("v_45_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_45_end_0 = const()[name = tensor("v_45_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_45_end_mask_0 = const()[name = tensor("v_45_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_45_squeeze_mask_0 = const()[name = tensor("v_45_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_45_cast_fp16 = slice_by_index(begin = v_45_begin_0, end = v_45_end_0, end_mask = v_45_end_mask_0, squeeze_mask = v_45_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("v_45_cast_fp16")]; + tensor var_5457 = const()[name = tensor("op_5457"), val = tensor([256, 16, 32, 2])]; + tensor pairs_89_cast_fp16 = reshape(shape = var_5457, x = x_89_cast_fp16)[name = tensor("pairs_89_cast_fp16")]; + tensor xr_89_begin_0 = const()[name = tensor("xr_89_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_89_end_0 = const()[name = tensor("xr_89_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_89_end_mask_0 = const()[name = tensor("xr_89_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_89_squeeze_mask_0 = const()[name = tensor("xr_89_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_89_cast_fp16 = slice_by_index(begin = xr_89_begin_0, end = xr_89_end_0, end_mask = xr_89_end_mask_0, squeeze_mask = xr_89_squeeze_mask_0, x = pairs_89_cast_fp16)[name = tensor("xr_89_cast_fp16")]; + tensor xi_89_begin_0 = const()[name = tensor("xi_89_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_89_end_0 = const()[name = tensor("xi_89_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_89_end_mask_0 = const()[name = tensor("xi_89_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_89_squeeze_mask_0 = const()[name = tensor("xi_89_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_89_cast_fp16 = slice_by_index(begin = xi_89_begin_0, end = xi_89_end_0, end_mask = xi_89_end_mask_0, squeeze_mask = xi_89_squeeze_mask_0, x = pairs_89_cast_fp16)[name = tensor("xi_89_cast_fp16")]; + tensor var_5465_cast_fp16 = mul(x = xr_89_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5465_cast_fp16")]; + tensor var_5466_cast_fp16 = mul(x = xi_89_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5466_cast_fp16")]; + tensor xor__89_cast_fp16 = sub(x = var_5465_cast_fp16, y = var_5466_cast_fp16)[name = tensor("xor__89_cast_fp16")]; + tensor var_5469_cast_fp16 = mul(x = xr_89_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5469_cast_fp16")]; + tensor var_5470_cast_fp16 = mul(x = xi_89_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5470_cast_fp16")]; + tensor xoi_89_cast_fp16 = add(x = var_5469_cast_fp16, y = var_5470_cast_fp16)[name = tensor("xoi_89_cast_fp16")]; + tensor out_89_axis_0 = const()[name = tensor("out_89_axis_0"), val = tensor(-1)]; + tensor out_89_cast_fp16 = stack(axis = out_89_axis_0, values = (xor__89_cast_fp16, xoi_89_cast_fp16))[name = tensor("out_89_cast_fp16")]; + tensor var_5480 = const()[name = tensor("op_5480"), val = tensor([1, 256, 16, 64])]; + tensor q_45_cast_fp16 = reshape(shape = var_5480, x = out_89_cast_fp16)[name = tensor("q_45_cast_fp16")]; + tensor var_5486 = const()[name = tensor("op_5486"), val = tensor([256, 16, 32, 2])]; + tensor pairs_91_cast_fp16 = reshape(shape = var_5486, x = x_91_cast_fp16)[name = tensor("pairs_91_cast_fp16")]; + tensor xr_91_begin_0 = const()[name = tensor("xr_91_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_91_end_0 = const()[name = tensor("xr_91_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_91_end_mask_0 = const()[name = tensor("xr_91_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_91_squeeze_mask_0 = const()[name = tensor("xr_91_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_91_cast_fp16 = slice_by_index(begin = xr_91_begin_0, end = xr_91_end_0, end_mask = xr_91_end_mask_0, squeeze_mask = xr_91_squeeze_mask_0, x = pairs_91_cast_fp16)[name = tensor("xr_91_cast_fp16")]; + tensor xi_91_begin_0 = const()[name = tensor("xi_91_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_91_end_0 = const()[name = tensor("xi_91_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_91_end_mask_0 = const()[name = tensor("xi_91_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_91_squeeze_mask_0 = const()[name = tensor("xi_91_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_91_cast_fp16 = slice_by_index(begin = xi_91_begin_0, end = xi_91_end_0, end_mask = xi_91_end_mask_0, squeeze_mask = xi_91_squeeze_mask_0, x = pairs_91_cast_fp16)[name = tensor("xi_91_cast_fp16")]; + tensor var_5494_cast_fp16 = mul(x = xr_91_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5494_cast_fp16")]; + tensor var_5495_cast_fp16 = mul(x = xi_91_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5495_cast_fp16")]; + tensor xor__91_cast_fp16 = sub(x = var_5494_cast_fp16, y = var_5495_cast_fp16)[name = tensor("xor__91_cast_fp16")]; + tensor var_5498_cast_fp16 = mul(x = xr_91_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5498_cast_fp16")]; + tensor var_5499_cast_fp16 = mul(x = xi_91_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5499_cast_fp16")]; + tensor xoi_91_cast_fp16 = add(x = var_5498_cast_fp16, y = var_5499_cast_fp16)[name = tensor("xoi_91_cast_fp16")]; + tensor out_91_axis_0 = const()[name = tensor("out_91_axis_0"), val = tensor(-1)]; + tensor out_91_cast_fp16 = stack(axis = out_91_axis_0, values = (xor__91_cast_fp16, xoi_91_cast_fp16))[name = tensor("out_91_cast_fp16")]; + tensor var_5516 = const()[name = tensor("op_5516"), val = tensor([256, 1024])]; + tensor var_5517_cast_fp16 = reshape(shape = var_5516, x = out_91_cast_fp16)[name = tensor("op_5517_cast_fp16")]; + tensor written_k_45_transpose_x_1 = const()[name = tensor("written_k_45_transpose_x_1"), val = tensor(true)]; + tensor written_k_45_transpose_y_1 = const()[name = tensor("written_k_45_transpose_y_1"), val = tensor(false)]; + tensor written_k_45_cast_fp16 = matmul(transpose_x = written_k_45_transpose_x_1, transpose_y = written_k_45_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_5517_cast_fp16)[name = tensor("written_k_45_cast_fp16")]; + tensor var_5524 = const()[name = tensor("op_5524"), val = tensor([256, 1024])]; + tensor var_5525_cast_fp16 = reshape(shape = var_5524, x = v_45_cast_fp16)[name = tensor("op_5525_cast_fp16")]; + tensor written_v_45_transpose_x_1 = const()[name = tensor("written_v_45_transpose_x_1"), val = tensor(true)]; + tensor written_v_45_transpose_y_1 = const()[name = tensor("written_v_45_transpose_y_1"), val = tensor(false)]; + tensor written_v_45_cast_fp16 = matmul(transpose_x = written_v_45_transpose_x_1, transpose_y = written_v_45_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_5525_cast_fp16)[name = tensor("written_v_45_cast_fp16")]; + tensor k_cache22_to_fp16_dtype_0 = const()[name = tensor("k_cache22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache22_to_fp16 = cast(dtype = k_cache22_to_fp16_dtype_0, x = k_cache22)[name = tensor("cast_333")]; + tensor var_5536_cast_fp16 = mul(x = k_cache22_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_5536_cast_fp16")]; + tensor var_5541 = const()[name = tensor("op_5541"), val = tensor([1, 512, 16, 64])]; + tensor var_5542_cast_fp16 = reshape(shape = var_5541, x = written_k_45_cast_fp16)[name = tensor("op_5542_cast_fp16")]; + tensor new_k_cache_45_cast_fp16 = add(x = var_5536_cast_fp16, y = var_5542_cast_fp16)[name = tensor("new_k_cache_45_cast_fp16")]; + tensor new_k_cache_45_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_45_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache22_to_fp16_dtype_0 = const()[name = tensor("v_cache22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache22_to_fp16 = cast(dtype = v_cache22_to_fp16_dtype_0, x = v_cache22)[name = tensor("cast_331")]; + tensor var_5545_cast_fp16 = mul(x = v_cache22_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_5545_cast_fp16")]; + tensor var_5550 = const()[name = tensor("op_5550"), val = tensor([1, 512, 16, 64])]; + tensor var_5551_cast_fp16 = reshape(shape = var_5550, x = written_v_45_cast_fp16)[name = tensor("op_5551_cast_fp16")]; + tensor new_v_cache_45_cast_fp16 = add(x = var_5545_cast_fp16, y = var_5551_cast_fp16)[name = tensor("new_v_cache_45_cast_fp16")]; + tensor new_v_cache_45_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_45_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_5564 = const()[name = tensor("op_5564"), val = tensor([0, 2, 1, 3])]; + tensor var_5575_transpose_x_0 = const()[name = tensor("op_5575_transpose_x_0"), val = tensor(false)]; + tensor var_5575_transpose_y_0 = const()[name = tensor("op_5575_transpose_y_0"), val = tensor(false)]; + tensor transpose_113_perm_0 = const()[name = tensor("transpose_113_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_114_perm_0 = const()[name = tensor("transpose_114_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_114 = transpose(perm = transpose_114_perm_0, x = new_k_cache_45_cast_fp16)[name = tensor("transpose_116")]; + tensor transpose_113 = transpose(perm = transpose_113_perm_0, x = q_45_cast_fp16)[name = tensor("transpose_117")]; + tensor var_5575_cast_fp16 = matmul(transpose_x = var_5575_transpose_x_0, transpose_y = var_5575_transpose_y_0, x = transpose_113, y = transpose_114)[name = tensor("op_5575_cast_fp16")]; + tensor var_5576_to_fp16 = const()[name = tensor("op_5576_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_89_cast_fp16 = mul(x = var_5575_cast_fp16, y = var_5576_to_fp16)[name = tensor("scores_89_cast_fp16")]; + tensor scores_91_cast_fp16 = add(x = scores_89_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_91_cast_fp16")]; + tensor var_5586 = const()[name = tensor("op_5586"), val = tensor(-1)]; + tensor weights_45_cast_fp16 = softmax(axis = var_5586, x = scores_91_cast_fp16)[name = tensor("weights_45_cast_fp16")]; + tensor attn_45_transpose_x_0 = const()[name = tensor("attn_45_transpose_x_0"), val = tensor(false)]; + tensor attn_45_transpose_y_0 = const()[name = tensor("attn_45_transpose_y_0"), val = tensor(false)]; + tensor values_45_cast_fp16 = transpose(perm = var_5564, x = new_v_cache_45_cast_fp16)[name = tensor("transpose_118")]; + tensor attn_45_cast_fp16 = matmul(transpose_x = attn_45_transpose_x_0, transpose_y = attn_45_transpose_y_0, x = weights_45_cast_fp16, y = values_45_cast_fp16)[name = tensor("attn_45_cast_fp16")]; + tensor var_5594 = const()[name = tensor("op_5594"), val = tensor([0, 2, 1, 3])]; + tensor var_5599 = const()[name = tensor("op_5599"), val = tensor([1, 256, 1024])]; + tensor var_5595_cast_fp16 = transpose(perm = var_5594, x = attn_45_cast_fp16)[name = tensor("transpose_115")]; + tensor input_269_cast_fp16 = reshape(shape = var_5599, x = var_5595_cast_fp16)[name = tensor("input_269_cast_fp16")]; + tensor attn22_out_proj_weight_to_fp16 = const()[name = tensor("attn22_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(560154368)))]; + tensor linear_89_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn22_out_proj_weight_to_fp16, x = input_269_cast_fp16)[name = tensor("linear_89_cast_fp16")]; + tensor position22_to_fp16_dtype_0 = const()[name = tensor("position22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position22_to_fp16 = cast(dtype = position22_to_fp16_dtype_0, x = position22)[name = tensor("cast_329")]; + tensor var_5608_cast_fp16 = add(x = position22_to_fp16, y = valid_len_to_fp16)[name = tensor("op_5608_cast_fp16")]; + tensor var_5608_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5608_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_271_cast_fp16 = add(x = input_265_cast_fp16, y = linear_89_cast_fp16)[name = tensor("input_271_cast_fp16")]; + tensor input_273_axes_0 = const()[name = tensor("input_273_axes_0"), val = tensor([-1])]; + tensor norm22_2_weight_to_fp16 = const()[name = tensor("norm22_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(562251584)))]; + tensor norm22_2_bias_to_fp16 = const()[name = tensor("norm22_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(562253696)))]; + tensor var_5612_to_fp16 = const()[name = tensor("op_5612_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_273_cast_fp16 = layer_norm(axes = input_273_axes_0, beta = norm22_2_bias_to_fp16, epsilon = var_5612_to_fp16, gamma = norm22_2_weight_to_fp16, x = input_271_cast_fp16)[name = tensor("input_273_cast_fp16")]; + tensor linear22_1_weight_to_fp16 = const()[name = tensor("linear22_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(562255808)))]; + tensor linear_90_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear22_1_weight_to_fp16, x = input_273_cast_fp16)[name = tensor("linear_90_cast_fp16")]; + tensor input_275_mode_0 = const()[name = tensor("input_275_mode_0"), val = tensor("EXACT")]; + tensor input_275_cast_fp16 = gelu(mode = input_275_mode_0, x = linear_90_cast_fp16)[name = tensor("input_275_cast_fp16")]; + tensor linear22_2_weight_to_fp16 = const()[name = tensor("linear22_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(570644480)))]; + tensor linear_91_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear22_2_weight_to_fp16, x = input_275_cast_fp16)[name = tensor("linear_91_cast_fp16")]; + tensor input_277_cast_fp16 = add(x = input_271_cast_fp16, y = linear_91_cast_fp16)[name = tensor("input_277_cast_fp16")]; + tensor input_279_axes_0 = const()[name = tensor("input_279_axes_0"), val = tensor([-1])]; + tensor norm23_1_weight_to_fp16 = const()[name = tensor("norm23_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(579033152)))]; + tensor norm23_1_bias_to_fp16 = const()[name = tensor("norm23_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(579035264)))]; + tensor var_5629_to_fp16 = const()[name = tensor("op_5629_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_279_cast_fp16 = layer_norm(axes = input_279_axes_0, beta = norm23_1_bias_to_fp16, epsilon = var_5629_to_fp16, gamma = norm23_1_weight_to_fp16, x = input_277_cast_fp16)[name = tensor("input_279_cast_fp16")]; + tensor attn23_in_proj_weight_to_fp16 = const()[name = tensor("attn23_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(579037376)))]; + tensor linear_92_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn23_in_proj_weight_to_fp16, x = input_279_cast_fp16)[name = tensor("linear_92_cast_fp16")]; + tensor var_5643 = const()[name = tensor("op_5643"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_cast_fp16 = reshape(shape = var_5643, x = linear_92_cast_fp16)[name = tensor("qkv_cast_fp16")]; + tensor x_begin_0 = const()[name = tensor("x_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_end_0 = const()[name = tensor("x_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_end_mask_0 = const()[name = tensor("x_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_squeeze_mask_0 = const()[name = tensor("x_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_cast_fp16 = slice_by_index(begin = x_begin_0, end = x_end_0, end_mask = x_end_mask_0, squeeze_mask = x_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("x_cast_fp16")]; + tensor v_begin_0 = const()[name = tensor("v_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_end_0 = const()[name = tensor("v_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_end_mask_0 = const()[name = tensor("v_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_squeeze_mask_0 = const()[name = tensor("v_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_cast_fp16 = slice_by_index(begin = v_begin_0, end = v_end_0, end_mask = v_end_mask_0, squeeze_mask = v_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("v_cast_fp16")]; + tensor var_5675 = const()[name = tensor("op_5675"), val = tensor([256, 16, 32, 2])]; + tensor pairs_cast_fp16 = reshape(shape = var_5675, x = x_cast_fp16)[name = tensor("pairs_cast_fp16")]; + tensor xr_begin_0 = const()[name = tensor("xr_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_end_0 = const()[name = tensor("xr_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_end_mask_0 = const()[name = tensor("xr_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_squeeze_mask_0 = const()[name = tensor("xr_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_cast_fp16 = slice_by_index(begin = xr_begin_0, end = xr_end_0, end_mask = xr_end_mask_0, squeeze_mask = xr_squeeze_mask_0, x = pairs_cast_fp16)[name = tensor("xr_cast_fp16")]; + tensor xi_begin_0 = const()[name = tensor("xi_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_end_0 = const()[name = tensor("xi_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_end_mask_0 = const()[name = tensor("xi_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_squeeze_mask_0 = const()[name = tensor("xi_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_cast_fp16 = slice_by_index(begin = xi_begin_0, end = xi_end_0, end_mask = xi_end_mask_0, squeeze_mask = xi_squeeze_mask_0, x = pairs_cast_fp16)[name = tensor("xi_cast_fp16")]; + tensor var_5683_cast_fp16 = mul(x = xr_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5683_cast_fp16")]; + tensor var_5684_cast_fp16 = mul(x = xi_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5684_cast_fp16")]; + tensor xor__cast_fp16 = sub(x = var_5683_cast_fp16, y = var_5684_cast_fp16)[name = tensor("xor__cast_fp16")]; + tensor var_5687_cast_fp16 = mul(x = xr_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5687_cast_fp16")]; + tensor var_5688_cast_fp16 = mul(x = xi_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5688_cast_fp16")]; + tensor xoi_cast_fp16 = add(x = var_5687_cast_fp16, y = var_5688_cast_fp16)[name = tensor("xoi_cast_fp16")]; + tensor out_axis_0 = const()[name = tensor("out_axis_0"), val = tensor(-1)]; + tensor out_cast_fp16 = stack(axis = out_axis_0, values = (xor__cast_fp16, xoi_cast_fp16))[name = tensor("out_cast_fp16")]; + tensor var_5705 = const()[name = tensor("op_5705"), val = tensor([256, 1024])]; + tensor var_5706_cast_fp16 = reshape(shape = var_5705, x = out_cast_fp16)[name = tensor("op_5706_cast_fp16")]; + tensor written_k_transpose_x_1 = const()[name = tensor("written_k_transpose_x_1"), val = tensor(true)]; + tensor written_k_transpose_y_1 = const()[name = tensor("written_k_transpose_y_1"), val = tensor(false)]; + tensor written_k_cast_fp16 = matmul(transpose_x = written_k_transpose_x_1, transpose_y = written_k_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_5706_cast_fp16)[name = tensor("written_k_cast_fp16")]; + tensor var_5713 = const()[name = tensor("op_5713"), val = tensor([256, 1024])]; + tensor var_5714_cast_fp16 = reshape(shape = var_5713, x = v_cast_fp16)[name = tensor("op_5714_cast_fp16")]; + tensor written_v_transpose_x_1 = const()[name = tensor("written_v_transpose_x_1"), val = tensor(true)]; + tensor written_v_transpose_y_1 = const()[name = tensor("written_v_transpose_y_1"), val = tensor(false)]; + tensor written_v_cast_fp16 = matmul(transpose_x = written_v_transpose_x_1, transpose_y = written_v_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_5714_cast_fp16)[name = tensor("written_v_cast_fp16")]; + tensor k_cache23_to_fp16_dtype_0 = const()[name = tensor("k_cache23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache23_to_fp16 = cast(dtype = k_cache23_to_fp16_dtype_0, x = k_cache23)[name = tensor("cast_327")]; + tensor var_5725_cast_fp16 = mul(x = k_cache23_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_5725_cast_fp16")]; + tensor var_5730 = const()[name = tensor("op_5730"), val = tensor([1, 512, 16, 64])]; + tensor var_5731_cast_fp16 = reshape(shape = var_5730, x = written_k_cast_fp16)[name = tensor("op_5731_cast_fp16")]; + tensor new_k_cache_cast_fp16 = add(x = var_5725_cast_fp16, y = var_5731_cast_fp16)[name = tensor("new_k_cache_cast_fp16")]; + tensor new_k_cache_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache23_to_fp16_dtype_0 = const()[name = tensor("v_cache23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache23_to_fp16 = cast(dtype = v_cache23_to_fp16_dtype_0, x = v_cache23)[name = tensor("cast_325")]; + tensor var_5734_cast_fp16 = mul(x = v_cache23_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_5734_cast_fp16")]; + tensor var_5739 = const()[name = tensor("op_5739"), val = tensor([1, 512, 16, 64])]; + tensor var_5740_cast_fp16 = reshape(shape = var_5739, x = written_v_cast_fp16)[name = tensor("op_5740_cast_fp16")]; + tensor new_v_cache_cast_fp16 = add(x = var_5734_cast_fp16, y = var_5740_cast_fp16)[name = tensor("new_v_cache_cast_fp16")]; + tensor new_v_cache_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor position23_to_fp16_dtype_0 = const()[name = tensor("position23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position23_to_fp16 = cast(dtype = position23_to_fp16_dtype_0, x = position23)[name = tensor("cast_323")]; + tensor var_5747_cast_fp16 = add(x = position23_to_fp16, y = valid_len_to_fp16)[name = tensor("op_5747_cast_fp16")]; + tensor var_5747_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5747_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor new_position23 = cast(dtype = var_5747_cast_fp16_to_fp32_dtype_0, x = var_5747_cast_fp16)[name = tensor("cast_322")]; + tensor new_v_cache23 = cast(dtype = new_v_cache_cast_fp16_to_fp32_dtype_0, x = new_v_cache_cast_fp16)[name = tensor("cast_324")]; + tensor new_k_cache23 = cast(dtype = new_k_cache_cast_fp16_to_fp32_dtype_0, x = new_k_cache_cast_fp16)[name = tensor("cast_326")]; + tensor new_position22 = cast(dtype = var_5608_cast_fp16_to_fp32_dtype_0, x = var_5608_cast_fp16)[name = tensor("cast_328")]; + tensor new_v_cache22 = cast(dtype = new_v_cache_45_cast_fp16_to_fp32_dtype_0, x = new_v_cache_45_cast_fp16)[name = tensor("cast_330")]; + tensor new_k_cache22 = cast(dtype = new_k_cache_45_cast_fp16_to_fp32_dtype_0, x = new_k_cache_45_cast_fp16)[name = tensor("cast_332")]; + tensor new_position21 = cast(dtype = var_5377_cast_fp16_to_fp32_dtype_0, x = var_5377_cast_fp16)[name = tensor("cast_334")]; + tensor new_v_cache21 = cast(dtype = new_v_cache_43_cast_fp16_to_fp32_dtype_0, x = new_v_cache_43_cast_fp16)[name = tensor("cast_336")]; + tensor new_k_cache21 = cast(dtype = new_k_cache_43_cast_fp16_to_fp32_dtype_0, x = new_k_cache_43_cast_fp16)[name = tensor("cast_338")]; + tensor new_position20 = cast(dtype = var_5146_cast_fp16_to_fp32_dtype_0, x = var_5146_cast_fp16)[name = tensor("cast_340")]; + tensor new_v_cache20 = cast(dtype = new_v_cache_41_cast_fp16_to_fp32_dtype_0, x = new_v_cache_41_cast_fp16)[name = tensor("cast_342")]; + tensor new_k_cache20 = cast(dtype = new_k_cache_41_cast_fp16_to_fp32_dtype_0, x = new_k_cache_41_cast_fp16)[name = tensor("cast_344")]; + tensor new_position19 = cast(dtype = var_4915_cast_fp16_to_fp32_dtype_0, x = var_4915_cast_fp16)[name = tensor("cast_346")]; + tensor new_v_cache19 = cast(dtype = new_v_cache_39_cast_fp16_to_fp32_dtype_0, x = new_v_cache_39_cast_fp16)[name = tensor("cast_348")]; + tensor new_k_cache19 = cast(dtype = new_k_cache_39_cast_fp16_to_fp32_dtype_0, x = new_k_cache_39_cast_fp16)[name = tensor("cast_350")]; + tensor new_position18 = cast(dtype = var_4684_cast_fp16_to_fp32_dtype_0, x = var_4684_cast_fp16)[name = tensor("cast_352")]; + tensor new_v_cache18 = cast(dtype = new_v_cache_37_cast_fp16_to_fp32_dtype_0, x = new_v_cache_37_cast_fp16)[name = tensor("cast_354")]; + tensor new_k_cache18 = cast(dtype = new_k_cache_37_cast_fp16_to_fp32_dtype_0, x = new_k_cache_37_cast_fp16)[name = tensor("cast_356")]; + tensor new_position17 = cast(dtype = var_4453_cast_fp16_to_fp32_dtype_0, x = var_4453_cast_fp16)[name = tensor("cast_358")]; + tensor new_v_cache17 = cast(dtype = new_v_cache_35_cast_fp16_to_fp32_dtype_0, x = new_v_cache_35_cast_fp16)[name = tensor("cast_360")]; + tensor new_k_cache17 = cast(dtype = new_k_cache_35_cast_fp16_to_fp32_dtype_0, x = new_k_cache_35_cast_fp16)[name = tensor("cast_362")]; + tensor new_position16 = cast(dtype = var_4222_cast_fp16_to_fp32_dtype_0, x = var_4222_cast_fp16)[name = tensor("cast_364")]; + tensor new_v_cache16 = cast(dtype = new_v_cache_33_cast_fp16_to_fp32_dtype_0, x = new_v_cache_33_cast_fp16)[name = tensor("cast_366")]; + tensor new_k_cache16 = cast(dtype = new_k_cache_33_cast_fp16_to_fp32_dtype_0, x = new_k_cache_33_cast_fp16)[name = tensor("cast_368")]; + tensor new_position15 = cast(dtype = var_3991_cast_fp16_to_fp32_dtype_0, x = var_3991_cast_fp16)[name = tensor("cast_370")]; + tensor new_v_cache15 = cast(dtype = new_v_cache_31_cast_fp16_to_fp32_dtype_0, x = new_v_cache_31_cast_fp16)[name = tensor("cast_372")]; + tensor new_k_cache15 = cast(dtype = new_k_cache_31_cast_fp16_to_fp32_dtype_0, x = new_k_cache_31_cast_fp16)[name = tensor("cast_374")]; + tensor new_position14 = cast(dtype = var_3760_cast_fp16_to_fp32_dtype_0, x = var_3760_cast_fp16)[name = tensor("cast_376")]; + tensor new_v_cache14 = cast(dtype = new_v_cache_29_cast_fp16_to_fp32_dtype_0, x = new_v_cache_29_cast_fp16)[name = tensor("cast_378")]; + tensor new_k_cache14 = cast(dtype = new_k_cache_29_cast_fp16_to_fp32_dtype_0, x = new_k_cache_29_cast_fp16)[name = tensor("cast_380")]; + tensor new_position13 = cast(dtype = var_3529_cast_fp16_to_fp32_dtype_0, x = var_3529_cast_fp16)[name = tensor("cast_382")]; + tensor new_v_cache13 = cast(dtype = new_v_cache_27_cast_fp16_to_fp32_dtype_0, x = new_v_cache_27_cast_fp16)[name = tensor("cast_384")]; + tensor new_k_cache13 = cast(dtype = new_k_cache_27_cast_fp16_to_fp32_dtype_0, x = new_k_cache_27_cast_fp16)[name = tensor("cast_386")]; + tensor new_position12 = cast(dtype = var_3298_cast_fp16_to_fp32_dtype_0, x = var_3298_cast_fp16)[name = tensor("cast_388")]; + tensor new_v_cache12 = cast(dtype = new_v_cache_25_cast_fp16_to_fp32_dtype_0, x = new_v_cache_25_cast_fp16)[name = tensor("cast_390")]; + tensor new_k_cache12 = cast(dtype = new_k_cache_25_cast_fp16_to_fp32_dtype_0, x = new_k_cache_25_cast_fp16)[name = tensor("cast_392")]; + tensor new_position11 = cast(dtype = var_3067_cast_fp16_to_fp32_dtype_0, x = var_3067_cast_fp16)[name = tensor("cast_394")]; + tensor new_v_cache11 = cast(dtype = new_v_cache_23_cast_fp16_to_fp32_dtype_0, x = new_v_cache_23_cast_fp16)[name = tensor("cast_396")]; + tensor new_k_cache11 = cast(dtype = new_k_cache_23_cast_fp16_to_fp32_dtype_0, x = new_k_cache_23_cast_fp16)[name = tensor("cast_398")]; + tensor new_position10 = cast(dtype = var_2836_cast_fp16_to_fp32_dtype_0, x = var_2836_cast_fp16)[name = tensor("cast_400")]; + tensor new_v_cache10 = cast(dtype = new_v_cache_21_cast_fp16_to_fp32_dtype_0, x = new_v_cache_21_cast_fp16)[name = tensor("cast_402")]; + tensor new_k_cache10 = cast(dtype = new_k_cache_21_cast_fp16_to_fp32_dtype_0, x = new_k_cache_21_cast_fp16)[name = tensor("cast_404")]; + tensor new_position9 = cast(dtype = var_2605_cast_fp16_to_fp32_dtype_0, x = var_2605_cast_fp16)[name = tensor("cast_406")]; + tensor new_v_cache9 = cast(dtype = new_v_cache_19_cast_fp16_to_fp32_dtype_0, x = new_v_cache_19_cast_fp16)[name = tensor("cast_408")]; + tensor new_k_cache9 = cast(dtype = new_k_cache_19_cast_fp16_to_fp32_dtype_0, x = new_k_cache_19_cast_fp16)[name = tensor("cast_410")]; + tensor new_position8 = cast(dtype = var_2374_cast_fp16_to_fp32_dtype_0, x = var_2374_cast_fp16)[name = tensor("cast_412")]; + tensor new_v_cache8 = cast(dtype = new_v_cache_17_cast_fp16_to_fp32_dtype_0, x = new_v_cache_17_cast_fp16)[name = tensor("cast_414")]; + tensor new_k_cache8 = cast(dtype = new_k_cache_17_cast_fp16_to_fp32_dtype_0, x = new_k_cache_17_cast_fp16)[name = tensor("cast_416")]; + tensor new_position7 = cast(dtype = var_2143_cast_fp16_to_fp32_dtype_0, x = var_2143_cast_fp16)[name = tensor("cast_418")]; + tensor new_v_cache7 = cast(dtype = new_v_cache_15_cast_fp16_to_fp32_dtype_0, x = new_v_cache_15_cast_fp16)[name = tensor("cast_420")]; + tensor new_k_cache7 = cast(dtype = new_k_cache_15_cast_fp16_to_fp32_dtype_0, x = new_k_cache_15_cast_fp16)[name = tensor("cast_422")]; + tensor new_position6 = cast(dtype = var_1912_cast_fp16_to_fp32_dtype_0, x = var_1912_cast_fp16)[name = tensor("cast_424")]; + tensor new_v_cache6 = cast(dtype = new_v_cache_13_cast_fp16_to_fp32_dtype_0, x = new_v_cache_13_cast_fp16)[name = tensor("cast_426")]; + tensor new_k_cache6 = cast(dtype = new_k_cache_13_cast_fp16_to_fp32_dtype_0, x = new_k_cache_13_cast_fp16)[name = tensor("cast_428")]; + tensor new_position5 = cast(dtype = var_1681_cast_fp16_to_fp32_dtype_0, x = var_1681_cast_fp16)[name = tensor("cast_430")]; + tensor new_v_cache5 = cast(dtype = new_v_cache_11_cast_fp16_to_fp32_dtype_0, x = new_v_cache_11_cast_fp16)[name = tensor("cast_432")]; + tensor new_k_cache5 = cast(dtype = new_k_cache_11_cast_fp16_to_fp32_dtype_0, x = new_k_cache_11_cast_fp16)[name = tensor("cast_434")]; + tensor new_position4 = cast(dtype = var_1450_cast_fp16_to_fp32_dtype_0, x = var_1450_cast_fp16)[name = tensor("cast_436")]; + tensor new_v_cache4 = cast(dtype = new_v_cache_9_cast_fp16_to_fp32_dtype_0, x = new_v_cache_9_cast_fp16)[name = tensor("cast_438")]; + tensor new_k_cache4 = cast(dtype = new_k_cache_9_cast_fp16_to_fp32_dtype_0, x = new_k_cache_9_cast_fp16)[name = tensor("cast_440")]; + tensor new_position3 = cast(dtype = var_1219_cast_fp16_to_fp32_dtype_0, x = var_1219_cast_fp16)[name = tensor("cast_442")]; + tensor new_v_cache3 = cast(dtype = new_v_cache_7_cast_fp16_to_fp32_dtype_0, x = new_v_cache_7_cast_fp16)[name = tensor("cast_444")]; + tensor new_k_cache3 = cast(dtype = new_k_cache_7_cast_fp16_to_fp32_dtype_0, x = new_k_cache_7_cast_fp16)[name = tensor("cast_446")]; + tensor new_position2 = cast(dtype = var_988_cast_fp16_to_fp32_dtype_0, x = var_988_cast_fp16)[name = tensor("cast_448")]; + tensor new_v_cache2 = cast(dtype = new_v_cache_5_cast_fp16_to_fp32_dtype_0, x = new_v_cache_5_cast_fp16)[name = tensor("cast_450")]; + tensor new_k_cache2 = cast(dtype = new_k_cache_5_cast_fp16_to_fp32_dtype_0, x = new_k_cache_5_cast_fp16)[name = tensor("cast_452")]; + tensor new_position1 = cast(dtype = var_757_cast_fp16_to_fp32_dtype_0, x = var_757_cast_fp16)[name = tensor("cast_454")]; + tensor new_v_cache1 = cast(dtype = new_v_cache_3_cast_fp16_to_fp32_dtype_0, x = new_v_cache_3_cast_fp16)[name = tensor("cast_456")]; + tensor new_k_cache1 = cast(dtype = new_k_cache_3_cast_fp16_to_fp32_dtype_0, x = new_k_cache_3_cast_fp16)[name = tensor("cast_458")]; + tensor new_position0 = cast(dtype = var_526_cast_fp16_to_fp32_dtype_0, x = var_297_cast_fp16)[name = tensor("cast_460")]; + tensor new_v_cache0 = cast(dtype = new_v_cache_1_cast_fp16_to_fp32_dtype_0, x = new_v_cache_1_cast_fp16)[name = tensor("cast_461")]; + tensor new_k_cache0 = cast(dtype = new_k_cache_1_cast_fp16_to_fp32_dtype_0, x = new_k_cache_1_cast_fp16)[name = tensor("cast_463")]; + } -> (new_k_cache0, new_v_cache0, new_position0, new_k_cache1, new_v_cache1, new_position1, new_k_cache2, new_v_cache2, new_position2, new_k_cache3, new_v_cache3, new_position3, new_k_cache4, new_v_cache4, new_position4, new_k_cache5, new_v_cache5, new_position5, new_k_cache6, new_v_cache6, new_position6, new_k_cache7, new_v_cache7, new_position7, new_k_cache8, new_v_cache8, new_position8, new_k_cache9, new_v_cache9, new_position9, new_k_cache10, new_v_cache10, new_position10, new_k_cache11, new_v_cache11, new_position11, new_k_cache12, new_v_cache12, new_position12, new_k_cache13, new_v_cache13, new_position13, new_k_cache14, new_v_cache14, new_position14, new_k_cache15, new_v_cache15, new_position15, new_k_cache16, new_v_cache16, new_position16, new_k_cache17, new_v_cache17, new_position17, new_k_cache18, new_v_cache18, new_position18, new_k_cache19, new_v_cache19, new_position19, new_k_cache20, new_v_cache20, new_position20, new_k_cache21, new_v_cache21, new_position21, new_k_cache22, new_v_cache22, new_position22, new_k_cache23, new_v_cache23, new_position23); +} \ No newline at end of file diff --git a/v2.1/portuguese_24l/cond_prefill_ane.mlmodelc/weights/weight.bin b/v2.1/portuguese_24l/cond_prefill_ane.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..128b17d06d025c6fb8fbdb215d12b1289713f28b --- /dev/null +++ b/v2.1/portuguese_24l/cond_prefill_ane.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cda2cf8b07678449891b4038fa88e5a2ecd3f5f7dbb63adbed0e90e2c1d3a6ce +size 585328896 diff --git a/v2.1/portuguese_24l/cond_prefill_ane.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/portuguese_24l/cond_prefill_ane.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..17dbb7154c7a2b4fb5191de4b2049a8ca337a79f --- /dev/null +++ b/v2.1/portuguese_24l/cond_prefill_ane.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c093591e33948494088ba9c34174843af56ef2b9f3d87f263976e6c0bb59763d +size 483717 diff --git a/v2.1/portuguese_24l/cond_prefill_ane.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/portuguese_24l/cond_prefill_ane.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..128b17d06d025c6fb8fbdb215d12b1289713f28b --- /dev/null +++ b/v2.1/portuguese_24l/cond_prefill_ane.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cda2cf8b07678449891b4038fa88e5a2ecd3f5f7dbb63adbed0e90e2c1d3a6ce +size 585328896 diff --git a/v2.1/portuguese_24l/cond_prefill_ane.mlpackage/Manifest.json b/v2.1/portuguese_24l/cond_prefill_ane.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..9851c7b353a03b5cbff606cb2b8151fcb7eb63b8 --- /dev/null +++ b/v2.1/portuguese_24l/cond_prefill_ane.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "7EBE7357-107B-440D-BD51-42B1C98812FB": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "A9CCCF0B-81CA-4DBC-A3E9-11963D2FC643": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "7EBE7357-107B-440D-BD51-42B1C98812FB" +} diff --git a/v2.1/portuguese_24l/flowlm_step_ane.mlmodelc/analytics/coremldata.bin b/v2.1/portuguese_24l/flowlm_step_ane.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..584c2be7e0abef8809e47679708da2bf58dc27fe --- /dev/null +++ b/v2.1/portuguese_24l/flowlm_step_ane.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8cc82afc3f4b3e565c33313755e5b146f0b5b0d62f224ef0491dcc992edd27b2 +size 243 diff --git a/v2.1/portuguese_24l/flowlm_step_ane.mlmodelc/coremldata.bin b/v2.1/portuguese_24l/flowlm_step_ane.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..e9dde1d6832648c390c766a3cbb2252463829f53 --- /dev/null +++ b/v2.1/portuguese_24l/flowlm_step_ane.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:86c44575e42ca18b300d312cd46338b94f806688386428c869b34e2d28a76d96 +size 4522 diff --git a/v2.1/portuguese_24l/flowlm_step_ane.mlmodelc/model.mil b/v2.1/portuguese_24l/flowlm_step_ane.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..81c0fcd7cbde107f32ed2697e8759e67c7922399 --- /dev/null +++ b/v2.1/portuguese_24l/flowlm_step_ane.mlmodelc/model.mil @@ -0,0 +1,3699 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.9.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor k_cache0, tensor k_cache1, tensor k_cache10, tensor k_cache11, tensor k_cache12, tensor k_cache13, tensor k_cache14, tensor k_cache15, tensor k_cache16, tensor k_cache17, tensor k_cache18, tensor k_cache19, tensor k_cache2, tensor k_cache20, tensor k_cache21, tensor k_cache22, tensor k_cache23, tensor k_cache3, tensor k_cache4, tensor k_cache5, tensor k_cache6, tensor k_cache7, tensor k_cache8, tensor k_cache9, tensor position0, tensor position1, tensor position10, tensor position11, tensor position12, tensor position13, tensor position14, tensor position15, tensor position16, tensor position17, tensor position18, tensor position19, tensor position2, tensor position20, tensor position21, tensor position22, tensor position23, tensor position3, tensor position4, tensor position5, tensor position6, tensor position7, tensor position8, tensor position9, tensor sequence, tensor v_cache0, tensor v_cache1, tensor v_cache10, tensor v_cache11, tensor v_cache12, tensor v_cache13, tensor v_cache14, tensor v_cache15, tensor v_cache16, tensor v_cache17, tensor v_cache18, tensor v_cache19, tensor v_cache2, tensor v_cache20, tensor v_cache21, tensor v_cache22, tensor v_cache23, tensor v_cache3, tensor v_cache4, tensor v_cache5, tensor v_cache6, tensor v_cache7, tensor v_cache8, tensor v_cache9) { + tensor sequence_to_fp16_dtype_0 = const()[name = tensor("sequence_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor input_linear_weight_to_fp16 = const()[name = tensor("input_linear_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(65664)))]; + tensor sequence_to_fp16 = cast(dtype = sequence_to_fp16_dtype_0, x = sequence)[name = tensor("cast_578")]; + tensor linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = input_linear_weight_to_fp16, x = sequence_to_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; + tensor norm0_1_weight_to_fp16 = const()[name = tensor("norm0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67776)))]; + tensor norm0_1_bias_to_fp16 = const()[name = tensor("norm0_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(69888)))]; + tensor var_227_to_fp16 = const()[name = tensor("op_227_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_5_cast_fp16 = layer_norm(axes = input_5_axes_0, beta = norm0_1_bias_to_fp16, epsilon = var_227_to_fp16, gamma = norm0_1_weight_to_fp16, x = linear_0_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor attn0_in_proj_weight_to_fp16 = const()[name = tensor("attn0_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(72000)))]; + tensor linear_1_bias_0_to_fp16 = const()[name = tensor("linear_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6363520)))]; + tensor linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn0_in_proj_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("linear_1_cast_fp16")]; + tensor var_240 = const()[name = tensor("op_240"), val = tensor([1, 3, 16, 64])]; + tensor qkv_1_cast_fp16 = reshape(shape = var_240, x = linear_1_cast_fp16)[name = tensor("qkv_1_cast_fp16")]; + tensor x_1_begin_0 = const()[name = tensor("x_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_1_end_0 = const()[name = tensor("x_1_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_1_end_mask_0 = const()[name = tensor("x_1_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_1_squeeze_mask_0 = const()[name = tensor("x_1_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_1_cast_fp16 = slice_by_index(begin = x_1_begin_0, end = x_1_end_0, end_mask = x_1_end_mask_0, squeeze_mask = x_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("x_1_cast_fp16")]; + tensor x_3_begin_0 = const()[name = tensor("x_3_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_3_end_0 = const()[name = tensor("x_3_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_3_end_mask_0 = const()[name = tensor("x_3_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_3_squeeze_mask_0 = const()[name = tensor("x_3_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_3_cast_fp16 = slice_by_index(begin = x_3_begin_0, end = x_3_end_0, end_mask = x_3_end_mask_0, squeeze_mask = x_3_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("x_3_cast_fp16")]; + tensor v_1_begin_0 = const()[name = tensor("v_1_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_1_end_0 = const()[name = tensor("v_1_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_1_end_mask_0 = const()[name = tensor("v_1_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_1_squeeze_mask_0 = const()[name = tensor("v_1_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_1_cast_fp16 = slice_by_index(begin = v_1_begin_0, end = v_1_end_0, end_mask = v_1_end_mask_0, squeeze_mask = v_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("v_1_cast_fp16")]; + tensor var_269 = const()[name = tensor("op_269"), val = tensor([1, 1, 1])]; + tensor position0_to_fp16_dtype_0 = const()[name = tensor("position0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position0_to_fp16 = cast(dtype = position0_to_fp16_dtype_0, x = position0)[name = tensor("cast_577")]; + tensor var_270_cast_fp16 = reshape(shape = var_269, x = position0_to_fp16)[name = tensor("op_270_cast_fp16")]; + tensor rope_freqs_to_fp16 = const()[name = tensor("rope_freqs_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6369728)))]; + tensor angles_1_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_270_cast_fp16)[name = tensor("angles_1_cast_fp16")]; + tensor rotr_1_cast_fp16 = cos(x = angles_1_cast_fp16)[name = tensor("rotr_1_cast_fp16")]; + tensor roti_1_cast_fp16 = sin(x = angles_1_cast_fp16)[name = tensor("roti_1_cast_fp16")]; + tensor var_278 = const()[name = tensor("op_278"), val = tensor([1, 16, 32, 2])]; + tensor pairs_1_cast_fp16 = reshape(shape = var_278, x = x_1_cast_fp16)[name = tensor("pairs_1_cast_fp16")]; + tensor xr_1_begin_0 = const()[name = tensor("xr_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_1_end_0 = const()[name = tensor("xr_1_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_1_end_mask_0 = const()[name = tensor("xr_1_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_1_squeeze_mask_0 = const()[name = tensor("xr_1_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_1_cast_fp16 = slice_by_index(begin = xr_1_begin_0, end = xr_1_end_0, end_mask = xr_1_end_mask_0, squeeze_mask = xr_1_squeeze_mask_0, x = pairs_1_cast_fp16)[name = tensor("xr_1_cast_fp16")]; + tensor xi_1_begin_0 = const()[name = tensor("xi_1_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_1_end_0 = const()[name = tensor("xi_1_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_1_end_mask_0 = const()[name = tensor("xi_1_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_1_squeeze_mask_0 = const()[name = tensor("xi_1_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_1_cast_fp16 = slice_by_index(begin = xi_1_begin_0, end = xi_1_end_0, end_mask = xi_1_end_mask_0, squeeze_mask = xi_1_squeeze_mask_0, x = pairs_1_cast_fp16)[name = tensor("xi_1_cast_fp16")]; + tensor var_286_cast_fp16 = mul(x = xr_1_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_286_cast_fp16")]; + tensor var_287_cast_fp16 = mul(x = xi_1_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_287_cast_fp16")]; + tensor xor__1_cast_fp16 = sub(x = var_286_cast_fp16, y = var_287_cast_fp16)[name = tensor("xor__1_cast_fp16")]; + tensor var_290_cast_fp16 = mul(x = xr_1_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_290_cast_fp16")]; + tensor var_291_cast_fp16 = mul(x = xi_1_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_291_cast_fp16")]; + tensor xoi_1_cast_fp16 = add(x = var_290_cast_fp16, y = var_291_cast_fp16)[name = tensor("xoi_1_cast_fp16")]; + tensor out_1_axis_0 = const()[name = tensor("out_1_axis_0"), val = tensor(-1)]; + tensor out_1_cast_fp16 = stack(axis = out_1_axis_0, values = (xor__1_cast_fp16, xoi_1_cast_fp16))[name = tensor("out_1_cast_fp16")]; + tensor var_314 = const()[name = tensor("op_314"), val = tensor([1, 16, 32, 2])]; + tensor pairs_3_cast_fp16 = reshape(shape = var_314, x = x_3_cast_fp16)[name = tensor("pairs_3_cast_fp16")]; + tensor xr_3_begin_0 = const()[name = tensor("xr_3_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_3_end_0 = const()[name = tensor("xr_3_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_3_end_mask_0 = const()[name = tensor("xr_3_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_3_squeeze_mask_0 = const()[name = tensor("xr_3_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_3_cast_fp16 = slice_by_index(begin = xr_3_begin_0, end = xr_3_end_0, end_mask = xr_3_end_mask_0, squeeze_mask = xr_3_squeeze_mask_0, x = pairs_3_cast_fp16)[name = tensor("xr_3_cast_fp16")]; + tensor xi_3_begin_0 = const()[name = tensor("xi_3_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_3_end_0 = const()[name = tensor("xi_3_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_3_end_mask_0 = const()[name = tensor("xi_3_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_3_squeeze_mask_0 = const()[name = tensor("xi_3_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_3_cast_fp16 = slice_by_index(begin = xi_3_begin_0, end = xi_3_end_0, end_mask = xi_3_end_mask_0, squeeze_mask = xi_3_squeeze_mask_0, x = pairs_3_cast_fp16)[name = tensor("xi_3_cast_fp16")]; + tensor var_322_cast_fp16 = mul(x = xr_3_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_322_cast_fp16")]; + tensor var_323_cast_fp16 = mul(x = xi_3_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_323_cast_fp16")]; + tensor xor__3_cast_fp16 = sub(x = var_322_cast_fp16, y = var_323_cast_fp16)[name = tensor("xor__3_cast_fp16")]; + tensor var_326_cast_fp16 = mul(x = xr_3_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_326_cast_fp16")]; + tensor var_327_cast_fp16 = mul(x = xi_3_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_327_cast_fp16")]; + tensor xoi_3_cast_fp16 = add(x = var_326_cast_fp16, y = var_327_cast_fp16)[name = tensor("xoi_3_cast_fp16")]; + tensor out_3_axis_0 = const()[name = tensor("out_3_axis_0"), val = tensor(-1)]; + tensor out_3_cast_fp16 = stack(axis = out_3_axis_0, values = (xor__3_cast_fp16, xoi_3_cast_fp16))[name = tensor("out_3_cast_fp16")]; + tensor _inversed_339_y_0_to_fp16 = const()[name = tensor("_inversed_339_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_339_cast_fp16 = mul(x = position0_to_fp16, y = _inversed_339_y_0_to_fp16)[name = tensor("_inversed_339_cast_fp16")]; + tensor var_340_cast_fp16 = floor(x = _inversed_339_cast_fp16)[name = tensor("op_340_cast_fp16")]; + tensor var_341_to_fp16 = const()[name = tensor("op_341_to_fp16"), val = tensor(0x1p+9)]; + tensor var_342_cast_fp16 = mul(x = var_340_cast_fp16, y = var_341_to_fp16)[name = tensor("op_342_cast_fp16")]; + tensor write_idx_1_cast_fp16 = sub(x = position0_to_fp16, y = var_342_cast_fp16)[name = tensor("write_idx_1_cast_fp16")]; + tensor slot_indices_to_fp16 = const()[name = tensor("slot_indices_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6369856)))]; + tensor var_345_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_1_cast_fp16)[name = tensor("op_345_cast_fp16")]; + tensor var_355 = const()[name = tensor("op_355"), val = tensor([1, 512, 1, 1])]; + tensor onehot_1_to_fp16_dtype_0 = const()[name = tensor("onehot_1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_345_cast_fp16_to_fp16 = cast(dtype = onehot_1_to_fp16_dtype_0, x = var_345_cast_fp16)[name = tensor("cast_576")]; + tensor onehot_3_cast_fp16 = reshape(shape = var_355, x = var_345_cast_fp16_to_fp16)[name = tensor("onehot_3_cast_fp16")]; + tensor var_357_to_fp16 = const()[name = tensor("op_357_to_fp16"), val = tensor(0x1p+0)]; + tensor var_359_cast_fp16 = sub(x = var_357_to_fp16, y = onehot_3_cast_fp16)[name = tensor("op_359_cast_fp16")]; + tensor k_cache0_to_fp16_dtype_0 = const()[name = tensor("k_cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache0_to_fp16 = cast(dtype = k_cache0_to_fp16_dtype_0, x = k_cache0)[name = tensor("cast_575")]; + tensor var_360_cast_fp16 = mul(x = k_cache0_to_fp16, y = var_359_cast_fp16)[name = tensor("op_360_cast_fp16")]; + tensor var_365 = const()[name = tensor("op_365"), val = tensor([1, 1, 16, 64])]; + tensor var_366_cast_fp16 = reshape(shape = var_365, x = out_3_cast_fp16)[name = tensor("op_366_cast_fp16")]; + tensor var_367_cast_fp16 = mul(x = var_366_cast_fp16, y = onehot_3_cast_fp16)[name = tensor("op_367_cast_fp16")]; + tensor new_k_cache_1_cast_fp16 = add(x = var_360_cast_fp16, y = var_367_cast_fp16)[name = tensor("new_k_cache_1_cast_fp16")]; + tensor new_k_cache_1_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_1_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache0_to_fp16_dtype_0 = const()[name = tensor("v_cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache0_to_fp16 = cast(dtype = v_cache0_to_fp16_dtype_0, x = v_cache0)[name = tensor("cast_573")]; + tensor var_373_cast_fp16 = mul(x = v_cache0_to_fp16, y = var_359_cast_fp16)[name = tensor("op_373_cast_fp16")]; + tensor var_378 = const()[name = tensor("op_378"), val = tensor([1, 1, 16, 64])]; + tensor var_379_cast_fp16 = reshape(shape = var_378, x = v_1_cast_fp16)[name = tensor("op_379_cast_fp16")]; + tensor var_380_cast_fp16 = mul(x = var_379_cast_fp16, y = onehot_3_cast_fp16)[name = tensor("op_380_cast_fp16")]; + tensor new_v_cache_1_cast_fp16 = add(x = var_373_cast_fp16, y = var_380_cast_fp16)[name = tensor("new_v_cache_1_cast_fp16")]; + tensor new_v_cache_1_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_1_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_0_perm_0 = const()[name = tensor("transpose_0_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_393 = const()[name = tensor("op_393"), val = tensor([0, 2, 1, 3])]; + tensor var_399 = const()[name = tensor("op_399"), val = tensor([1, 16, 1, 64])]; + tensor q4_1_cast_fp16 = reshape(shape = var_399, x = out_1_cast_fp16)[name = tensor("q4_1_cast_fp16")]; + tensor var_404_transpose_x_0 = const()[name = tensor("op_404_transpose_x_0"), val = tensor(false)]; + tensor var_404_transpose_y_0 = const()[name = tensor("op_404_transpose_y_0"), val = tensor(false)]; + tensor transpose_0_cast_fp16 = transpose(perm = transpose_0_perm_0, x = new_k_cache_1_cast_fp16)[name = tensor("transpose_95")]; + tensor var_404_cast_fp16 = matmul(transpose_x = var_404_transpose_x_0, transpose_y = var_404_transpose_y_0, x = q4_1_cast_fp16, y = transpose_0_cast_fp16)[name = tensor("op_404_cast_fp16")]; + tensor var_405_to_fp16 = const()[name = tensor("op_405_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_1_cast_fp16 = mul(x = var_404_cast_fp16, y = var_405_to_fp16)[name = tensor("scores_1_cast_fp16")]; + tensor var_409 = const()[name = tensor("op_409"), val = tensor([1, 1])]; + tensor var_410_cast_fp16 = reshape(shape = var_409, x = position0_to_fp16)[name = tensor("op_410_cast_fp16")]; + tensor var_411_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_410_cast_fp16)[name = tensor("op_411_cast_fp16")]; + tensor mask_1_to_fp16_dtype_0 = const()[name = tensor("mask_1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_418_to_fp16 = const()[name = tensor("op_418_to_fp16"), val = tensor(0x1p+0)]; + tensor var_411_cast_fp16_to_fp16 = cast(dtype = mask_1_to_fp16_dtype_0, x = var_411_cast_fp16)[name = tensor("cast_571")]; + tensor var_419_cast_fp16 = sub(x = var_411_cast_fp16_to_fp16, y = var_418_to_fp16)[name = tensor("op_419_cast_fp16")]; + tensor var_424 = const()[name = tensor("op_424"), val = tensor([1, 1, 1, 512])]; + tensor var_425_cast_fp16 = reshape(shape = var_424, x = var_419_cast_fp16)[name = tensor("op_425_cast_fp16")]; + tensor var_426_to_fp16 = const()[name = tensor("op_426_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_427_cast_fp16 = mul(x = var_425_cast_fp16, y = var_426_to_fp16)[name = tensor("op_427_cast_fp16")]; + tensor scores_3_cast_fp16 = add(x = scores_1_cast_fp16, y = var_427_cast_fp16)[name = tensor("scores_3_cast_fp16")]; + tensor var_430 = const()[name = tensor("op_430"), val = tensor(-1)]; + tensor weights_1_cast_fp16 = softmax(axis = var_430, x = scores_3_cast_fp16)[name = tensor("weights_1_cast_fp16")]; + tensor attn_1_transpose_x_0 = const()[name = tensor("attn_1_transpose_x_0"), val = tensor(false)]; + tensor attn_1_transpose_y_0 = const()[name = tensor("attn_1_transpose_y_0"), val = tensor(false)]; + tensor values_1_cast_fp16 = transpose(perm = var_393, x = new_v_cache_1_cast_fp16)[name = tensor("transpose_94")]; + tensor attn_1_cast_fp16 = matmul(transpose_x = attn_1_transpose_x_0, transpose_y = attn_1_transpose_y_0, x = weights_1_cast_fp16, y = values_1_cast_fp16)[name = tensor("attn_1_cast_fp16")]; + tensor var_438 = const()[name = tensor("op_438"), val = tensor([0, 2, 1, 3])]; + tensor var_443 = const()[name = tensor("op_443"), val = tensor([1, 1, 1024])]; + tensor var_439_cast_fp16 = transpose(perm = var_438, x = attn_1_cast_fp16)[name = tensor("transpose_93")]; + tensor input_7_cast_fp16 = reshape(shape = var_443, x = var_439_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor attn0_out_proj_weight_to_fp16 = const()[name = tensor("attn0_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6370944)))]; + tensor linear_2_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn0_out_proj_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor var_449_to_fp16 = const()[name = tensor("op_449_to_fp16"), val = tensor(0x1p+0)]; + tensor var_450_cast_fp16 = add(x = position0_to_fp16, y = var_449_to_fp16)[name = tensor("op_450_cast_fp16")]; + tensor var_450_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_450_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_9_cast_fp16 = add(x = linear_0_cast_fp16, y = linear_2_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor input_11_axes_0 = const()[name = tensor("input_11_axes_0"), val = tensor([-1])]; + tensor norm0_2_weight_to_fp16 = const()[name = tensor("norm0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8468160)))]; + tensor norm0_2_bias_to_fp16 = const()[name = tensor("norm0_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8470272)))]; + tensor var_454_to_fp16 = const()[name = tensor("op_454_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_11_cast_fp16 = layer_norm(axes = input_11_axes_0, beta = norm0_2_bias_to_fp16, epsilon = var_454_to_fp16, gamma = norm0_2_weight_to_fp16, x = input_9_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor linear0_1_weight_to_fp16 = const()[name = tensor("linear0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8472384)))]; + tensor linear_3_bias_0_to_fp16 = const()[name = tensor("linear_3_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16861056)))]; + tensor linear_3_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear0_1_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("linear_3_cast_fp16")]; + tensor input_13_mode_0 = const()[name = tensor("input_13_mode_0"), val = tensor("EXACT")]; + tensor input_13_cast_fp16 = gelu(mode = input_13_mode_0, x = linear_3_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor linear0_2_weight_to_fp16 = const()[name = tensor("linear0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16869312)))]; + tensor linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear0_2_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor input_15_cast_fp16 = add(x = input_9_cast_fp16, y = linear_4_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor input_17_axes_0 = const()[name = tensor("input_17_axes_0"), val = tensor([-1])]; + tensor norm1_1_weight_to_fp16 = const()[name = tensor("norm1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25257984)))]; + tensor norm1_1_bias_to_fp16 = const()[name = tensor("norm1_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25260096)))]; + tensor var_471_to_fp16 = const()[name = tensor("op_471_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_17_cast_fp16 = layer_norm(axes = input_17_axes_0, beta = norm1_1_bias_to_fp16, epsilon = var_471_to_fp16, gamma = norm1_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor attn1_in_proj_weight_to_fp16 = const()[name = tensor("attn1_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25262208)))]; + tensor linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn1_in_proj_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor var_484 = const()[name = tensor("op_484"), val = tensor([1, 3, 16, 64])]; + tensor qkv_3_cast_fp16 = reshape(shape = var_484, x = linear_5_cast_fp16)[name = tensor("qkv_3_cast_fp16")]; + tensor x_5_begin_0 = const()[name = tensor("x_5_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_5_end_0 = const()[name = tensor("x_5_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_5_end_mask_0 = const()[name = tensor("x_5_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_5_squeeze_mask_0 = const()[name = tensor("x_5_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_5_cast_fp16 = slice_by_index(begin = x_5_begin_0, end = x_5_end_0, end_mask = x_5_end_mask_0, squeeze_mask = x_5_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("x_5_cast_fp16")]; + tensor x_7_begin_0 = const()[name = tensor("x_7_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_7_end_0 = const()[name = tensor("x_7_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_7_end_mask_0 = const()[name = tensor("x_7_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_7_squeeze_mask_0 = const()[name = tensor("x_7_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_7_cast_fp16 = slice_by_index(begin = x_7_begin_0, end = x_7_end_0, end_mask = x_7_end_mask_0, squeeze_mask = x_7_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("x_7_cast_fp16")]; + tensor v_3_begin_0 = const()[name = tensor("v_3_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_3_end_0 = const()[name = tensor("v_3_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_3_end_mask_0 = const()[name = tensor("v_3_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_3_squeeze_mask_0 = const()[name = tensor("v_3_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_3_cast_fp16 = slice_by_index(begin = v_3_begin_0, end = v_3_end_0, end_mask = v_3_end_mask_0, squeeze_mask = v_3_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("v_3_cast_fp16")]; + tensor var_513 = const()[name = tensor("op_513"), val = tensor([1, 1, 1])]; + tensor position1_to_fp16_dtype_0 = const()[name = tensor("position1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position1_to_fp16 = cast(dtype = position1_to_fp16_dtype_0, x = position1)[name = tensor("cast_569")]; + tensor var_514_cast_fp16 = reshape(shape = var_513, x = position1_to_fp16)[name = tensor("op_514_cast_fp16")]; + tensor angles_5_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_514_cast_fp16)[name = tensor("angles_5_cast_fp16")]; + tensor rotr_5_cast_fp16 = cos(x = angles_5_cast_fp16)[name = tensor("rotr_5_cast_fp16")]; + tensor roti_5_cast_fp16 = sin(x = angles_5_cast_fp16)[name = tensor("roti_5_cast_fp16")]; + tensor var_522 = const()[name = tensor("op_522"), val = tensor([1, 16, 32, 2])]; + tensor pairs_5_cast_fp16 = reshape(shape = var_522, x = x_5_cast_fp16)[name = tensor("pairs_5_cast_fp16")]; + tensor xr_5_begin_0 = const()[name = tensor("xr_5_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_5_end_0 = const()[name = tensor("xr_5_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_5_end_mask_0 = const()[name = tensor("xr_5_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_5_squeeze_mask_0 = const()[name = tensor("xr_5_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_5_cast_fp16 = slice_by_index(begin = xr_5_begin_0, end = xr_5_end_0, end_mask = xr_5_end_mask_0, squeeze_mask = xr_5_squeeze_mask_0, x = pairs_5_cast_fp16)[name = tensor("xr_5_cast_fp16")]; + tensor xi_5_begin_0 = const()[name = tensor("xi_5_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_5_end_0 = const()[name = tensor("xi_5_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_5_end_mask_0 = const()[name = tensor("xi_5_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_5_squeeze_mask_0 = const()[name = tensor("xi_5_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_5_cast_fp16 = slice_by_index(begin = xi_5_begin_0, end = xi_5_end_0, end_mask = xi_5_end_mask_0, squeeze_mask = xi_5_squeeze_mask_0, x = pairs_5_cast_fp16)[name = tensor("xi_5_cast_fp16")]; + tensor var_530_cast_fp16 = mul(x = xr_5_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_530_cast_fp16")]; + tensor var_531_cast_fp16 = mul(x = xi_5_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_531_cast_fp16")]; + tensor xor__5_cast_fp16 = sub(x = var_530_cast_fp16, y = var_531_cast_fp16)[name = tensor("xor__5_cast_fp16")]; + tensor var_534_cast_fp16 = mul(x = xr_5_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_534_cast_fp16")]; + tensor var_535_cast_fp16 = mul(x = xi_5_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_535_cast_fp16")]; + tensor xoi_5_cast_fp16 = add(x = var_534_cast_fp16, y = var_535_cast_fp16)[name = tensor("xoi_5_cast_fp16")]; + tensor out_5_axis_0 = const()[name = tensor("out_5_axis_0"), val = tensor(-1)]; + tensor out_5_cast_fp16 = stack(axis = out_5_axis_0, values = (xor__5_cast_fp16, xoi_5_cast_fp16))[name = tensor("out_5_cast_fp16")]; + tensor var_558 = const()[name = tensor("op_558"), val = tensor([1, 16, 32, 2])]; + tensor pairs_7_cast_fp16 = reshape(shape = var_558, x = x_7_cast_fp16)[name = tensor("pairs_7_cast_fp16")]; + tensor xr_7_begin_0 = const()[name = tensor("xr_7_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_7_end_0 = const()[name = tensor("xr_7_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_7_end_mask_0 = const()[name = tensor("xr_7_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_7_squeeze_mask_0 = const()[name = tensor("xr_7_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_7_cast_fp16 = slice_by_index(begin = xr_7_begin_0, end = xr_7_end_0, end_mask = xr_7_end_mask_0, squeeze_mask = xr_7_squeeze_mask_0, x = pairs_7_cast_fp16)[name = tensor("xr_7_cast_fp16")]; + tensor xi_7_begin_0 = const()[name = tensor("xi_7_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_7_end_0 = const()[name = tensor("xi_7_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_7_end_mask_0 = const()[name = tensor("xi_7_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_7_squeeze_mask_0 = const()[name = tensor("xi_7_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_7_cast_fp16 = slice_by_index(begin = xi_7_begin_0, end = xi_7_end_0, end_mask = xi_7_end_mask_0, squeeze_mask = xi_7_squeeze_mask_0, x = pairs_7_cast_fp16)[name = tensor("xi_7_cast_fp16")]; + tensor var_566_cast_fp16 = mul(x = xr_7_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_566_cast_fp16")]; + tensor var_567_cast_fp16 = mul(x = xi_7_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_567_cast_fp16")]; + tensor xor__7_cast_fp16 = sub(x = var_566_cast_fp16, y = var_567_cast_fp16)[name = tensor("xor__7_cast_fp16")]; + tensor var_570_cast_fp16 = mul(x = xr_7_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_570_cast_fp16")]; + tensor var_571_cast_fp16 = mul(x = xi_7_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_571_cast_fp16")]; + tensor xoi_7_cast_fp16 = add(x = var_570_cast_fp16, y = var_571_cast_fp16)[name = tensor("xoi_7_cast_fp16")]; + tensor out_7_axis_0 = const()[name = tensor("out_7_axis_0"), val = tensor(-1)]; + tensor out_7_cast_fp16 = stack(axis = out_7_axis_0, values = (xor__7_cast_fp16, xoi_7_cast_fp16))[name = tensor("out_7_cast_fp16")]; + tensor _inversed_583_y_0_to_fp16 = const()[name = tensor("_inversed_583_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_583_cast_fp16 = mul(x = position1_to_fp16, y = _inversed_583_y_0_to_fp16)[name = tensor("_inversed_583_cast_fp16")]; + tensor var_584_cast_fp16 = floor(x = _inversed_583_cast_fp16)[name = tensor("op_584_cast_fp16")]; + tensor var_585_to_fp16 = const()[name = tensor("op_585_to_fp16"), val = tensor(0x1p+9)]; + tensor var_586_cast_fp16 = mul(x = var_584_cast_fp16, y = var_585_to_fp16)[name = tensor("op_586_cast_fp16")]; + tensor write_idx_3_cast_fp16 = sub(x = position1_to_fp16, y = var_586_cast_fp16)[name = tensor("write_idx_3_cast_fp16")]; + tensor var_589_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_3_cast_fp16)[name = tensor("op_589_cast_fp16")]; + tensor var_599 = const()[name = tensor("op_599"), val = tensor([1, 512, 1, 1])]; + tensor onehot_5_to_fp16_dtype_0 = const()[name = tensor("onehot_5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_589_cast_fp16_to_fp16 = cast(dtype = onehot_5_to_fp16_dtype_0, x = var_589_cast_fp16)[name = tensor("cast_568")]; + tensor onehot_7_cast_fp16 = reshape(shape = var_599, x = var_589_cast_fp16_to_fp16)[name = tensor("onehot_7_cast_fp16")]; + tensor var_601_to_fp16 = const()[name = tensor("op_601_to_fp16"), val = tensor(0x1p+0)]; + tensor var_603_cast_fp16 = sub(x = var_601_to_fp16, y = onehot_7_cast_fp16)[name = tensor("op_603_cast_fp16")]; + tensor k_cache1_to_fp16_dtype_0 = const()[name = tensor("k_cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache1_to_fp16 = cast(dtype = k_cache1_to_fp16_dtype_0, x = k_cache1)[name = tensor("cast_567")]; + tensor var_604_cast_fp16 = mul(x = k_cache1_to_fp16, y = var_603_cast_fp16)[name = tensor("op_604_cast_fp16")]; + tensor var_609 = const()[name = tensor("op_609"), val = tensor([1, 1, 16, 64])]; + tensor var_610_cast_fp16 = reshape(shape = var_609, x = out_7_cast_fp16)[name = tensor("op_610_cast_fp16")]; + tensor var_611_cast_fp16 = mul(x = var_610_cast_fp16, y = onehot_7_cast_fp16)[name = tensor("op_611_cast_fp16")]; + tensor new_k_cache_3_cast_fp16 = add(x = var_604_cast_fp16, y = var_611_cast_fp16)[name = tensor("new_k_cache_3_cast_fp16")]; + tensor new_k_cache_3_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_3_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache1_to_fp16_dtype_0 = const()[name = tensor("v_cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache1_to_fp16 = cast(dtype = v_cache1_to_fp16_dtype_0, x = v_cache1)[name = tensor("cast_565")]; + tensor var_617_cast_fp16 = mul(x = v_cache1_to_fp16, y = var_603_cast_fp16)[name = tensor("op_617_cast_fp16")]; + tensor var_622 = const()[name = tensor("op_622"), val = tensor([1, 1, 16, 64])]; + tensor var_623_cast_fp16 = reshape(shape = var_622, x = v_3_cast_fp16)[name = tensor("op_623_cast_fp16")]; + tensor var_624_cast_fp16 = mul(x = var_623_cast_fp16, y = onehot_7_cast_fp16)[name = tensor("op_624_cast_fp16")]; + tensor new_v_cache_3_cast_fp16 = add(x = var_617_cast_fp16, y = var_624_cast_fp16)[name = tensor("new_v_cache_3_cast_fp16")]; + tensor new_v_cache_3_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_3_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_1_perm_0 = const()[name = tensor("transpose_1_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_637 = const()[name = tensor("op_637"), val = tensor([0, 2, 1, 3])]; + tensor var_643 = const()[name = tensor("op_643"), val = tensor([1, 16, 1, 64])]; + tensor q4_3_cast_fp16 = reshape(shape = var_643, x = out_5_cast_fp16)[name = tensor("q4_3_cast_fp16")]; + tensor var_648_transpose_x_0 = const()[name = tensor("op_648_transpose_x_0"), val = tensor(false)]; + tensor var_648_transpose_y_0 = const()[name = tensor("op_648_transpose_y_0"), val = tensor(false)]; + tensor transpose_1_cast_fp16 = transpose(perm = transpose_1_perm_0, x = new_k_cache_3_cast_fp16)[name = tensor("transpose_92")]; + tensor var_648_cast_fp16 = matmul(transpose_x = var_648_transpose_x_0, transpose_y = var_648_transpose_y_0, x = q4_3_cast_fp16, y = transpose_1_cast_fp16)[name = tensor("op_648_cast_fp16")]; + tensor var_649_to_fp16 = const()[name = tensor("op_649_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_5_cast_fp16 = mul(x = var_648_cast_fp16, y = var_649_to_fp16)[name = tensor("scores_5_cast_fp16")]; + tensor var_653 = const()[name = tensor("op_653"), val = tensor([1, 1])]; + tensor var_654_cast_fp16 = reshape(shape = var_653, x = position1_to_fp16)[name = tensor("op_654_cast_fp16")]; + tensor var_655_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_654_cast_fp16)[name = tensor("op_655_cast_fp16")]; + tensor mask_3_to_fp16_dtype_0 = const()[name = tensor("mask_3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_662_to_fp16 = const()[name = tensor("op_662_to_fp16"), val = tensor(0x1p+0)]; + tensor var_655_cast_fp16_to_fp16 = cast(dtype = mask_3_to_fp16_dtype_0, x = var_655_cast_fp16)[name = tensor("cast_563")]; + tensor var_663_cast_fp16 = sub(x = var_655_cast_fp16_to_fp16, y = var_662_to_fp16)[name = tensor("op_663_cast_fp16")]; + tensor var_668 = const()[name = tensor("op_668"), val = tensor([1, 1, 1, 512])]; + tensor var_669_cast_fp16 = reshape(shape = var_668, x = var_663_cast_fp16)[name = tensor("op_669_cast_fp16")]; + tensor var_670_to_fp16 = const()[name = tensor("op_670_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_671_cast_fp16 = mul(x = var_669_cast_fp16, y = var_670_to_fp16)[name = tensor("op_671_cast_fp16")]; + tensor scores_7_cast_fp16 = add(x = scores_5_cast_fp16, y = var_671_cast_fp16)[name = tensor("scores_7_cast_fp16")]; + tensor var_674 = const()[name = tensor("op_674"), val = tensor(-1)]; + tensor weights_3_cast_fp16 = softmax(axis = var_674, x = scores_7_cast_fp16)[name = tensor("weights_3_cast_fp16")]; + tensor attn_3_transpose_x_0 = const()[name = tensor("attn_3_transpose_x_0"), val = tensor(false)]; + tensor attn_3_transpose_y_0 = const()[name = tensor("attn_3_transpose_y_0"), val = tensor(false)]; + tensor values_3_cast_fp16 = transpose(perm = var_637, x = new_v_cache_3_cast_fp16)[name = tensor("transpose_91")]; + tensor attn_3_cast_fp16 = matmul(transpose_x = attn_3_transpose_x_0, transpose_y = attn_3_transpose_y_0, x = weights_3_cast_fp16, y = values_3_cast_fp16)[name = tensor("attn_3_cast_fp16")]; + tensor var_682 = const()[name = tensor("op_682"), val = tensor([0, 2, 1, 3])]; + tensor var_687 = const()[name = tensor("op_687"), val = tensor([1, 1, 1024])]; + tensor var_683_cast_fp16 = transpose(perm = var_682, x = attn_3_cast_fp16)[name = tensor("transpose_90")]; + tensor input_19_cast_fp16 = reshape(shape = var_687, x = var_683_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor attn1_out_proj_weight_to_fp16 = const()[name = tensor("attn1_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31553728)))]; + tensor linear_6_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn1_out_proj_weight_to_fp16, x = input_19_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor var_693_to_fp16 = const()[name = tensor("op_693_to_fp16"), val = tensor(0x1p+0)]; + tensor var_694_cast_fp16 = add(x = position1_to_fp16, y = var_693_to_fp16)[name = tensor("op_694_cast_fp16")]; + tensor var_694_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_694_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_21_cast_fp16 = add(x = input_15_cast_fp16, y = linear_6_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor input_23_axes_0 = const()[name = tensor("input_23_axes_0"), val = tensor([-1])]; + tensor norm1_2_weight_to_fp16 = const()[name = tensor("norm1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33650944)))]; + tensor norm1_2_bias_to_fp16 = const()[name = tensor("norm1_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33653056)))]; + tensor var_698_to_fp16 = const()[name = tensor("op_698_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_23_cast_fp16 = layer_norm(axes = input_23_axes_0, beta = norm1_2_bias_to_fp16, epsilon = var_698_to_fp16, gamma = norm1_2_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor linear1_1_weight_to_fp16 = const()[name = tensor("linear1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33655168)))]; + tensor linear_7_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear1_1_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_25_mode_0 = const()[name = tensor("input_25_mode_0"), val = tensor("EXACT")]; + tensor input_25_cast_fp16 = gelu(mode = input_25_mode_0, x = linear_7_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor linear1_2_weight_to_fp16 = const()[name = tensor("linear1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42043840)))]; + tensor linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear1_2_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor input_27_cast_fp16 = add(x = input_21_cast_fp16, y = linear_8_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor input_29_axes_0 = const()[name = tensor("input_29_axes_0"), val = tensor([-1])]; + tensor norm2_1_weight_to_fp16 = const()[name = tensor("norm2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50432512)))]; + tensor norm2_1_bias_to_fp16 = const()[name = tensor("norm2_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50434624)))]; + tensor var_715_to_fp16 = const()[name = tensor("op_715_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_29_cast_fp16 = layer_norm(axes = input_29_axes_0, beta = norm2_1_bias_to_fp16, epsilon = var_715_to_fp16, gamma = norm2_1_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor attn2_in_proj_weight_to_fp16 = const()[name = tensor("attn2_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50436736)))]; + tensor linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn2_in_proj_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor var_728 = const()[name = tensor("op_728"), val = tensor([1, 3, 16, 64])]; + tensor qkv_5_cast_fp16 = reshape(shape = var_728, x = linear_9_cast_fp16)[name = tensor("qkv_5_cast_fp16")]; + tensor x_9_begin_0 = const()[name = tensor("x_9_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_9_end_0 = const()[name = tensor("x_9_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_9_end_mask_0 = const()[name = tensor("x_9_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_9_squeeze_mask_0 = const()[name = tensor("x_9_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_9_cast_fp16 = slice_by_index(begin = x_9_begin_0, end = x_9_end_0, end_mask = x_9_end_mask_0, squeeze_mask = x_9_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("x_9_cast_fp16")]; + tensor x_11_begin_0 = const()[name = tensor("x_11_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_11_end_0 = const()[name = tensor("x_11_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_11_end_mask_0 = const()[name = tensor("x_11_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_11_squeeze_mask_0 = const()[name = tensor("x_11_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_11_cast_fp16 = slice_by_index(begin = x_11_begin_0, end = x_11_end_0, end_mask = x_11_end_mask_0, squeeze_mask = x_11_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("x_11_cast_fp16")]; + tensor v_5_begin_0 = const()[name = tensor("v_5_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_5_end_0 = const()[name = tensor("v_5_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_5_end_mask_0 = const()[name = tensor("v_5_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_5_squeeze_mask_0 = const()[name = tensor("v_5_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_5_cast_fp16 = slice_by_index(begin = v_5_begin_0, end = v_5_end_0, end_mask = v_5_end_mask_0, squeeze_mask = v_5_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("v_5_cast_fp16")]; + tensor var_757 = const()[name = tensor("op_757"), val = tensor([1, 1, 1])]; + tensor position2_to_fp16_dtype_0 = const()[name = tensor("position2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position2_to_fp16 = cast(dtype = position2_to_fp16_dtype_0, x = position2)[name = tensor("cast_561")]; + tensor var_758_cast_fp16 = reshape(shape = var_757, x = position2_to_fp16)[name = tensor("op_758_cast_fp16")]; + tensor angles_9_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_758_cast_fp16)[name = tensor("angles_9_cast_fp16")]; + tensor rotr_9_cast_fp16 = cos(x = angles_9_cast_fp16)[name = tensor("rotr_9_cast_fp16")]; + tensor roti_9_cast_fp16 = sin(x = angles_9_cast_fp16)[name = tensor("roti_9_cast_fp16")]; + tensor var_766 = const()[name = tensor("op_766"), val = tensor([1, 16, 32, 2])]; + tensor pairs_9_cast_fp16 = reshape(shape = var_766, x = x_9_cast_fp16)[name = tensor("pairs_9_cast_fp16")]; + tensor xr_9_begin_0 = const()[name = tensor("xr_9_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_9_end_0 = const()[name = tensor("xr_9_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_9_end_mask_0 = const()[name = tensor("xr_9_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_9_squeeze_mask_0 = const()[name = tensor("xr_9_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_9_cast_fp16 = slice_by_index(begin = xr_9_begin_0, end = xr_9_end_0, end_mask = xr_9_end_mask_0, squeeze_mask = xr_9_squeeze_mask_0, x = pairs_9_cast_fp16)[name = tensor("xr_9_cast_fp16")]; + tensor xi_9_begin_0 = const()[name = tensor("xi_9_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_9_end_0 = const()[name = tensor("xi_9_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_9_end_mask_0 = const()[name = tensor("xi_9_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_9_squeeze_mask_0 = const()[name = tensor("xi_9_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_9_cast_fp16 = slice_by_index(begin = xi_9_begin_0, end = xi_9_end_0, end_mask = xi_9_end_mask_0, squeeze_mask = xi_9_squeeze_mask_0, x = pairs_9_cast_fp16)[name = tensor("xi_9_cast_fp16")]; + tensor var_774_cast_fp16 = mul(x = xr_9_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_774_cast_fp16")]; + tensor var_775_cast_fp16 = mul(x = xi_9_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_775_cast_fp16")]; + tensor xor__9_cast_fp16 = sub(x = var_774_cast_fp16, y = var_775_cast_fp16)[name = tensor("xor__9_cast_fp16")]; + tensor var_778_cast_fp16 = mul(x = xr_9_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_778_cast_fp16")]; + tensor var_779_cast_fp16 = mul(x = xi_9_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_779_cast_fp16")]; + tensor xoi_9_cast_fp16 = add(x = var_778_cast_fp16, y = var_779_cast_fp16)[name = tensor("xoi_9_cast_fp16")]; + tensor out_9_axis_0 = const()[name = tensor("out_9_axis_0"), val = tensor(-1)]; + tensor out_9_cast_fp16 = stack(axis = out_9_axis_0, values = (xor__9_cast_fp16, xoi_9_cast_fp16))[name = tensor("out_9_cast_fp16")]; + tensor var_802 = const()[name = tensor("op_802"), val = tensor([1, 16, 32, 2])]; + tensor pairs_11_cast_fp16 = reshape(shape = var_802, x = x_11_cast_fp16)[name = tensor("pairs_11_cast_fp16")]; + tensor xr_11_begin_0 = const()[name = tensor("xr_11_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_11_end_0 = const()[name = tensor("xr_11_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_11_end_mask_0 = const()[name = tensor("xr_11_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_11_squeeze_mask_0 = const()[name = tensor("xr_11_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_11_cast_fp16 = slice_by_index(begin = xr_11_begin_0, end = xr_11_end_0, end_mask = xr_11_end_mask_0, squeeze_mask = xr_11_squeeze_mask_0, x = pairs_11_cast_fp16)[name = tensor("xr_11_cast_fp16")]; + tensor xi_11_begin_0 = const()[name = tensor("xi_11_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_11_end_0 = const()[name = tensor("xi_11_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_11_end_mask_0 = const()[name = tensor("xi_11_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_11_squeeze_mask_0 = const()[name = tensor("xi_11_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_11_cast_fp16 = slice_by_index(begin = xi_11_begin_0, end = xi_11_end_0, end_mask = xi_11_end_mask_0, squeeze_mask = xi_11_squeeze_mask_0, x = pairs_11_cast_fp16)[name = tensor("xi_11_cast_fp16")]; + tensor var_810_cast_fp16 = mul(x = xr_11_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_810_cast_fp16")]; + tensor var_811_cast_fp16 = mul(x = xi_11_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_811_cast_fp16")]; + tensor xor__11_cast_fp16 = sub(x = var_810_cast_fp16, y = var_811_cast_fp16)[name = tensor("xor__11_cast_fp16")]; + tensor var_814_cast_fp16 = mul(x = xr_11_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_814_cast_fp16")]; + tensor var_815_cast_fp16 = mul(x = xi_11_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_815_cast_fp16")]; + tensor xoi_11_cast_fp16 = add(x = var_814_cast_fp16, y = var_815_cast_fp16)[name = tensor("xoi_11_cast_fp16")]; + tensor out_11_axis_0 = const()[name = tensor("out_11_axis_0"), val = tensor(-1)]; + tensor out_11_cast_fp16 = stack(axis = out_11_axis_0, values = (xor__11_cast_fp16, xoi_11_cast_fp16))[name = tensor("out_11_cast_fp16")]; + tensor _inversed_827_y_0_to_fp16 = const()[name = tensor("_inversed_827_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_827_cast_fp16 = mul(x = position2_to_fp16, y = _inversed_827_y_0_to_fp16)[name = tensor("_inversed_827_cast_fp16")]; + tensor var_828_cast_fp16 = floor(x = _inversed_827_cast_fp16)[name = tensor("op_828_cast_fp16")]; + tensor var_829_to_fp16 = const()[name = tensor("op_829_to_fp16"), val = tensor(0x1p+9)]; + tensor var_830_cast_fp16 = mul(x = var_828_cast_fp16, y = var_829_to_fp16)[name = tensor("op_830_cast_fp16")]; + tensor write_idx_5_cast_fp16 = sub(x = position2_to_fp16, y = var_830_cast_fp16)[name = tensor("write_idx_5_cast_fp16")]; + tensor var_833_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_5_cast_fp16)[name = tensor("op_833_cast_fp16")]; + tensor var_843 = const()[name = tensor("op_843"), val = tensor([1, 512, 1, 1])]; + tensor onehot_9_to_fp16_dtype_0 = const()[name = tensor("onehot_9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_833_cast_fp16_to_fp16 = cast(dtype = onehot_9_to_fp16_dtype_0, x = var_833_cast_fp16)[name = tensor("cast_560")]; + tensor onehot_11_cast_fp16 = reshape(shape = var_843, x = var_833_cast_fp16_to_fp16)[name = tensor("onehot_11_cast_fp16")]; + tensor var_845_to_fp16 = const()[name = tensor("op_845_to_fp16"), val = tensor(0x1p+0)]; + tensor var_847_cast_fp16 = sub(x = var_845_to_fp16, y = onehot_11_cast_fp16)[name = tensor("op_847_cast_fp16")]; + tensor k_cache2_to_fp16_dtype_0 = const()[name = tensor("k_cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache2_to_fp16 = cast(dtype = k_cache2_to_fp16_dtype_0, x = k_cache2)[name = tensor("cast_559")]; + tensor var_848_cast_fp16 = mul(x = k_cache2_to_fp16, y = var_847_cast_fp16)[name = tensor("op_848_cast_fp16")]; + tensor var_853 = const()[name = tensor("op_853"), val = tensor([1, 1, 16, 64])]; + tensor var_854_cast_fp16 = reshape(shape = var_853, x = out_11_cast_fp16)[name = tensor("op_854_cast_fp16")]; + tensor var_855_cast_fp16 = mul(x = var_854_cast_fp16, y = onehot_11_cast_fp16)[name = tensor("op_855_cast_fp16")]; + tensor new_k_cache_5_cast_fp16 = add(x = var_848_cast_fp16, y = var_855_cast_fp16)[name = tensor("new_k_cache_5_cast_fp16")]; + tensor new_k_cache_5_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_5_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache2_to_fp16_dtype_0 = const()[name = tensor("v_cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache2_to_fp16 = cast(dtype = v_cache2_to_fp16_dtype_0, x = v_cache2)[name = tensor("cast_557")]; + tensor var_861_cast_fp16 = mul(x = v_cache2_to_fp16, y = var_847_cast_fp16)[name = tensor("op_861_cast_fp16")]; + tensor var_866 = const()[name = tensor("op_866"), val = tensor([1, 1, 16, 64])]; + tensor var_867_cast_fp16 = reshape(shape = var_866, x = v_5_cast_fp16)[name = tensor("op_867_cast_fp16")]; + tensor var_868_cast_fp16 = mul(x = var_867_cast_fp16, y = onehot_11_cast_fp16)[name = tensor("op_868_cast_fp16")]; + tensor new_v_cache_5_cast_fp16 = add(x = var_861_cast_fp16, y = var_868_cast_fp16)[name = tensor("new_v_cache_5_cast_fp16")]; + tensor new_v_cache_5_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_5_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_2_perm_0 = const()[name = tensor("transpose_2_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_881 = const()[name = tensor("op_881"), val = tensor([0, 2, 1, 3])]; + tensor var_887 = const()[name = tensor("op_887"), val = tensor([1, 16, 1, 64])]; + tensor q4_5_cast_fp16 = reshape(shape = var_887, x = out_9_cast_fp16)[name = tensor("q4_5_cast_fp16")]; + tensor var_892_transpose_x_0 = const()[name = tensor("op_892_transpose_x_0"), val = tensor(false)]; + tensor var_892_transpose_y_0 = const()[name = tensor("op_892_transpose_y_0"), val = tensor(false)]; + tensor transpose_2_cast_fp16 = transpose(perm = transpose_2_perm_0, x = new_k_cache_5_cast_fp16)[name = tensor("transpose_89")]; + tensor var_892_cast_fp16 = matmul(transpose_x = var_892_transpose_x_0, transpose_y = var_892_transpose_y_0, x = q4_5_cast_fp16, y = transpose_2_cast_fp16)[name = tensor("op_892_cast_fp16")]; + tensor var_893_to_fp16 = const()[name = tensor("op_893_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_9_cast_fp16 = mul(x = var_892_cast_fp16, y = var_893_to_fp16)[name = tensor("scores_9_cast_fp16")]; + tensor var_897 = const()[name = tensor("op_897"), val = tensor([1, 1])]; + tensor var_898_cast_fp16 = reshape(shape = var_897, x = position2_to_fp16)[name = tensor("op_898_cast_fp16")]; + tensor var_899_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_898_cast_fp16)[name = tensor("op_899_cast_fp16")]; + tensor mask_5_to_fp16_dtype_0 = const()[name = tensor("mask_5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_906_to_fp16 = const()[name = tensor("op_906_to_fp16"), val = tensor(0x1p+0)]; + tensor var_899_cast_fp16_to_fp16 = cast(dtype = mask_5_to_fp16_dtype_0, x = var_899_cast_fp16)[name = tensor("cast_555")]; + tensor var_907_cast_fp16 = sub(x = var_899_cast_fp16_to_fp16, y = var_906_to_fp16)[name = tensor("op_907_cast_fp16")]; + tensor var_912 = const()[name = tensor("op_912"), val = tensor([1, 1, 1, 512])]; + tensor var_913_cast_fp16 = reshape(shape = var_912, x = var_907_cast_fp16)[name = tensor("op_913_cast_fp16")]; + tensor var_914_to_fp16 = const()[name = tensor("op_914_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_915_cast_fp16 = mul(x = var_913_cast_fp16, y = var_914_to_fp16)[name = tensor("op_915_cast_fp16")]; + tensor scores_11_cast_fp16 = add(x = scores_9_cast_fp16, y = var_915_cast_fp16)[name = tensor("scores_11_cast_fp16")]; + tensor var_918 = const()[name = tensor("op_918"), val = tensor(-1)]; + tensor weights_5_cast_fp16 = softmax(axis = var_918, x = scores_11_cast_fp16)[name = tensor("weights_5_cast_fp16")]; + tensor attn_5_transpose_x_0 = const()[name = tensor("attn_5_transpose_x_0"), val = tensor(false)]; + tensor attn_5_transpose_y_0 = const()[name = tensor("attn_5_transpose_y_0"), val = tensor(false)]; + tensor values_5_cast_fp16 = transpose(perm = var_881, x = new_v_cache_5_cast_fp16)[name = tensor("transpose_88")]; + tensor attn_5_cast_fp16 = matmul(transpose_x = attn_5_transpose_x_0, transpose_y = attn_5_transpose_y_0, x = weights_5_cast_fp16, y = values_5_cast_fp16)[name = tensor("attn_5_cast_fp16")]; + tensor var_926 = const()[name = tensor("op_926"), val = tensor([0, 2, 1, 3])]; + tensor var_931 = const()[name = tensor("op_931"), val = tensor([1, 1, 1024])]; + tensor var_927_cast_fp16 = transpose(perm = var_926, x = attn_5_cast_fp16)[name = tensor("transpose_87")]; + tensor input_31_cast_fp16 = reshape(shape = var_931, x = var_927_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor attn2_out_proj_weight_to_fp16 = const()[name = tensor("attn2_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(56728256)))]; + tensor linear_10_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn2_out_proj_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor var_937_to_fp16 = const()[name = tensor("op_937_to_fp16"), val = tensor(0x1p+0)]; + tensor var_938_cast_fp16 = add(x = position2_to_fp16, y = var_937_to_fp16)[name = tensor("op_938_cast_fp16")]; + tensor var_938_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_938_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_33_cast_fp16 = add(x = input_27_cast_fp16, y = linear_10_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor input_35_axes_0 = const()[name = tensor("input_35_axes_0"), val = tensor([-1])]; + tensor norm2_2_weight_to_fp16 = const()[name = tensor("norm2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58825472)))]; + tensor norm2_2_bias_to_fp16 = const()[name = tensor("norm2_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58827584)))]; + tensor var_942_to_fp16 = const()[name = tensor("op_942_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_35_cast_fp16 = layer_norm(axes = input_35_axes_0, beta = norm2_2_bias_to_fp16, epsilon = var_942_to_fp16, gamma = norm2_2_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor linear2_1_weight_to_fp16 = const()[name = tensor("linear2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58829696)))]; + tensor linear_11_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear2_1_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor input_37_mode_0 = const()[name = tensor("input_37_mode_0"), val = tensor("EXACT")]; + tensor input_37_cast_fp16 = gelu(mode = input_37_mode_0, x = linear_11_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor linear2_2_weight_to_fp16 = const()[name = tensor("linear2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67218368)))]; + tensor linear_12_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear2_2_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor input_39_cast_fp16 = add(x = input_33_cast_fp16, y = linear_12_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor input_41_axes_0 = const()[name = tensor("input_41_axes_0"), val = tensor([-1])]; + tensor norm3_1_weight_to_fp16 = const()[name = tensor("norm3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75607040)))]; + tensor norm3_1_bias_to_fp16 = const()[name = tensor("norm3_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75609152)))]; + tensor var_959_to_fp16 = const()[name = tensor("op_959_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_41_cast_fp16 = layer_norm(axes = input_41_axes_0, beta = norm3_1_bias_to_fp16, epsilon = var_959_to_fp16, gamma = norm3_1_weight_to_fp16, x = input_39_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor attn3_in_proj_weight_to_fp16 = const()[name = tensor("attn3_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75611264)))]; + tensor linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn3_in_proj_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor var_972 = const()[name = tensor("op_972"), val = tensor([1, 3, 16, 64])]; + tensor qkv_7_cast_fp16 = reshape(shape = var_972, x = linear_13_cast_fp16)[name = tensor("qkv_7_cast_fp16")]; + tensor x_13_begin_0 = const()[name = tensor("x_13_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_13_end_0 = const()[name = tensor("x_13_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_13_end_mask_0 = const()[name = tensor("x_13_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_13_squeeze_mask_0 = const()[name = tensor("x_13_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_13_cast_fp16 = slice_by_index(begin = x_13_begin_0, end = x_13_end_0, end_mask = x_13_end_mask_0, squeeze_mask = x_13_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("x_13_cast_fp16")]; + tensor x_15_begin_0 = const()[name = tensor("x_15_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_15_end_0 = const()[name = tensor("x_15_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_15_end_mask_0 = const()[name = tensor("x_15_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_15_squeeze_mask_0 = const()[name = tensor("x_15_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_15_cast_fp16 = slice_by_index(begin = x_15_begin_0, end = x_15_end_0, end_mask = x_15_end_mask_0, squeeze_mask = x_15_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("x_15_cast_fp16")]; + tensor v_7_begin_0 = const()[name = tensor("v_7_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_7_end_0 = const()[name = tensor("v_7_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_7_end_mask_0 = const()[name = tensor("v_7_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_7_squeeze_mask_0 = const()[name = tensor("v_7_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_7_cast_fp16 = slice_by_index(begin = v_7_begin_0, end = v_7_end_0, end_mask = v_7_end_mask_0, squeeze_mask = v_7_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("v_7_cast_fp16")]; + tensor var_1001 = const()[name = tensor("op_1001"), val = tensor([1, 1, 1])]; + tensor position3_to_fp16_dtype_0 = const()[name = tensor("position3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position3_to_fp16 = cast(dtype = position3_to_fp16_dtype_0, x = position3)[name = tensor("cast_553")]; + tensor var_1002_cast_fp16 = reshape(shape = var_1001, x = position3_to_fp16)[name = tensor("op_1002_cast_fp16")]; + tensor angles_13_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_1002_cast_fp16)[name = tensor("angles_13_cast_fp16")]; + tensor rotr_13_cast_fp16 = cos(x = angles_13_cast_fp16)[name = tensor("rotr_13_cast_fp16")]; + tensor roti_13_cast_fp16 = sin(x = angles_13_cast_fp16)[name = tensor("roti_13_cast_fp16")]; + tensor var_1010 = const()[name = tensor("op_1010"), val = tensor([1, 16, 32, 2])]; + tensor pairs_13_cast_fp16 = reshape(shape = var_1010, x = x_13_cast_fp16)[name = tensor("pairs_13_cast_fp16")]; + tensor xr_13_begin_0 = const()[name = tensor("xr_13_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_13_end_0 = const()[name = tensor("xr_13_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_13_end_mask_0 = const()[name = tensor("xr_13_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_13_squeeze_mask_0 = const()[name = tensor("xr_13_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_13_cast_fp16 = slice_by_index(begin = xr_13_begin_0, end = xr_13_end_0, end_mask = xr_13_end_mask_0, squeeze_mask = xr_13_squeeze_mask_0, x = pairs_13_cast_fp16)[name = tensor("xr_13_cast_fp16")]; + tensor xi_13_begin_0 = const()[name = tensor("xi_13_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_13_end_0 = const()[name = tensor("xi_13_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_13_end_mask_0 = const()[name = tensor("xi_13_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_13_squeeze_mask_0 = const()[name = tensor("xi_13_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_13_cast_fp16 = slice_by_index(begin = xi_13_begin_0, end = xi_13_end_0, end_mask = xi_13_end_mask_0, squeeze_mask = xi_13_squeeze_mask_0, x = pairs_13_cast_fp16)[name = tensor("xi_13_cast_fp16")]; + tensor var_1018_cast_fp16 = mul(x = xr_13_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_1018_cast_fp16")]; + tensor var_1019_cast_fp16 = mul(x = xi_13_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_1019_cast_fp16")]; + tensor xor__13_cast_fp16 = sub(x = var_1018_cast_fp16, y = var_1019_cast_fp16)[name = tensor("xor__13_cast_fp16")]; + tensor var_1022_cast_fp16 = mul(x = xr_13_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_1022_cast_fp16")]; + tensor var_1023_cast_fp16 = mul(x = xi_13_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_1023_cast_fp16")]; + tensor xoi_13_cast_fp16 = add(x = var_1022_cast_fp16, y = var_1023_cast_fp16)[name = tensor("xoi_13_cast_fp16")]; + tensor out_13_axis_0 = const()[name = tensor("out_13_axis_0"), val = tensor(-1)]; + tensor out_13_cast_fp16 = stack(axis = out_13_axis_0, values = (xor__13_cast_fp16, xoi_13_cast_fp16))[name = tensor("out_13_cast_fp16")]; + tensor var_1046 = const()[name = tensor("op_1046"), val = tensor([1, 16, 32, 2])]; + tensor pairs_15_cast_fp16 = reshape(shape = var_1046, x = x_15_cast_fp16)[name = tensor("pairs_15_cast_fp16")]; + tensor xr_15_begin_0 = const()[name = tensor("xr_15_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_15_end_0 = const()[name = tensor("xr_15_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_15_end_mask_0 = const()[name = tensor("xr_15_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_15_squeeze_mask_0 = const()[name = tensor("xr_15_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_15_cast_fp16 = slice_by_index(begin = xr_15_begin_0, end = xr_15_end_0, end_mask = xr_15_end_mask_0, squeeze_mask = xr_15_squeeze_mask_0, x = pairs_15_cast_fp16)[name = tensor("xr_15_cast_fp16")]; + tensor xi_15_begin_0 = const()[name = tensor("xi_15_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_15_end_0 = const()[name = tensor("xi_15_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_15_end_mask_0 = const()[name = tensor("xi_15_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_15_squeeze_mask_0 = const()[name = tensor("xi_15_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_15_cast_fp16 = slice_by_index(begin = xi_15_begin_0, end = xi_15_end_0, end_mask = xi_15_end_mask_0, squeeze_mask = xi_15_squeeze_mask_0, x = pairs_15_cast_fp16)[name = tensor("xi_15_cast_fp16")]; + tensor var_1054_cast_fp16 = mul(x = xr_15_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_1054_cast_fp16")]; + tensor var_1055_cast_fp16 = mul(x = xi_15_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_1055_cast_fp16")]; + tensor xor__15_cast_fp16 = sub(x = var_1054_cast_fp16, y = var_1055_cast_fp16)[name = tensor("xor__15_cast_fp16")]; + tensor var_1058_cast_fp16 = mul(x = xr_15_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_1058_cast_fp16")]; + tensor var_1059_cast_fp16 = mul(x = xi_15_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_1059_cast_fp16")]; + tensor xoi_15_cast_fp16 = add(x = var_1058_cast_fp16, y = var_1059_cast_fp16)[name = tensor("xoi_15_cast_fp16")]; + tensor out_15_axis_0 = const()[name = tensor("out_15_axis_0"), val = tensor(-1)]; + tensor out_15_cast_fp16 = stack(axis = out_15_axis_0, values = (xor__15_cast_fp16, xoi_15_cast_fp16))[name = tensor("out_15_cast_fp16")]; + tensor _inversed_1071_y_0_to_fp16 = const()[name = tensor("_inversed_1071_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1071_cast_fp16 = mul(x = position3_to_fp16, y = _inversed_1071_y_0_to_fp16)[name = tensor("_inversed_1071_cast_fp16")]; + tensor var_1072_cast_fp16 = floor(x = _inversed_1071_cast_fp16)[name = tensor("op_1072_cast_fp16")]; + tensor var_1073_to_fp16 = const()[name = tensor("op_1073_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1074_cast_fp16 = mul(x = var_1072_cast_fp16, y = var_1073_to_fp16)[name = tensor("op_1074_cast_fp16")]; + tensor write_idx_7_cast_fp16 = sub(x = position3_to_fp16, y = var_1074_cast_fp16)[name = tensor("write_idx_7_cast_fp16")]; + tensor var_1077_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_7_cast_fp16)[name = tensor("op_1077_cast_fp16")]; + tensor var_1087 = const()[name = tensor("op_1087"), val = tensor([1, 512, 1, 1])]; + tensor onehot_13_to_fp16_dtype_0 = const()[name = tensor("onehot_13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_1077_cast_fp16_to_fp16 = cast(dtype = onehot_13_to_fp16_dtype_0, x = var_1077_cast_fp16)[name = tensor("cast_552")]; + tensor onehot_15_cast_fp16 = reshape(shape = var_1087, x = var_1077_cast_fp16_to_fp16)[name = tensor("onehot_15_cast_fp16")]; + tensor var_1089_to_fp16 = const()[name = tensor("op_1089_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1091_cast_fp16 = sub(x = var_1089_to_fp16, y = onehot_15_cast_fp16)[name = tensor("op_1091_cast_fp16")]; + tensor k_cache3_to_fp16_dtype_0 = const()[name = tensor("k_cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache3_to_fp16 = cast(dtype = k_cache3_to_fp16_dtype_0, x = k_cache3)[name = tensor("cast_551")]; + tensor var_1092_cast_fp16 = mul(x = k_cache3_to_fp16, y = var_1091_cast_fp16)[name = tensor("op_1092_cast_fp16")]; + tensor var_1097 = const()[name = tensor("op_1097"), val = tensor([1, 1, 16, 64])]; + tensor var_1098_cast_fp16 = reshape(shape = var_1097, x = out_15_cast_fp16)[name = tensor("op_1098_cast_fp16")]; + tensor var_1099_cast_fp16 = mul(x = var_1098_cast_fp16, y = onehot_15_cast_fp16)[name = tensor("op_1099_cast_fp16")]; + tensor new_k_cache_7_cast_fp16 = add(x = var_1092_cast_fp16, y = var_1099_cast_fp16)[name = tensor("new_k_cache_7_cast_fp16")]; + tensor new_k_cache_7_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_7_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache3_to_fp16_dtype_0 = const()[name = tensor("v_cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache3_to_fp16 = cast(dtype = v_cache3_to_fp16_dtype_0, x = v_cache3)[name = tensor("cast_549")]; + tensor var_1105_cast_fp16 = mul(x = v_cache3_to_fp16, y = var_1091_cast_fp16)[name = tensor("op_1105_cast_fp16")]; + tensor var_1110 = const()[name = tensor("op_1110"), val = tensor([1, 1, 16, 64])]; + tensor var_1111_cast_fp16 = reshape(shape = var_1110, x = v_7_cast_fp16)[name = tensor("op_1111_cast_fp16")]; + tensor var_1112_cast_fp16 = mul(x = var_1111_cast_fp16, y = onehot_15_cast_fp16)[name = tensor("op_1112_cast_fp16")]; + tensor new_v_cache_7_cast_fp16 = add(x = var_1105_cast_fp16, y = var_1112_cast_fp16)[name = tensor("new_v_cache_7_cast_fp16")]; + tensor new_v_cache_7_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_7_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_3_perm_0 = const()[name = tensor("transpose_3_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_1125 = const()[name = tensor("op_1125"), val = tensor([0, 2, 1, 3])]; + tensor var_1131 = const()[name = tensor("op_1131"), val = tensor([1, 16, 1, 64])]; + tensor q4_7_cast_fp16 = reshape(shape = var_1131, x = out_13_cast_fp16)[name = tensor("q4_7_cast_fp16")]; + tensor var_1136_transpose_x_0 = const()[name = tensor("op_1136_transpose_x_0"), val = tensor(false)]; + tensor var_1136_transpose_y_0 = const()[name = tensor("op_1136_transpose_y_0"), val = tensor(false)]; + tensor transpose_3_cast_fp16 = transpose(perm = transpose_3_perm_0, x = new_k_cache_7_cast_fp16)[name = tensor("transpose_86")]; + tensor var_1136_cast_fp16 = matmul(transpose_x = var_1136_transpose_x_0, transpose_y = var_1136_transpose_y_0, x = q4_7_cast_fp16, y = transpose_3_cast_fp16)[name = tensor("op_1136_cast_fp16")]; + tensor var_1137_to_fp16 = const()[name = tensor("op_1137_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_13_cast_fp16 = mul(x = var_1136_cast_fp16, y = var_1137_to_fp16)[name = tensor("scores_13_cast_fp16")]; + tensor var_1141 = const()[name = tensor("op_1141"), val = tensor([1, 1])]; + tensor var_1142_cast_fp16 = reshape(shape = var_1141, x = position3_to_fp16)[name = tensor("op_1142_cast_fp16")]; + tensor var_1143_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_1142_cast_fp16)[name = tensor("op_1143_cast_fp16")]; + tensor mask_7_to_fp16_dtype_0 = const()[name = tensor("mask_7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_1150_to_fp16 = const()[name = tensor("op_1150_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1143_cast_fp16_to_fp16 = cast(dtype = mask_7_to_fp16_dtype_0, x = var_1143_cast_fp16)[name = tensor("cast_547")]; + tensor var_1151_cast_fp16 = sub(x = var_1143_cast_fp16_to_fp16, y = var_1150_to_fp16)[name = tensor("op_1151_cast_fp16")]; + tensor var_1156 = const()[name = tensor("op_1156"), val = tensor([1, 1, 1, 512])]; + tensor var_1157_cast_fp16 = reshape(shape = var_1156, x = var_1151_cast_fp16)[name = tensor("op_1157_cast_fp16")]; + tensor var_1158_to_fp16 = const()[name = tensor("op_1158_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_1159_cast_fp16 = mul(x = var_1157_cast_fp16, y = var_1158_to_fp16)[name = tensor("op_1159_cast_fp16")]; + tensor scores_15_cast_fp16 = add(x = scores_13_cast_fp16, y = var_1159_cast_fp16)[name = tensor("scores_15_cast_fp16")]; + tensor var_1162 = const()[name = tensor("op_1162"), val = tensor(-1)]; + tensor weights_7_cast_fp16 = softmax(axis = var_1162, x = scores_15_cast_fp16)[name = tensor("weights_7_cast_fp16")]; + tensor attn_7_transpose_x_0 = const()[name = tensor("attn_7_transpose_x_0"), val = tensor(false)]; + tensor attn_7_transpose_y_0 = const()[name = tensor("attn_7_transpose_y_0"), val = tensor(false)]; + tensor values_7_cast_fp16 = transpose(perm = var_1125, x = new_v_cache_7_cast_fp16)[name = tensor("transpose_85")]; + tensor attn_7_cast_fp16 = matmul(transpose_x = attn_7_transpose_x_0, transpose_y = attn_7_transpose_y_0, x = weights_7_cast_fp16, y = values_7_cast_fp16)[name = tensor("attn_7_cast_fp16")]; + tensor var_1170 = const()[name = tensor("op_1170"), val = tensor([0, 2, 1, 3])]; + tensor var_1175 = const()[name = tensor("op_1175"), val = tensor([1, 1, 1024])]; + tensor var_1171_cast_fp16 = transpose(perm = var_1170, x = attn_7_cast_fp16)[name = tensor("transpose_84")]; + tensor input_43_cast_fp16 = reshape(shape = var_1175, x = var_1171_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor attn3_out_proj_weight_to_fp16 = const()[name = tensor("attn3_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(81902784)))]; + tensor linear_14_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn3_out_proj_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor var_1181_to_fp16 = const()[name = tensor("op_1181_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1182_cast_fp16 = add(x = position3_to_fp16, y = var_1181_to_fp16)[name = tensor("op_1182_cast_fp16")]; + tensor var_1182_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1182_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_45_cast_fp16 = add(x = input_39_cast_fp16, y = linear_14_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor input_47_axes_0 = const()[name = tensor("input_47_axes_0"), val = tensor([-1])]; + tensor norm3_2_weight_to_fp16 = const()[name = tensor("norm3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84000000)))]; + tensor norm3_2_bias_to_fp16 = const()[name = tensor("norm3_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84002112)))]; + tensor var_1186_to_fp16 = const()[name = tensor("op_1186_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_47_cast_fp16 = layer_norm(axes = input_47_axes_0, beta = norm3_2_bias_to_fp16, epsilon = var_1186_to_fp16, gamma = norm3_2_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor linear3_1_weight_to_fp16 = const()[name = tensor("linear3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84004224)))]; + tensor linear_15_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear3_1_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor input_49_mode_0 = const()[name = tensor("input_49_mode_0"), val = tensor("EXACT")]; + tensor input_49_cast_fp16 = gelu(mode = input_49_mode_0, x = linear_15_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor linear3_2_weight_to_fp16 = const()[name = tensor("linear3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(92392896)))]; + tensor linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear3_2_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor input_51_cast_fp16 = add(x = input_45_cast_fp16, y = linear_16_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor input_53_axes_0 = const()[name = tensor("input_53_axes_0"), val = tensor([-1])]; + tensor norm4_1_weight_to_fp16 = const()[name = tensor("norm4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100781568)))]; + tensor norm4_1_bias_to_fp16 = const()[name = tensor("norm4_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100783680)))]; + tensor var_1203_to_fp16 = const()[name = tensor("op_1203_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_53_cast_fp16 = layer_norm(axes = input_53_axes_0, beta = norm4_1_bias_to_fp16, epsilon = var_1203_to_fp16, gamma = norm4_1_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor attn4_in_proj_weight_to_fp16 = const()[name = tensor("attn4_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100785792)))]; + tensor linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn4_in_proj_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor var_1216 = const()[name = tensor("op_1216"), val = tensor([1, 3, 16, 64])]; + tensor qkv_9_cast_fp16 = reshape(shape = var_1216, x = linear_17_cast_fp16)[name = tensor("qkv_9_cast_fp16")]; + tensor x_17_begin_0 = const()[name = tensor("x_17_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_17_end_0 = const()[name = tensor("x_17_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_17_end_mask_0 = const()[name = tensor("x_17_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_17_squeeze_mask_0 = const()[name = tensor("x_17_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_17_cast_fp16 = slice_by_index(begin = x_17_begin_0, end = x_17_end_0, end_mask = x_17_end_mask_0, squeeze_mask = x_17_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("x_17_cast_fp16")]; + tensor x_19_begin_0 = const()[name = tensor("x_19_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_19_end_0 = const()[name = tensor("x_19_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_19_end_mask_0 = const()[name = tensor("x_19_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_19_squeeze_mask_0 = const()[name = tensor("x_19_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_19_cast_fp16 = slice_by_index(begin = x_19_begin_0, end = x_19_end_0, end_mask = x_19_end_mask_0, squeeze_mask = x_19_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("x_19_cast_fp16")]; + tensor v_9_begin_0 = const()[name = tensor("v_9_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_9_end_0 = const()[name = tensor("v_9_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_9_end_mask_0 = const()[name = tensor("v_9_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_9_squeeze_mask_0 = const()[name = tensor("v_9_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_9_cast_fp16 = slice_by_index(begin = v_9_begin_0, end = v_9_end_0, end_mask = v_9_end_mask_0, squeeze_mask = v_9_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("v_9_cast_fp16")]; + tensor var_1245 = const()[name = tensor("op_1245"), val = tensor([1, 1, 1])]; + tensor position4_to_fp16_dtype_0 = const()[name = tensor("position4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position4_to_fp16 = cast(dtype = position4_to_fp16_dtype_0, x = position4)[name = tensor("cast_545")]; + tensor var_1246_cast_fp16 = reshape(shape = var_1245, x = position4_to_fp16)[name = tensor("op_1246_cast_fp16")]; + tensor angles_17_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_1246_cast_fp16)[name = tensor("angles_17_cast_fp16")]; + tensor rotr_17_cast_fp16 = cos(x = angles_17_cast_fp16)[name = tensor("rotr_17_cast_fp16")]; + tensor roti_17_cast_fp16 = sin(x = angles_17_cast_fp16)[name = tensor("roti_17_cast_fp16")]; + tensor var_1254 = const()[name = tensor("op_1254"), val = tensor([1, 16, 32, 2])]; + tensor pairs_17_cast_fp16 = reshape(shape = var_1254, x = x_17_cast_fp16)[name = tensor("pairs_17_cast_fp16")]; + tensor xr_17_begin_0 = const()[name = tensor("xr_17_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_17_end_0 = const()[name = tensor("xr_17_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_17_end_mask_0 = const()[name = tensor("xr_17_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_17_squeeze_mask_0 = const()[name = tensor("xr_17_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_17_cast_fp16 = slice_by_index(begin = xr_17_begin_0, end = xr_17_end_0, end_mask = xr_17_end_mask_0, squeeze_mask = xr_17_squeeze_mask_0, x = pairs_17_cast_fp16)[name = tensor("xr_17_cast_fp16")]; + tensor xi_17_begin_0 = const()[name = tensor("xi_17_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_17_end_0 = const()[name = tensor("xi_17_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_17_end_mask_0 = const()[name = tensor("xi_17_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_17_squeeze_mask_0 = const()[name = tensor("xi_17_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_17_cast_fp16 = slice_by_index(begin = xi_17_begin_0, end = xi_17_end_0, end_mask = xi_17_end_mask_0, squeeze_mask = xi_17_squeeze_mask_0, x = pairs_17_cast_fp16)[name = tensor("xi_17_cast_fp16")]; + tensor var_1262_cast_fp16 = mul(x = xr_17_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_1262_cast_fp16")]; + tensor var_1263_cast_fp16 = mul(x = xi_17_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_1263_cast_fp16")]; + tensor xor__17_cast_fp16 = sub(x = var_1262_cast_fp16, y = var_1263_cast_fp16)[name = tensor("xor__17_cast_fp16")]; + tensor var_1266_cast_fp16 = mul(x = xr_17_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_1266_cast_fp16")]; + tensor var_1267_cast_fp16 = mul(x = xi_17_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_1267_cast_fp16")]; + tensor xoi_17_cast_fp16 = add(x = var_1266_cast_fp16, y = var_1267_cast_fp16)[name = tensor("xoi_17_cast_fp16")]; + tensor out_17_axis_0 = const()[name = tensor("out_17_axis_0"), val = tensor(-1)]; + tensor out_17_cast_fp16 = stack(axis = out_17_axis_0, values = (xor__17_cast_fp16, xoi_17_cast_fp16))[name = tensor("out_17_cast_fp16")]; + tensor var_1290 = const()[name = tensor("op_1290"), val = tensor([1, 16, 32, 2])]; + tensor pairs_19_cast_fp16 = reshape(shape = var_1290, x = x_19_cast_fp16)[name = tensor("pairs_19_cast_fp16")]; + tensor xr_19_begin_0 = const()[name = tensor("xr_19_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_19_end_0 = const()[name = tensor("xr_19_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_19_end_mask_0 = const()[name = tensor("xr_19_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_19_squeeze_mask_0 = const()[name = tensor("xr_19_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_19_cast_fp16 = slice_by_index(begin = xr_19_begin_0, end = xr_19_end_0, end_mask = xr_19_end_mask_0, squeeze_mask = xr_19_squeeze_mask_0, x = pairs_19_cast_fp16)[name = tensor("xr_19_cast_fp16")]; + tensor xi_19_begin_0 = const()[name = tensor("xi_19_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_19_end_0 = const()[name = tensor("xi_19_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_19_end_mask_0 = const()[name = tensor("xi_19_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_19_squeeze_mask_0 = const()[name = tensor("xi_19_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_19_cast_fp16 = slice_by_index(begin = xi_19_begin_0, end = xi_19_end_0, end_mask = xi_19_end_mask_0, squeeze_mask = xi_19_squeeze_mask_0, x = pairs_19_cast_fp16)[name = tensor("xi_19_cast_fp16")]; + tensor var_1298_cast_fp16 = mul(x = xr_19_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_1298_cast_fp16")]; + tensor var_1299_cast_fp16 = mul(x = xi_19_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_1299_cast_fp16")]; + tensor xor__19_cast_fp16 = sub(x = var_1298_cast_fp16, y = var_1299_cast_fp16)[name = tensor("xor__19_cast_fp16")]; + tensor var_1302_cast_fp16 = mul(x = xr_19_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_1302_cast_fp16")]; + tensor var_1303_cast_fp16 = mul(x = xi_19_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_1303_cast_fp16")]; + tensor xoi_19_cast_fp16 = add(x = var_1302_cast_fp16, y = var_1303_cast_fp16)[name = tensor("xoi_19_cast_fp16")]; + tensor out_19_axis_0 = const()[name = tensor("out_19_axis_0"), val = tensor(-1)]; + tensor out_19_cast_fp16 = stack(axis = out_19_axis_0, values = (xor__19_cast_fp16, xoi_19_cast_fp16))[name = tensor("out_19_cast_fp16")]; + tensor _inversed_1315_y_0_to_fp16 = const()[name = tensor("_inversed_1315_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1315_cast_fp16 = mul(x = position4_to_fp16, y = _inversed_1315_y_0_to_fp16)[name = tensor("_inversed_1315_cast_fp16")]; + tensor var_1316_cast_fp16 = floor(x = _inversed_1315_cast_fp16)[name = tensor("op_1316_cast_fp16")]; + tensor var_1317_to_fp16 = const()[name = tensor("op_1317_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1318_cast_fp16 = mul(x = var_1316_cast_fp16, y = var_1317_to_fp16)[name = tensor("op_1318_cast_fp16")]; + tensor write_idx_9_cast_fp16 = sub(x = position4_to_fp16, y = var_1318_cast_fp16)[name = tensor("write_idx_9_cast_fp16")]; + tensor var_1321_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_9_cast_fp16)[name = tensor("op_1321_cast_fp16")]; + tensor var_1331 = const()[name = tensor("op_1331"), val = tensor([1, 512, 1, 1])]; + tensor onehot_17_to_fp16_dtype_0 = const()[name = tensor("onehot_17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_1321_cast_fp16_to_fp16 = cast(dtype = onehot_17_to_fp16_dtype_0, x = var_1321_cast_fp16)[name = tensor("cast_544")]; + tensor onehot_19_cast_fp16 = reshape(shape = var_1331, x = var_1321_cast_fp16_to_fp16)[name = tensor("onehot_19_cast_fp16")]; + tensor var_1333_to_fp16 = const()[name = tensor("op_1333_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1335_cast_fp16 = sub(x = var_1333_to_fp16, y = onehot_19_cast_fp16)[name = tensor("op_1335_cast_fp16")]; + tensor k_cache4_to_fp16_dtype_0 = const()[name = tensor("k_cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache4_to_fp16 = cast(dtype = k_cache4_to_fp16_dtype_0, x = k_cache4)[name = tensor("cast_543")]; + tensor var_1336_cast_fp16 = mul(x = k_cache4_to_fp16, y = var_1335_cast_fp16)[name = tensor("op_1336_cast_fp16")]; + tensor var_1341 = const()[name = tensor("op_1341"), val = tensor([1, 1, 16, 64])]; + tensor var_1342_cast_fp16 = reshape(shape = var_1341, x = out_19_cast_fp16)[name = tensor("op_1342_cast_fp16")]; + tensor var_1343_cast_fp16 = mul(x = var_1342_cast_fp16, y = onehot_19_cast_fp16)[name = tensor("op_1343_cast_fp16")]; + tensor new_k_cache_9_cast_fp16 = add(x = var_1336_cast_fp16, y = var_1343_cast_fp16)[name = tensor("new_k_cache_9_cast_fp16")]; + tensor new_k_cache_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache4_to_fp16_dtype_0 = const()[name = tensor("v_cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache4_to_fp16 = cast(dtype = v_cache4_to_fp16_dtype_0, x = v_cache4)[name = tensor("cast_541")]; + tensor var_1349_cast_fp16 = mul(x = v_cache4_to_fp16, y = var_1335_cast_fp16)[name = tensor("op_1349_cast_fp16")]; + tensor var_1354 = const()[name = tensor("op_1354"), val = tensor([1, 1, 16, 64])]; + tensor var_1355_cast_fp16 = reshape(shape = var_1354, x = v_9_cast_fp16)[name = tensor("op_1355_cast_fp16")]; + tensor var_1356_cast_fp16 = mul(x = var_1355_cast_fp16, y = onehot_19_cast_fp16)[name = tensor("op_1356_cast_fp16")]; + tensor new_v_cache_9_cast_fp16 = add(x = var_1349_cast_fp16, y = var_1356_cast_fp16)[name = tensor("new_v_cache_9_cast_fp16")]; + tensor new_v_cache_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_4_perm_0 = const()[name = tensor("transpose_4_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_1369 = const()[name = tensor("op_1369"), val = tensor([0, 2, 1, 3])]; + tensor var_1375 = const()[name = tensor("op_1375"), val = tensor([1, 16, 1, 64])]; + tensor q4_9_cast_fp16 = reshape(shape = var_1375, x = out_17_cast_fp16)[name = tensor("q4_9_cast_fp16")]; + tensor var_1380_transpose_x_0 = const()[name = tensor("op_1380_transpose_x_0"), val = tensor(false)]; + tensor var_1380_transpose_y_0 = const()[name = tensor("op_1380_transpose_y_0"), val = tensor(false)]; + tensor transpose_4_cast_fp16 = transpose(perm = transpose_4_perm_0, x = new_k_cache_9_cast_fp16)[name = tensor("transpose_83")]; + tensor var_1380_cast_fp16 = matmul(transpose_x = var_1380_transpose_x_0, transpose_y = var_1380_transpose_y_0, x = q4_9_cast_fp16, y = transpose_4_cast_fp16)[name = tensor("op_1380_cast_fp16")]; + tensor var_1381_to_fp16 = const()[name = tensor("op_1381_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_17_cast_fp16 = mul(x = var_1380_cast_fp16, y = var_1381_to_fp16)[name = tensor("scores_17_cast_fp16")]; + tensor var_1385 = const()[name = tensor("op_1385"), val = tensor([1, 1])]; + tensor var_1386_cast_fp16 = reshape(shape = var_1385, x = position4_to_fp16)[name = tensor("op_1386_cast_fp16")]; + tensor var_1387_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_1386_cast_fp16)[name = tensor("op_1387_cast_fp16")]; + tensor mask_9_to_fp16_dtype_0 = const()[name = tensor("mask_9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_1394_to_fp16 = const()[name = tensor("op_1394_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1387_cast_fp16_to_fp16 = cast(dtype = mask_9_to_fp16_dtype_0, x = var_1387_cast_fp16)[name = tensor("cast_539")]; + tensor var_1395_cast_fp16 = sub(x = var_1387_cast_fp16_to_fp16, y = var_1394_to_fp16)[name = tensor("op_1395_cast_fp16")]; + tensor var_1400 = const()[name = tensor("op_1400"), val = tensor([1, 1, 1, 512])]; + tensor var_1401_cast_fp16 = reshape(shape = var_1400, x = var_1395_cast_fp16)[name = tensor("op_1401_cast_fp16")]; + tensor var_1402_to_fp16 = const()[name = tensor("op_1402_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_1403_cast_fp16 = mul(x = var_1401_cast_fp16, y = var_1402_to_fp16)[name = tensor("op_1403_cast_fp16")]; + tensor scores_19_cast_fp16 = add(x = scores_17_cast_fp16, y = var_1403_cast_fp16)[name = tensor("scores_19_cast_fp16")]; + tensor var_1406 = const()[name = tensor("op_1406"), val = tensor(-1)]; + tensor weights_9_cast_fp16 = softmax(axis = var_1406, x = scores_19_cast_fp16)[name = tensor("weights_9_cast_fp16")]; + tensor attn_9_transpose_x_0 = const()[name = tensor("attn_9_transpose_x_0"), val = tensor(false)]; + tensor attn_9_transpose_y_0 = const()[name = tensor("attn_9_transpose_y_0"), val = tensor(false)]; + tensor values_9_cast_fp16 = transpose(perm = var_1369, x = new_v_cache_9_cast_fp16)[name = tensor("transpose_82")]; + tensor attn_9_cast_fp16 = matmul(transpose_x = attn_9_transpose_x_0, transpose_y = attn_9_transpose_y_0, x = weights_9_cast_fp16, y = values_9_cast_fp16)[name = tensor("attn_9_cast_fp16")]; + tensor var_1414 = const()[name = tensor("op_1414"), val = tensor([0, 2, 1, 3])]; + tensor var_1419 = const()[name = tensor("op_1419"), val = tensor([1, 1, 1024])]; + tensor var_1415_cast_fp16 = transpose(perm = var_1414, x = attn_9_cast_fp16)[name = tensor("transpose_81")]; + tensor input_55_cast_fp16 = reshape(shape = var_1419, x = var_1415_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor attn4_out_proj_weight_to_fp16 = const()[name = tensor("attn4_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(107077312)))]; + tensor linear_18_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn4_out_proj_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor var_1425_to_fp16 = const()[name = tensor("op_1425_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1426_cast_fp16 = add(x = position4_to_fp16, y = var_1425_to_fp16)[name = tensor("op_1426_cast_fp16")]; + tensor var_1426_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1426_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_57_cast_fp16 = add(x = input_51_cast_fp16, y = linear_18_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor input_59_axes_0 = const()[name = tensor("input_59_axes_0"), val = tensor([-1])]; + tensor norm4_2_weight_to_fp16 = const()[name = tensor("norm4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109174528)))]; + tensor norm4_2_bias_to_fp16 = const()[name = tensor("norm4_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109176640)))]; + tensor var_1430_to_fp16 = const()[name = tensor("op_1430_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_59_cast_fp16 = layer_norm(axes = input_59_axes_0, beta = norm4_2_bias_to_fp16, epsilon = var_1430_to_fp16, gamma = norm4_2_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor linear4_1_weight_to_fp16 = const()[name = tensor("linear4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109178752)))]; + tensor linear_19_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear4_1_weight_to_fp16, x = input_59_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_61_mode_0 = const()[name = tensor("input_61_mode_0"), val = tensor("EXACT")]; + tensor input_61_cast_fp16 = gelu(mode = input_61_mode_0, x = linear_19_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor linear4_2_weight_to_fp16 = const()[name = tensor("linear4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(117567424)))]; + tensor linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear4_2_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor input_63_cast_fp16 = add(x = input_57_cast_fp16, y = linear_20_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor input_65_axes_0 = const()[name = tensor("input_65_axes_0"), val = tensor([-1])]; + tensor norm5_1_weight_to_fp16 = const()[name = tensor("norm5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125956096)))]; + tensor norm5_1_bias_to_fp16 = const()[name = tensor("norm5_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125958208)))]; + tensor var_1447_to_fp16 = const()[name = tensor("op_1447_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_65_cast_fp16 = layer_norm(axes = input_65_axes_0, beta = norm5_1_bias_to_fp16, epsilon = var_1447_to_fp16, gamma = norm5_1_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("input_65_cast_fp16")]; + tensor attn5_in_proj_weight_to_fp16 = const()[name = tensor("attn5_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125960320)))]; + tensor linear_21_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn5_in_proj_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("linear_21_cast_fp16")]; + tensor var_1460 = const()[name = tensor("op_1460"), val = tensor([1, 3, 16, 64])]; + tensor qkv_11_cast_fp16 = reshape(shape = var_1460, x = linear_21_cast_fp16)[name = tensor("qkv_11_cast_fp16")]; + tensor x_21_begin_0 = const()[name = tensor("x_21_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_21_end_0 = const()[name = tensor("x_21_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_21_end_mask_0 = const()[name = tensor("x_21_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_21_squeeze_mask_0 = const()[name = tensor("x_21_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_21_cast_fp16 = slice_by_index(begin = x_21_begin_0, end = x_21_end_0, end_mask = x_21_end_mask_0, squeeze_mask = x_21_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("x_21_cast_fp16")]; + tensor x_23_begin_0 = const()[name = tensor("x_23_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_23_end_0 = const()[name = tensor("x_23_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_23_end_mask_0 = const()[name = tensor("x_23_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_23_squeeze_mask_0 = const()[name = tensor("x_23_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_23_cast_fp16 = slice_by_index(begin = x_23_begin_0, end = x_23_end_0, end_mask = x_23_end_mask_0, squeeze_mask = x_23_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("x_23_cast_fp16")]; + tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_11_cast_fp16 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("v_11_cast_fp16")]; + tensor var_1489 = const()[name = tensor("op_1489"), val = tensor([1, 1, 1])]; + tensor position5_to_fp16_dtype_0 = const()[name = tensor("position5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position5_to_fp16 = cast(dtype = position5_to_fp16_dtype_0, x = position5)[name = tensor("cast_537")]; + tensor var_1490_cast_fp16 = reshape(shape = var_1489, x = position5_to_fp16)[name = tensor("op_1490_cast_fp16")]; + tensor angles_21_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_1490_cast_fp16)[name = tensor("angles_21_cast_fp16")]; + tensor rotr_21_cast_fp16 = cos(x = angles_21_cast_fp16)[name = tensor("rotr_21_cast_fp16")]; + tensor roti_21_cast_fp16 = sin(x = angles_21_cast_fp16)[name = tensor("roti_21_cast_fp16")]; + tensor var_1498 = const()[name = tensor("op_1498"), val = tensor([1, 16, 32, 2])]; + tensor pairs_21_cast_fp16 = reshape(shape = var_1498, x = x_21_cast_fp16)[name = tensor("pairs_21_cast_fp16")]; + tensor xr_21_begin_0 = const()[name = tensor("xr_21_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_21_end_0 = const()[name = tensor("xr_21_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_21_end_mask_0 = const()[name = tensor("xr_21_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_21_squeeze_mask_0 = const()[name = tensor("xr_21_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_21_cast_fp16 = slice_by_index(begin = xr_21_begin_0, end = xr_21_end_0, end_mask = xr_21_end_mask_0, squeeze_mask = xr_21_squeeze_mask_0, x = pairs_21_cast_fp16)[name = tensor("xr_21_cast_fp16")]; + tensor xi_21_begin_0 = const()[name = tensor("xi_21_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_21_end_0 = const()[name = tensor("xi_21_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_21_end_mask_0 = const()[name = tensor("xi_21_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_21_squeeze_mask_0 = const()[name = tensor("xi_21_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_21_cast_fp16 = slice_by_index(begin = xi_21_begin_0, end = xi_21_end_0, end_mask = xi_21_end_mask_0, squeeze_mask = xi_21_squeeze_mask_0, x = pairs_21_cast_fp16)[name = tensor("xi_21_cast_fp16")]; + tensor var_1506_cast_fp16 = mul(x = xr_21_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_1506_cast_fp16")]; + tensor var_1507_cast_fp16 = mul(x = xi_21_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_1507_cast_fp16")]; + tensor xor__21_cast_fp16 = sub(x = var_1506_cast_fp16, y = var_1507_cast_fp16)[name = tensor("xor__21_cast_fp16")]; + tensor var_1510_cast_fp16 = mul(x = xr_21_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_1510_cast_fp16")]; + tensor var_1511_cast_fp16 = mul(x = xi_21_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_1511_cast_fp16")]; + tensor xoi_21_cast_fp16 = add(x = var_1510_cast_fp16, y = var_1511_cast_fp16)[name = tensor("xoi_21_cast_fp16")]; + tensor out_21_axis_0 = const()[name = tensor("out_21_axis_0"), val = tensor(-1)]; + tensor out_21_cast_fp16 = stack(axis = out_21_axis_0, values = (xor__21_cast_fp16, xoi_21_cast_fp16))[name = tensor("out_21_cast_fp16")]; + tensor var_1534 = const()[name = tensor("op_1534"), val = tensor([1, 16, 32, 2])]; + tensor pairs_23_cast_fp16 = reshape(shape = var_1534, x = x_23_cast_fp16)[name = tensor("pairs_23_cast_fp16")]; + tensor xr_23_begin_0 = const()[name = tensor("xr_23_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_23_end_0 = const()[name = tensor("xr_23_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_23_end_mask_0 = const()[name = tensor("xr_23_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_23_squeeze_mask_0 = const()[name = tensor("xr_23_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_23_cast_fp16 = slice_by_index(begin = xr_23_begin_0, end = xr_23_end_0, end_mask = xr_23_end_mask_0, squeeze_mask = xr_23_squeeze_mask_0, x = pairs_23_cast_fp16)[name = tensor("xr_23_cast_fp16")]; + tensor xi_23_begin_0 = const()[name = tensor("xi_23_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_23_end_0 = const()[name = tensor("xi_23_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_23_end_mask_0 = const()[name = tensor("xi_23_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_23_squeeze_mask_0 = const()[name = tensor("xi_23_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_23_cast_fp16 = slice_by_index(begin = xi_23_begin_0, end = xi_23_end_0, end_mask = xi_23_end_mask_0, squeeze_mask = xi_23_squeeze_mask_0, x = pairs_23_cast_fp16)[name = tensor("xi_23_cast_fp16")]; + tensor var_1542_cast_fp16 = mul(x = xr_23_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_1542_cast_fp16")]; + tensor var_1543_cast_fp16 = mul(x = xi_23_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_1543_cast_fp16")]; + tensor xor__23_cast_fp16 = sub(x = var_1542_cast_fp16, y = var_1543_cast_fp16)[name = tensor("xor__23_cast_fp16")]; + tensor var_1546_cast_fp16 = mul(x = xr_23_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_1546_cast_fp16")]; + tensor var_1547_cast_fp16 = mul(x = xi_23_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_1547_cast_fp16")]; + tensor xoi_23_cast_fp16 = add(x = var_1546_cast_fp16, y = var_1547_cast_fp16)[name = tensor("xoi_23_cast_fp16")]; + tensor out_23_axis_0 = const()[name = tensor("out_23_axis_0"), val = tensor(-1)]; + tensor out_23_cast_fp16 = stack(axis = out_23_axis_0, values = (xor__23_cast_fp16, xoi_23_cast_fp16))[name = tensor("out_23_cast_fp16")]; + tensor _inversed_1559_y_0_to_fp16 = const()[name = tensor("_inversed_1559_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1559_cast_fp16 = mul(x = position5_to_fp16, y = _inversed_1559_y_0_to_fp16)[name = tensor("_inversed_1559_cast_fp16")]; + tensor var_1560_cast_fp16 = floor(x = _inversed_1559_cast_fp16)[name = tensor("op_1560_cast_fp16")]; + tensor var_1561_to_fp16 = const()[name = tensor("op_1561_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1562_cast_fp16 = mul(x = var_1560_cast_fp16, y = var_1561_to_fp16)[name = tensor("op_1562_cast_fp16")]; + tensor write_idx_11_cast_fp16 = sub(x = position5_to_fp16, y = var_1562_cast_fp16)[name = tensor("write_idx_11_cast_fp16")]; + tensor var_1565_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_11_cast_fp16)[name = tensor("op_1565_cast_fp16")]; + tensor var_1575 = const()[name = tensor("op_1575"), val = tensor([1, 512, 1, 1])]; + tensor onehot_21_to_fp16_dtype_0 = const()[name = tensor("onehot_21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_1565_cast_fp16_to_fp16 = cast(dtype = onehot_21_to_fp16_dtype_0, x = var_1565_cast_fp16)[name = tensor("cast_536")]; + tensor onehot_23_cast_fp16 = reshape(shape = var_1575, x = var_1565_cast_fp16_to_fp16)[name = tensor("onehot_23_cast_fp16")]; + tensor var_1577_to_fp16 = const()[name = tensor("op_1577_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1579_cast_fp16 = sub(x = var_1577_to_fp16, y = onehot_23_cast_fp16)[name = tensor("op_1579_cast_fp16")]; + tensor k_cache5_to_fp16_dtype_0 = const()[name = tensor("k_cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache5_to_fp16 = cast(dtype = k_cache5_to_fp16_dtype_0, x = k_cache5)[name = tensor("cast_535")]; + tensor var_1580_cast_fp16 = mul(x = k_cache5_to_fp16, y = var_1579_cast_fp16)[name = tensor("op_1580_cast_fp16")]; + tensor var_1585 = const()[name = tensor("op_1585"), val = tensor([1, 1, 16, 64])]; + tensor var_1586_cast_fp16 = reshape(shape = var_1585, x = out_23_cast_fp16)[name = tensor("op_1586_cast_fp16")]; + tensor var_1587_cast_fp16 = mul(x = var_1586_cast_fp16, y = onehot_23_cast_fp16)[name = tensor("op_1587_cast_fp16")]; + tensor new_k_cache_11_cast_fp16 = add(x = var_1580_cast_fp16, y = var_1587_cast_fp16)[name = tensor("new_k_cache_11_cast_fp16")]; + tensor new_k_cache_11_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_11_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache5_to_fp16_dtype_0 = const()[name = tensor("v_cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache5_to_fp16 = cast(dtype = v_cache5_to_fp16_dtype_0, x = v_cache5)[name = tensor("cast_533")]; + tensor var_1593_cast_fp16 = mul(x = v_cache5_to_fp16, y = var_1579_cast_fp16)[name = tensor("op_1593_cast_fp16")]; + tensor var_1598 = const()[name = tensor("op_1598"), val = tensor([1, 1, 16, 64])]; + tensor var_1599_cast_fp16 = reshape(shape = var_1598, x = v_11_cast_fp16)[name = tensor("op_1599_cast_fp16")]; + tensor var_1600_cast_fp16 = mul(x = var_1599_cast_fp16, y = onehot_23_cast_fp16)[name = tensor("op_1600_cast_fp16")]; + tensor new_v_cache_11_cast_fp16 = add(x = var_1593_cast_fp16, y = var_1600_cast_fp16)[name = tensor("new_v_cache_11_cast_fp16")]; + tensor new_v_cache_11_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_11_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_5_perm_0 = const()[name = tensor("transpose_5_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_1613 = const()[name = tensor("op_1613"), val = tensor([0, 2, 1, 3])]; + tensor var_1619 = const()[name = tensor("op_1619"), val = tensor([1, 16, 1, 64])]; + tensor q4_11_cast_fp16 = reshape(shape = var_1619, x = out_21_cast_fp16)[name = tensor("q4_11_cast_fp16")]; + tensor var_1624_transpose_x_0 = const()[name = tensor("op_1624_transpose_x_0"), val = tensor(false)]; + tensor var_1624_transpose_y_0 = const()[name = tensor("op_1624_transpose_y_0"), val = tensor(false)]; + tensor transpose_5_cast_fp16 = transpose(perm = transpose_5_perm_0, x = new_k_cache_11_cast_fp16)[name = tensor("transpose_80")]; + tensor var_1624_cast_fp16 = matmul(transpose_x = var_1624_transpose_x_0, transpose_y = var_1624_transpose_y_0, x = q4_11_cast_fp16, y = transpose_5_cast_fp16)[name = tensor("op_1624_cast_fp16")]; + tensor var_1625_to_fp16 = const()[name = tensor("op_1625_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_21_cast_fp16 = mul(x = var_1624_cast_fp16, y = var_1625_to_fp16)[name = tensor("scores_21_cast_fp16")]; + tensor var_1629 = const()[name = tensor("op_1629"), val = tensor([1, 1])]; + tensor var_1630_cast_fp16 = reshape(shape = var_1629, x = position5_to_fp16)[name = tensor("op_1630_cast_fp16")]; + tensor var_1631_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_1630_cast_fp16)[name = tensor("op_1631_cast_fp16")]; + tensor mask_11_to_fp16_dtype_0 = const()[name = tensor("mask_11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_1638_to_fp16 = const()[name = tensor("op_1638_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1631_cast_fp16_to_fp16 = cast(dtype = mask_11_to_fp16_dtype_0, x = var_1631_cast_fp16)[name = tensor("cast_531")]; + tensor var_1639_cast_fp16 = sub(x = var_1631_cast_fp16_to_fp16, y = var_1638_to_fp16)[name = tensor("op_1639_cast_fp16")]; + tensor var_1644 = const()[name = tensor("op_1644"), val = tensor([1, 1, 1, 512])]; + tensor var_1645_cast_fp16 = reshape(shape = var_1644, x = var_1639_cast_fp16)[name = tensor("op_1645_cast_fp16")]; + tensor var_1646_to_fp16 = const()[name = tensor("op_1646_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_1647_cast_fp16 = mul(x = var_1645_cast_fp16, y = var_1646_to_fp16)[name = tensor("op_1647_cast_fp16")]; + tensor scores_23_cast_fp16 = add(x = scores_21_cast_fp16, y = var_1647_cast_fp16)[name = tensor("scores_23_cast_fp16")]; + tensor var_1650 = const()[name = tensor("op_1650"), val = tensor(-1)]; + tensor weights_11_cast_fp16 = softmax(axis = var_1650, x = scores_23_cast_fp16)[name = tensor("weights_11_cast_fp16")]; + tensor attn_11_transpose_x_0 = const()[name = tensor("attn_11_transpose_x_0"), val = tensor(false)]; + tensor attn_11_transpose_y_0 = const()[name = tensor("attn_11_transpose_y_0"), val = tensor(false)]; + tensor values_11_cast_fp16 = transpose(perm = var_1613, x = new_v_cache_11_cast_fp16)[name = tensor("transpose_79")]; + tensor attn_11_cast_fp16 = matmul(transpose_x = attn_11_transpose_x_0, transpose_y = attn_11_transpose_y_0, x = weights_11_cast_fp16, y = values_11_cast_fp16)[name = tensor("attn_11_cast_fp16")]; + tensor var_1658 = const()[name = tensor("op_1658"), val = tensor([0, 2, 1, 3])]; + tensor var_1663 = const()[name = tensor("op_1663"), val = tensor([1, 1, 1024])]; + tensor var_1659_cast_fp16 = transpose(perm = var_1658, x = attn_11_cast_fp16)[name = tensor("transpose_78")]; + tensor input_67_cast_fp16 = reshape(shape = var_1663, x = var_1659_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor attn5_out_proj_weight_to_fp16 = const()[name = tensor("attn5_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132251840)))]; + tensor linear_22_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn5_out_proj_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("linear_22_cast_fp16")]; + tensor var_1669_to_fp16 = const()[name = tensor("op_1669_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1670_cast_fp16 = add(x = position5_to_fp16, y = var_1669_to_fp16)[name = tensor("op_1670_cast_fp16")]; + tensor var_1670_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1670_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_69_cast_fp16 = add(x = input_63_cast_fp16, y = linear_22_cast_fp16)[name = tensor("input_69_cast_fp16")]; + tensor input_71_axes_0 = const()[name = tensor("input_71_axes_0"), val = tensor([-1])]; + tensor norm5_2_weight_to_fp16 = const()[name = tensor("norm5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134349056)))]; + tensor norm5_2_bias_to_fp16 = const()[name = tensor("norm5_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134351168)))]; + tensor var_1674_to_fp16 = const()[name = tensor("op_1674_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_71_cast_fp16 = layer_norm(axes = input_71_axes_0, beta = norm5_2_bias_to_fp16, epsilon = var_1674_to_fp16, gamma = norm5_2_weight_to_fp16, x = input_69_cast_fp16)[name = tensor("input_71_cast_fp16")]; + tensor linear5_1_weight_to_fp16 = const()[name = tensor("linear5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134353280)))]; + tensor linear_23_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear5_1_weight_to_fp16, x = input_71_cast_fp16)[name = tensor("linear_23_cast_fp16")]; + tensor input_73_mode_0 = const()[name = tensor("input_73_mode_0"), val = tensor("EXACT")]; + tensor input_73_cast_fp16 = gelu(mode = input_73_mode_0, x = linear_23_cast_fp16)[name = tensor("input_73_cast_fp16")]; + tensor linear5_2_weight_to_fp16 = const()[name = tensor("linear5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142741952)))]; + tensor linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear5_2_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("linear_24_cast_fp16")]; + tensor input_75_cast_fp16 = add(x = input_69_cast_fp16, y = linear_24_cast_fp16)[name = tensor("input_75_cast_fp16")]; + tensor input_77_axes_0 = const()[name = tensor("input_77_axes_0"), val = tensor([-1])]; + tensor norm6_1_weight_to_fp16 = const()[name = tensor("norm6_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151130624)))]; + tensor norm6_1_bias_to_fp16 = const()[name = tensor("norm6_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151132736)))]; + tensor var_1691_to_fp16 = const()[name = tensor("op_1691_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_77_cast_fp16 = layer_norm(axes = input_77_axes_0, beta = norm6_1_bias_to_fp16, epsilon = var_1691_to_fp16, gamma = norm6_1_weight_to_fp16, x = input_75_cast_fp16)[name = tensor("input_77_cast_fp16")]; + tensor attn6_in_proj_weight_to_fp16 = const()[name = tensor("attn6_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151134848)))]; + tensor linear_25_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn6_in_proj_weight_to_fp16, x = input_77_cast_fp16)[name = tensor("linear_25_cast_fp16")]; + tensor var_1704 = const()[name = tensor("op_1704"), val = tensor([1, 3, 16, 64])]; + tensor qkv_13_cast_fp16 = reshape(shape = var_1704, x = linear_25_cast_fp16)[name = tensor("qkv_13_cast_fp16")]; + tensor x_25_begin_0 = const()[name = tensor("x_25_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_25_end_0 = const()[name = tensor("x_25_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_25_end_mask_0 = const()[name = tensor("x_25_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_25_squeeze_mask_0 = const()[name = tensor("x_25_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_25_cast_fp16 = slice_by_index(begin = x_25_begin_0, end = x_25_end_0, end_mask = x_25_end_mask_0, squeeze_mask = x_25_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("x_25_cast_fp16")]; + tensor x_27_begin_0 = const()[name = tensor("x_27_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_27_end_0 = const()[name = tensor("x_27_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_27_end_mask_0 = const()[name = tensor("x_27_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_27_squeeze_mask_0 = const()[name = tensor("x_27_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_27_cast_fp16 = slice_by_index(begin = x_27_begin_0, end = x_27_end_0, end_mask = x_27_end_mask_0, squeeze_mask = x_27_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("x_27_cast_fp16")]; + tensor v_13_begin_0 = const()[name = tensor("v_13_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_13_end_0 = const()[name = tensor("v_13_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_13_end_mask_0 = const()[name = tensor("v_13_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_13_squeeze_mask_0 = const()[name = tensor("v_13_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_13_cast_fp16 = slice_by_index(begin = v_13_begin_0, end = v_13_end_0, end_mask = v_13_end_mask_0, squeeze_mask = v_13_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("v_13_cast_fp16")]; + tensor var_1733 = const()[name = tensor("op_1733"), val = tensor([1, 1, 1])]; + tensor position6_to_fp16_dtype_0 = const()[name = tensor("position6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position6_to_fp16 = cast(dtype = position6_to_fp16_dtype_0, x = position6)[name = tensor("cast_529")]; + tensor var_1734_cast_fp16 = reshape(shape = var_1733, x = position6_to_fp16)[name = tensor("op_1734_cast_fp16")]; + tensor angles_25_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_1734_cast_fp16)[name = tensor("angles_25_cast_fp16")]; + tensor rotr_25_cast_fp16 = cos(x = angles_25_cast_fp16)[name = tensor("rotr_25_cast_fp16")]; + tensor roti_25_cast_fp16 = sin(x = angles_25_cast_fp16)[name = tensor("roti_25_cast_fp16")]; + tensor var_1742 = const()[name = tensor("op_1742"), val = tensor([1, 16, 32, 2])]; + tensor pairs_25_cast_fp16 = reshape(shape = var_1742, x = x_25_cast_fp16)[name = tensor("pairs_25_cast_fp16")]; + tensor xr_25_begin_0 = const()[name = tensor("xr_25_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_25_end_0 = const()[name = tensor("xr_25_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_25_end_mask_0 = const()[name = tensor("xr_25_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_25_squeeze_mask_0 = const()[name = tensor("xr_25_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_25_cast_fp16 = slice_by_index(begin = xr_25_begin_0, end = xr_25_end_0, end_mask = xr_25_end_mask_0, squeeze_mask = xr_25_squeeze_mask_0, x = pairs_25_cast_fp16)[name = tensor("xr_25_cast_fp16")]; + tensor xi_25_begin_0 = const()[name = tensor("xi_25_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_25_end_0 = const()[name = tensor("xi_25_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_25_end_mask_0 = const()[name = tensor("xi_25_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_25_squeeze_mask_0 = const()[name = tensor("xi_25_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_25_cast_fp16 = slice_by_index(begin = xi_25_begin_0, end = xi_25_end_0, end_mask = xi_25_end_mask_0, squeeze_mask = xi_25_squeeze_mask_0, x = pairs_25_cast_fp16)[name = tensor("xi_25_cast_fp16")]; + tensor var_1750_cast_fp16 = mul(x = xr_25_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_1750_cast_fp16")]; + tensor var_1751_cast_fp16 = mul(x = xi_25_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_1751_cast_fp16")]; + tensor xor__25_cast_fp16 = sub(x = var_1750_cast_fp16, y = var_1751_cast_fp16)[name = tensor("xor__25_cast_fp16")]; + tensor var_1754_cast_fp16 = mul(x = xr_25_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_1754_cast_fp16")]; + tensor var_1755_cast_fp16 = mul(x = xi_25_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_1755_cast_fp16")]; + tensor xoi_25_cast_fp16 = add(x = var_1754_cast_fp16, y = var_1755_cast_fp16)[name = tensor("xoi_25_cast_fp16")]; + tensor out_25_axis_0 = const()[name = tensor("out_25_axis_0"), val = tensor(-1)]; + tensor out_25_cast_fp16 = stack(axis = out_25_axis_0, values = (xor__25_cast_fp16, xoi_25_cast_fp16))[name = tensor("out_25_cast_fp16")]; + tensor var_1778 = const()[name = tensor("op_1778"), val = tensor([1, 16, 32, 2])]; + tensor pairs_27_cast_fp16 = reshape(shape = var_1778, x = x_27_cast_fp16)[name = tensor("pairs_27_cast_fp16")]; + tensor xr_27_begin_0 = const()[name = tensor("xr_27_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_27_end_0 = const()[name = tensor("xr_27_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_27_end_mask_0 = const()[name = tensor("xr_27_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_27_squeeze_mask_0 = const()[name = tensor("xr_27_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_27_cast_fp16 = slice_by_index(begin = xr_27_begin_0, end = xr_27_end_0, end_mask = xr_27_end_mask_0, squeeze_mask = xr_27_squeeze_mask_0, x = pairs_27_cast_fp16)[name = tensor("xr_27_cast_fp16")]; + tensor xi_27_begin_0 = const()[name = tensor("xi_27_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_27_end_0 = const()[name = tensor("xi_27_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_27_end_mask_0 = const()[name = tensor("xi_27_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_27_squeeze_mask_0 = const()[name = tensor("xi_27_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_27_cast_fp16 = slice_by_index(begin = xi_27_begin_0, end = xi_27_end_0, end_mask = xi_27_end_mask_0, squeeze_mask = xi_27_squeeze_mask_0, x = pairs_27_cast_fp16)[name = tensor("xi_27_cast_fp16")]; + tensor var_1786_cast_fp16 = mul(x = xr_27_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_1786_cast_fp16")]; + tensor var_1787_cast_fp16 = mul(x = xi_27_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_1787_cast_fp16")]; + tensor xor__27_cast_fp16 = sub(x = var_1786_cast_fp16, y = var_1787_cast_fp16)[name = tensor("xor__27_cast_fp16")]; + tensor var_1790_cast_fp16 = mul(x = xr_27_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_1790_cast_fp16")]; + tensor var_1791_cast_fp16 = mul(x = xi_27_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_1791_cast_fp16")]; + tensor xoi_27_cast_fp16 = add(x = var_1790_cast_fp16, y = var_1791_cast_fp16)[name = tensor("xoi_27_cast_fp16")]; + tensor out_27_axis_0 = const()[name = tensor("out_27_axis_0"), val = tensor(-1)]; + tensor out_27_cast_fp16 = stack(axis = out_27_axis_0, values = (xor__27_cast_fp16, xoi_27_cast_fp16))[name = tensor("out_27_cast_fp16")]; + tensor _inversed_1803_y_0_to_fp16 = const()[name = tensor("_inversed_1803_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1803_cast_fp16 = mul(x = position6_to_fp16, y = _inversed_1803_y_0_to_fp16)[name = tensor("_inversed_1803_cast_fp16")]; + tensor var_1804_cast_fp16 = floor(x = _inversed_1803_cast_fp16)[name = tensor("op_1804_cast_fp16")]; + tensor var_1805_to_fp16 = const()[name = tensor("op_1805_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1806_cast_fp16 = mul(x = var_1804_cast_fp16, y = var_1805_to_fp16)[name = tensor("op_1806_cast_fp16")]; + tensor write_idx_13_cast_fp16 = sub(x = position6_to_fp16, y = var_1806_cast_fp16)[name = tensor("write_idx_13_cast_fp16")]; + tensor var_1809_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_13_cast_fp16)[name = tensor("op_1809_cast_fp16")]; + tensor var_1819 = const()[name = tensor("op_1819"), val = tensor([1, 512, 1, 1])]; + tensor onehot_25_to_fp16_dtype_0 = const()[name = tensor("onehot_25_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_1809_cast_fp16_to_fp16 = cast(dtype = onehot_25_to_fp16_dtype_0, x = var_1809_cast_fp16)[name = tensor("cast_528")]; + tensor onehot_27_cast_fp16 = reshape(shape = var_1819, x = var_1809_cast_fp16_to_fp16)[name = tensor("onehot_27_cast_fp16")]; + tensor var_1821_to_fp16 = const()[name = tensor("op_1821_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1823_cast_fp16 = sub(x = var_1821_to_fp16, y = onehot_27_cast_fp16)[name = tensor("op_1823_cast_fp16")]; + tensor k_cache6_to_fp16_dtype_0 = const()[name = tensor("k_cache6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache6_to_fp16 = cast(dtype = k_cache6_to_fp16_dtype_0, x = k_cache6)[name = tensor("cast_527")]; + tensor var_1824_cast_fp16 = mul(x = k_cache6_to_fp16, y = var_1823_cast_fp16)[name = tensor("op_1824_cast_fp16")]; + tensor var_1829 = const()[name = tensor("op_1829"), val = tensor([1, 1, 16, 64])]; + tensor var_1830_cast_fp16 = reshape(shape = var_1829, x = out_27_cast_fp16)[name = tensor("op_1830_cast_fp16")]; + tensor var_1831_cast_fp16 = mul(x = var_1830_cast_fp16, y = onehot_27_cast_fp16)[name = tensor("op_1831_cast_fp16")]; + tensor new_k_cache_13_cast_fp16 = add(x = var_1824_cast_fp16, y = var_1831_cast_fp16)[name = tensor("new_k_cache_13_cast_fp16")]; + tensor new_k_cache_13_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_13_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache6_to_fp16_dtype_0 = const()[name = tensor("v_cache6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache6_to_fp16 = cast(dtype = v_cache6_to_fp16_dtype_0, x = v_cache6)[name = tensor("cast_525")]; + tensor var_1837_cast_fp16 = mul(x = v_cache6_to_fp16, y = var_1823_cast_fp16)[name = tensor("op_1837_cast_fp16")]; + tensor var_1842 = const()[name = tensor("op_1842"), val = tensor([1, 1, 16, 64])]; + tensor var_1843_cast_fp16 = reshape(shape = var_1842, x = v_13_cast_fp16)[name = tensor("op_1843_cast_fp16")]; + tensor var_1844_cast_fp16 = mul(x = var_1843_cast_fp16, y = onehot_27_cast_fp16)[name = tensor("op_1844_cast_fp16")]; + tensor new_v_cache_13_cast_fp16 = add(x = var_1837_cast_fp16, y = var_1844_cast_fp16)[name = tensor("new_v_cache_13_cast_fp16")]; + tensor new_v_cache_13_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_13_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_6_perm_0 = const()[name = tensor("transpose_6_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_1857 = const()[name = tensor("op_1857"), val = tensor([0, 2, 1, 3])]; + tensor var_1863 = const()[name = tensor("op_1863"), val = tensor([1, 16, 1, 64])]; + tensor q4_13_cast_fp16 = reshape(shape = var_1863, x = out_25_cast_fp16)[name = tensor("q4_13_cast_fp16")]; + tensor var_1868_transpose_x_0 = const()[name = tensor("op_1868_transpose_x_0"), val = tensor(false)]; + tensor var_1868_transpose_y_0 = const()[name = tensor("op_1868_transpose_y_0"), val = tensor(false)]; + tensor transpose_6_cast_fp16 = transpose(perm = transpose_6_perm_0, x = new_k_cache_13_cast_fp16)[name = tensor("transpose_77")]; + tensor var_1868_cast_fp16 = matmul(transpose_x = var_1868_transpose_x_0, transpose_y = var_1868_transpose_y_0, x = q4_13_cast_fp16, y = transpose_6_cast_fp16)[name = tensor("op_1868_cast_fp16")]; + tensor var_1869_to_fp16 = const()[name = tensor("op_1869_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_25_cast_fp16 = mul(x = var_1868_cast_fp16, y = var_1869_to_fp16)[name = tensor("scores_25_cast_fp16")]; + tensor var_1873 = const()[name = tensor("op_1873"), val = tensor([1, 1])]; + tensor var_1874_cast_fp16 = reshape(shape = var_1873, x = position6_to_fp16)[name = tensor("op_1874_cast_fp16")]; + tensor var_1875_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_1874_cast_fp16)[name = tensor("op_1875_cast_fp16")]; + tensor mask_13_to_fp16_dtype_0 = const()[name = tensor("mask_13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_1882_to_fp16 = const()[name = tensor("op_1882_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1875_cast_fp16_to_fp16 = cast(dtype = mask_13_to_fp16_dtype_0, x = var_1875_cast_fp16)[name = tensor("cast_523")]; + tensor var_1883_cast_fp16 = sub(x = var_1875_cast_fp16_to_fp16, y = var_1882_to_fp16)[name = tensor("op_1883_cast_fp16")]; + tensor var_1888 = const()[name = tensor("op_1888"), val = tensor([1, 1, 1, 512])]; + tensor var_1889_cast_fp16 = reshape(shape = var_1888, x = var_1883_cast_fp16)[name = tensor("op_1889_cast_fp16")]; + tensor var_1890_to_fp16 = const()[name = tensor("op_1890_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_1891_cast_fp16 = mul(x = var_1889_cast_fp16, y = var_1890_to_fp16)[name = tensor("op_1891_cast_fp16")]; + tensor scores_27_cast_fp16 = add(x = scores_25_cast_fp16, y = var_1891_cast_fp16)[name = tensor("scores_27_cast_fp16")]; + tensor var_1894 = const()[name = tensor("op_1894"), val = tensor(-1)]; + tensor weights_13_cast_fp16 = softmax(axis = var_1894, x = scores_27_cast_fp16)[name = tensor("weights_13_cast_fp16")]; + tensor attn_13_transpose_x_0 = const()[name = tensor("attn_13_transpose_x_0"), val = tensor(false)]; + tensor attn_13_transpose_y_0 = const()[name = tensor("attn_13_transpose_y_0"), val = tensor(false)]; + tensor values_13_cast_fp16 = transpose(perm = var_1857, x = new_v_cache_13_cast_fp16)[name = tensor("transpose_76")]; + tensor attn_13_cast_fp16 = matmul(transpose_x = attn_13_transpose_x_0, transpose_y = attn_13_transpose_y_0, x = weights_13_cast_fp16, y = values_13_cast_fp16)[name = tensor("attn_13_cast_fp16")]; + tensor var_1902 = const()[name = tensor("op_1902"), val = tensor([0, 2, 1, 3])]; + tensor var_1907 = const()[name = tensor("op_1907"), val = tensor([1, 1, 1024])]; + tensor var_1903_cast_fp16 = transpose(perm = var_1902, x = attn_13_cast_fp16)[name = tensor("transpose_75")]; + tensor input_79_cast_fp16 = reshape(shape = var_1907, x = var_1903_cast_fp16)[name = tensor("input_79_cast_fp16")]; + tensor attn6_out_proj_weight_to_fp16 = const()[name = tensor("attn6_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(157426368)))]; + tensor linear_26_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn6_out_proj_weight_to_fp16, x = input_79_cast_fp16)[name = tensor("linear_26_cast_fp16")]; + tensor var_1913_to_fp16 = const()[name = tensor("op_1913_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1914_cast_fp16 = add(x = position6_to_fp16, y = var_1913_to_fp16)[name = tensor("op_1914_cast_fp16")]; + tensor var_1914_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1914_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_81_cast_fp16 = add(x = input_75_cast_fp16, y = linear_26_cast_fp16)[name = tensor("input_81_cast_fp16")]; + tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([-1])]; + tensor norm6_2_weight_to_fp16 = const()[name = tensor("norm6_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159523584)))]; + tensor norm6_2_bias_to_fp16 = const()[name = tensor("norm6_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159525696)))]; + tensor var_1918_to_fp16 = const()[name = tensor("op_1918_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_83_cast_fp16 = layer_norm(axes = input_83_axes_0, beta = norm6_2_bias_to_fp16, epsilon = var_1918_to_fp16, gamma = norm6_2_weight_to_fp16, x = input_81_cast_fp16)[name = tensor("input_83_cast_fp16")]; + tensor linear6_1_weight_to_fp16 = const()[name = tensor("linear6_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159527808)))]; + tensor linear_27_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear6_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_27_cast_fp16")]; + tensor input_85_mode_0 = const()[name = tensor("input_85_mode_0"), val = tensor("EXACT")]; + tensor input_85_cast_fp16 = gelu(mode = input_85_mode_0, x = linear_27_cast_fp16)[name = tensor("input_85_cast_fp16")]; + tensor linear6_2_weight_to_fp16 = const()[name = tensor("linear6_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(167916480)))]; + tensor linear_28_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear6_2_weight_to_fp16, x = input_85_cast_fp16)[name = tensor("linear_28_cast_fp16")]; + tensor input_87_cast_fp16 = add(x = input_81_cast_fp16, y = linear_28_cast_fp16)[name = tensor("input_87_cast_fp16")]; + tensor input_89_axes_0 = const()[name = tensor("input_89_axes_0"), val = tensor([-1])]; + tensor norm7_1_weight_to_fp16 = const()[name = tensor("norm7_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176305152)))]; + tensor norm7_1_bias_to_fp16 = const()[name = tensor("norm7_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176307264)))]; + tensor var_1935_to_fp16 = const()[name = tensor("op_1935_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_89_cast_fp16 = layer_norm(axes = input_89_axes_0, beta = norm7_1_bias_to_fp16, epsilon = var_1935_to_fp16, gamma = norm7_1_weight_to_fp16, x = input_87_cast_fp16)[name = tensor("input_89_cast_fp16")]; + tensor attn7_in_proj_weight_to_fp16 = const()[name = tensor("attn7_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176309376)))]; + tensor linear_29_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn7_in_proj_weight_to_fp16, x = input_89_cast_fp16)[name = tensor("linear_29_cast_fp16")]; + tensor var_1948 = const()[name = tensor("op_1948"), val = tensor([1, 3, 16, 64])]; + tensor qkv_15_cast_fp16 = reshape(shape = var_1948, x = linear_29_cast_fp16)[name = tensor("qkv_15_cast_fp16")]; + tensor x_29_begin_0 = const()[name = tensor("x_29_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_29_end_0 = const()[name = tensor("x_29_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_29_end_mask_0 = const()[name = tensor("x_29_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_29_squeeze_mask_0 = const()[name = tensor("x_29_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_29_cast_fp16 = slice_by_index(begin = x_29_begin_0, end = x_29_end_0, end_mask = x_29_end_mask_0, squeeze_mask = x_29_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("x_29_cast_fp16")]; + tensor x_31_begin_0 = const()[name = tensor("x_31_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_31_end_0 = const()[name = tensor("x_31_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_31_end_mask_0 = const()[name = tensor("x_31_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_31_squeeze_mask_0 = const()[name = tensor("x_31_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_31_cast_fp16 = slice_by_index(begin = x_31_begin_0, end = x_31_end_0, end_mask = x_31_end_mask_0, squeeze_mask = x_31_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("x_31_cast_fp16")]; + tensor v_15_begin_0 = const()[name = tensor("v_15_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_15_end_0 = const()[name = tensor("v_15_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_15_end_mask_0 = const()[name = tensor("v_15_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_15_squeeze_mask_0 = const()[name = tensor("v_15_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_15_cast_fp16 = slice_by_index(begin = v_15_begin_0, end = v_15_end_0, end_mask = v_15_end_mask_0, squeeze_mask = v_15_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("v_15_cast_fp16")]; + tensor var_1977 = const()[name = tensor("op_1977"), val = tensor([1, 1, 1])]; + tensor position7_to_fp16_dtype_0 = const()[name = tensor("position7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position7_to_fp16 = cast(dtype = position7_to_fp16_dtype_0, x = position7)[name = tensor("cast_521")]; + tensor var_1978_cast_fp16 = reshape(shape = var_1977, x = position7_to_fp16)[name = tensor("op_1978_cast_fp16")]; + tensor angles_29_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_1978_cast_fp16)[name = tensor("angles_29_cast_fp16")]; + tensor rotr_29_cast_fp16 = cos(x = angles_29_cast_fp16)[name = tensor("rotr_29_cast_fp16")]; + tensor roti_29_cast_fp16 = sin(x = angles_29_cast_fp16)[name = tensor("roti_29_cast_fp16")]; + tensor var_1986 = const()[name = tensor("op_1986"), val = tensor([1, 16, 32, 2])]; + tensor pairs_29_cast_fp16 = reshape(shape = var_1986, x = x_29_cast_fp16)[name = tensor("pairs_29_cast_fp16")]; + tensor xr_29_begin_0 = const()[name = tensor("xr_29_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_29_end_0 = const()[name = tensor("xr_29_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_29_end_mask_0 = const()[name = tensor("xr_29_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_29_squeeze_mask_0 = const()[name = tensor("xr_29_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_29_cast_fp16 = slice_by_index(begin = xr_29_begin_0, end = xr_29_end_0, end_mask = xr_29_end_mask_0, squeeze_mask = xr_29_squeeze_mask_0, x = pairs_29_cast_fp16)[name = tensor("xr_29_cast_fp16")]; + tensor xi_29_begin_0 = const()[name = tensor("xi_29_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_29_end_0 = const()[name = tensor("xi_29_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_29_end_mask_0 = const()[name = tensor("xi_29_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_29_squeeze_mask_0 = const()[name = tensor("xi_29_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_29_cast_fp16 = slice_by_index(begin = xi_29_begin_0, end = xi_29_end_0, end_mask = xi_29_end_mask_0, squeeze_mask = xi_29_squeeze_mask_0, x = pairs_29_cast_fp16)[name = tensor("xi_29_cast_fp16")]; + tensor var_1994_cast_fp16 = mul(x = xr_29_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_1994_cast_fp16")]; + tensor var_1995_cast_fp16 = mul(x = xi_29_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_1995_cast_fp16")]; + tensor xor__29_cast_fp16 = sub(x = var_1994_cast_fp16, y = var_1995_cast_fp16)[name = tensor("xor__29_cast_fp16")]; + tensor var_1998_cast_fp16 = mul(x = xr_29_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_1998_cast_fp16")]; + tensor var_1999_cast_fp16 = mul(x = xi_29_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_1999_cast_fp16")]; + tensor xoi_29_cast_fp16 = add(x = var_1998_cast_fp16, y = var_1999_cast_fp16)[name = tensor("xoi_29_cast_fp16")]; + tensor out_29_axis_0 = const()[name = tensor("out_29_axis_0"), val = tensor(-1)]; + tensor out_29_cast_fp16 = stack(axis = out_29_axis_0, values = (xor__29_cast_fp16, xoi_29_cast_fp16))[name = tensor("out_29_cast_fp16")]; + tensor var_2022 = const()[name = tensor("op_2022"), val = tensor([1, 16, 32, 2])]; + tensor pairs_31_cast_fp16 = reshape(shape = var_2022, x = x_31_cast_fp16)[name = tensor("pairs_31_cast_fp16")]; + tensor xr_31_begin_0 = const()[name = tensor("xr_31_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_31_end_0 = const()[name = tensor("xr_31_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_31_end_mask_0 = const()[name = tensor("xr_31_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_31_squeeze_mask_0 = const()[name = tensor("xr_31_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_31_cast_fp16 = slice_by_index(begin = xr_31_begin_0, end = xr_31_end_0, end_mask = xr_31_end_mask_0, squeeze_mask = xr_31_squeeze_mask_0, x = pairs_31_cast_fp16)[name = tensor("xr_31_cast_fp16")]; + tensor xi_31_begin_0 = const()[name = tensor("xi_31_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_31_end_0 = const()[name = tensor("xi_31_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_31_end_mask_0 = const()[name = tensor("xi_31_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_31_squeeze_mask_0 = const()[name = tensor("xi_31_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_31_cast_fp16 = slice_by_index(begin = xi_31_begin_0, end = xi_31_end_0, end_mask = xi_31_end_mask_0, squeeze_mask = xi_31_squeeze_mask_0, x = pairs_31_cast_fp16)[name = tensor("xi_31_cast_fp16")]; + tensor var_2030_cast_fp16 = mul(x = xr_31_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_2030_cast_fp16")]; + tensor var_2031_cast_fp16 = mul(x = xi_31_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_2031_cast_fp16")]; + tensor xor__31_cast_fp16 = sub(x = var_2030_cast_fp16, y = var_2031_cast_fp16)[name = tensor("xor__31_cast_fp16")]; + tensor var_2034_cast_fp16 = mul(x = xr_31_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_2034_cast_fp16")]; + tensor var_2035_cast_fp16 = mul(x = xi_31_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_2035_cast_fp16")]; + tensor xoi_31_cast_fp16 = add(x = var_2034_cast_fp16, y = var_2035_cast_fp16)[name = tensor("xoi_31_cast_fp16")]; + tensor out_31_axis_0 = const()[name = tensor("out_31_axis_0"), val = tensor(-1)]; + tensor out_31_cast_fp16 = stack(axis = out_31_axis_0, values = (xor__31_cast_fp16, xoi_31_cast_fp16))[name = tensor("out_31_cast_fp16")]; + tensor _inversed_2047_y_0_to_fp16 = const()[name = tensor("_inversed_2047_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2047_cast_fp16 = mul(x = position7_to_fp16, y = _inversed_2047_y_0_to_fp16)[name = tensor("_inversed_2047_cast_fp16")]; + tensor var_2048_cast_fp16 = floor(x = _inversed_2047_cast_fp16)[name = tensor("op_2048_cast_fp16")]; + tensor var_2049_to_fp16 = const()[name = tensor("op_2049_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2050_cast_fp16 = mul(x = var_2048_cast_fp16, y = var_2049_to_fp16)[name = tensor("op_2050_cast_fp16")]; + tensor write_idx_15_cast_fp16 = sub(x = position7_to_fp16, y = var_2050_cast_fp16)[name = tensor("write_idx_15_cast_fp16")]; + tensor var_2053_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_15_cast_fp16)[name = tensor("op_2053_cast_fp16")]; + tensor var_2063 = const()[name = tensor("op_2063"), val = tensor([1, 512, 1, 1])]; + tensor onehot_29_to_fp16_dtype_0 = const()[name = tensor("onehot_29_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_2053_cast_fp16_to_fp16 = cast(dtype = onehot_29_to_fp16_dtype_0, x = var_2053_cast_fp16)[name = tensor("cast_520")]; + tensor onehot_31_cast_fp16 = reshape(shape = var_2063, x = var_2053_cast_fp16_to_fp16)[name = tensor("onehot_31_cast_fp16")]; + tensor var_2065_to_fp16 = const()[name = tensor("op_2065_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2067_cast_fp16 = sub(x = var_2065_to_fp16, y = onehot_31_cast_fp16)[name = tensor("op_2067_cast_fp16")]; + tensor k_cache7_to_fp16_dtype_0 = const()[name = tensor("k_cache7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache7_to_fp16 = cast(dtype = k_cache7_to_fp16_dtype_0, x = k_cache7)[name = tensor("cast_519")]; + tensor var_2068_cast_fp16 = mul(x = k_cache7_to_fp16, y = var_2067_cast_fp16)[name = tensor("op_2068_cast_fp16")]; + tensor var_2073 = const()[name = tensor("op_2073"), val = tensor([1, 1, 16, 64])]; + tensor var_2074_cast_fp16 = reshape(shape = var_2073, x = out_31_cast_fp16)[name = tensor("op_2074_cast_fp16")]; + tensor var_2075_cast_fp16 = mul(x = var_2074_cast_fp16, y = onehot_31_cast_fp16)[name = tensor("op_2075_cast_fp16")]; + tensor new_k_cache_15_cast_fp16 = add(x = var_2068_cast_fp16, y = var_2075_cast_fp16)[name = tensor("new_k_cache_15_cast_fp16")]; + tensor new_k_cache_15_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_15_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache7_to_fp16_dtype_0 = const()[name = tensor("v_cache7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache7_to_fp16 = cast(dtype = v_cache7_to_fp16_dtype_0, x = v_cache7)[name = tensor("cast_517")]; + tensor var_2081_cast_fp16 = mul(x = v_cache7_to_fp16, y = var_2067_cast_fp16)[name = tensor("op_2081_cast_fp16")]; + tensor var_2086 = const()[name = tensor("op_2086"), val = tensor([1, 1, 16, 64])]; + tensor var_2087_cast_fp16 = reshape(shape = var_2086, x = v_15_cast_fp16)[name = tensor("op_2087_cast_fp16")]; + tensor var_2088_cast_fp16 = mul(x = var_2087_cast_fp16, y = onehot_31_cast_fp16)[name = tensor("op_2088_cast_fp16")]; + tensor new_v_cache_15_cast_fp16 = add(x = var_2081_cast_fp16, y = var_2088_cast_fp16)[name = tensor("new_v_cache_15_cast_fp16")]; + tensor new_v_cache_15_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_15_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_7_perm_0 = const()[name = tensor("transpose_7_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_2101 = const()[name = tensor("op_2101"), val = tensor([0, 2, 1, 3])]; + tensor var_2107 = const()[name = tensor("op_2107"), val = tensor([1, 16, 1, 64])]; + tensor q4_15_cast_fp16 = reshape(shape = var_2107, x = out_29_cast_fp16)[name = tensor("q4_15_cast_fp16")]; + tensor var_2112_transpose_x_0 = const()[name = tensor("op_2112_transpose_x_0"), val = tensor(false)]; + tensor var_2112_transpose_y_0 = const()[name = tensor("op_2112_transpose_y_0"), val = tensor(false)]; + tensor transpose_7_cast_fp16 = transpose(perm = transpose_7_perm_0, x = new_k_cache_15_cast_fp16)[name = tensor("transpose_74")]; + tensor var_2112_cast_fp16 = matmul(transpose_x = var_2112_transpose_x_0, transpose_y = var_2112_transpose_y_0, x = q4_15_cast_fp16, y = transpose_7_cast_fp16)[name = tensor("op_2112_cast_fp16")]; + tensor var_2113_to_fp16 = const()[name = tensor("op_2113_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_29_cast_fp16 = mul(x = var_2112_cast_fp16, y = var_2113_to_fp16)[name = tensor("scores_29_cast_fp16")]; + tensor var_2117 = const()[name = tensor("op_2117"), val = tensor([1, 1])]; + tensor var_2118_cast_fp16 = reshape(shape = var_2117, x = position7_to_fp16)[name = tensor("op_2118_cast_fp16")]; + tensor var_2119_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_2118_cast_fp16)[name = tensor("op_2119_cast_fp16")]; + tensor mask_15_to_fp16_dtype_0 = const()[name = tensor("mask_15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_2126_to_fp16 = const()[name = tensor("op_2126_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2119_cast_fp16_to_fp16 = cast(dtype = mask_15_to_fp16_dtype_0, x = var_2119_cast_fp16)[name = tensor("cast_515")]; + tensor var_2127_cast_fp16 = sub(x = var_2119_cast_fp16_to_fp16, y = var_2126_to_fp16)[name = tensor("op_2127_cast_fp16")]; + tensor var_2132 = const()[name = tensor("op_2132"), val = tensor([1, 1, 1, 512])]; + tensor var_2133_cast_fp16 = reshape(shape = var_2132, x = var_2127_cast_fp16)[name = tensor("op_2133_cast_fp16")]; + tensor var_2134_to_fp16 = const()[name = tensor("op_2134_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_2135_cast_fp16 = mul(x = var_2133_cast_fp16, y = var_2134_to_fp16)[name = tensor("op_2135_cast_fp16")]; + tensor scores_31_cast_fp16 = add(x = scores_29_cast_fp16, y = var_2135_cast_fp16)[name = tensor("scores_31_cast_fp16")]; + tensor var_2138 = const()[name = tensor("op_2138"), val = tensor(-1)]; + tensor weights_15_cast_fp16 = softmax(axis = var_2138, x = scores_31_cast_fp16)[name = tensor("weights_15_cast_fp16")]; + tensor attn_15_transpose_x_0 = const()[name = tensor("attn_15_transpose_x_0"), val = tensor(false)]; + tensor attn_15_transpose_y_0 = const()[name = tensor("attn_15_transpose_y_0"), val = tensor(false)]; + tensor values_15_cast_fp16 = transpose(perm = var_2101, x = new_v_cache_15_cast_fp16)[name = tensor("transpose_73")]; + tensor attn_15_cast_fp16 = matmul(transpose_x = attn_15_transpose_x_0, transpose_y = attn_15_transpose_y_0, x = weights_15_cast_fp16, y = values_15_cast_fp16)[name = tensor("attn_15_cast_fp16")]; + tensor var_2146 = const()[name = tensor("op_2146"), val = tensor([0, 2, 1, 3])]; + tensor var_2151 = const()[name = tensor("op_2151"), val = tensor([1, 1, 1024])]; + tensor var_2147_cast_fp16 = transpose(perm = var_2146, x = attn_15_cast_fp16)[name = tensor("transpose_72")]; + tensor input_91_cast_fp16 = reshape(shape = var_2151, x = var_2147_cast_fp16)[name = tensor("input_91_cast_fp16")]; + tensor attn7_out_proj_weight_to_fp16 = const()[name = tensor("attn7_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(182600896)))]; + tensor linear_30_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn7_out_proj_weight_to_fp16, x = input_91_cast_fp16)[name = tensor("linear_30_cast_fp16")]; + tensor var_2157_to_fp16 = const()[name = tensor("op_2157_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2158_cast_fp16 = add(x = position7_to_fp16, y = var_2157_to_fp16)[name = tensor("op_2158_cast_fp16")]; + tensor var_2158_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2158_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_93_cast_fp16 = add(x = input_87_cast_fp16, y = linear_30_cast_fp16)[name = tensor("input_93_cast_fp16")]; + tensor input_95_axes_0 = const()[name = tensor("input_95_axes_0"), val = tensor([-1])]; + tensor norm7_2_weight_to_fp16 = const()[name = tensor("norm7_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184698112)))]; + tensor norm7_2_bias_to_fp16 = const()[name = tensor("norm7_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184700224)))]; + tensor var_2162_to_fp16 = const()[name = tensor("op_2162_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_95_cast_fp16 = layer_norm(axes = input_95_axes_0, beta = norm7_2_bias_to_fp16, epsilon = var_2162_to_fp16, gamma = norm7_2_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("input_95_cast_fp16")]; + tensor linear7_1_weight_to_fp16 = const()[name = tensor("linear7_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184702336)))]; + tensor linear_31_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear7_1_weight_to_fp16, x = input_95_cast_fp16)[name = tensor("linear_31_cast_fp16")]; + tensor input_97_mode_0 = const()[name = tensor("input_97_mode_0"), val = tensor("EXACT")]; + tensor input_97_cast_fp16 = gelu(mode = input_97_mode_0, x = linear_31_cast_fp16)[name = tensor("input_97_cast_fp16")]; + tensor linear7_2_weight_to_fp16 = const()[name = tensor("linear7_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(193091008)))]; + tensor linear_32_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear7_2_weight_to_fp16, x = input_97_cast_fp16)[name = tensor("linear_32_cast_fp16")]; + tensor input_99_cast_fp16 = add(x = input_93_cast_fp16, y = linear_32_cast_fp16)[name = tensor("input_99_cast_fp16")]; + tensor input_101_axes_0 = const()[name = tensor("input_101_axes_0"), val = tensor([-1])]; + tensor norm8_1_weight_to_fp16 = const()[name = tensor("norm8_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201479680)))]; + tensor norm8_1_bias_to_fp16 = const()[name = tensor("norm8_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201481792)))]; + tensor var_2179_to_fp16 = const()[name = tensor("op_2179_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_101_cast_fp16 = layer_norm(axes = input_101_axes_0, beta = norm8_1_bias_to_fp16, epsilon = var_2179_to_fp16, gamma = norm8_1_weight_to_fp16, x = input_99_cast_fp16)[name = tensor("input_101_cast_fp16")]; + tensor attn8_in_proj_weight_to_fp16 = const()[name = tensor("attn8_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201483904)))]; + tensor linear_33_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn8_in_proj_weight_to_fp16, x = input_101_cast_fp16)[name = tensor("linear_33_cast_fp16")]; + tensor var_2192 = const()[name = tensor("op_2192"), val = tensor([1, 3, 16, 64])]; + tensor qkv_17_cast_fp16 = reshape(shape = var_2192, x = linear_33_cast_fp16)[name = tensor("qkv_17_cast_fp16")]; + tensor x_33_begin_0 = const()[name = tensor("x_33_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_33_end_0 = const()[name = tensor("x_33_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_33_end_mask_0 = const()[name = tensor("x_33_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_33_squeeze_mask_0 = const()[name = tensor("x_33_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_33_cast_fp16 = slice_by_index(begin = x_33_begin_0, end = x_33_end_0, end_mask = x_33_end_mask_0, squeeze_mask = x_33_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("x_33_cast_fp16")]; + tensor x_35_begin_0 = const()[name = tensor("x_35_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_35_end_0 = const()[name = tensor("x_35_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_35_end_mask_0 = const()[name = tensor("x_35_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_35_squeeze_mask_0 = const()[name = tensor("x_35_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_35_cast_fp16 = slice_by_index(begin = x_35_begin_0, end = x_35_end_0, end_mask = x_35_end_mask_0, squeeze_mask = x_35_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("x_35_cast_fp16")]; + tensor v_17_begin_0 = const()[name = tensor("v_17_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_17_end_0 = const()[name = tensor("v_17_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_17_end_mask_0 = const()[name = tensor("v_17_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_17_squeeze_mask_0 = const()[name = tensor("v_17_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_17_cast_fp16 = slice_by_index(begin = v_17_begin_0, end = v_17_end_0, end_mask = v_17_end_mask_0, squeeze_mask = v_17_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("v_17_cast_fp16")]; + tensor var_2221 = const()[name = tensor("op_2221"), val = tensor([1, 1, 1])]; + tensor position8_to_fp16_dtype_0 = const()[name = tensor("position8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position8_to_fp16 = cast(dtype = position8_to_fp16_dtype_0, x = position8)[name = tensor("cast_513")]; + tensor var_2222_cast_fp16 = reshape(shape = var_2221, x = position8_to_fp16)[name = tensor("op_2222_cast_fp16")]; + tensor angles_33_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_2222_cast_fp16)[name = tensor("angles_33_cast_fp16")]; + tensor rotr_33_cast_fp16 = cos(x = angles_33_cast_fp16)[name = tensor("rotr_33_cast_fp16")]; + tensor roti_33_cast_fp16 = sin(x = angles_33_cast_fp16)[name = tensor("roti_33_cast_fp16")]; + tensor var_2230 = const()[name = tensor("op_2230"), val = tensor([1, 16, 32, 2])]; + tensor pairs_33_cast_fp16 = reshape(shape = var_2230, x = x_33_cast_fp16)[name = tensor("pairs_33_cast_fp16")]; + tensor xr_33_begin_0 = const()[name = tensor("xr_33_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_33_end_0 = const()[name = tensor("xr_33_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_33_end_mask_0 = const()[name = tensor("xr_33_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_33_squeeze_mask_0 = const()[name = tensor("xr_33_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_33_cast_fp16 = slice_by_index(begin = xr_33_begin_0, end = xr_33_end_0, end_mask = xr_33_end_mask_0, squeeze_mask = xr_33_squeeze_mask_0, x = pairs_33_cast_fp16)[name = tensor("xr_33_cast_fp16")]; + tensor xi_33_begin_0 = const()[name = tensor("xi_33_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_33_end_0 = const()[name = tensor("xi_33_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_33_end_mask_0 = const()[name = tensor("xi_33_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_33_squeeze_mask_0 = const()[name = tensor("xi_33_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_33_cast_fp16 = slice_by_index(begin = xi_33_begin_0, end = xi_33_end_0, end_mask = xi_33_end_mask_0, squeeze_mask = xi_33_squeeze_mask_0, x = pairs_33_cast_fp16)[name = tensor("xi_33_cast_fp16")]; + tensor var_2238_cast_fp16 = mul(x = xr_33_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_2238_cast_fp16")]; + tensor var_2239_cast_fp16 = mul(x = xi_33_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_2239_cast_fp16")]; + tensor xor__33_cast_fp16 = sub(x = var_2238_cast_fp16, y = var_2239_cast_fp16)[name = tensor("xor__33_cast_fp16")]; + tensor var_2242_cast_fp16 = mul(x = xr_33_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_2242_cast_fp16")]; + tensor var_2243_cast_fp16 = mul(x = xi_33_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_2243_cast_fp16")]; + tensor xoi_33_cast_fp16 = add(x = var_2242_cast_fp16, y = var_2243_cast_fp16)[name = tensor("xoi_33_cast_fp16")]; + tensor out_33_axis_0 = const()[name = tensor("out_33_axis_0"), val = tensor(-1)]; + tensor out_33_cast_fp16 = stack(axis = out_33_axis_0, values = (xor__33_cast_fp16, xoi_33_cast_fp16))[name = tensor("out_33_cast_fp16")]; + tensor var_2266 = const()[name = tensor("op_2266"), val = tensor([1, 16, 32, 2])]; + tensor pairs_35_cast_fp16 = reshape(shape = var_2266, x = x_35_cast_fp16)[name = tensor("pairs_35_cast_fp16")]; + tensor xr_35_begin_0 = const()[name = tensor("xr_35_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_35_end_0 = const()[name = tensor("xr_35_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_35_end_mask_0 = const()[name = tensor("xr_35_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_35_squeeze_mask_0 = const()[name = tensor("xr_35_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_35_cast_fp16 = slice_by_index(begin = xr_35_begin_0, end = xr_35_end_0, end_mask = xr_35_end_mask_0, squeeze_mask = xr_35_squeeze_mask_0, x = pairs_35_cast_fp16)[name = tensor("xr_35_cast_fp16")]; + tensor xi_35_begin_0 = const()[name = tensor("xi_35_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_35_end_0 = const()[name = tensor("xi_35_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_35_end_mask_0 = const()[name = tensor("xi_35_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_35_squeeze_mask_0 = const()[name = tensor("xi_35_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_35_cast_fp16 = slice_by_index(begin = xi_35_begin_0, end = xi_35_end_0, end_mask = xi_35_end_mask_0, squeeze_mask = xi_35_squeeze_mask_0, x = pairs_35_cast_fp16)[name = tensor("xi_35_cast_fp16")]; + tensor var_2274_cast_fp16 = mul(x = xr_35_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_2274_cast_fp16")]; + tensor var_2275_cast_fp16 = mul(x = xi_35_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_2275_cast_fp16")]; + tensor xor__35_cast_fp16 = sub(x = var_2274_cast_fp16, y = var_2275_cast_fp16)[name = tensor("xor__35_cast_fp16")]; + tensor var_2278_cast_fp16 = mul(x = xr_35_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_2278_cast_fp16")]; + tensor var_2279_cast_fp16 = mul(x = xi_35_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_2279_cast_fp16")]; + tensor xoi_35_cast_fp16 = add(x = var_2278_cast_fp16, y = var_2279_cast_fp16)[name = tensor("xoi_35_cast_fp16")]; + tensor out_35_axis_0 = const()[name = tensor("out_35_axis_0"), val = tensor(-1)]; + tensor out_35_cast_fp16 = stack(axis = out_35_axis_0, values = (xor__35_cast_fp16, xoi_35_cast_fp16))[name = tensor("out_35_cast_fp16")]; + tensor _inversed_2291_y_0_to_fp16 = const()[name = tensor("_inversed_2291_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2291_cast_fp16 = mul(x = position8_to_fp16, y = _inversed_2291_y_0_to_fp16)[name = tensor("_inversed_2291_cast_fp16")]; + tensor var_2292_cast_fp16 = floor(x = _inversed_2291_cast_fp16)[name = tensor("op_2292_cast_fp16")]; + tensor var_2293_to_fp16 = const()[name = tensor("op_2293_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2294_cast_fp16 = mul(x = var_2292_cast_fp16, y = var_2293_to_fp16)[name = tensor("op_2294_cast_fp16")]; + tensor write_idx_17_cast_fp16 = sub(x = position8_to_fp16, y = var_2294_cast_fp16)[name = tensor("write_idx_17_cast_fp16")]; + tensor var_2297_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_17_cast_fp16)[name = tensor("op_2297_cast_fp16")]; + tensor var_2307 = const()[name = tensor("op_2307"), val = tensor([1, 512, 1, 1])]; + tensor onehot_33_to_fp16_dtype_0 = const()[name = tensor("onehot_33_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_2297_cast_fp16_to_fp16 = cast(dtype = onehot_33_to_fp16_dtype_0, x = var_2297_cast_fp16)[name = tensor("cast_512")]; + tensor onehot_35_cast_fp16 = reshape(shape = var_2307, x = var_2297_cast_fp16_to_fp16)[name = tensor("onehot_35_cast_fp16")]; + tensor var_2309_to_fp16 = const()[name = tensor("op_2309_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2311_cast_fp16 = sub(x = var_2309_to_fp16, y = onehot_35_cast_fp16)[name = tensor("op_2311_cast_fp16")]; + tensor k_cache8_to_fp16_dtype_0 = const()[name = tensor("k_cache8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache8_to_fp16 = cast(dtype = k_cache8_to_fp16_dtype_0, x = k_cache8)[name = tensor("cast_511")]; + tensor var_2312_cast_fp16 = mul(x = k_cache8_to_fp16, y = var_2311_cast_fp16)[name = tensor("op_2312_cast_fp16")]; + tensor var_2317 = const()[name = tensor("op_2317"), val = tensor([1, 1, 16, 64])]; + tensor var_2318_cast_fp16 = reshape(shape = var_2317, x = out_35_cast_fp16)[name = tensor("op_2318_cast_fp16")]; + tensor var_2319_cast_fp16 = mul(x = var_2318_cast_fp16, y = onehot_35_cast_fp16)[name = tensor("op_2319_cast_fp16")]; + tensor new_k_cache_17_cast_fp16 = add(x = var_2312_cast_fp16, y = var_2319_cast_fp16)[name = tensor("new_k_cache_17_cast_fp16")]; + tensor new_k_cache_17_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_17_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache8_to_fp16_dtype_0 = const()[name = tensor("v_cache8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache8_to_fp16 = cast(dtype = v_cache8_to_fp16_dtype_0, x = v_cache8)[name = tensor("cast_509")]; + tensor var_2325_cast_fp16 = mul(x = v_cache8_to_fp16, y = var_2311_cast_fp16)[name = tensor("op_2325_cast_fp16")]; + tensor var_2330 = const()[name = tensor("op_2330"), val = tensor([1, 1, 16, 64])]; + tensor var_2331_cast_fp16 = reshape(shape = var_2330, x = v_17_cast_fp16)[name = tensor("op_2331_cast_fp16")]; + tensor var_2332_cast_fp16 = mul(x = var_2331_cast_fp16, y = onehot_35_cast_fp16)[name = tensor("op_2332_cast_fp16")]; + tensor new_v_cache_17_cast_fp16 = add(x = var_2325_cast_fp16, y = var_2332_cast_fp16)[name = tensor("new_v_cache_17_cast_fp16")]; + tensor new_v_cache_17_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_17_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_8_perm_0 = const()[name = tensor("transpose_8_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_2345 = const()[name = tensor("op_2345"), val = tensor([0, 2, 1, 3])]; + tensor var_2351 = const()[name = tensor("op_2351"), val = tensor([1, 16, 1, 64])]; + tensor q4_17_cast_fp16 = reshape(shape = var_2351, x = out_33_cast_fp16)[name = tensor("q4_17_cast_fp16")]; + tensor var_2356_transpose_x_0 = const()[name = tensor("op_2356_transpose_x_0"), val = tensor(false)]; + tensor var_2356_transpose_y_0 = const()[name = tensor("op_2356_transpose_y_0"), val = tensor(false)]; + tensor transpose_8_cast_fp16 = transpose(perm = transpose_8_perm_0, x = new_k_cache_17_cast_fp16)[name = tensor("transpose_71")]; + tensor var_2356_cast_fp16 = matmul(transpose_x = var_2356_transpose_x_0, transpose_y = var_2356_transpose_y_0, x = q4_17_cast_fp16, y = transpose_8_cast_fp16)[name = tensor("op_2356_cast_fp16")]; + tensor var_2357_to_fp16 = const()[name = tensor("op_2357_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_33_cast_fp16 = mul(x = var_2356_cast_fp16, y = var_2357_to_fp16)[name = tensor("scores_33_cast_fp16")]; + tensor var_2361 = const()[name = tensor("op_2361"), val = tensor([1, 1])]; + tensor var_2362_cast_fp16 = reshape(shape = var_2361, x = position8_to_fp16)[name = tensor("op_2362_cast_fp16")]; + tensor var_2363_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_2362_cast_fp16)[name = tensor("op_2363_cast_fp16")]; + tensor mask_17_to_fp16_dtype_0 = const()[name = tensor("mask_17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_2370_to_fp16 = const()[name = tensor("op_2370_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2363_cast_fp16_to_fp16 = cast(dtype = mask_17_to_fp16_dtype_0, x = var_2363_cast_fp16)[name = tensor("cast_507")]; + tensor var_2371_cast_fp16 = sub(x = var_2363_cast_fp16_to_fp16, y = var_2370_to_fp16)[name = tensor("op_2371_cast_fp16")]; + tensor var_2376 = const()[name = tensor("op_2376"), val = tensor([1, 1, 1, 512])]; + tensor var_2377_cast_fp16 = reshape(shape = var_2376, x = var_2371_cast_fp16)[name = tensor("op_2377_cast_fp16")]; + tensor var_2378_to_fp16 = const()[name = tensor("op_2378_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_2379_cast_fp16 = mul(x = var_2377_cast_fp16, y = var_2378_to_fp16)[name = tensor("op_2379_cast_fp16")]; + tensor scores_35_cast_fp16 = add(x = scores_33_cast_fp16, y = var_2379_cast_fp16)[name = tensor("scores_35_cast_fp16")]; + tensor var_2382 = const()[name = tensor("op_2382"), val = tensor(-1)]; + tensor weights_17_cast_fp16 = softmax(axis = var_2382, x = scores_35_cast_fp16)[name = tensor("weights_17_cast_fp16")]; + tensor attn_17_transpose_x_0 = const()[name = tensor("attn_17_transpose_x_0"), val = tensor(false)]; + tensor attn_17_transpose_y_0 = const()[name = tensor("attn_17_transpose_y_0"), val = tensor(false)]; + tensor values_17_cast_fp16 = transpose(perm = var_2345, x = new_v_cache_17_cast_fp16)[name = tensor("transpose_70")]; + tensor attn_17_cast_fp16 = matmul(transpose_x = attn_17_transpose_x_0, transpose_y = attn_17_transpose_y_0, x = weights_17_cast_fp16, y = values_17_cast_fp16)[name = tensor("attn_17_cast_fp16")]; + tensor var_2390 = const()[name = tensor("op_2390"), val = tensor([0, 2, 1, 3])]; + tensor var_2395 = const()[name = tensor("op_2395"), val = tensor([1, 1, 1024])]; + tensor var_2391_cast_fp16 = transpose(perm = var_2390, x = attn_17_cast_fp16)[name = tensor("transpose_69")]; + tensor input_103_cast_fp16 = reshape(shape = var_2395, x = var_2391_cast_fp16)[name = tensor("input_103_cast_fp16")]; + tensor attn8_out_proj_weight_to_fp16 = const()[name = tensor("attn8_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(207775424)))]; + tensor linear_34_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn8_out_proj_weight_to_fp16, x = input_103_cast_fp16)[name = tensor("linear_34_cast_fp16")]; + tensor var_2401_to_fp16 = const()[name = tensor("op_2401_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2402_cast_fp16 = add(x = position8_to_fp16, y = var_2401_to_fp16)[name = tensor("op_2402_cast_fp16")]; + tensor var_2402_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2402_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_105_cast_fp16 = add(x = input_99_cast_fp16, y = linear_34_cast_fp16)[name = tensor("input_105_cast_fp16")]; + tensor input_107_axes_0 = const()[name = tensor("input_107_axes_0"), val = tensor([-1])]; + tensor norm8_2_weight_to_fp16 = const()[name = tensor("norm8_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209872640)))]; + tensor norm8_2_bias_to_fp16 = const()[name = tensor("norm8_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209874752)))]; + tensor var_2406_to_fp16 = const()[name = tensor("op_2406_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_107_cast_fp16 = layer_norm(axes = input_107_axes_0, beta = norm8_2_bias_to_fp16, epsilon = var_2406_to_fp16, gamma = norm8_2_weight_to_fp16, x = input_105_cast_fp16)[name = tensor("input_107_cast_fp16")]; + tensor linear8_1_weight_to_fp16 = const()[name = tensor("linear8_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209876864)))]; + tensor linear_35_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear8_1_weight_to_fp16, x = input_107_cast_fp16)[name = tensor("linear_35_cast_fp16")]; + tensor input_109_mode_0 = const()[name = tensor("input_109_mode_0"), val = tensor("EXACT")]; + tensor input_109_cast_fp16 = gelu(mode = input_109_mode_0, x = linear_35_cast_fp16)[name = tensor("input_109_cast_fp16")]; + tensor linear8_2_weight_to_fp16 = const()[name = tensor("linear8_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(218265536)))]; + tensor linear_36_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear8_2_weight_to_fp16, x = input_109_cast_fp16)[name = tensor("linear_36_cast_fp16")]; + tensor input_111_cast_fp16 = add(x = input_105_cast_fp16, y = linear_36_cast_fp16)[name = tensor("input_111_cast_fp16")]; + tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([-1])]; + tensor norm9_1_weight_to_fp16 = const()[name = tensor("norm9_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226654208)))]; + tensor norm9_1_bias_to_fp16 = const()[name = tensor("norm9_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226656320)))]; + tensor var_2423_to_fp16 = const()[name = tensor("op_2423_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_113_cast_fp16 = layer_norm(axes = input_113_axes_0, beta = norm9_1_bias_to_fp16, epsilon = var_2423_to_fp16, gamma = norm9_1_weight_to_fp16, x = input_111_cast_fp16)[name = tensor("input_113_cast_fp16")]; + tensor attn9_in_proj_weight_to_fp16 = const()[name = tensor("attn9_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226658432)))]; + tensor linear_37_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn9_in_proj_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("linear_37_cast_fp16")]; + tensor var_2436 = const()[name = tensor("op_2436"), val = tensor([1, 3, 16, 64])]; + tensor qkv_19_cast_fp16 = reshape(shape = var_2436, x = linear_37_cast_fp16)[name = tensor("qkv_19_cast_fp16")]; + tensor x_37_begin_0 = const()[name = tensor("x_37_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_37_end_0 = const()[name = tensor("x_37_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_37_end_mask_0 = const()[name = tensor("x_37_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_37_squeeze_mask_0 = const()[name = tensor("x_37_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_37_cast_fp16 = slice_by_index(begin = x_37_begin_0, end = x_37_end_0, end_mask = x_37_end_mask_0, squeeze_mask = x_37_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("x_37_cast_fp16")]; + tensor x_39_begin_0 = const()[name = tensor("x_39_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_39_end_0 = const()[name = tensor("x_39_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_39_end_mask_0 = const()[name = tensor("x_39_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_39_squeeze_mask_0 = const()[name = tensor("x_39_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_39_cast_fp16 = slice_by_index(begin = x_39_begin_0, end = x_39_end_0, end_mask = x_39_end_mask_0, squeeze_mask = x_39_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("x_39_cast_fp16")]; + tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_19_cast_fp16 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("v_19_cast_fp16")]; + tensor var_2465 = const()[name = tensor("op_2465"), val = tensor([1, 1, 1])]; + tensor position9_to_fp16_dtype_0 = const()[name = tensor("position9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position9_to_fp16 = cast(dtype = position9_to_fp16_dtype_0, x = position9)[name = tensor("cast_505")]; + tensor var_2466_cast_fp16 = reshape(shape = var_2465, x = position9_to_fp16)[name = tensor("op_2466_cast_fp16")]; + tensor angles_37_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_2466_cast_fp16)[name = tensor("angles_37_cast_fp16")]; + tensor rotr_37_cast_fp16 = cos(x = angles_37_cast_fp16)[name = tensor("rotr_37_cast_fp16")]; + tensor roti_37_cast_fp16 = sin(x = angles_37_cast_fp16)[name = tensor("roti_37_cast_fp16")]; + tensor var_2474 = const()[name = tensor("op_2474"), val = tensor([1, 16, 32, 2])]; + tensor pairs_37_cast_fp16 = reshape(shape = var_2474, x = x_37_cast_fp16)[name = tensor("pairs_37_cast_fp16")]; + tensor xr_37_begin_0 = const()[name = tensor("xr_37_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_37_end_0 = const()[name = tensor("xr_37_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_37_end_mask_0 = const()[name = tensor("xr_37_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_37_squeeze_mask_0 = const()[name = tensor("xr_37_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_37_cast_fp16 = slice_by_index(begin = xr_37_begin_0, end = xr_37_end_0, end_mask = xr_37_end_mask_0, squeeze_mask = xr_37_squeeze_mask_0, x = pairs_37_cast_fp16)[name = tensor("xr_37_cast_fp16")]; + tensor xi_37_begin_0 = const()[name = tensor("xi_37_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_37_end_0 = const()[name = tensor("xi_37_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_37_end_mask_0 = const()[name = tensor("xi_37_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_37_squeeze_mask_0 = const()[name = tensor("xi_37_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_37_cast_fp16 = slice_by_index(begin = xi_37_begin_0, end = xi_37_end_0, end_mask = xi_37_end_mask_0, squeeze_mask = xi_37_squeeze_mask_0, x = pairs_37_cast_fp16)[name = tensor("xi_37_cast_fp16")]; + tensor var_2482_cast_fp16 = mul(x = xr_37_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_2482_cast_fp16")]; + tensor var_2483_cast_fp16 = mul(x = xi_37_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_2483_cast_fp16")]; + tensor xor__37_cast_fp16 = sub(x = var_2482_cast_fp16, y = var_2483_cast_fp16)[name = tensor("xor__37_cast_fp16")]; + tensor var_2486_cast_fp16 = mul(x = xr_37_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_2486_cast_fp16")]; + tensor var_2487_cast_fp16 = mul(x = xi_37_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_2487_cast_fp16")]; + tensor xoi_37_cast_fp16 = add(x = var_2486_cast_fp16, y = var_2487_cast_fp16)[name = tensor("xoi_37_cast_fp16")]; + tensor out_37_axis_0 = const()[name = tensor("out_37_axis_0"), val = tensor(-1)]; + tensor out_37_cast_fp16 = stack(axis = out_37_axis_0, values = (xor__37_cast_fp16, xoi_37_cast_fp16))[name = tensor("out_37_cast_fp16")]; + tensor var_2510 = const()[name = tensor("op_2510"), val = tensor([1, 16, 32, 2])]; + tensor pairs_39_cast_fp16 = reshape(shape = var_2510, x = x_39_cast_fp16)[name = tensor("pairs_39_cast_fp16")]; + tensor xr_39_begin_0 = const()[name = tensor("xr_39_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_39_end_0 = const()[name = tensor("xr_39_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_39_end_mask_0 = const()[name = tensor("xr_39_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_39_squeeze_mask_0 = const()[name = tensor("xr_39_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_39_cast_fp16 = slice_by_index(begin = xr_39_begin_0, end = xr_39_end_0, end_mask = xr_39_end_mask_0, squeeze_mask = xr_39_squeeze_mask_0, x = pairs_39_cast_fp16)[name = tensor("xr_39_cast_fp16")]; + tensor xi_39_begin_0 = const()[name = tensor("xi_39_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_39_end_0 = const()[name = tensor("xi_39_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_39_end_mask_0 = const()[name = tensor("xi_39_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_39_squeeze_mask_0 = const()[name = tensor("xi_39_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_39_cast_fp16 = slice_by_index(begin = xi_39_begin_0, end = xi_39_end_0, end_mask = xi_39_end_mask_0, squeeze_mask = xi_39_squeeze_mask_0, x = pairs_39_cast_fp16)[name = tensor("xi_39_cast_fp16")]; + tensor var_2518_cast_fp16 = mul(x = xr_39_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_2518_cast_fp16")]; + tensor var_2519_cast_fp16 = mul(x = xi_39_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_2519_cast_fp16")]; + tensor xor__39_cast_fp16 = sub(x = var_2518_cast_fp16, y = var_2519_cast_fp16)[name = tensor("xor__39_cast_fp16")]; + tensor var_2522_cast_fp16 = mul(x = xr_39_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_2522_cast_fp16")]; + tensor var_2523_cast_fp16 = mul(x = xi_39_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_2523_cast_fp16")]; + tensor xoi_39_cast_fp16 = add(x = var_2522_cast_fp16, y = var_2523_cast_fp16)[name = tensor("xoi_39_cast_fp16")]; + tensor out_39_axis_0 = const()[name = tensor("out_39_axis_0"), val = tensor(-1)]; + tensor out_39_cast_fp16 = stack(axis = out_39_axis_0, values = (xor__39_cast_fp16, xoi_39_cast_fp16))[name = tensor("out_39_cast_fp16")]; + tensor _inversed_2535_y_0_to_fp16 = const()[name = tensor("_inversed_2535_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2535_cast_fp16 = mul(x = position9_to_fp16, y = _inversed_2535_y_0_to_fp16)[name = tensor("_inversed_2535_cast_fp16")]; + tensor var_2536_cast_fp16 = floor(x = _inversed_2535_cast_fp16)[name = tensor("op_2536_cast_fp16")]; + tensor var_2537_to_fp16 = const()[name = tensor("op_2537_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2538_cast_fp16 = mul(x = var_2536_cast_fp16, y = var_2537_to_fp16)[name = tensor("op_2538_cast_fp16")]; + tensor write_idx_19_cast_fp16 = sub(x = position9_to_fp16, y = var_2538_cast_fp16)[name = tensor("write_idx_19_cast_fp16")]; + tensor var_2541_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_19_cast_fp16)[name = tensor("op_2541_cast_fp16")]; + tensor var_2551 = const()[name = tensor("op_2551"), val = tensor([1, 512, 1, 1])]; + tensor onehot_37_to_fp16_dtype_0 = const()[name = tensor("onehot_37_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_2541_cast_fp16_to_fp16 = cast(dtype = onehot_37_to_fp16_dtype_0, x = var_2541_cast_fp16)[name = tensor("cast_504")]; + tensor onehot_39_cast_fp16 = reshape(shape = var_2551, x = var_2541_cast_fp16_to_fp16)[name = tensor("onehot_39_cast_fp16")]; + tensor var_2553_to_fp16 = const()[name = tensor("op_2553_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2555_cast_fp16 = sub(x = var_2553_to_fp16, y = onehot_39_cast_fp16)[name = tensor("op_2555_cast_fp16")]; + tensor k_cache9_to_fp16_dtype_0 = const()[name = tensor("k_cache9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache9_to_fp16 = cast(dtype = k_cache9_to_fp16_dtype_0, x = k_cache9)[name = tensor("cast_503")]; + tensor var_2556_cast_fp16 = mul(x = k_cache9_to_fp16, y = var_2555_cast_fp16)[name = tensor("op_2556_cast_fp16")]; + tensor var_2561 = const()[name = tensor("op_2561"), val = tensor([1, 1, 16, 64])]; + tensor var_2562_cast_fp16 = reshape(shape = var_2561, x = out_39_cast_fp16)[name = tensor("op_2562_cast_fp16")]; + tensor var_2563_cast_fp16 = mul(x = var_2562_cast_fp16, y = onehot_39_cast_fp16)[name = tensor("op_2563_cast_fp16")]; + tensor new_k_cache_19_cast_fp16 = add(x = var_2556_cast_fp16, y = var_2563_cast_fp16)[name = tensor("new_k_cache_19_cast_fp16")]; + tensor new_k_cache_19_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_19_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache9_to_fp16_dtype_0 = const()[name = tensor("v_cache9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache9_to_fp16 = cast(dtype = v_cache9_to_fp16_dtype_0, x = v_cache9)[name = tensor("cast_501")]; + tensor var_2569_cast_fp16 = mul(x = v_cache9_to_fp16, y = var_2555_cast_fp16)[name = tensor("op_2569_cast_fp16")]; + tensor var_2574 = const()[name = tensor("op_2574"), val = tensor([1, 1, 16, 64])]; + tensor var_2575_cast_fp16 = reshape(shape = var_2574, x = v_19_cast_fp16)[name = tensor("op_2575_cast_fp16")]; + tensor var_2576_cast_fp16 = mul(x = var_2575_cast_fp16, y = onehot_39_cast_fp16)[name = tensor("op_2576_cast_fp16")]; + tensor new_v_cache_19_cast_fp16 = add(x = var_2569_cast_fp16, y = var_2576_cast_fp16)[name = tensor("new_v_cache_19_cast_fp16")]; + tensor new_v_cache_19_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_19_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_9_perm_0 = const()[name = tensor("transpose_9_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_2589 = const()[name = tensor("op_2589"), val = tensor([0, 2, 1, 3])]; + tensor var_2595 = const()[name = tensor("op_2595"), val = tensor([1, 16, 1, 64])]; + tensor q4_19_cast_fp16 = reshape(shape = var_2595, x = out_37_cast_fp16)[name = tensor("q4_19_cast_fp16")]; + tensor var_2600_transpose_x_0 = const()[name = tensor("op_2600_transpose_x_0"), val = tensor(false)]; + tensor var_2600_transpose_y_0 = const()[name = tensor("op_2600_transpose_y_0"), val = tensor(false)]; + tensor transpose_9_cast_fp16 = transpose(perm = transpose_9_perm_0, x = new_k_cache_19_cast_fp16)[name = tensor("transpose_68")]; + tensor var_2600_cast_fp16 = matmul(transpose_x = var_2600_transpose_x_0, transpose_y = var_2600_transpose_y_0, x = q4_19_cast_fp16, y = transpose_9_cast_fp16)[name = tensor("op_2600_cast_fp16")]; + tensor var_2601_to_fp16 = const()[name = tensor("op_2601_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_37_cast_fp16 = mul(x = var_2600_cast_fp16, y = var_2601_to_fp16)[name = tensor("scores_37_cast_fp16")]; + tensor var_2605 = const()[name = tensor("op_2605"), val = tensor([1, 1])]; + tensor var_2606_cast_fp16 = reshape(shape = var_2605, x = position9_to_fp16)[name = tensor("op_2606_cast_fp16")]; + tensor var_2607_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_2606_cast_fp16)[name = tensor("op_2607_cast_fp16")]; + tensor mask_19_to_fp16_dtype_0 = const()[name = tensor("mask_19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_2614_to_fp16 = const()[name = tensor("op_2614_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2607_cast_fp16_to_fp16 = cast(dtype = mask_19_to_fp16_dtype_0, x = var_2607_cast_fp16)[name = tensor("cast_499")]; + tensor var_2615_cast_fp16 = sub(x = var_2607_cast_fp16_to_fp16, y = var_2614_to_fp16)[name = tensor("op_2615_cast_fp16")]; + tensor var_2620 = const()[name = tensor("op_2620"), val = tensor([1, 1, 1, 512])]; + tensor var_2621_cast_fp16 = reshape(shape = var_2620, x = var_2615_cast_fp16)[name = tensor("op_2621_cast_fp16")]; + tensor var_2622_to_fp16 = const()[name = tensor("op_2622_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_2623_cast_fp16 = mul(x = var_2621_cast_fp16, y = var_2622_to_fp16)[name = tensor("op_2623_cast_fp16")]; + tensor scores_39_cast_fp16 = add(x = scores_37_cast_fp16, y = var_2623_cast_fp16)[name = tensor("scores_39_cast_fp16")]; + tensor var_2626 = const()[name = tensor("op_2626"), val = tensor(-1)]; + tensor weights_19_cast_fp16 = softmax(axis = var_2626, x = scores_39_cast_fp16)[name = tensor("weights_19_cast_fp16")]; + tensor attn_19_transpose_x_0 = const()[name = tensor("attn_19_transpose_x_0"), val = tensor(false)]; + tensor attn_19_transpose_y_0 = const()[name = tensor("attn_19_transpose_y_0"), val = tensor(false)]; + tensor values_19_cast_fp16 = transpose(perm = var_2589, x = new_v_cache_19_cast_fp16)[name = tensor("transpose_67")]; + tensor attn_19_cast_fp16 = matmul(transpose_x = attn_19_transpose_x_0, transpose_y = attn_19_transpose_y_0, x = weights_19_cast_fp16, y = values_19_cast_fp16)[name = tensor("attn_19_cast_fp16")]; + tensor var_2634 = const()[name = tensor("op_2634"), val = tensor([0, 2, 1, 3])]; + tensor var_2639 = const()[name = tensor("op_2639"), val = tensor([1, 1, 1024])]; + tensor var_2635_cast_fp16 = transpose(perm = var_2634, x = attn_19_cast_fp16)[name = tensor("transpose_66")]; + tensor input_115_cast_fp16 = reshape(shape = var_2639, x = var_2635_cast_fp16)[name = tensor("input_115_cast_fp16")]; + tensor attn9_out_proj_weight_to_fp16 = const()[name = tensor("attn9_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(232949952)))]; + tensor linear_38_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn9_out_proj_weight_to_fp16, x = input_115_cast_fp16)[name = tensor("linear_38_cast_fp16")]; + tensor var_2645_to_fp16 = const()[name = tensor("op_2645_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2646_cast_fp16 = add(x = position9_to_fp16, y = var_2645_to_fp16)[name = tensor("op_2646_cast_fp16")]; + tensor var_2646_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2646_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_117_cast_fp16 = add(x = input_111_cast_fp16, y = linear_38_cast_fp16)[name = tensor("input_117_cast_fp16")]; + tensor input_119_axes_0 = const()[name = tensor("input_119_axes_0"), val = tensor([-1])]; + tensor norm9_2_weight_to_fp16 = const()[name = tensor("norm9_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(235047168)))]; + tensor norm9_2_bias_to_fp16 = const()[name = tensor("norm9_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(235049280)))]; + tensor var_2650_to_fp16 = const()[name = tensor("op_2650_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_119_cast_fp16 = layer_norm(axes = input_119_axes_0, beta = norm9_2_bias_to_fp16, epsilon = var_2650_to_fp16, gamma = norm9_2_weight_to_fp16, x = input_117_cast_fp16)[name = tensor("input_119_cast_fp16")]; + tensor linear9_1_weight_to_fp16 = const()[name = tensor("linear9_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(235051392)))]; + tensor linear_39_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear9_1_weight_to_fp16, x = input_119_cast_fp16)[name = tensor("linear_39_cast_fp16")]; + tensor input_121_mode_0 = const()[name = tensor("input_121_mode_0"), val = tensor("EXACT")]; + tensor input_121_cast_fp16 = gelu(mode = input_121_mode_0, x = linear_39_cast_fp16)[name = tensor("input_121_cast_fp16")]; + tensor linear9_2_weight_to_fp16 = const()[name = tensor("linear9_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(243440064)))]; + tensor linear_40_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear9_2_weight_to_fp16, x = input_121_cast_fp16)[name = tensor("linear_40_cast_fp16")]; + tensor input_123_cast_fp16 = add(x = input_117_cast_fp16, y = linear_40_cast_fp16)[name = tensor("input_123_cast_fp16")]; + tensor input_125_axes_0 = const()[name = tensor("input_125_axes_0"), val = tensor([-1])]; + tensor norm10_1_weight_to_fp16 = const()[name = tensor("norm10_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(251828736)))]; + tensor norm10_1_bias_to_fp16 = const()[name = tensor("norm10_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(251830848)))]; + tensor var_2667_to_fp16 = const()[name = tensor("op_2667_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_125_cast_fp16 = layer_norm(axes = input_125_axes_0, beta = norm10_1_bias_to_fp16, epsilon = var_2667_to_fp16, gamma = norm10_1_weight_to_fp16, x = input_123_cast_fp16)[name = tensor("input_125_cast_fp16")]; + tensor attn10_in_proj_weight_to_fp16 = const()[name = tensor("attn10_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(251832960)))]; + tensor linear_41_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn10_in_proj_weight_to_fp16, x = input_125_cast_fp16)[name = tensor("linear_41_cast_fp16")]; + tensor var_2680 = const()[name = tensor("op_2680"), val = tensor([1, 3, 16, 64])]; + tensor qkv_21_cast_fp16 = reshape(shape = var_2680, x = linear_41_cast_fp16)[name = tensor("qkv_21_cast_fp16")]; + tensor x_41_begin_0 = const()[name = tensor("x_41_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_41_end_0 = const()[name = tensor("x_41_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_41_end_mask_0 = const()[name = tensor("x_41_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_41_squeeze_mask_0 = const()[name = tensor("x_41_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_41_cast_fp16 = slice_by_index(begin = x_41_begin_0, end = x_41_end_0, end_mask = x_41_end_mask_0, squeeze_mask = x_41_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("x_41_cast_fp16")]; + tensor x_43_begin_0 = const()[name = tensor("x_43_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_43_end_0 = const()[name = tensor("x_43_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_43_end_mask_0 = const()[name = tensor("x_43_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_43_squeeze_mask_0 = const()[name = tensor("x_43_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_43_cast_fp16 = slice_by_index(begin = x_43_begin_0, end = x_43_end_0, end_mask = x_43_end_mask_0, squeeze_mask = x_43_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("x_43_cast_fp16")]; + tensor v_21_begin_0 = const()[name = tensor("v_21_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_21_end_0 = const()[name = tensor("v_21_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_21_end_mask_0 = const()[name = tensor("v_21_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_21_squeeze_mask_0 = const()[name = tensor("v_21_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_21_cast_fp16 = slice_by_index(begin = v_21_begin_0, end = v_21_end_0, end_mask = v_21_end_mask_0, squeeze_mask = v_21_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("v_21_cast_fp16")]; + tensor var_2709 = const()[name = tensor("op_2709"), val = tensor([1, 1, 1])]; + tensor position10_to_fp16_dtype_0 = const()[name = tensor("position10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position10_to_fp16 = cast(dtype = position10_to_fp16_dtype_0, x = position10)[name = tensor("cast_497")]; + tensor var_2710_cast_fp16 = reshape(shape = var_2709, x = position10_to_fp16)[name = tensor("op_2710_cast_fp16")]; + tensor angles_41_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_2710_cast_fp16)[name = tensor("angles_41_cast_fp16")]; + tensor rotr_41_cast_fp16 = cos(x = angles_41_cast_fp16)[name = tensor("rotr_41_cast_fp16")]; + tensor roti_41_cast_fp16 = sin(x = angles_41_cast_fp16)[name = tensor("roti_41_cast_fp16")]; + tensor var_2718 = const()[name = tensor("op_2718"), val = tensor([1, 16, 32, 2])]; + tensor pairs_41_cast_fp16 = reshape(shape = var_2718, x = x_41_cast_fp16)[name = tensor("pairs_41_cast_fp16")]; + tensor xr_41_begin_0 = const()[name = tensor("xr_41_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_41_end_0 = const()[name = tensor("xr_41_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_41_end_mask_0 = const()[name = tensor("xr_41_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_41_squeeze_mask_0 = const()[name = tensor("xr_41_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_41_cast_fp16 = slice_by_index(begin = xr_41_begin_0, end = xr_41_end_0, end_mask = xr_41_end_mask_0, squeeze_mask = xr_41_squeeze_mask_0, x = pairs_41_cast_fp16)[name = tensor("xr_41_cast_fp16")]; + tensor xi_41_begin_0 = const()[name = tensor("xi_41_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_41_end_0 = const()[name = tensor("xi_41_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_41_end_mask_0 = const()[name = tensor("xi_41_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_41_squeeze_mask_0 = const()[name = tensor("xi_41_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_41_cast_fp16 = slice_by_index(begin = xi_41_begin_0, end = xi_41_end_0, end_mask = xi_41_end_mask_0, squeeze_mask = xi_41_squeeze_mask_0, x = pairs_41_cast_fp16)[name = tensor("xi_41_cast_fp16")]; + tensor var_2726_cast_fp16 = mul(x = xr_41_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_2726_cast_fp16")]; + tensor var_2727_cast_fp16 = mul(x = xi_41_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_2727_cast_fp16")]; + tensor xor__41_cast_fp16 = sub(x = var_2726_cast_fp16, y = var_2727_cast_fp16)[name = tensor("xor__41_cast_fp16")]; + tensor var_2730_cast_fp16 = mul(x = xr_41_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_2730_cast_fp16")]; + tensor var_2731_cast_fp16 = mul(x = xi_41_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_2731_cast_fp16")]; + tensor xoi_41_cast_fp16 = add(x = var_2730_cast_fp16, y = var_2731_cast_fp16)[name = tensor("xoi_41_cast_fp16")]; + tensor out_41_axis_0 = const()[name = tensor("out_41_axis_0"), val = tensor(-1)]; + tensor out_41_cast_fp16 = stack(axis = out_41_axis_0, values = (xor__41_cast_fp16, xoi_41_cast_fp16))[name = tensor("out_41_cast_fp16")]; + tensor var_2754 = const()[name = tensor("op_2754"), val = tensor([1, 16, 32, 2])]; + tensor pairs_43_cast_fp16 = reshape(shape = var_2754, x = x_43_cast_fp16)[name = tensor("pairs_43_cast_fp16")]; + tensor xr_43_begin_0 = const()[name = tensor("xr_43_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_43_end_0 = const()[name = tensor("xr_43_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_43_end_mask_0 = const()[name = tensor("xr_43_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_43_squeeze_mask_0 = const()[name = tensor("xr_43_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_43_cast_fp16 = slice_by_index(begin = xr_43_begin_0, end = xr_43_end_0, end_mask = xr_43_end_mask_0, squeeze_mask = xr_43_squeeze_mask_0, x = pairs_43_cast_fp16)[name = tensor("xr_43_cast_fp16")]; + tensor xi_43_begin_0 = const()[name = tensor("xi_43_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_43_end_0 = const()[name = tensor("xi_43_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_43_end_mask_0 = const()[name = tensor("xi_43_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_43_squeeze_mask_0 = const()[name = tensor("xi_43_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_43_cast_fp16 = slice_by_index(begin = xi_43_begin_0, end = xi_43_end_0, end_mask = xi_43_end_mask_0, squeeze_mask = xi_43_squeeze_mask_0, x = pairs_43_cast_fp16)[name = tensor("xi_43_cast_fp16")]; + tensor var_2762_cast_fp16 = mul(x = xr_43_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_2762_cast_fp16")]; + tensor var_2763_cast_fp16 = mul(x = xi_43_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_2763_cast_fp16")]; + tensor xor__43_cast_fp16 = sub(x = var_2762_cast_fp16, y = var_2763_cast_fp16)[name = tensor("xor__43_cast_fp16")]; + tensor var_2766_cast_fp16 = mul(x = xr_43_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_2766_cast_fp16")]; + tensor var_2767_cast_fp16 = mul(x = xi_43_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_2767_cast_fp16")]; + tensor xoi_43_cast_fp16 = add(x = var_2766_cast_fp16, y = var_2767_cast_fp16)[name = tensor("xoi_43_cast_fp16")]; + tensor out_43_axis_0 = const()[name = tensor("out_43_axis_0"), val = tensor(-1)]; + tensor out_43_cast_fp16 = stack(axis = out_43_axis_0, values = (xor__43_cast_fp16, xoi_43_cast_fp16))[name = tensor("out_43_cast_fp16")]; + tensor _inversed_2779_y_0_to_fp16 = const()[name = tensor("_inversed_2779_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2779_cast_fp16 = mul(x = position10_to_fp16, y = _inversed_2779_y_0_to_fp16)[name = tensor("_inversed_2779_cast_fp16")]; + tensor var_2780_cast_fp16 = floor(x = _inversed_2779_cast_fp16)[name = tensor("op_2780_cast_fp16")]; + tensor var_2781_to_fp16 = const()[name = tensor("op_2781_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2782_cast_fp16 = mul(x = var_2780_cast_fp16, y = var_2781_to_fp16)[name = tensor("op_2782_cast_fp16")]; + tensor write_idx_21_cast_fp16 = sub(x = position10_to_fp16, y = var_2782_cast_fp16)[name = tensor("write_idx_21_cast_fp16")]; + tensor var_2785_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_21_cast_fp16)[name = tensor("op_2785_cast_fp16")]; + tensor var_2795 = const()[name = tensor("op_2795"), val = tensor([1, 512, 1, 1])]; + tensor onehot_41_to_fp16_dtype_0 = const()[name = tensor("onehot_41_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_2785_cast_fp16_to_fp16 = cast(dtype = onehot_41_to_fp16_dtype_0, x = var_2785_cast_fp16)[name = tensor("cast_496")]; + tensor onehot_43_cast_fp16 = reshape(shape = var_2795, x = var_2785_cast_fp16_to_fp16)[name = tensor("onehot_43_cast_fp16")]; + tensor var_2797_to_fp16 = const()[name = tensor("op_2797_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2799_cast_fp16 = sub(x = var_2797_to_fp16, y = onehot_43_cast_fp16)[name = tensor("op_2799_cast_fp16")]; + tensor k_cache10_to_fp16_dtype_0 = const()[name = tensor("k_cache10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache10_to_fp16 = cast(dtype = k_cache10_to_fp16_dtype_0, x = k_cache10)[name = tensor("cast_495")]; + tensor var_2800_cast_fp16 = mul(x = k_cache10_to_fp16, y = var_2799_cast_fp16)[name = tensor("op_2800_cast_fp16")]; + tensor var_2805 = const()[name = tensor("op_2805"), val = tensor([1, 1, 16, 64])]; + tensor var_2806_cast_fp16 = reshape(shape = var_2805, x = out_43_cast_fp16)[name = tensor("op_2806_cast_fp16")]; + tensor var_2807_cast_fp16 = mul(x = var_2806_cast_fp16, y = onehot_43_cast_fp16)[name = tensor("op_2807_cast_fp16")]; + tensor new_k_cache_21_cast_fp16 = add(x = var_2800_cast_fp16, y = var_2807_cast_fp16)[name = tensor("new_k_cache_21_cast_fp16")]; + tensor new_k_cache_21_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_21_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache10_to_fp16_dtype_0 = const()[name = tensor("v_cache10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache10_to_fp16 = cast(dtype = v_cache10_to_fp16_dtype_0, x = v_cache10)[name = tensor("cast_493")]; + tensor var_2813_cast_fp16 = mul(x = v_cache10_to_fp16, y = var_2799_cast_fp16)[name = tensor("op_2813_cast_fp16")]; + tensor var_2818 = const()[name = tensor("op_2818"), val = tensor([1, 1, 16, 64])]; + tensor var_2819_cast_fp16 = reshape(shape = var_2818, x = v_21_cast_fp16)[name = tensor("op_2819_cast_fp16")]; + tensor var_2820_cast_fp16 = mul(x = var_2819_cast_fp16, y = onehot_43_cast_fp16)[name = tensor("op_2820_cast_fp16")]; + tensor new_v_cache_21_cast_fp16 = add(x = var_2813_cast_fp16, y = var_2820_cast_fp16)[name = tensor("new_v_cache_21_cast_fp16")]; + tensor new_v_cache_21_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_21_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_10_perm_0 = const()[name = tensor("transpose_10_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_2833 = const()[name = tensor("op_2833"), val = tensor([0, 2, 1, 3])]; + tensor var_2839 = const()[name = tensor("op_2839"), val = tensor([1, 16, 1, 64])]; + tensor q4_21_cast_fp16 = reshape(shape = var_2839, x = out_41_cast_fp16)[name = tensor("q4_21_cast_fp16")]; + tensor var_2844_transpose_x_0 = const()[name = tensor("op_2844_transpose_x_0"), val = tensor(false)]; + tensor var_2844_transpose_y_0 = const()[name = tensor("op_2844_transpose_y_0"), val = tensor(false)]; + tensor transpose_10_cast_fp16 = transpose(perm = transpose_10_perm_0, x = new_k_cache_21_cast_fp16)[name = tensor("transpose_65")]; + tensor var_2844_cast_fp16 = matmul(transpose_x = var_2844_transpose_x_0, transpose_y = var_2844_transpose_y_0, x = q4_21_cast_fp16, y = transpose_10_cast_fp16)[name = tensor("op_2844_cast_fp16")]; + tensor var_2845_to_fp16 = const()[name = tensor("op_2845_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_41_cast_fp16 = mul(x = var_2844_cast_fp16, y = var_2845_to_fp16)[name = tensor("scores_41_cast_fp16")]; + tensor var_2849 = const()[name = tensor("op_2849"), val = tensor([1, 1])]; + tensor var_2850_cast_fp16 = reshape(shape = var_2849, x = position10_to_fp16)[name = tensor("op_2850_cast_fp16")]; + tensor var_2851_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_2850_cast_fp16)[name = tensor("op_2851_cast_fp16")]; + tensor mask_21_to_fp16_dtype_0 = const()[name = tensor("mask_21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_2858_to_fp16 = const()[name = tensor("op_2858_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2851_cast_fp16_to_fp16 = cast(dtype = mask_21_to_fp16_dtype_0, x = var_2851_cast_fp16)[name = tensor("cast_491")]; + tensor var_2859_cast_fp16 = sub(x = var_2851_cast_fp16_to_fp16, y = var_2858_to_fp16)[name = tensor("op_2859_cast_fp16")]; + tensor var_2864 = const()[name = tensor("op_2864"), val = tensor([1, 1, 1, 512])]; + tensor var_2865_cast_fp16 = reshape(shape = var_2864, x = var_2859_cast_fp16)[name = tensor("op_2865_cast_fp16")]; + tensor var_2866_to_fp16 = const()[name = tensor("op_2866_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_2867_cast_fp16 = mul(x = var_2865_cast_fp16, y = var_2866_to_fp16)[name = tensor("op_2867_cast_fp16")]; + tensor scores_43_cast_fp16 = add(x = scores_41_cast_fp16, y = var_2867_cast_fp16)[name = tensor("scores_43_cast_fp16")]; + tensor var_2870 = const()[name = tensor("op_2870"), val = tensor(-1)]; + tensor weights_21_cast_fp16 = softmax(axis = var_2870, x = scores_43_cast_fp16)[name = tensor("weights_21_cast_fp16")]; + tensor attn_21_transpose_x_0 = const()[name = tensor("attn_21_transpose_x_0"), val = tensor(false)]; + tensor attn_21_transpose_y_0 = const()[name = tensor("attn_21_transpose_y_0"), val = tensor(false)]; + tensor values_21_cast_fp16 = transpose(perm = var_2833, x = new_v_cache_21_cast_fp16)[name = tensor("transpose_64")]; + tensor attn_21_cast_fp16 = matmul(transpose_x = attn_21_transpose_x_0, transpose_y = attn_21_transpose_y_0, x = weights_21_cast_fp16, y = values_21_cast_fp16)[name = tensor("attn_21_cast_fp16")]; + tensor var_2878 = const()[name = tensor("op_2878"), val = tensor([0, 2, 1, 3])]; + tensor var_2883 = const()[name = tensor("op_2883"), val = tensor([1, 1, 1024])]; + tensor var_2879_cast_fp16 = transpose(perm = var_2878, x = attn_21_cast_fp16)[name = tensor("transpose_63")]; + tensor input_127_cast_fp16 = reshape(shape = var_2883, x = var_2879_cast_fp16)[name = tensor("input_127_cast_fp16")]; + tensor attn10_out_proj_weight_to_fp16 = const()[name = tensor("attn10_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(258124480)))]; + tensor linear_42_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn10_out_proj_weight_to_fp16, x = input_127_cast_fp16)[name = tensor("linear_42_cast_fp16")]; + tensor var_2889_to_fp16 = const()[name = tensor("op_2889_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2890_cast_fp16 = add(x = position10_to_fp16, y = var_2889_to_fp16)[name = tensor("op_2890_cast_fp16")]; + tensor var_2890_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2890_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_129_cast_fp16 = add(x = input_123_cast_fp16, y = linear_42_cast_fp16)[name = tensor("input_129_cast_fp16")]; + tensor input_131_axes_0 = const()[name = tensor("input_131_axes_0"), val = tensor([-1])]; + tensor norm10_2_weight_to_fp16 = const()[name = tensor("norm10_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260221696)))]; + tensor norm10_2_bias_to_fp16 = const()[name = tensor("norm10_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260223808)))]; + tensor var_2894_to_fp16 = const()[name = tensor("op_2894_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_131_cast_fp16 = layer_norm(axes = input_131_axes_0, beta = norm10_2_bias_to_fp16, epsilon = var_2894_to_fp16, gamma = norm10_2_weight_to_fp16, x = input_129_cast_fp16)[name = tensor("input_131_cast_fp16")]; + tensor linear10_1_weight_to_fp16 = const()[name = tensor("linear10_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260225920)))]; + tensor linear_43_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear10_1_weight_to_fp16, x = input_131_cast_fp16)[name = tensor("linear_43_cast_fp16")]; + tensor input_133_mode_0 = const()[name = tensor("input_133_mode_0"), val = tensor("EXACT")]; + tensor input_133_cast_fp16 = gelu(mode = input_133_mode_0, x = linear_43_cast_fp16)[name = tensor("input_133_cast_fp16")]; + tensor linear10_2_weight_to_fp16 = const()[name = tensor("linear10_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(268614592)))]; + tensor linear_44_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear10_2_weight_to_fp16, x = input_133_cast_fp16)[name = tensor("linear_44_cast_fp16")]; + tensor input_135_cast_fp16 = add(x = input_129_cast_fp16, y = linear_44_cast_fp16)[name = tensor("input_135_cast_fp16")]; + tensor input_137_axes_0 = const()[name = tensor("input_137_axes_0"), val = tensor([-1])]; + tensor norm11_1_weight_to_fp16 = const()[name = tensor("norm11_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277003264)))]; + tensor norm11_1_bias_to_fp16 = const()[name = tensor("norm11_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277005376)))]; + tensor var_2911_to_fp16 = const()[name = tensor("op_2911_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_137_cast_fp16 = layer_norm(axes = input_137_axes_0, beta = norm11_1_bias_to_fp16, epsilon = var_2911_to_fp16, gamma = norm11_1_weight_to_fp16, x = input_135_cast_fp16)[name = tensor("input_137_cast_fp16")]; + tensor attn11_in_proj_weight_to_fp16 = const()[name = tensor("attn11_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277007488)))]; + tensor linear_45_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn11_in_proj_weight_to_fp16, x = input_137_cast_fp16)[name = tensor("linear_45_cast_fp16")]; + tensor var_2924 = const()[name = tensor("op_2924"), val = tensor([1, 3, 16, 64])]; + tensor qkv_23_cast_fp16 = reshape(shape = var_2924, x = linear_45_cast_fp16)[name = tensor("qkv_23_cast_fp16")]; + tensor x_45_begin_0 = const()[name = tensor("x_45_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_45_end_0 = const()[name = tensor("x_45_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_45_end_mask_0 = const()[name = tensor("x_45_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_45_squeeze_mask_0 = const()[name = tensor("x_45_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_45_cast_fp16 = slice_by_index(begin = x_45_begin_0, end = x_45_end_0, end_mask = x_45_end_mask_0, squeeze_mask = x_45_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("x_45_cast_fp16")]; + tensor x_47_begin_0 = const()[name = tensor("x_47_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_47_end_0 = const()[name = tensor("x_47_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_47_end_mask_0 = const()[name = tensor("x_47_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_47_squeeze_mask_0 = const()[name = tensor("x_47_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_47_cast_fp16 = slice_by_index(begin = x_47_begin_0, end = x_47_end_0, end_mask = x_47_end_mask_0, squeeze_mask = x_47_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("x_47_cast_fp16")]; + tensor v_23_begin_0 = const()[name = tensor("v_23_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_23_end_0 = const()[name = tensor("v_23_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_23_end_mask_0 = const()[name = tensor("v_23_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_23_squeeze_mask_0 = const()[name = tensor("v_23_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_23_cast_fp16 = slice_by_index(begin = v_23_begin_0, end = v_23_end_0, end_mask = v_23_end_mask_0, squeeze_mask = v_23_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("v_23_cast_fp16")]; + tensor var_2953 = const()[name = tensor("op_2953"), val = tensor([1, 1, 1])]; + tensor position11_to_fp16_dtype_0 = const()[name = tensor("position11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position11_to_fp16 = cast(dtype = position11_to_fp16_dtype_0, x = position11)[name = tensor("cast_489")]; + tensor var_2954_cast_fp16 = reshape(shape = var_2953, x = position11_to_fp16)[name = tensor("op_2954_cast_fp16")]; + tensor angles_45_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_2954_cast_fp16)[name = tensor("angles_45_cast_fp16")]; + tensor rotr_45_cast_fp16 = cos(x = angles_45_cast_fp16)[name = tensor("rotr_45_cast_fp16")]; + tensor roti_45_cast_fp16 = sin(x = angles_45_cast_fp16)[name = tensor("roti_45_cast_fp16")]; + tensor var_2962 = const()[name = tensor("op_2962"), val = tensor([1, 16, 32, 2])]; + tensor pairs_45_cast_fp16 = reshape(shape = var_2962, x = x_45_cast_fp16)[name = tensor("pairs_45_cast_fp16")]; + tensor xr_45_begin_0 = const()[name = tensor("xr_45_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_45_end_0 = const()[name = tensor("xr_45_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_45_end_mask_0 = const()[name = tensor("xr_45_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_45_squeeze_mask_0 = const()[name = tensor("xr_45_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_45_cast_fp16 = slice_by_index(begin = xr_45_begin_0, end = xr_45_end_0, end_mask = xr_45_end_mask_0, squeeze_mask = xr_45_squeeze_mask_0, x = pairs_45_cast_fp16)[name = tensor("xr_45_cast_fp16")]; + tensor xi_45_begin_0 = const()[name = tensor("xi_45_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_45_end_0 = const()[name = tensor("xi_45_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_45_end_mask_0 = const()[name = tensor("xi_45_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_45_squeeze_mask_0 = const()[name = tensor("xi_45_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_45_cast_fp16 = slice_by_index(begin = xi_45_begin_0, end = xi_45_end_0, end_mask = xi_45_end_mask_0, squeeze_mask = xi_45_squeeze_mask_0, x = pairs_45_cast_fp16)[name = tensor("xi_45_cast_fp16")]; + tensor var_2970_cast_fp16 = mul(x = xr_45_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_2970_cast_fp16")]; + tensor var_2971_cast_fp16 = mul(x = xi_45_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_2971_cast_fp16")]; + tensor xor__45_cast_fp16 = sub(x = var_2970_cast_fp16, y = var_2971_cast_fp16)[name = tensor("xor__45_cast_fp16")]; + tensor var_2974_cast_fp16 = mul(x = xr_45_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_2974_cast_fp16")]; + tensor var_2975_cast_fp16 = mul(x = xi_45_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_2975_cast_fp16")]; + tensor xoi_45_cast_fp16 = add(x = var_2974_cast_fp16, y = var_2975_cast_fp16)[name = tensor("xoi_45_cast_fp16")]; + tensor out_45_axis_0 = const()[name = tensor("out_45_axis_0"), val = tensor(-1)]; + tensor out_45_cast_fp16 = stack(axis = out_45_axis_0, values = (xor__45_cast_fp16, xoi_45_cast_fp16))[name = tensor("out_45_cast_fp16")]; + tensor var_2998 = const()[name = tensor("op_2998"), val = tensor([1, 16, 32, 2])]; + tensor pairs_47_cast_fp16 = reshape(shape = var_2998, x = x_47_cast_fp16)[name = tensor("pairs_47_cast_fp16")]; + tensor xr_47_begin_0 = const()[name = tensor("xr_47_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_47_end_0 = const()[name = tensor("xr_47_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_47_end_mask_0 = const()[name = tensor("xr_47_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_47_squeeze_mask_0 = const()[name = tensor("xr_47_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_47_cast_fp16 = slice_by_index(begin = xr_47_begin_0, end = xr_47_end_0, end_mask = xr_47_end_mask_0, squeeze_mask = xr_47_squeeze_mask_0, x = pairs_47_cast_fp16)[name = tensor("xr_47_cast_fp16")]; + tensor xi_47_begin_0 = const()[name = tensor("xi_47_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_47_end_0 = const()[name = tensor("xi_47_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_47_end_mask_0 = const()[name = tensor("xi_47_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_47_squeeze_mask_0 = const()[name = tensor("xi_47_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_47_cast_fp16 = slice_by_index(begin = xi_47_begin_0, end = xi_47_end_0, end_mask = xi_47_end_mask_0, squeeze_mask = xi_47_squeeze_mask_0, x = pairs_47_cast_fp16)[name = tensor("xi_47_cast_fp16")]; + tensor var_3006_cast_fp16 = mul(x = xr_47_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_3006_cast_fp16")]; + tensor var_3007_cast_fp16 = mul(x = xi_47_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_3007_cast_fp16")]; + tensor xor__47_cast_fp16 = sub(x = var_3006_cast_fp16, y = var_3007_cast_fp16)[name = tensor("xor__47_cast_fp16")]; + tensor var_3010_cast_fp16 = mul(x = xr_47_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_3010_cast_fp16")]; + tensor var_3011_cast_fp16 = mul(x = xi_47_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_3011_cast_fp16")]; + tensor xoi_47_cast_fp16 = add(x = var_3010_cast_fp16, y = var_3011_cast_fp16)[name = tensor("xoi_47_cast_fp16")]; + tensor out_47_axis_0 = const()[name = tensor("out_47_axis_0"), val = tensor(-1)]; + tensor out_47_cast_fp16 = stack(axis = out_47_axis_0, values = (xor__47_cast_fp16, xoi_47_cast_fp16))[name = tensor("out_47_cast_fp16")]; + tensor _inversed_3023_y_0_to_fp16 = const()[name = tensor("_inversed_3023_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3023_cast_fp16 = mul(x = position11_to_fp16, y = _inversed_3023_y_0_to_fp16)[name = tensor("_inversed_3023_cast_fp16")]; + tensor var_3024_cast_fp16 = floor(x = _inversed_3023_cast_fp16)[name = tensor("op_3024_cast_fp16")]; + tensor var_3025_to_fp16 = const()[name = tensor("op_3025_to_fp16"), val = tensor(0x1p+9)]; + tensor var_3026_cast_fp16 = mul(x = var_3024_cast_fp16, y = var_3025_to_fp16)[name = tensor("op_3026_cast_fp16")]; + tensor write_idx_23_cast_fp16 = sub(x = position11_to_fp16, y = var_3026_cast_fp16)[name = tensor("write_idx_23_cast_fp16")]; + tensor var_3029_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_23_cast_fp16)[name = tensor("op_3029_cast_fp16")]; + tensor var_3039 = const()[name = tensor("op_3039"), val = tensor([1, 512, 1, 1])]; + tensor onehot_45_to_fp16_dtype_0 = const()[name = tensor("onehot_45_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_3029_cast_fp16_to_fp16 = cast(dtype = onehot_45_to_fp16_dtype_0, x = var_3029_cast_fp16)[name = tensor("cast_488")]; + tensor onehot_47_cast_fp16 = reshape(shape = var_3039, x = var_3029_cast_fp16_to_fp16)[name = tensor("onehot_47_cast_fp16")]; + tensor var_3041_to_fp16 = const()[name = tensor("op_3041_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3043_cast_fp16 = sub(x = var_3041_to_fp16, y = onehot_47_cast_fp16)[name = tensor("op_3043_cast_fp16")]; + tensor k_cache11_to_fp16_dtype_0 = const()[name = tensor("k_cache11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache11_to_fp16 = cast(dtype = k_cache11_to_fp16_dtype_0, x = k_cache11)[name = tensor("cast_487")]; + tensor var_3044_cast_fp16 = mul(x = k_cache11_to_fp16, y = var_3043_cast_fp16)[name = tensor("op_3044_cast_fp16")]; + tensor var_3049 = const()[name = tensor("op_3049"), val = tensor([1, 1, 16, 64])]; + tensor var_3050_cast_fp16 = reshape(shape = var_3049, x = out_47_cast_fp16)[name = tensor("op_3050_cast_fp16")]; + tensor var_3051_cast_fp16 = mul(x = var_3050_cast_fp16, y = onehot_47_cast_fp16)[name = tensor("op_3051_cast_fp16")]; + tensor new_k_cache_23_cast_fp16 = add(x = var_3044_cast_fp16, y = var_3051_cast_fp16)[name = tensor("new_k_cache_23_cast_fp16")]; + tensor new_k_cache_23_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_23_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache11_to_fp16_dtype_0 = const()[name = tensor("v_cache11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache11_to_fp16 = cast(dtype = v_cache11_to_fp16_dtype_0, x = v_cache11)[name = tensor("cast_485")]; + tensor var_3057_cast_fp16 = mul(x = v_cache11_to_fp16, y = var_3043_cast_fp16)[name = tensor("op_3057_cast_fp16")]; + tensor var_3062 = const()[name = tensor("op_3062"), val = tensor([1, 1, 16, 64])]; + tensor var_3063_cast_fp16 = reshape(shape = var_3062, x = v_23_cast_fp16)[name = tensor("op_3063_cast_fp16")]; + tensor var_3064_cast_fp16 = mul(x = var_3063_cast_fp16, y = onehot_47_cast_fp16)[name = tensor("op_3064_cast_fp16")]; + tensor new_v_cache_23_cast_fp16 = add(x = var_3057_cast_fp16, y = var_3064_cast_fp16)[name = tensor("new_v_cache_23_cast_fp16")]; + tensor new_v_cache_23_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_23_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_11_perm_0 = const()[name = tensor("transpose_11_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_3077 = const()[name = tensor("op_3077"), val = tensor([0, 2, 1, 3])]; + tensor var_3083 = const()[name = tensor("op_3083"), val = tensor([1, 16, 1, 64])]; + tensor q4_23_cast_fp16 = reshape(shape = var_3083, x = out_45_cast_fp16)[name = tensor("q4_23_cast_fp16")]; + tensor var_3088_transpose_x_0 = const()[name = tensor("op_3088_transpose_x_0"), val = tensor(false)]; + tensor var_3088_transpose_y_0 = const()[name = tensor("op_3088_transpose_y_0"), val = tensor(false)]; + tensor transpose_11_cast_fp16 = transpose(perm = transpose_11_perm_0, x = new_k_cache_23_cast_fp16)[name = tensor("transpose_62")]; + tensor var_3088_cast_fp16 = matmul(transpose_x = var_3088_transpose_x_0, transpose_y = var_3088_transpose_y_0, x = q4_23_cast_fp16, y = transpose_11_cast_fp16)[name = tensor("op_3088_cast_fp16")]; + tensor var_3089_to_fp16 = const()[name = tensor("op_3089_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_45_cast_fp16 = mul(x = var_3088_cast_fp16, y = var_3089_to_fp16)[name = tensor("scores_45_cast_fp16")]; + tensor var_3093 = const()[name = tensor("op_3093"), val = tensor([1, 1])]; + tensor var_3094_cast_fp16 = reshape(shape = var_3093, x = position11_to_fp16)[name = tensor("op_3094_cast_fp16")]; + tensor var_3095_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_3094_cast_fp16)[name = tensor("op_3095_cast_fp16")]; + tensor mask_23_to_fp16_dtype_0 = const()[name = tensor("mask_23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_3102_to_fp16 = const()[name = tensor("op_3102_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3095_cast_fp16_to_fp16 = cast(dtype = mask_23_to_fp16_dtype_0, x = var_3095_cast_fp16)[name = tensor("cast_483")]; + tensor var_3103_cast_fp16 = sub(x = var_3095_cast_fp16_to_fp16, y = var_3102_to_fp16)[name = tensor("op_3103_cast_fp16")]; + tensor var_3108 = const()[name = tensor("op_3108"), val = tensor([1, 1, 1, 512])]; + tensor var_3109_cast_fp16 = reshape(shape = var_3108, x = var_3103_cast_fp16)[name = tensor("op_3109_cast_fp16")]; + tensor var_3110_to_fp16 = const()[name = tensor("op_3110_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_3111_cast_fp16 = mul(x = var_3109_cast_fp16, y = var_3110_to_fp16)[name = tensor("op_3111_cast_fp16")]; + tensor scores_47_cast_fp16 = add(x = scores_45_cast_fp16, y = var_3111_cast_fp16)[name = tensor("scores_47_cast_fp16")]; + tensor var_3114 = const()[name = tensor("op_3114"), val = tensor(-1)]; + tensor weights_23_cast_fp16 = softmax(axis = var_3114, x = scores_47_cast_fp16)[name = tensor("weights_23_cast_fp16")]; + tensor attn_23_transpose_x_0 = const()[name = tensor("attn_23_transpose_x_0"), val = tensor(false)]; + tensor attn_23_transpose_y_0 = const()[name = tensor("attn_23_transpose_y_0"), val = tensor(false)]; + tensor values_23_cast_fp16 = transpose(perm = var_3077, x = new_v_cache_23_cast_fp16)[name = tensor("transpose_61")]; + tensor attn_23_cast_fp16 = matmul(transpose_x = attn_23_transpose_x_0, transpose_y = attn_23_transpose_y_0, x = weights_23_cast_fp16, y = values_23_cast_fp16)[name = tensor("attn_23_cast_fp16")]; + tensor var_3122 = const()[name = tensor("op_3122"), val = tensor([0, 2, 1, 3])]; + tensor var_3127 = const()[name = tensor("op_3127"), val = tensor([1, 1, 1024])]; + tensor var_3123_cast_fp16 = transpose(perm = var_3122, x = attn_23_cast_fp16)[name = tensor("transpose_60")]; + tensor input_139_cast_fp16 = reshape(shape = var_3127, x = var_3123_cast_fp16)[name = tensor("input_139_cast_fp16")]; + tensor attn11_out_proj_weight_to_fp16 = const()[name = tensor("attn11_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(283299008)))]; + tensor linear_46_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn11_out_proj_weight_to_fp16, x = input_139_cast_fp16)[name = tensor("linear_46_cast_fp16")]; + tensor var_3133_to_fp16 = const()[name = tensor("op_3133_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3134_cast_fp16 = add(x = position11_to_fp16, y = var_3133_to_fp16)[name = tensor("op_3134_cast_fp16")]; + tensor var_3134_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3134_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_141_cast_fp16 = add(x = input_135_cast_fp16, y = linear_46_cast_fp16)[name = tensor("input_141_cast_fp16")]; + tensor input_143_axes_0 = const()[name = tensor("input_143_axes_0"), val = tensor([-1])]; + tensor norm11_2_weight_to_fp16 = const()[name = tensor("norm11_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285396224)))]; + tensor norm11_2_bias_to_fp16 = const()[name = tensor("norm11_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285398336)))]; + tensor var_3138_to_fp16 = const()[name = tensor("op_3138_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_143_cast_fp16 = layer_norm(axes = input_143_axes_0, beta = norm11_2_bias_to_fp16, epsilon = var_3138_to_fp16, gamma = norm11_2_weight_to_fp16, x = input_141_cast_fp16)[name = tensor("input_143_cast_fp16")]; + tensor linear11_1_weight_to_fp16 = const()[name = tensor("linear11_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285400448)))]; + tensor linear_47_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear11_1_weight_to_fp16, x = input_143_cast_fp16)[name = tensor("linear_47_cast_fp16")]; + tensor input_145_mode_0 = const()[name = tensor("input_145_mode_0"), val = tensor("EXACT")]; + tensor input_145_cast_fp16 = gelu(mode = input_145_mode_0, x = linear_47_cast_fp16)[name = tensor("input_145_cast_fp16")]; + tensor linear11_2_weight_to_fp16 = const()[name = tensor("linear11_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293789120)))]; + tensor linear_48_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear11_2_weight_to_fp16, x = input_145_cast_fp16)[name = tensor("linear_48_cast_fp16")]; + tensor input_147_cast_fp16 = add(x = input_141_cast_fp16, y = linear_48_cast_fp16)[name = tensor("input_147_cast_fp16")]; + tensor input_149_axes_0 = const()[name = tensor("input_149_axes_0"), val = tensor([-1])]; + tensor norm12_1_weight_to_fp16 = const()[name = tensor("norm12_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302177792)))]; + tensor norm12_1_bias_to_fp16 = const()[name = tensor("norm12_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302179904)))]; + tensor var_3155_to_fp16 = const()[name = tensor("op_3155_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_149_cast_fp16 = layer_norm(axes = input_149_axes_0, beta = norm12_1_bias_to_fp16, epsilon = var_3155_to_fp16, gamma = norm12_1_weight_to_fp16, x = input_147_cast_fp16)[name = tensor("input_149_cast_fp16")]; + tensor attn12_in_proj_weight_to_fp16 = const()[name = tensor("attn12_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302182016)))]; + tensor linear_49_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn12_in_proj_weight_to_fp16, x = input_149_cast_fp16)[name = tensor("linear_49_cast_fp16")]; + tensor var_3168 = const()[name = tensor("op_3168"), val = tensor([1, 3, 16, 64])]; + tensor qkv_25_cast_fp16 = reshape(shape = var_3168, x = linear_49_cast_fp16)[name = tensor("qkv_25_cast_fp16")]; + tensor x_49_begin_0 = const()[name = tensor("x_49_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_49_end_0 = const()[name = tensor("x_49_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_49_end_mask_0 = const()[name = tensor("x_49_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_49_squeeze_mask_0 = const()[name = tensor("x_49_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_49_cast_fp16 = slice_by_index(begin = x_49_begin_0, end = x_49_end_0, end_mask = x_49_end_mask_0, squeeze_mask = x_49_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("x_49_cast_fp16")]; + tensor x_51_begin_0 = const()[name = tensor("x_51_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_51_end_0 = const()[name = tensor("x_51_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_51_end_mask_0 = const()[name = tensor("x_51_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_51_squeeze_mask_0 = const()[name = tensor("x_51_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_51_cast_fp16 = slice_by_index(begin = x_51_begin_0, end = x_51_end_0, end_mask = x_51_end_mask_0, squeeze_mask = x_51_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("x_51_cast_fp16")]; + tensor v_25_begin_0 = const()[name = tensor("v_25_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_25_end_0 = const()[name = tensor("v_25_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_25_end_mask_0 = const()[name = tensor("v_25_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_25_squeeze_mask_0 = const()[name = tensor("v_25_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_25_cast_fp16 = slice_by_index(begin = v_25_begin_0, end = v_25_end_0, end_mask = v_25_end_mask_0, squeeze_mask = v_25_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("v_25_cast_fp16")]; + tensor var_3197 = const()[name = tensor("op_3197"), val = tensor([1, 1, 1])]; + tensor position12_to_fp16_dtype_0 = const()[name = tensor("position12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position12_to_fp16 = cast(dtype = position12_to_fp16_dtype_0, x = position12)[name = tensor("cast_481")]; + tensor var_3198_cast_fp16 = reshape(shape = var_3197, x = position12_to_fp16)[name = tensor("op_3198_cast_fp16")]; + tensor angles_49_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_3198_cast_fp16)[name = tensor("angles_49_cast_fp16")]; + tensor rotr_49_cast_fp16 = cos(x = angles_49_cast_fp16)[name = tensor("rotr_49_cast_fp16")]; + tensor roti_49_cast_fp16 = sin(x = angles_49_cast_fp16)[name = tensor("roti_49_cast_fp16")]; + tensor var_3206 = const()[name = tensor("op_3206"), val = tensor([1, 16, 32, 2])]; + tensor pairs_49_cast_fp16 = reshape(shape = var_3206, x = x_49_cast_fp16)[name = tensor("pairs_49_cast_fp16")]; + tensor xr_49_begin_0 = const()[name = tensor("xr_49_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_49_end_0 = const()[name = tensor("xr_49_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_49_end_mask_0 = const()[name = tensor("xr_49_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_49_squeeze_mask_0 = const()[name = tensor("xr_49_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_49_cast_fp16 = slice_by_index(begin = xr_49_begin_0, end = xr_49_end_0, end_mask = xr_49_end_mask_0, squeeze_mask = xr_49_squeeze_mask_0, x = pairs_49_cast_fp16)[name = tensor("xr_49_cast_fp16")]; + tensor xi_49_begin_0 = const()[name = tensor("xi_49_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_49_end_0 = const()[name = tensor("xi_49_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_49_end_mask_0 = const()[name = tensor("xi_49_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_49_squeeze_mask_0 = const()[name = tensor("xi_49_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_49_cast_fp16 = slice_by_index(begin = xi_49_begin_0, end = xi_49_end_0, end_mask = xi_49_end_mask_0, squeeze_mask = xi_49_squeeze_mask_0, x = pairs_49_cast_fp16)[name = tensor("xi_49_cast_fp16")]; + tensor var_3214_cast_fp16 = mul(x = xr_49_cast_fp16, y = rotr_49_cast_fp16)[name = tensor("op_3214_cast_fp16")]; + tensor var_3215_cast_fp16 = mul(x = xi_49_cast_fp16, y = roti_49_cast_fp16)[name = tensor("op_3215_cast_fp16")]; + tensor xor__49_cast_fp16 = sub(x = var_3214_cast_fp16, y = var_3215_cast_fp16)[name = tensor("xor__49_cast_fp16")]; + tensor var_3218_cast_fp16 = mul(x = xr_49_cast_fp16, y = roti_49_cast_fp16)[name = tensor("op_3218_cast_fp16")]; + tensor var_3219_cast_fp16 = mul(x = xi_49_cast_fp16, y = rotr_49_cast_fp16)[name = tensor("op_3219_cast_fp16")]; + tensor xoi_49_cast_fp16 = add(x = var_3218_cast_fp16, y = var_3219_cast_fp16)[name = tensor("xoi_49_cast_fp16")]; + tensor out_49_axis_0 = const()[name = tensor("out_49_axis_0"), val = tensor(-1)]; + tensor out_49_cast_fp16 = stack(axis = out_49_axis_0, values = (xor__49_cast_fp16, xoi_49_cast_fp16))[name = tensor("out_49_cast_fp16")]; + tensor var_3242 = const()[name = tensor("op_3242"), val = tensor([1, 16, 32, 2])]; + tensor pairs_51_cast_fp16 = reshape(shape = var_3242, x = x_51_cast_fp16)[name = tensor("pairs_51_cast_fp16")]; + tensor xr_51_begin_0 = const()[name = tensor("xr_51_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_51_end_0 = const()[name = tensor("xr_51_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_51_end_mask_0 = const()[name = tensor("xr_51_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_51_squeeze_mask_0 = const()[name = tensor("xr_51_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_51_cast_fp16 = slice_by_index(begin = xr_51_begin_0, end = xr_51_end_0, end_mask = xr_51_end_mask_0, squeeze_mask = xr_51_squeeze_mask_0, x = pairs_51_cast_fp16)[name = tensor("xr_51_cast_fp16")]; + tensor xi_51_begin_0 = const()[name = tensor("xi_51_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_51_end_0 = const()[name = tensor("xi_51_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_51_end_mask_0 = const()[name = tensor("xi_51_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_51_squeeze_mask_0 = const()[name = tensor("xi_51_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_51_cast_fp16 = slice_by_index(begin = xi_51_begin_0, end = xi_51_end_0, end_mask = xi_51_end_mask_0, squeeze_mask = xi_51_squeeze_mask_0, x = pairs_51_cast_fp16)[name = tensor("xi_51_cast_fp16")]; + tensor var_3250_cast_fp16 = mul(x = xr_51_cast_fp16, y = rotr_49_cast_fp16)[name = tensor("op_3250_cast_fp16")]; + tensor var_3251_cast_fp16 = mul(x = xi_51_cast_fp16, y = roti_49_cast_fp16)[name = tensor("op_3251_cast_fp16")]; + tensor xor__51_cast_fp16 = sub(x = var_3250_cast_fp16, y = var_3251_cast_fp16)[name = tensor("xor__51_cast_fp16")]; + tensor var_3254_cast_fp16 = mul(x = xr_51_cast_fp16, y = roti_49_cast_fp16)[name = tensor("op_3254_cast_fp16")]; + tensor var_3255_cast_fp16 = mul(x = xi_51_cast_fp16, y = rotr_49_cast_fp16)[name = tensor("op_3255_cast_fp16")]; + tensor xoi_51_cast_fp16 = add(x = var_3254_cast_fp16, y = var_3255_cast_fp16)[name = tensor("xoi_51_cast_fp16")]; + tensor out_51_axis_0 = const()[name = tensor("out_51_axis_0"), val = tensor(-1)]; + tensor out_51_cast_fp16 = stack(axis = out_51_axis_0, values = (xor__51_cast_fp16, xoi_51_cast_fp16))[name = tensor("out_51_cast_fp16")]; + tensor _inversed_3267_y_0_to_fp16 = const()[name = tensor("_inversed_3267_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3267_cast_fp16 = mul(x = position12_to_fp16, y = _inversed_3267_y_0_to_fp16)[name = tensor("_inversed_3267_cast_fp16")]; + tensor var_3268_cast_fp16 = floor(x = _inversed_3267_cast_fp16)[name = tensor("op_3268_cast_fp16")]; + tensor var_3269_to_fp16 = const()[name = tensor("op_3269_to_fp16"), val = tensor(0x1p+9)]; + tensor var_3270_cast_fp16 = mul(x = var_3268_cast_fp16, y = var_3269_to_fp16)[name = tensor("op_3270_cast_fp16")]; + tensor write_idx_25_cast_fp16 = sub(x = position12_to_fp16, y = var_3270_cast_fp16)[name = tensor("write_idx_25_cast_fp16")]; + tensor var_3273_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_25_cast_fp16)[name = tensor("op_3273_cast_fp16")]; + tensor var_3283 = const()[name = tensor("op_3283"), val = tensor([1, 512, 1, 1])]; + tensor onehot_49_to_fp16_dtype_0 = const()[name = tensor("onehot_49_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_3273_cast_fp16_to_fp16 = cast(dtype = onehot_49_to_fp16_dtype_0, x = var_3273_cast_fp16)[name = tensor("cast_480")]; + tensor onehot_51_cast_fp16 = reshape(shape = var_3283, x = var_3273_cast_fp16_to_fp16)[name = tensor("onehot_51_cast_fp16")]; + tensor var_3285_to_fp16 = const()[name = tensor("op_3285_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3287_cast_fp16 = sub(x = var_3285_to_fp16, y = onehot_51_cast_fp16)[name = tensor("op_3287_cast_fp16")]; + tensor k_cache12_to_fp16_dtype_0 = const()[name = tensor("k_cache12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache12_to_fp16 = cast(dtype = k_cache12_to_fp16_dtype_0, x = k_cache12)[name = tensor("cast_479")]; + tensor var_3288_cast_fp16 = mul(x = k_cache12_to_fp16, y = var_3287_cast_fp16)[name = tensor("op_3288_cast_fp16")]; + tensor var_3293 = const()[name = tensor("op_3293"), val = tensor([1, 1, 16, 64])]; + tensor var_3294_cast_fp16 = reshape(shape = var_3293, x = out_51_cast_fp16)[name = tensor("op_3294_cast_fp16")]; + tensor var_3295_cast_fp16 = mul(x = var_3294_cast_fp16, y = onehot_51_cast_fp16)[name = tensor("op_3295_cast_fp16")]; + tensor new_k_cache_25_cast_fp16 = add(x = var_3288_cast_fp16, y = var_3295_cast_fp16)[name = tensor("new_k_cache_25_cast_fp16")]; + tensor new_k_cache_25_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_25_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache12_to_fp16_dtype_0 = const()[name = tensor("v_cache12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache12_to_fp16 = cast(dtype = v_cache12_to_fp16_dtype_0, x = v_cache12)[name = tensor("cast_477")]; + tensor var_3301_cast_fp16 = mul(x = v_cache12_to_fp16, y = var_3287_cast_fp16)[name = tensor("op_3301_cast_fp16")]; + tensor var_3306 = const()[name = tensor("op_3306"), val = tensor([1, 1, 16, 64])]; + tensor var_3307_cast_fp16 = reshape(shape = var_3306, x = v_25_cast_fp16)[name = tensor("op_3307_cast_fp16")]; + tensor var_3308_cast_fp16 = mul(x = var_3307_cast_fp16, y = onehot_51_cast_fp16)[name = tensor("op_3308_cast_fp16")]; + tensor new_v_cache_25_cast_fp16 = add(x = var_3301_cast_fp16, y = var_3308_cast_fp16)[name = tensor("new_v_cache_25_cast_fp16")]; + tensor new_v_cache_25_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_25_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_12_perm_0 = const()[name = tensor("transpose_12_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_3321 = const()[name = tensor("op_3321"), val = tensor([0, 2, 1, 3])]; + tensor var_3327 = const()[name = tensor("op_3327"), val = tensor([1, 16, 1, 64])]; + tensor q4_25_cast_fp16 = reshape(shape = var_3327, x = out_49_cast_fp16)[name = tensor("q4_25_cast_fp16")]; + tensor var_3332_transpose_x_0 = const()[name = tensor("op_3332_transpose_x_0"), val = tensor(false)]; + tensor var_3332_transpose_y_0 = const()[name = tensor("op_3332_transpose_y_0"), val = tensor(false)]; + tensor transpose_12_cast_fp16 = transpose(perm = transpose_12_perm_0, x = new_k_cache_25_cast_fp16)[name = tensor("transpose_59")]; + tensor var_3332_cast_fp16 = matmul(transpose_x = var_3332_transpose_x_0, transpose_y = var_3332_transpose_y_0, x = q4_25_cast_fp16, y = transpose_12_cast_fp16)[name = tensor("op_3332_cast_fp16")]; + tensor var_3333_to_fp16 = const()[name = tensor("op_3333_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_49_cast_fp16 = mul(x = var_3332_cast_fp16, y = var_3333_to_fp16)[name = tensor("scores_49_cast_fp16")]; + tensor var_3337 = const()[name = tensor("op_3337"), val = tensor([1, 1])]; + tensor var_3338_cast_fp16 = reshape(shape = var_3337, x = position12_to_fp16)[name = tensor("op_3338_cast_fp16")]; + tensor var_3339_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_3338_cast_fp16)[name = tensor("op_3339_cast_fp16")]; + tensor mask_25_to_fp16_dtype_0 = const()[name = tensor("mask_25_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_3346_to_fp16 = const()[name = tensor("op_3346_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3339_cast_fp16_to_fp16 = cast(dtype = mask_25_to_fp16_dtype_0, x = var_3339_cast_fp16)[name = tensor("cast_475")]; + tensor var_3347_cast_fp16 = sub(x = var_3339_cast_fp16_to_fp16, y = var_3346_to_fp16)[name = tensor("op_3347_cast_fp16")]; + tensor var_3352 = const()[name = tensor("op_3352"), val = tensor([1, 1, 1, 512])]; + tensor var_3353_cast_fp16 = reshape(shape = var_3352, x = var_3347_cast_fp16)[name = tensor("op_3353_cast_fp16")]; + tensor var_3354_to_fp16 = const()[name = tensor("op_3354_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_3355_cast_fp16 = mul(x = var_3353_cast_fp16, y = var_3354_to_fp16)[name = tensor("op_3355_cast_fp16")]; + tensor scores_51_cast_fp16 = add(x = scores_49_cast_fp16, y = var_3355_cast_fp16)[name = tensor("scores_51_cast_fp16")]; + tensor var_3358 = const()[name = tensor("op_3358"), val = tensor(-1)]; + tensor weights_25_cast_fp16 = softmax(axis = var_3358, x = scores_51_cast_fp16)[name = tensor("weights_25_cast_fp16")]; + tensor attn_25_transpose_x_0 = const()[name = tensor("attn_25_transpose_x_0"), val = tensor(false)]; + tensor attn_25_transpose_y_0 = const()[name = tensor("attn_25_transpose_y_0"), val = tensor(false)]; + tensor values_25_cast_fp16 = transpose(perm = var_3321, x = new_v_cache_25_cast_fp16)[name = tensor("transpose_58")]; + tensor attn_25_cast_fp16 = matmul(transpose_x = attn_25_transpose_x_0, transpose_y = attn_25_transpose_y_0, x = weights_25_cast_fp16, y = values_25_cast_fp16)[name = tensor("attn_25_cast_fp16")]; + tensor var_3366 = const()[name = tensor("op_3366"), val = tensor([0, 2, 1, 3])]; + tensor var_3371 = const()[name = tensor("op_3371"), val = tensor([1, 1, 1024])]; + tensor var_3367_cast_fp16 = transpose(perm = var_3366, x = attn_25_cast_fp16)[name = tensor("transpose_57")]; + tensor input_151_cast_fp16 = reshape(shape = var_3371, x = var_3367_cast_fp16)[name = tensor("input_151_cast_fp16")]; + tensor attn12_out_proj_weight_to_fp16 = const()[name = tensor("attn12_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(308473536)))]; + tensor linear_50_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn12_out_proj_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_50_cast_fp16")]; + tensor var_3377_to_fp16 = const()[name = tensor("op_3377_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3378_cast_fp16 = add(x = position12_to_fp16, y = var_3377_to_fp16)[name = tensor("op_3378_cast_fp16")]; + tensor var_3378_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3378_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_153_cast_fp16 = add(x = input_147_cast_fp16, y = linear_50_cast_fp16)[name = tensor("input_153_cast_fp16")]; + tensor input_155_axes_0 = const()[name = tensor("input_155_axes_0"), val = tensor([-1])]; + tensor norm12_2_weight_to_fp16 = const()[name = tensor("norm12_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(310570752)))]; + tensor norm12_2_bias_to_fp16 = const()[name = tensor("norm12_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(310572864)))]; + tensor var_3382_to_fp16 = const()[name = tensor("op_3382_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_155_cast_fp16 = layer_norm(axes = input_155_axes_0, beta = norm12_2_bias_to_fp16, epsilon = var_3382_to_fp16, gamma = norm12_2_weight_to_fp16, x = input_153_cast_fp16)[name = tensor("input_155_cast_fp16")]; + tensor linear12_1_weight_to_fp16 = const()[name = tensor("linear12_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(310574976)))]; + tensor linear_51_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear12_1_weight_to_fp16, x = input_155_cast_fp16)[name = tensor("linear_51_cast_fp16")]; + tensor input_157_mode_0 = const()[name = tensor("input_157_mode_0"), val = tensor("EXACT")]; + tensor input_157_cast_fp16 = gelu(mode = input_157_mode_0, x = linear_51_cast_fp16)[name = tensor("input_157_cast_fp16")]; + tensor linear12_2_weight_to_fp16 = const()[name = tensor("linear12_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(318963648)))]; + tensor linear_52_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear12_2_weight_to_fp16, x = input_157_cast_fp16)[name = tensor("linear_52_cast_fp16")]; + tensor input_159_cast_fp16 = add(x = input_153_cast_fp16, y = linear_52_cast_fp16)[name = tensor("input_159_cast_fp16")]; + tensor input_161_axes_0 = const()[name = tensor("input_161_axes_0"), val = tensor([-1])]; + tensor norm13_1_weight_to_fp16 = const()[name = tensor("norm13_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(327352320)))]; + tensor norm13_1_bias_to_fp16 = const()[name = tensor("norm13_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(327354432)))]; + tensor var_3399_to_fp16 = const()[name = tensor("op_3399_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_161_cast_fp16 = layer_norm(axes = input_161_axes_0, beta = norm13_1_bias_to_fp16, epsilon = var_3399_to_fp16, gamma = norm13_1_weight_to_fp16, x = input_159_cast_fp16)[name = tensor("input_161_cast_fp16")]; + tensor attn13_in_proj_weight_to_fp16 = const()[name = tensor("attn13_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(327356544)))]; + tensor linear_53_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn13_in_proj_weight_to_fp16, x = input_161_cast_fp16)[name = tensor("linear_53_cast_fp16")]; + tensor var_3412 = const()[name = tensor("op_3412"), val = tensor([1, 3, 16, 64])]; + tensor qkv_27_cast_fp16 = reshape(shape = var_3412, x = linear_53_cast_fp16)[name = tensor("qkv_27_cast_fp16")]; + tensor x_53_begin_0 = const()[name = tensor("x_53_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_53_end_0 = const()[name = tensor("x_53_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_53_end_mask_0 = const()[name = tensor("x_53_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_53_squeeze_mask_0 = const()[name = tensor("x_53_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_53_cast_fp16 = slice_by_index(begin = x_53_begin_0, end = x_53_end_0, end_mask = x_53_end_mask_0, squeeze_mask = x_53_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("x_53_cast_fp16")]; + tensor x_55_begin_0 = const()[name = tensor("x_55_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_55_end_0 = const()[name = tensor("x_55_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_55_end_mask_0 = const()[name = tensor("x_55_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_55_squeeze_mask_0 = const()[name = tensor("x_55_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_55_cast_fp16 = slice_by_index(begin = x_55_begin_0, end = x_55_end_0, end_mask = x_55_end_mask_0, squeeze_mask = x_55_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("x_55_cast_fp16")]; + tensor v_27_begin_0 = const()[name = tensor("v_27_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_27_end_0 = const()[name = tensor("v_27_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_27_end_mask_0 = const()[name = tensor("v_27_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_27_squeeze_mask_0 = const()[name = tensor("v_27_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_27_cast_fp16 = slice_by_index(begin = v_27_begin_0, end = v_27_end_0, end_mask = v_27_end_mask_0, squeeze_mask = v_27_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("v_27_cast_fp16")]; + tensor var_3441 = const()[name = tensor("op_3441"), val = tensor([1, 1, 1])]; + tensor position13_to_fp16_dtype_0 = const()[name = tensor("position13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position13_to_fp16 = cast(dtype = position13_to_fp16_dtype_0, x = position13)[name = tensor("cast_473")]; + tensor var_3442_cast_fp16 = reshape(shape = var_3441, x = position13_to_fp16)[name = tensor("op_3442_cast_fp16")]; + tensor angles_53_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_3442_cast_fp16)[name = tensor("angles_53_cast_fp16")]; + tensor rotr_53_cast_fp16 = cos(x = angles_53_cast_fp16)[name = tensor("rotr_53_cast_fp16")]; + tensor roti_53_cast_fp16 = sin(x = angles_53_cast_fp16)[name = tensor("roti_53_cast_fp16")]; + tensor var_3450 = const()[name = tensor("op_3450"), val = tensor([1, 16, 32, 2])]; + tensor pairs_53_cast_fp16 = reshape(shape = var_3450, x = x_53_cast_fp16)[name = tensor("pairs_53_cast_fp16")]; + tensor xr_53_begin_0 = const()[name = tensor("xr_53_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_53_end_0 = const()[name = tensor("xr_53_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_53_end_mask_0 = const()[name = tensor("xr_53_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_53_squeeze_mask_0 = const()[name = tensor("xr_53_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_53_cast_fp16 = slice_by_index(begin = xr_53_begin_0, end = xr_53_end_0, end_mask = xr_53_end_mask_0, squeeze_mask = xr_53_squeeze_mask_0, x = pairs_53_cast_fp16)[name = tensor("xr_53_cast_fp16")]; + tensor xi_53_begin_0 = const()[name = tensor("xi_53_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_53_end_0 = const()[name = tensor("xi_53_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_53_end_mask_0 = const()[name = tensor("xi_53_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_53_squeeze_mask_0 = const()[name = tensor("xi_53_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_53_cast_fp16 = slice_by_index(begin = xi_53_begin_0, end = xi_53_end_0, end_mask = xi_53_end_mask_0, squeeze_mask = xi_53_squeeze_mask_0, x = pairs_53_cast_fp16)[name = tensor("xi_53_cast_fp16")]; + tensor var_3458_cast_fp16 = mul(x = xr_53_cast_fp16, y = rotr_53_cast_fp16)[name = tensor("op_3458_cast_fp16")]; + tensor var_3459_cast_fp16 = mul(x = xi_53_cast_fp16, y = roti_53_cast_fp16)[name = tensor("op_3459_cast_fp16")]; + tensor xor__53_cast_fp16 = sub(x = var_3458_cast_fp16, y = var_3459_cast_fp16)[name = tensor("xor__53_cast_fp16")]; + tensor var_3462_cast_fp16 = mul(x = xr_53_cast_fp16, y = roti_53_cast_fp16)[name = tensor("op_3462_cast_fp16")]; + tensor var_3463_cast_fp16 = mul(x = xi_53_cast_fp16, y = rotr_53_cast_fp16)[name = tensor("op_3463_cast_fp16")]; + tensor xoi_53_cast_fp16 = add(x = var_3462_cast_fp16, y = var_3463_cast_fp16)[name = tensor("xoi_53_cast_fp16")]; + tensor out_53_axis_0 = const()[name = tensor("out_53_axis_0"), val = tensor(-1)]; + tensor out_53_cast_fp16 = stack(axis = out_53_axis_0, values = (xor__53_cast_fp16, xoi_53_cast_fp16))[name = tensor("out_53_cast_fp16")]; + tensor var_3486 = const()[name = tensor("op_3486"), val = tensor([1, 16, 32, 2])]; + tensor pairs_55_cast_fp16 = reshape(shape = var_3486, x = x_55_cast_fp16)[name = tensor("pairs_55_cast_fp16")]; + tensor xr_55_begin_0 = const()[name = tensor("xr_55_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_55_end_0 = const()[name = tensor("xr_55_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_55_end_mask_0 = const()[name = tensor("xr_55_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_55_squeeze_mask_0 = const()[name = tensor("xr_55_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_55_cast_fp16 = slice_by_index(begin = xr_55_begin_0, end = xr_55_end_0, end_mask = xr_55_end_mask_0, squeeze_mask = xr_55_squeeze_mask_0, x = pairs_55_cast_fp16)[name = tensor("xr_55_cast_fp16")]; + tensor xi_55_begin_0 = const()[name = tensor("xi_55_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_55_end_0 = const()[name = tensor("xi_55_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_55_end_mask_0 = const()[name = tensor("xi_55_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_55_squeeze_mask_0 = const()[name = tensor("xi_55_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_55_cast_fp16 = slice_by_index(begin = xi_55_begin_0, end = xi_55_end_0, end_mask = xi_55_end_mask_0, squeeze_mask = xi_55_squeeze_mask_0, x = pairs_55_cast_fp16)[name = tensor("xi_55_cast_fp16")]; + tensor var_3494_cast_fp16 = mul(x = xr_55_cast_fp16, y = rotr_53_cast_fp16)[name = tensor("op_3494_cast_fp16")]; + tensor var_3495_cast_fp16 = mul(x = xi_55_cast_fp16, y = roti_53_cast_fp16)[name = tensor("op_3495_cast_fp16")]; + tensor xor__55_cast_fp16 = sub(x = var_3494_cast_fp16, y = var_3495_cast_fp16)[name = tensor("xor__55_cast_fp16")]; + tensor var_3498_cast_fp16 = mul(x = xr_55_cast_fp16, y = roti_53_cast_fp16)[name = tensor("op_3498_cast_fp16")]; + tensor var_3499_cast_fp16 = mul(x = xi_55_cast_fp16, y = rotr_53_cast_fp16)[name = tensor("op_3499_cast_fp16")]; + tensor xoi_55_cast_fp16 = add(x = var_3498_cast_fp16, y = var_3499_cast_fp16)[name = tensor("xoi_55_cast_fp16")]; + tensor out_55_axis_0 = const()[name = tensor("out_55_axis_0"), val = tensor(-1)]; + tensor out_55_cast_fp16 = stack(axis = out_55_axis_0, values = (xor__55_cast_fp16, xoi_55_cast_fp16))[name = tensor("out_55_cast_fp16")]; + tensor _inversed_3511_y_0_to_fp16 = const()[name = tensor("_inversed_3511_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3511_cast_fp16 = mul(x = position13_to_fp16, y = _inversed_3511_y_0_to_fp16)[name = tensor("_inversed_3511_cast_fp16")]; + tensor var_3512_cast_fp16 = floor(x = _inversed_3511_cast_fp16)[name = tensor("op_3512_cast_fp16")]; + tensor var_3513_to_fp16 = const()[name = tensor("op_3513_to_fp16"), val = tensor(0x1p+9)]; + tensor var_3514_cast_fp16 = mul(x = var_3512_cast_fp16, y = var_3513_to_fp16)[name = tensor("op_3514_cast_fp16")]; + tensor write_idx_27_cast_fp16 = sub(x = position13_to_fp16, y = var_3514_cast_fp16)[name = tensor("write_idx_27_cast_fp16")]; + tensor var_3517_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_27_cast_fp16)[name = tensor("op_3517_cast_fp16")]; + tensor var_3527 = const()[name = tensor("op_3527"), val = tensor([1, 512, 1, 1])]; + tensor onehot_53_to_fp16_dtype_0 = const()[name = tensor("onehot_53_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_3517_cast_fp16_to_fp16 = cast(dtype = onehot_53_to_fp16_dtype_0, x = var_3517_cast_fp16)[name = tensor("cast_472")]; + tensor onehot_55_cast_fp16 = reshape(shape = var_3527, x = var_3517_cast_fp16_to_fp16)[name = tensor("onehot_55_cast_fp16")]; + tensor var_3529_to_fp16 = const()[name = tensor("op_3529_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3531_cast_fp16 = sub(x = var_3529_to_fp16, y = onehot_55_cast_fp16)[name = tensor("op_3531_cast_fp16")]; + tensor k_cache13_to_fp16_dtype_0 = const()[name = tensor("k_cache13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache13_to_fp16 = cast(dtype = k_cache13_to_fp16_dtype_0, x = k_cache13)[name = tensor("cast_471")]; + tensor var_3532_cast_fp16 = mul(x = k_cache13_to_fp16, y = var_3531_cast_fp16)[name = tensor("op_3532_cast_fp16")]; + tensor var_3537 = const()[name = tensor("op_3537"), val = tensor([1, 1, 16, 64])]; + tensor var_3538_cast_fp16 = reshape(shape = var_3537, x = out_55_cast_fp16)[name = tensor("op_3538_cast_fp16")]; + tensor var_3539_cast_fp16 = mul(x = var_3538_cast_fp16, y = onehot_55_cast_fp16)[name = tensor("op_3539_cast_fp16")]; + tensor new_k_cache_27_cast_fp16 = add(x = var_3532_cast_fp16, y = var_3539_cast_fp16)[name = tensor("new_k_cache_27_cast_fp16")]; + tensor new_k_cache_27_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_27_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache13_to_fp16_dtype_0 = const()[name = tensor("v_cache13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache13_to_fp16 = cast(dtype = v_cache13_to_fp16_dtype_0, x = v_cache13)[name = tensor("cast_469")]; + tensor var_3545_cast_fp16 = mul(x = v_cache13_to_fp16, y = var_3531_cast_fp16)[name = tensor("op_3545_cast_fp16")]; + tensor var_3550 = const()[name = tensor("op_3550"), val = tensor([1, 1, 16, 64])]; + tensor var_3551_cast_fp16 = reshape(shape = var_3550, x = v_27_cast_fp16)[name = tensor("op_3551_cast_fp16")]; + tensor var_3552_cast_fp16 = mul(x = var_3551_cast_fp16, y = onehot_55_cast_fp16)[name = tensor("op_3552_cast_fp16")]; + tensor new_v_cache_27_cast_fp16 = add(x = var_3545_cast_fp16, y = var_3552_cast_fp16)[name = tensor("new_v_cache_27_cast_fp16")]; + tensor new_v_cache_27_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_27_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_13_perm_0 = const()[name = tensor("transpose_13_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_3565 = const()[name = tensor("op_3565"), val = tensor([0, 2, 1, 3])]; + tensor var_3571 = const()[name = tensor("op_3571"), val = tensor([1, 16, 1, 64])]; + tensor q4_27_cast_fp16 = reshape(shape = var_3571, x = out_53_cast_fp16)[name = tensor("q4_27_cast_fp16")]; + tensor var_3576_transpose_x_0 = const()[name = tensor("op_3576_transpose_x_0"), val = tensor(false)]; + tensor var_3576_transpose_y_0 = const()[name = tensor("op_3576_transpose_y_0"), val = tensor(false)]; + tensor transpose_13_cast_fp16 = transpose(perm = transpose_13_perm_0, x = new_k_cache_27_cast_fp16)[name = tensor("transpose_56")]; + tensor var_3576_cast_fp16 = matmul(transpose_x = var_3576_transpose_x_0, transpose_y = var_3576_transpose_y_0, x = q4_27_cast_fp16, y = transpose_13_cast_fp16)[name = tensor("op_3576_cast_fp16")]; + tensor var_3577_to_fp16 = const()[name = tensor("op_3577_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_53_cast_fp16 = mul(x = var_3576_cast_fp16, y = var_3577_to_fp16)[name = tensor("scores_53_cast_fp16")]; + tensor var_3581 = const()[name = tensor("op_3581"), val = tensor([1, 1])]; + tensor var_3582_cast_fp16 = reshape(shape = var_3581, x = position13_to_fp16)[name = tensor("op_3582_cast_fp16")]; + tensor var_3583_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_3582_cast_fp16)[name = tensor("op_3583_cast_fp16")]; + tensor mask_27_to_fp16_dtype_0 = const()[name = tensor("mask_27_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_3590_to_fp16 = const()[name = tensor("op_3590_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3583_cast_fp16_to_fp16 = cast(dtype = mask_27_to_fp16_dtype_0, x = var_3583_cast_fp16)[name = tensor("cast_467")]; + tensor var_3591_cast_fp16 = sub(x = var_3583_cast_fp16_to_fp16, y = var_3590_to_fp16)[name = tensor("op_3591_cast_fp16")]; + tensor var_3596 = const()[name = tensor("op_3596"), val = tensor([1, 1, 1, 512])]; + tensor var_3597_cast_fp16 = reshape(shape = var_3596, x = var_3591_cast_fp16)[name = tensor("op_3597_cast_fp16")]; + tensor var_3598_to_fp16 = const()[name = tensor("op_3598_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_3599_cast_fp16 = mul(x = var_3597_cast_fp16, y = var_3598_to_fp16)[name = tensor("op_3599_cast_fp16")]; + tensor scores_55_cast_fp16 = add(x = scores_53_cast_fp16, y = var_3599_cast_fp16)[name = tensor("scores_55_cast_fp16")]; + tensor var_3602 = const()[name = tensor("op_3602"), val = tensor(-1)]; + tensor weights_27_cast_fp16 = softmax(axis = var_3602, x = scores_55_cast_fp16)[name = tensor("weights_27_cast_fp16")]; + tensor attn_27_transpose_x_0 = const()[name = tensor("attn_27_transpose_x_0"), val = tensor(false)]; + tensor attn_27_transpose_y_0 = const()[name = tensor("attn_27_transpose_y_0"), val = tensor(false)]; + tensor values_27_cast_fp16 = transpose(perm = var_3565, x = new_v_cache_27_cast_fp16)[name = tensor("transpose_55")]; + tensor attn_27_cast_fp16 = matmul(transpose_x = attn_27_transpose_x_0, transpose_y = attn_27_transpose_y_0, x = weights_27_cast_fp16, y = values_27_cast_fp16)[name = tensor("attn_27_cast_fp16")]; + tensor var_3610 = const()[name = tensor("op_3610"), val = tensor([0, 2, 1, 3])]; + tensor var_3615 = const()[name = tensor("op_3615"), val = tensor([1, 1, 1024])]; + tensor var_3611_cast_fp16 = transpose(perm = var_3610, x = attn_27_cast_fp16)[name = tensor("transpose_54")]; + tensor input_163_cast_fp16 = reshape(shape = var_3615, x = var_3611_cast_fp16)[name = tensor("input_163_cast_fp16")]; + tensor attn13_out_proj_weight_to_fp16 = const()[name = tensor("attn13_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(333648064)))]; + tensor linear_54_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn13_out_proj_weight_to_fp16, x = input_163_cast_fp16)[name = tensor("linear_54_cast_fp16")]; + tensor var_3621_to_fp16 = const()[name = tensor("op_3621_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3622_cast_fp16 = add(x = position13_to_fp16, y = var_3621_to_fp16)[name = tensor("op_3622_cast_fp16")]; + tensor var_3622_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3622_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_165_cast_fp16 = add(x = input_159_cast_fp16, y = linear_54_cast_fp16)[name = tensor("input_165_cast_fp16")]; + tensor input_167_axes_0 = const()[name = tensor("input_167_axes_0"), val = tensor([-1])]; + tensor norm13_2_weight_to_fp16 = const()[name = tensor("norm13_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335745280)))]; + tensor norm13_2_bias_to_fp16 = const()[name = tensor("norm13_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335747392)))]; + tensor var_3626_to_fp16 = const()[name = tensor("op_3626_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_167_cast_fp16 = layer_norm(axes = input_167_axes_0, beta = norm13_2_bias_to_fp16, epsilon = var_3626_to_fp16, gamma = norm13_2_weight_to_fp16, x = input_165_cast_fp16)[name = tensor("input_167_cast_fp16")]; + tensor linear13_1_weight_to_fp16 = const()[name = tensor("linear13_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335749504)))]; + tensor linear_55_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear13_1_weight_to_fp16, x = input_167_cast_fp16)[name = tensor("linear_55_cast_fp16")]; + tensor input_169_mode_0 = const()[name = tensor("input_169_mode_0"), val = tensor("EXACT")]; + tensor input_169_cast_fp16 = gelu(mode = input_169_mode_0, x = linear_55_cast_fp16)[name = tensor("input_169_cast_fp16")]; + tensor linear13_2_weight_to_fp16 = const()[name = tensor("linear13_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(344138176)))]; + tensor linear_56_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear13_2_weight_to_fp16, x = input_169_cast_fp16)[name = tensor("linear_56_cast_fp16")]; + tensor input_171_cast_fp16 = add(x = input_165_cast_fp16, y = linear_56_cast_fp16)[name = tensor("input_171_cast_fp16")]; + tensor input_173_axes_0 = const()[name = tensor("input_173_axes_0"), val = tensor([-1])]; + tensor norm14_1_weight_to_fp16 = const()[name = tensor("norm14_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(352526848)))]; + tensor norm14_1_bias_to_fp16 = const()[name = tensor("norm14_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(352528960)))]; + tensor var_3643_to_fp16 = const()[name = tensor("op_3643_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_173_cast_fp16 = layer_norm(axes = input_173_axes_0, beta = norm14_1_bias_to_fp16, epsilon = var_3643_to_fp16, gamma = norm14_1_weight_to_fp16, x = input_171_cast_fp16)[name = tensor("input_173_cast_fp16")]; + tensor attn14_in_proj_weight_to_fp16 = const()[name = tensor("attn14_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(352531072)))]; + tensor linear_57_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn14_in_proj_weight_to_fp16, x = input_173_cast_fp16)[name = tensor("linear_57_cast_fp16")]; + tensor var_3656 = const()[name = tensor("op_3656"), val = tensor([1, 3, 16, 64])]; + tensor qkv_29_cast_fp16 = reshape(shape = var_3656, x = linear_57_cast_fp16)[name = tensor("qkv_29_cast_fp16")]; + tensor x_57_begin_0 = const()[name = tensor("x_57_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_57_end_0 = const()[name = tensor("x_57_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_57_end_mask_0 = const()[name = tensor("x_57_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_57_squeeze_mask_0 = const()[name = tensor("x_57_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_57_cast_fp16 = slice_by_index(begin = x_57_begin_0, end = x_57_end_0, end_mask = x_57_end_mask_0, squeeze_mask = x_57_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("x_57_cast_fp16")]; + tensor x_59_begin_0 = const()[name = tensor("x_59_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_59_end_0 = const()[name = tensor("x_59_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_59_end_mask_0 = const()[name = tensor("x_59_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_59_squeeze_mask_0 = const()[name = tensor("x_59_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_59_cast_fp16 = slice_by_index(begin = x_59_begin_0, end = x_59_end_0, end_mask = x_59_end_mask_0, squeeze_mask = x_59_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("x_59_cast_fp16")]; + tensor v_29_begin_0 = const()[name = tensor("v_29_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_29_end_0 = const()[name = tensor("v_29_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_29_end_mask_0 = const()[name = tensor("v_29_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_29_squeeze_mask_0 = const()[name = tensor("v_29_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_29_cast_fp16 = slice_by_index(begin = v_29_begin_0, end = v_29_end_0, end_mask = v_29_end_mask_0, squeeze_mask = v_29_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("v_29_cast_fp16")]; + tensor var_3685 = const()[name = tensor("op_3685"), val = tensor([1, 1, 1])]; + tensor position14_to_fp16_dtype_0 = const()[name = tensor("position14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position14_to_fp16 = cast(dtype = position14_to_fp16_dtype_0, x = position14)[name = tensor("cast_465")]; + tensor var_3686_cast_fp16 = reshape(shape = var_3685, x = position14_to_fp16)[name = tensor("op_3686_cast_fp16")]; + tensor angles_57_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_3686_cast_fp16)[name = tensor("angles_57_cast_fp16")]; + tensor rotr_57_cast_fp16 = cos(x = angles_57_cast_fp16)[name = tensor("rotr_57_cast_fp16")]; + tensor roti_57_cast_fp16 = sin(x = angles_57_cast_fp16)[name = tensor("roti_57_cast_fp16")]; + tensor var_3694 = const()[name = tensor("op_3694"), val = tensor([1, 16, 32, 2])]; + tensor pairs_57_cast_fp16 = reshape(shape = var_3694, x = x_57_cast_fp16)[name = tensor("pairs_57_cast_fp16")]; + tensor xr_57_begin_0 = const()[name = tensor("xr_57_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_57_end_0 = const()[name = tensor("xr_57_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_57_end_mask_0 = const()[name = tensor("xr_57_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_57_squeeze_mask_0 = const()[name = tensor("xr_57_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_57_cast_fp16 = slice_by_index(begin = xr_57_begin_0, end = xr_57_end_0, end_mask = xr_57_end_mask_0, squeeze_mask = xr_57_squeeze_mask_0, x = pairs_57_cast_fp16)[name = tensor("xr_57_cast_fp16")]; + tensor xi_57_begin_0 = const()[name = tensor("xi_57_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_57_end_0 = const()[name = tensor("xi_57_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_57_end_mask_0 = const()[name = tensor("xi_57_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_57_squeeze_mask_0 = const()[name = tensor("xi_57_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_57_cast_fp16 = slice_by_index(begin = xi_57_begin_0, end = xi_57_end_0, end_mask = xi_57_end_mask_0, squeeze_mask = xi_57_squeeze_mask_0, x = pairs_57_cast_fp16)[name = tensor("xi_57_cast_fp16")]; + tensor var_3702_cast_fp16 = mul(x = xr_57_cast_fp16, y = rotr_57_cast_fp16)[name = tensor("op_3702_cast_fp16")]; + tensor var_3703_cast_fp16 = mul(x = xi_57_cast_fp16, y = roti_57_cast_fp16)[name = tensor("op_3703_cast_fp16")]; + tensor xor__57_cast_fp16 = sub(x = var_3702_cast_fp16, y = var_3703_cast_fp16)[name = tensor("xor__57_cast_fp16")]; + tensor var_3706_cast_fp16 = mul(x = xr_57_cast_fp16, y = roti_57_cast_fp16)[name = tensor("op_3706_cast_fp16")]; + tensor var_3707_cast_fp16 = mul(x = xi_57_cast_fp16, y = rotr_57_cast_fp16)[name = tensor("op_3707_cast_fp16")]; + tensor xoi_57_cast_fp16 = add(x = var_3706_cast_fp16, y = var_3707_cast_fp16)[name = tensor("xoi_57_cast_fp16")]; + tensor out_57_axis_0 = const()[name = tensor("out_57_axis_0"), val = tensor(-1)]; + tensor out_57_cast_fp16 = stack(axis = out_57_axis_0, values = (xor__57_cast_fp16, xoi_57_cast_fp16))[name = tensor("out_57_cast_fp16")]; + tensor var_3730 = const()[name = tensor("op_3730"), val = tensor([1, 16, 32, 2])]; + tensor pairs_59_cast_fp16 = reshape(shape = var_3730, x = x_59_cast_fp16)[name = tensor("pairs_59_cast_fp16")]; + tensor xr_59_begin_0 = const()[name = tensor("xr_59_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_59_end_0 = const()[name = tensor("xr_59_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_59_end_mask_0 = const()[name = tensor("xr_59_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_59_squeeze_mask_0 = const()[name = tensor("xr_59_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_59_cast_fp16 = slice_by_index(begin = xr_59_begin_0, end = xr_59_end_0, end_mask = xr_59_end_mask_0, squeeze_mask = xr_59_squeeze_mask_0, x = pairs_59_cast_fp16)[name = tensor("xr_59_cast_fp16")]; + tensor xi_59_begin_0 = const()[name = tensor("xi_59_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_59_end_0 = const()[name = tensor("xi_59_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_59_end_mask_0 = const()[name = tensor("xi_59_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_59_squeeze_mask_0 = const()[name = tensor("xi_59_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_59_cast_fp16 = slice_by_index(begin = xi_59_begin_0, end = xi_59_end_0, end_mask = xi_59_end_mask_0, squeeze_mask = xi_59_squeeze_mask_0, x = pairs_59_cast_fp16)[name = tensor("xi_59_cast_fp16")]; + tensor var_3738_cast_fp16 = mul(x = xr_59_cast_fp16, y = rotr_57_cast_fp16)[name = tensor("op_3738_cast_fp16")]; + tensor var_3739_cast_fp16 = mul(x = xi_59_cast_fp16, y = roti_57_cast_fp16)[name = tensor("op_3739_cast_fp16")]; + tensor xor__59_cast_fp16 = sub(x = var_3738_cast_fp16, y = var_3739_cast_fp16)[name = tensor("xor__59_cast_fp16")]; + tensor var_3742_cast_fp16 = mul(x = xr_59_cast_fp16, y = roti_57_cast_fp16)[name = tensor("op_3742_cast_fp16")]; + tensor var_3743_cast_fp16 = mul(x = xi_59_cast_fp16, y = rotr_57_cast_fp16)[name = tensor("op_3743_cast_fp16")]; + tensor xoi_59_cast_fp16 = add(x = var_3742_cast_fp16, y = var_3743_cast_fp16)[name = tensor("xoi_59_cast_fp16")]; + tensor out_59_axis_0 = const()[name = tensor("out_59_axis_0"), val = tensor(-1)]; + tensor out_59_cast_fp16 = stack(axis = out_59_axis_0, values = (xor__59_cast_fp16, xoi_59_cast_fp16))[name = tensor("out_59_cast_fp16")]; + tensor _inversed_3755_y_0_to_fp16 = const()[name = tensor("_inversed_3755_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3755_cast_fp16 = mul(x = position14_to_fp16, y = _inversed_3755_y_0_to_fp16)[name = tensor("_inversed_3755_cast_fp16")]; + tensor var_3756_cast_fp16 = floor(x = _inversed_3755_cast_fp16)[name = tensor("op_3756_cast_fp16")]; + tensor var_3757_to_fp16 = const()[name = tensor("op_3757_to_fp16"), val = tensor(0x1p+9)]; + tensor var_3758_cast_fp16 = mul(x = var_3756_cast_fp16, y = var_3757_to_fp16)[name = tensor("op_3758_cast_fp16")]; + tensor write_idx_29_cast_fp16 = sub(x = position14_to_fp16, y = var_3758_cast_fp16)[name = tensor("write_idx_29_cast_fp16")]; + tensor var_3761_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_29_cast_fp16)[name = tensor("op_3761_cast_fp16")]; + tensor var_3771 = const()[name = tensor("op_3771"), val = tensor([1, 512, 1, 1])]; + tensor onehot_57_to_fp16_dtype_0 = const()[name = tensor("onehot_57_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_3761_cast_fp16_to_fp16 = cast(dtype = onehot_57_to_fp16_dtype_0, x = var_3761_cast_fp16)[name = tensor("cast_464")]; + tensor onehot_59_cast_fp16 = reshape(shape = var_3771, x = var_3761_cast_fp16_to_fp16)[name = tensor("onehot_59_cast_fp16")]; + tensor var_3773_to_fp16 = const()[name = tensor("op_3773_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3775_cast_fp16 = sub(x = var_3773_to_fp16, y = onehot_59_cast_fp16)[name = tensor("op_3775_cast_fp16")]; + tensor k_cache14_to_fp16_dtype_0 = const()[name = tensor("k_cache14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache14_to_fp16 = cast(dtype = k_cache14_to_fp16_dtype_0, x = k_cache14)[name = tensor("cast_463")]; + tensor var_3776_cast_fp16 = mul(x = k_cache14_to_fp16, y = var_3775_cast_fp16)[name = tensor("op_3776_cast_fp16")]; + tensor var_3781 = const()[name = tensor("op_3781"), val = tensor([1, 1, 16, 64])]; + tensor var_3782_cast_fp16 = reshape(shape = var_3781, x = out_59_cast_fp16)[name = tensor("op_3782_cast_fp16")]; + tensor var_3783_cast_fp16 = mul(x = var_3782_cast_fp16, y = onehot_59_cast_fp16)[name = tensor("op_3783_cast_fp16")]; + tensor new_k_cache_29_cast_fp16 = add(x = var_3776_cast_fp16, y = var_3783_cast_fp16)[name = tensor("new_k_cache_29_cast_fp16")]; + tensor new_k_cache_29_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_29_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache14_to_fp16_dtype_0 = const()[name = tensor("v_cache14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache14_to_fp16 = cast(dtype = v_cache14_to_fp16_dtype_0, x = v_cache14)[name = tensor("cast_461")]; + tensor var_3789_cast_fp16 = mul(x = v_cache14_to_fp16, y = var_3775_cast_fp16)[name = tensor("op_3789_cast_fp16")]; + tensor var_3794 = const()[name = tensor("op_3794"), val = tensor([1, 1, 16, 64])]; + tensor var_3795_cast_fp16 = reshape(shape = var_3794, x = v_29_cast_fp16)[name = tensor("op_3795_cast_fp16")]; + tensor var_3796_cast_fp16 = mul(x = var_3795_cast_fp16, y = onehot_59_cast_fp16)[name = tensor("op_3796_cast_fp16")]; + tensor new_v_cache_29_cast_fp16 = add(x = var_3789_cast_fp16, y = var_3796_cast_fp16)[name = tensor("new_v_cache_29_cast_fp16")]; + tensor new_v_cache_29_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_29_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_14_perm_0 = const()[name = tensor("transpose_14_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_3809 = const()[name = tensor("op_3809"), val = tensor([0, 2, 1, 3])]; + tensor var_3815 = const()[name = tensor("op_3815"), val = tensor([1, 16, 1, 64])]; + tensor q4_29_cast_fp16 = reshape(shape = var_3815, x = out_57_cast_fp16)[name = tensor("q4_29_cast_fp16")]; + tensor var_3820_transpose_x_0 = const()[name = tensor("op_3820_transpose_x_0"), val = tensor(false)]; + tensor var_3820_transpose_y_0 = const()[name = tensor("op_3820_transpose_y_0"), val = tensor(false)]; + tensor transpose_14_cast_fp16 = transpose(perm = transpose_14_perm_0, x = new_k_cache_29_cast_fp16)[name = tensor("transpose_53")]; + tensor var_3820_cast_fp16 = matmul(transpose_x = var_3820_transpose_x_0, transpose_y = var_3820_transpose_y_0, x = q4_29_cast_fp16, y = transpose_14_cast_fp16)[name = tensor("op_3820_cast_fp16")]; + tensor var_3821_to_fp16 = const()[name = tensor("op_3821_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_57_cast_fp16 = mul(x = var_3820_cast_fp16, y = var_3821_to_fp16)[name = tensor("scores_57_cast_fp16")]; + tensor var_3825 = const()[name = tensor("op_3825"), val = tensor([1, 1])]; + tensor var_3826_cast_fp16 = reshape(shape = var_3825, x = position14_to_fp16)[name = tensor("op_3826_cast_fp16")]; + tensor var_3827_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_3826_cast_fp16)[name = tensor("op_3827_cast_fp16")]; + tensor mask_29_to_fp16_dtype_0 = const()[name = tensor("mask_29_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_3834_to_fp16 = const()[name = tensor("op_3834_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3827_cast_fp16_to_fp16 = cast(dtype = mask_29_to_fp16_dtype_0, x = var_3827_cast_fp16)[name = tensor("cast_459")]; + tensor var_3835_cast_fp16 = sub(x = var_3827_cast_fp16_to_fp16, y = var_3834_to_fp16)[name = tensor("op_3835_cast_fp16")]; + tensor var_3840 = const()[name = tensor("op_3840"), val = tensor([1, 1, 1, 512])]; + tensor var_3841_cast_fp16 = reshape(shape = var_3840, x = var_3835_cast_fp16)[name = tensor("op_3841_cast_fp16")]; + tensor var_3842_to_fp16 = const()[name = tensor("op_3842_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_3843_cast_fp16 = mul(x = var_3841_cast_fp16, y = var_3842_to_fp16)[name = tensor("op_3843_cast_fp16")]; + tensor scores_59_cast_fp16 = add(x = scores_57_cast_fp16, y = var_3843_cast_fp16)[name = tensor("scores_59_cast_fp16")]; + tensor var_3846 = const()[name = tensor("op_3846"), val = tensor(-1)]; + tensor weights_29_cast_fp16 = softmax(axis = var_3846, x = scores_59_cast_fp16)[name = tensor("weights_29_cast_fp16")]; + tensor attn_29_transpose_x_0 = const()[name = tensor("attn_29_transpose_x_0"), val = tensor(false)]; + tensor attn_29_transpose_y_0 = const()[name = tensor("attn_29_transpose_y_0"), val = tensor(false)]; + tensor values_29_cast_fp16 = transpose(perm = var_3809, x = new_v_cache_29_cast_fp16)[name = tensor("transpose_52")]; + tensor attn_29_cast_fp16 = matmul(transpose_x = attn_29_transpose_x_0, transpose_y = attn_29_transpose_y_0, x = weights_29_cast_fp16, y = values_29_cast_fp16)[name = tensor("attn_29_cast_fp16")]; + tensor var_3854 = const()[name = tensor("op_3854"), val = tensor([0, 2, 1, 3])]; + tensor var_3859 = const()[name = tensor("op_3859"), val = tensor([1, 1, 1024])]; + tensor var_3855_cast_fp16 = transpose(perm = var_3854, x = attn_29_cast_fp16)[name = tensor("transpose_51")]; + tensor input_175_cast_fp16 = reshape(shape = var_3859, x = var_3855_cast_fp16)[name = tensor("input_175_cast_fp16")]; + tensor attn14_out_proj_weight_to_fp16 = const()[name = tensor("attn14_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(358822592)))]; + tensor linear_58_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn14_out_proj_weight_to_fp16, x = input_175_cast_fp16)[name = tensor("linear_58_cast_fp16")]; + tensor var_3865_to_fp16 = const()[name = tensor("op_3865_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3866_cast_fp16 = add(x = position14_to_fp16, y = var_3865_to_fp16)[name = tensor("op_3866_cast_fp16")]; + tensor var_3866_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3866_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_177_cast_fp16 = add(x = input_171_cast_fp16, y = linear_58_cast_fp16)[name = tensor("input_177_cast_fp16")]; + tensor input_179_axes_0 = const()[name = tensor("input_179_axes_0"), val = tensor([-1])]; + tensor norm14_2_weight_to_fp16 = const()[name = tensor("norm14_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(360919808)))]; + tensor norm14_2_bias_to_fp16 = const()[name = tensor("norm14_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(360921920)))]; + tensor var_3870_to_fp16 = const()[name = tensor("op_3870_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_179_cast_fp16 = layer_norm(axes = input_179_axes_0, beta = norm14_2_bias_to_fp16, epsilon = var_3870_to_fp16, gamma = norm14_2_weight_to_fp16, x = input_177_cast_fp16)[name = tensor("input_179_cast_fp16")]; + tensor linear14_1_weight_to_fp16 = const()[name = tensor("linear14_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(360924032)))]; + tensor linear_59_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear14_1_weight_to_fp16, x = input_179_cast_fp16)[name = tensor("linear_59_cast_fp16")]; + tensor input_181_mode_0 = const()[name = tensor("input_181_mode_0"), val = tensor("EXACT")]; + tensor input_181_cast_fp16 = gelu(mode = input_181_mode_0, x = linear_59_cast_fp16)[name = tensor("input_181_cast_fp16")]; + tensor linear14_2_weight_to_fp16 = const()[name = tensor("linear14_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(369312704)))]; + tensor linear_60_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear14_2_weight_to_fp16, x = input_181_cast_fp16)[name = tensor("linear_60_cast_fp16")]; + tensor input_183_cast_fp16 = add(x = input_177_cast_fp16, y = linear_60_cast_fp16)[name = tensor("input_183_cast_fp16")]; + tensor input_185_axes_0 = const()[name = tensor("input_185_axes_0"), val = tensor([-1])]; + tensor norm15_1_weight_to_fp16 = const()[name = tensor("norm15_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(377701376)))]; + tensor norm15_1_bias_to_fp16 = const()[name = tensor("norm15_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(377703488)))]; + tensor var_3887_to_fp16 = const()[name = tensor("op_3887_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_185_cast_fp16 = layer_norm(axes = input_185_axes_0, beta = norm15_1_bias_to_fp16, epsilon = var_3887_to_fp16, gamma = norm15_1_weight_to_fp16, x = input_183_cast_fp16)[name = tensor("input_185_cast_fp16")]; + tensor attn15_in_proj_weight_to_fp16 = const()[name = tensor("attn15_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(377705600)))]; + tensor linear_61_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn15_in_proj_weight_to_fp16, x = input_185_cast_fp16)[name = tensor("linear_61_cast_fp16")]; + tensor var_3900 = const()[name = tensor("op_3900"), val = tensor([1, 3, 16, 64])]; + tensor qkv_31_cast_fp16 = reshape(shape = var_3900, x = linear_61_cast_fp16)[name = tensor("qkv_31_cast_fp16")]; + tensor x_61_begin_0 = const()[name = tensor("x_61_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_61_end_0 = const()[name = tensor("x_61_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_61_end_mask_0 = const()[name = tensor("x_61_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_61_squeeze_mask_0 = const()[name = tensor("x_61_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_61_cast_fp16 = slice_by_index(begin = x_61_begin_0, end = x_61_end_0, end_mask = x_61_end_mask_0, squeeze_mask = x_61_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("x_61_cast_fp16")]; + tensor x_63_begin_0 = const()[name = tensor("x_63_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_63_end_0 = const()[name = tensor("x_63_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_63_end_mask_0 = const()[name = tensor("x_63_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_63_squeeze_mask_0 = const()[name = tensor("x_63_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_63_cast_fp16 = slice_by_index(begin = x_63_begin_0, end = x_63_end_0, end_mask = x_63_end_mask_0, squeeze_mask = x_63_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("x_63_cast_fp16")]; + tensor v_31_begin_0 = const()[name = tensor("v_31_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_31_end_0 = const()[name = tensor("v_31_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_31_end_mask_0 = const()[name = tensor("v_31_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_31_squeeze_mask_0 = const()[name = tensor("v_31_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_31_cast_fp16 = slice_by_index(begin = v_31_begin_0, end = v_31_end_0, end_mask = v_31_end_mask_0, squeeze_mask = v_31_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("v_31_cast_fp16")]; + tensor var_3929 = const()[name = tensor("op_3929"), val = tensor([1, 1, 1])]; + tensor position15_to_fp16_dtype_0 = const()[name = tensor("position15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position15_to_fp16 = cast(dtype = position15_to_fp16_dtype_0, x = position15)[name = tensor("cast_457")]; + tensor var_3930_cast_fp16 = reshape(shape = var_3929, x = position15_to_fp16)[name = tensor("op_3930_cast_fp16")]; + tensor angles_61_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_3930_cast_fp16)[name = tensor("angles_61_cast_fp16")]; + tensor rotr_61_cast_fp16 = cos(x = angles_61_cast_fp16)[name = tensor("rotr_61_cast_fp16")]; + tensor roti_61_cast_fp16 = sin(x = angles_61_cast_fp16)[name = tensor("roti_61_cast_fp16")]; + tensor var_3938 = const()[name = tensor("op_3938"), val = tensor([1, 16, 32, 2])]; + tensor pairs_61_cast_fp16 = reshape(shape = var_3938, x = x_61_cast_fp16)[name = tensor("pairs_61_cast_fp16")]; + tensor xr_61_begin_0 = const()[name = tensor("xr_61_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_61_end_0 = const()[name = tensor("xr_61_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_61_end_mask_0 = const()[name = tensor("xr_61_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_61_squeeze_mask_0 = const()[name = tensor("xr_61_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_61_cast_fp16 = slice_by_index(begin = xr_61_begin_0, end = xr_61_end_0, end_mask = xr_61_end_mask_0, squeeze_mask = xr_61_squeeze_mask_0, x = pairs_61_cast_fp16)[name = tensor("xr_61_cast_fp16")]; + tensor xi_61_begin_0 = const()[name = tensor("xi_61_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_61_end_0 = const()[name = tensor("xi_61_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_61_end_mask_0 = const()[name = tensor("xi_61_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_61_squeeze_mask_0 = const()[name = tensor("xi_61_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_61_cast_fp16 = slice_by_index(begin = xi_61_begin_0, end = xi_61_end_0, end_mask = xi_61_end_mask_0, squeeze_mask = xi_61_squeeze_mask_0, x = pairs_61_cast_fp16)[name = tensor("xi_61_cast_fp16")]; + tensor var_3946_cast_fp16 = mul(x = xr_61_cast_fp16, y = rotr_61_cast_fp16)[name = tensor("op_3946_cast_fp16")]; + tensor var_3947_cast_fp16 = mul(x = xi_61_cast_fp16, y = roti_61_cast_fp16)[name = tensor("op_3947_cast_fp16")]; + tensor xor__61_cast_fp16 = sub(x = var_3946_cast_fp16, y = var_3947_cast_fp16)[name = tensor("xor__61_cast_fp16")]; + tensor var_3950_cast_fp16 = mul(x = xr_61_cast_fp16, y = roti_61_cast_fp16)[name = tensor("op_3950_cast_fp16")]; + tensor var_3951_cast_fp16 = mul(x = xi_61_cast_fp16, y = rotr_61_cast_fp16)[name = tensor("op_3951_cast_fp16")]; + tensor xoi_61_cast_fp16 = add(x = var_3950_cast_fp16, y = var_3951_cast_fp16)[name = tensor("xoi_61_cast_fp16")]; + tensor out_61_axis_0 = const()[name = tensor("out_61_axis_0"), val = tensor(-1)]; + tensor out_61_cast_fp16 = stack(axis = out_61_axis_0, values = (xor__61_cast_fp16, xoi_61_cast_fp16))[name = tensor("out_61_cast_fp16")]; + tensor var_3974 = const()[name = tensor("op_3974"), val = tensor([1, 16, 32, 2])]; + tensor pairs_63_cast_fp16 = reshape(shape = var_3974, x = x_63_cast_fp16)[name = tensor("pairs_63_cast_fp16")]; + tensor xr_63_begin_0 = const()[name = tensor("xr_63_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_63_end_0 = const()[name = tensor("xr_63_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_63_end_mask_0 = const()[name = tensor("xr_63_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_63_squeeze_mask_0 = const()[name = tensor("xr_63_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_63_cast_fp16 = slice_by_index(begin = xr_63_begin_0, end = xr_63_end_0, end_mask = xr_63_end_mask_0, squeeze_mask = xr_63_squeeze_mask_0, x = pairs_63_cast_fp16)[name = tensor("xr_63_cast_fp16")]; + tensor xi_63_begin_0 = const()[name = tensor("xi_63_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_63_end_0 = const()[name = tensor("xi_63_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_63_end_mask_0 = const()[name = tensor("xi_63_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_63_squeeze_mask_0 = const()[name = tensor("xi_63_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_63_cast_fp16 = slice_by_index(begin = xi_63_begin_0, end = xi_63_end_0, end_mask = xi_63_end_mask_0, squeeze_mask = xi_63_squeeze_mask_0, x = pairs_63_cast_fp16)[name = tensor("xi_63_cast_fp16")]; + tensor var_3982_cast_fp16 = mul(x = xr_63_cast_fp16, y = rotr_61_cast_fp16)[name = tensor("op_3982_cast_fp16")]; + tensor var_3983_cast_fp16 = mul(x = xi_63_cast_fp16, y = roti_61_cast_fp16)[name = tensor("op_3983_cast_fp16")]; + tensor xor__63_cast_fp16 = sub(x = var_3982_cast_fp16, y = var_3983_cast_fp16)[name = tensor("xor__63_cast_fp16")]; + tensor var_3986_cast_fp16 = mul(x = xr_63_cast_fp16, y = roti_61_cast_fp16)[name = tensor("op_3986_cast_fp16")]; + tensor var_3987_cast_fp16 = mul(x = xi_63_cast_fp16, y = rotr_61_cast_fp16)[name = tensor("op_3987_cast_fp16")]; + tensor xoi_63_cast_fp16 = add(x = var_3986_cast_fp16, y = var_3987_cast_fp16)[name = tensor("xoi_63_cast_fp16")]; + tensor out_63_axis_0 = const()[name = tensor("out_63_axis_0"), val = tensor(-1)]; + tensor out_63_cast_fp16 = stack(axis = out_63_axis_0, values = (xor__63_cast_fp16, xoi_63_cast_fp16))[name = tensor("out_63_cast_fp16")]; + tensor _inversed_3999_y_0_to_fp16 = const()[name = tensor("_inversed_3999_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3999_cast_fp16 = mul(x = position15_to_fp16, y = _inversed_3999_y_0_to_fp16)[name = tensor("_inversed_3999_cast_fp16")]; + tensor var_4000_cast_fp16 = floor(x = _inversed_3999_cast_fp16)[name = tensor("op_4000_cast_fp16")]; + tensor var_4001_to_fp16 = const()[name = tensor("op_4001_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4002_cast_fp16 = mul(x = var_4000_cast_fp16, y = var_4001_to_fp16)[name = tensor("op_4002_cast_fp16")]; + tensor write_idx_31_cast_fp16 = sub(x = position15_to_fp16, y = var_4002_cast_fp16)[name = tensor("write_idx_31_cast_fp16")]; + tensor var_4005_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_31_cast_fp16)[name = tensor("op_4005_cast_fp16")]; + tensor var_4015 = const()[name = tensor("op_4015"), val = tensor([1, 512, 1, 1])]; + tensor onehot_61_to_fp16_dtype_0 = const()[name = tensor("onehot_61_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_4005_cast_fp16_to_fp16 = cast(dtype = onehot_61_to_fp16_dtype_0, x = var_4005_cast_fp16)[name = tensor("cast_456")]; + tensor onehot_63_cast_fp16 = reshape(shape = var_4015, x = var_4005_cast_fp16_to_fp16)[name = tensor("onehot_63_cast_fp16")]; + tensor var_4017_to_fp16 = const()[name = tensor("op_4017_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4019_cast_fp16 = sub(x = var_4017_to_fp16, y = onehot_63_cast_fp16)[name = tensor("op_4019_cast_fp16")]; + tensor k_cache15_to_fp16_dtype_0 = const()[name = tensor("k_cache15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache15_to_fp16 = cast(dtype = k_cache15_to_fp16_dtype_0, x = k_cache15)[name = tensor("cast_455")]; + tensor var_4020_cast_fp16 = mul(x = k_cache15_to_fp16, y = var_4019_cast_fp16)[name = tensor("op_4020_cast_fp16")]; + tensor var_4025 = const()[name = tensor("op_4025"), val = tensor([1, 1, 16, 64])]; + tensor var_4026_cast_fp16 = reshape(shape = var_4025, x = out_63_cast_fp16)[name = tensor("op_4026_cast_fp16")]; + tensor var_4027_cast_fp16 = mul(x = var_4026_cast_fp16, y = onehot_63_cast_fp16)[name = tensor("op_4027_cast_fp16")]; + tensor new_k_cache_31_cast_fp16 = add(x = var_4020_cast_fp16, y = var_4027_cast_fp16)[name = tensor("new_k_cache_31_cast_fp16")]; + tensor new_k_cache_31_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_31_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache15_to_fp16_dtype_0 = const()[name = tensor("v_cache15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache15_to_fp16 = cast(dtype = v_cache15_to_fp16_dtype_0, x = v_cache15)[name = tensor("cast_453")]; + tensor var_4033_cast_fp16 = mul(x = v_cache15_to_fp16, y = var_4019_cast_fp16)[name = tensor("op_4033_cast_fp16")]; + tensor var_4038 = const()[name = tensor("op_4038"), val = tensor([1, 1, 16, 64])]; + tensor var_4039_cast_fp16 = reshape(shape = var_4038, x = v_31_cast_fp16)[name = tensor("op_4039_cast_fp16")]; + tensor var_4040_cast_fp16 = mul(x = var_4039_cast_fp16, y = onehot_63_cast_fp16)[name = tensor("op_4040_cast_fp16")]; + tensor new_v_cache_31_cast_fp16 = add(x = var_4033_cast_fp16, y = var_4040_cast_fp16)[name = tensor("new_v_cache_31_cast_fp16")]; + tensor new_v_cache_31_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_31_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_15_perm_0 = const()[name = tensor("transpose_15_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_4053 = const()[name = tensor("op_4053"), val = tensor([0, 2, 1, 3])]; + tensor var_4059 = const()[name = tensor("op_4059"), val = tensor([1, 16, 1, 64])]; + tensor q4_31_cast_fp16 = reshape(shape = var_4059, x = out_61_cast_fp16)[name = tensor("q4_31_cast_fp16")]; + tensor var_4064_transpose_x_0 = const()[name = tensor("op_4064_transpose_x_0"), val = tensor(false)]; + tensor var_4064_transpose_y_0 = const()[name = tensor("op_4064_transpose_y_0"), val = tensor(false)]; + tensor transpose_15_cast_fp16 = transpose(perm = transpose_15_perm_0, x = new_k_cache_31_cast_fp16)[name = tensor("transpose_50")]; + tensor var_4064_cast_fp16 = matmul(transpose_x = var_4064_transpose_x_0, transpose_y = var_4064_transpose_y_0, x = q4_31_cast_fp16, y = transpose_15_cast_fp16)[name = tensor("op_4064_cast_fp16")]; + tensor var_4065_to_fp16 = const()[name = tensor("op_4065_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_61_cast_fp16 = mul(x = var_4064_cast_fp16, y = var_4065_to_fp16)[name = tensor("scores_61_cast_fp16")]; + tensor var_4069 = const()[name = tensor("op_4069"), val = tensor([1, 1])]; + tensor var_4070_cast_fp16 = reshape(shape = var_4069, x = position15_to_fp16)[name = tensor("op_4070_cast_fp16")]; + tensor var_4071_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_4070_cast_fp16)[name = tensor("op_4071_cast_fp16")]; + tensor mask_31_to_fp16_dtype_0 = const()[name = tensor("mask_31_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_4078_to_fp16 = const()[name = tensor("op_4078_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4071_cast_fp16_to_fp16 = cast(dtype = mask_31_to_fp16_dtype_0, x = var_4071_cast_fp16)[name = tensor("cast_451")]; + tensor var_4079_cast_fp16 = sub(x = var_4071_cast_fp16_to_fp16, y = var_4078_to_fp16)[name = tensor("op_4079_cast_fp16")]; + tensor var_4084 = const()[name = tensor("op_4084"), val = tensor([1, 1, 1, 512])]; + tensor var_4085_cast_fp16 = reshape(shape = var_4084, x = var_4079_cast_fp16)[name = tensor("op_4085_cast_fp16")]; + tensor var_4086_to_fp16 = const()[name = tensor("op_4086_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_4087_cast_fp16 = mul(x = var_4085_cast_fp16, y = var_4086_to_fp16)[name = tensor("op_4087_cast_fp16")]; + tensor scores_63_cast_fp16 = add(x = scores_61_cast_fp16, y = var_4087_cast_fp16)[name = tensor("scores_63_cast_fp16")]; + tensor var_4090 = const()[name = tensor("op_4090"), val = tensor(-1)]; + tensor weights_31_cast_fp16 = softmax(axis = var_4090, x = scores_63_cast_fp16)[name = tensor("weights_31_cast_fp16")]; + tensor attn_31_transpose_x_0 = const()[name = tensor("attn_31_transpose_x_0"), val = tensor(false)]; + tensor attn_31_transpose_y_0 = const()[name = tensor("attn_31_transpose_y_0"), val = tensor(false)]; + tensor values_31_cast_fp16 = transpose(perm = var_4053, x = new_v_cache_31_cast_fp16)[name = tensor("transpose_49")]; + tensor attn_31_cast_fp16 = matmul(transpose_x = attn_31_transpose_x_0, transpose_y = attn_31_transpose_y_0, x = weights_31_cast_fp16, y = values_31_cast_fp16)[name = tensor("attn_31_cast_fp16")]; + tensor var_4098 = const()[name = tensor("op_4098"), val = tensor([0, 2, 1, 3])]; + tensor var_4103 = const()[name = tensor("op_4103"), val = tensor([1, 1, 1024])]; + tensor var_4099_cast_fp16 = transpose(perm = var_4098, x = attn_31_cast_fp16)[name = tensor("transpose_48")]; + tensor input_187_cast_fp16 = reshape(shape = var_4103, x = var_4099_cast_fp16)[name = tensor("input_187_cast_fp16")]; + tensor attn15_out_proj_weight_to_fp16 = const()[name = tensor("attn15_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(383997120)))]; + tensor linear_62_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn15_out_proj_weight_to_fp16, x = input_187_cast_fp16)[name = tensor("linear_62_cast_fp16")]; + tensor var_4109_to_fp16 = const()[name = tensor("op_4109_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4110_cast_fp16 = add(x = position15_to_fp16, y = var_4109_to_fp16)[name = tensor("op_4110_cast_fp16")]; + tensor var_4110_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4110_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_189_cast_fp16 = add(x = input_183_cast_fp16, y = linear_62_cast_fp16)[name = tensor("input_189_cast_fp16")]; + tensor input_191_axes_0 = const()[name = tensor("input_191_axes_0"), val = tensor([-1])]; + tensor norm15_2_weight_to_fp16 = const()[name = tensor("norm15_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(386094336)))]; + tensor norm15_2_bias_to_fp16 = const()[name = tensor("norm15_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(386096448)))]; + tensor var_4114_to_fp16 = const()[name = tensor("op_4114_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_191_cast_fp16 = layer_norm(axes = input_191_axes_0, beta = norm15_2_bias_to_fp16, epsilon = var_4114_to_fp16, gamma = norm15_2_weight_to_fp16, x = input_189_cast_fp16)[name = tensor("input_191_cast_fp16")]; + tensor linear15_1_weight_to_fp16 = const()[name = tensor("linear15_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(386098560)))]; + tensor linear_63_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear15_1_weight_to_fp16, x = input_191_cast_fp16)[name = tensor("linear_63_cast_fp16")]; + tensor input_193_mode_0 = const()[name = tensor("input_193_mode_0"), val = tensor("EXACT")]; + tensor input_193_cast_fp16 = gelu(mode = input_193_mode_0, x = linear_63_cast_fp16)[name = tensor("input_193_cast_fp16")]; + tensor linear15_2_weight_to_fp16 = const()[name = tensor("linear15_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(394487232)))]; + tensor linear_64_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear15_2_weight_to_fp16, x = input_193_cast_fp16)[name = tensor("linear_64_cast_fp16")]; + tensor input_195_cast_fp16 = add(x = input_189_cast_fp16, y = linear_64_cast_fp16)[name = tensor("input_195_cast_fp16")]; + tensor input_197_axes_0 = const()[name = tensor("input_197_axes_0"), val = tensor([-1])]; + tensor norm16_1_weight_to_fp16 = const()[name = tensor("norm16_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(402875904)))]; + tensor norm16_1_bias_to_fp16 = const()[name = tensor("norm16_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(402878016)))]; + tensor var_4131_to_fp16 = const()[name = tensor("op_4131_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_197_cast_fp16 = layer_norm(axes = input_197_axes_0, beta = norm16_1_bias_to_fp16, epsilon = var_4131_to_fp16, gamma = norm16_1_weight_to_fp16, x = input_195_cast_fp16)[name = tensor("input_197_cast_fp16")]; + tensor attn16_in_proj_weight_to_fp16 = const()[name = tensor("attn16_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(402880128)))]; + tensor linear_65_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn16_in_proj_weight_to_fp16, x = input_197_cast_fp16)[name = tensor("linear_65_cast_fp16")]; + tensor var_4144 = const()[name = tensor("op_4144"), val = tensor([1, 3, 16, 64])]; + tensor qkv_33_cast_fp16 = reshape(shape = var_4144, x = linear_65_cast_fp16)[name = tensor("qkv_33_cast_fp16")]; + tensor x_65_begin_0 = const()[name = tensor("x_65_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_65_end_0 = const()[name = tensor("x_65_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_65_end_mask_0 = const()[name = tensor("x_65_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_65_squeeze_mask_0 = const()[name = tensor("x_65_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_65_cast_fp16 = slice_by_index(begin = x_65_begin_0, end = x_65_end_0, end_mask = x_65_end_mask_0, squeeze_mask = x_65_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("x_65_cast_fp16")]; + tensor x_67_begin_0 = const()[name = tensor("x_67_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_67_end_0 = const()[name = tensor("x_67_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_67_end_mask_0 = const()[name = tensor("x_67_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_67_squeeze_mask_0 = const()[name = tensor("x_67_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_67_cast_fp16 = slice_by_index(begin = x_67_begin_0, end = x_67_end_0, end_mask = x_67_end_mask_0, squeeze_mask = x_67_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("x_67_cast_fp16")]; + tensor v_33_begin_0 = const()[name = tensor("v_33_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_33_end_0 = const()[name = tensor("v_33_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_33_end_mask_0 = const()[name = tensor("v_33_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_33_squeeze_mask_0 = const()[name = tensor("v_33_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_33_cast_fp16 = slice_by_index(begin = v_33_begin_0, end = v_33_end_0, end_mask = v_33_end_mask_0, squeeze_mask = v_33_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("v_33_cast_fp16")]; + tensor var_4173 = const()[name = tensor("op_4173"), val = tensor([1, 1, 1])]; + tensor position16_to_fp16_dtype_0 = const()[name = tensor("position16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position16_to_fp16 = cast(dtype = position16_to_fp16_dtype_0, x = position16)[name = tensor("cast_449")]; + tensor var_4174_cast_fp16 = reshape(shape = var_4173, x = position16_to_fp16)[name = tensor("op_4174_cast_fp16")]; + tensor angles_65_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_4174_cast_fp16)[name = tensor("angles_65_cast_fp16")]; + tensor rotr_65_cast_fp16 = cos(x = angles_65_cast_fp16)[name = tensor("rotr_65_cast_fp16")]; + tensor roti_65_cast_fp16 = sin(x = angles_65_cast_fp16)[name = tensor("roti_65_cast_fp16")]; + tensor var_4182 = const()[name = tensor("op_4182"), val = tensor([1, 16, 32, 2])]; + tensor pairs_65_cast_fp16 = reshape(shape = var_4182, x = x_65_cast_fp16)[name = tensor("pairs_65_cast_fp16")]; + tensor xr_65_begin_0 = const()[name = tensor("xr_65_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_65_end_0 = const()[name = tensor("xr_65_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_65_end_mask_0 = const()[name = tensor("xr_65_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_65_squeeze_mask_0 = const()[name = tensor("xr_65_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_65_cast_fp16 = slice_by_index(begin = xr_65_begin_0, end = xr_65_end_0, end_mask = xr_65_end_mask_0, squeeze_mask = xr_65_squeeze_mask_0, x = pairs_65_cast_fp16)[name = tensor("xr_65_cast_fp16")]; + tensor xi_65_begin_0 = const()[name = tensor("xi_65_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_65_end_0 = const()[name = tensor("xi_65_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_65_end_mask_0 = const()[name = tensor("xi_65_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_65_squeeze_mask_0 = const()[name = tensor("xi_65_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_65_cast_fp16 = slice_by_index(begin = xi_65_begin_0, end = xi_65_end_0, end_mask = xi_65_end_mask_0, squeeze_mask = xi_65_squeeze_mask_0, x = pairs_65_cast_fp16)[name = tensor("xi_65_cast_fp16")]; + tensor var_4190_cast_fp16 = mul(x = xr_65_cast_fp16, y = rotr_65_cast_fp16)[name = tensor("op_4190_cast_fp16")]; + tensor var_4191_cast_fp16 = mul(x = xi_65_cast_fp16, y = roti_65_cast_fp16)[name = tensor("op_4191_cast_fp16")]; + tensor xor__65_cast_fp16 = sub(x = var_4190_cast_fp16, y = var_4191_cast_fp16)[name = tensor("xor__65_cast_fp16")]; + tensor var_4194_cast_fp16 = mul(x = xr_65_cast_fp16, y = roti_65_cast_fp16)[name = tensor("op_4194_cast_fp16")]; + tensor var_4195_cast_fp16 = mul(x = xi_65_cast_fp16, y = rotr_65_cast_fp16)[name = tensor("op_4195_cast_fp16")]; + tensor xoi_65_cast_fp16 = add(x = var_4194_cast_fp16, y = var_4195_cast_fp16)[name = tensor("xoi_65_cast_fp16")]; + tensor out_65_axis_0 = const()[name = tensor("out_65_axis_0"), val = tensor(-1)]; + tensor out_65_cast_fp16 = stack(axis = out_65_axis_0, values = (xor__65_cast_fp16, xoi_65_cast_fp16))[name = tensor("out_65_cast_fp16")]; + tensor var_4218 = const()[name = tensor("op_4218"), val = tensor([1, 16, 32, 2])]; + tensor pairs_67_cast_fp16 = reshape(shape = var_4218, x = x_67_cast_fp16)[name = tensor("pairs_67_cast_fp16")]; + tensor xr_67_begin_0 = const()[name = tensor("xr_67_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_67_end_0 = const()[name = tensor("xr_67_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_67_end_mask_0 = const()[name = tensor("xr_67_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_67_squeeze_mask_0 = const()[name = tensor("xr_67_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_67_cast_fp16 = slice_by_index(begin = xr_67_begin_0, end = xr_67_end_0, end_mask = xr_67_end_mask_0, squeeze_mask = xr_67_squeeze_mask_0, x = pairs_67_cast_fp16)[name = tensor("xr_67_cast_fp16")]; + tensor xi_67_begin_0 = const()[name = tensor("xi_67_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_67_end_0 = const()[name = tensor("xi_67_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_67_end_mask_0 = const()[name = tensor("xi_67_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_67_squeeze_mask_0 = const()[name = tensor("xi_67_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_67_cast_fp16 = slice_by_index(begin = xi_67_begin_0, end = xi_67_end_0, end_mask = xi_67_end_mask_0, squeeze_mask = xi_67_squeeze_mask_0, x = pairs_67_cast_fp16)[name = tensor("xi_67_cast_fp16")]; + tensor var_4226_cast_fp16 = mul(x = xr_67_cast_fp16, y = rotr_65_cast_fp16)[name = tensor("op_4226_cast_fp16")]; + tensor var_4227_cast_fp16 = mul(x = xi_67_cast_fp16, y = roti_65_cast_fp16)[name = tensor("op_4227_cast_fp16")]; + tensor xor__67_cast_fp16 = sub(x = var_4226_cast_fp16, y = var_4227_cast_fp16)[name = tensor("xor__67_cast_fp16")]; + tensor var_4230_cast_fp16 = mul(x = xr_67_cast_fp16, y = roti_65_cast_fp16)[name = tensor("op_4230_cast_fp16")]; + tensor var_4231_cast_fp16 = mul(x = xi_67_cast_fp16, y = rotr_65_cast_fp16)[name = tensor("op_4231_cast_fp16")]; + tensor xoi_67_cast_fp16 = add(x = var_4230_cast_fp16, y = var_4231_cast_fp16)[name = tensor("xoi_67_cast_fp16")]; + tensor out_67_axis_0 = const()[name = tensor("out_67_axis_0"), val = tensor(-1)]; + tensor out_67_cast_fp16 = stack(axis = out_67_axis_0, values = (xor__67_cast_fp16, xoi_67_cast_fp16))[name = tensor("out_67_cast_fp16")]; + tensor _inversed_4243_y_0_to_fp16 = const()[name = tensor("_inversed_4243_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_4243_cast_fp16 = mul(x = position16_to_fp16, y = _inversed_4243_y_0_to_fp16)[name = tensor("_inversed_4243_cast_fp16")]; + tensor var_4244_cast_fp16 = floor(x = _inversed_4243_cast_fp16)[name = tensor("op_4244_cast_fp16")]; + tensor var_4245_to_fp16 = const()[name = tensor("op_4245_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4246_cast_fp16 = mul(x = var_4244_cast_fp16, y = var_4245_to_fp16)[name = tensor("op_4246_cast_fp16")]; + tensor write_idx_33_cast_fp16 = sub(x = position16_to_fp16, y = var_4246_cast_fp16)[name = tensor("write_idx_33_cast_fp16")]; + tensor var_4249_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_33_cast_fp16)[name = tensor("op_4249_cast_fp16")]; + tensor var_4259 = const()[name = tensor("op_4259"), val = tensor([1, 512, 1, 1])]; + tensor onehot_65_to_fp16_dtype_0 = const()[name = tensor("onehot_65_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_4249_cast_fp16_to_fp16 = cast(dtype = onehot_65_to_fp16_dtype_0, x = var_4249_cast_fp16)[name = tensor("cast_448")]; + tensor onehot_67_cast_fp16 = reshape(shape = var_4259, x = var_4249_cast_fp16_to_fp16)[name = tensor("onehot_67_cast_fp16")]; + tensor var_4261_to_fp16 = const()[name = tensor("op_4261_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4263_cast_fp16 = sub(x = var_4261_to_fp16, y = onehot_67_cast_fp16)[name = tensor("op_4263_cast_fp16")]; + tensor k_cache16_to_fp16_dtype_0 = const()[name = tensor("k_cache16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache16_to_fp16 = cast(dtype = k_cache16_to_fp16_dtype_0, x = k_cache16)[name = tensor("cast_447")]; + tensor var_4264_cast_fp16 = mul(x = k_cache16_to_fp16, y = var_4263_cast_fp16)[name = tensor("op_4264_cast_fp16")]; + tensor var_4269 = const()[name = tensor("op_4269"), val = tensor([1, 1, 16, 64])]; + tensor var_4270_cast_fp16 = reshape(shape = var_4269, x = out_67_cast_fp16)[name = tensor("op_4270_cast_fp16")]; + tensor var_4271_cast_fp16 = mul(x = var_4270_cast_fp16, y = onehot_67_cast_fp16)[name = tensor("op_4271_cast_fp16")]; + tensor new_k_cache_33_cast_fp16 = add(x = var_4264_cast_fp16, y = var_4271_cast_fp16)[name = tensor("new_k_cache_33_cast_fp16")]; + tensor new_k_cache_33_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_33_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache16_to_fp16_dtype_0 = const()[name = tensor("v_cache16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache16_to_fp16 = cast(dtype = v_cache16_to_fp16_dtype_0, x = v_cache16)[name = tensor("cast_445")]; + tensor var_4277_cast_fp16 = mul(x = v_cache16_to_fp16, y = var_4263_cast_fp16)[name = tensor("op_4277_cast_fp16")]; + tensor var_4282 = const()[name = tensor("op_4282"), val = tensor([1, 1, 16, 64])]; + tensor var_4283_cast_fp16 = reshape(shape = var_4282, x = v_33_cast_fp16)[name = tensor("op_4283_cast_fp16")]; + tensor var_4284_cast_fp16 = mul(x = var_4283_cast_fp16, y = onehot_67_cast_fp16)[name = tensor("op_4284_cast_fp16")]; + tensor new_v_cache_33_cast_fp16 = add(x = var_4277_cast_fp16, y = var_4284_cast_fp16)[name = tensor("new_v_cache_33_cast_fp16")]; + tensor new_v_cache_33_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_33_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_16_perm_0 = const()[name = tensor("transpose_16_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_4297 = const()[name = tensor("op_4297"), val = tensor([0, 2, 1, 3])]; + tensor var_4303 = const()[name = tensor("op_4303"), val = tensor([1, 16, 1, 64])]; + tensor q4_33_cast_fp16 = reshape(shape = var_4303, x = out_65_cast_fp16)[name = tensor("q4_33_cast_fp16")]; + tensor var_4308_transpose_x_0 = const()[name = tensor("op_4308_transpose_x_0"), val = tensor(false)]; + tensor var_4308_transpose_y_0 = const()[name = tensor("op_4308_transpose_y_0"), val = tensor(false)]; + tensor transpose_16_cast_fp16 = transpose(perm = transpose_16_perm_0, x = new_k_cache_33_cast_fp16)[name = tensor("transpose_47")]; + tensor var_4308_cast_fp16 = matmul(transpose_x = var_4308_transpose_x_0, transpose_y = var_4308_transpose_y_0, x = q4_33_cast_fp16, y = transpose_16_cast_fp16)[name = tensor("op_4308_cast_fp16")]; + tensor var_4309_to_fp16 = const()[name = tensor("op_4309_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_65_cast_fp16 = mul(x = var_4308_cast_fp16, y = var_4309_to_fp16)[name = tensor("scores_65_cast_fp16")]; + tensor var_4313 = const()[name = tensor("op_4313"), val = tensor([1, 1])]; + tensor var_4314_cast_fp16 = reshape(shape = var_4313, x = position16_to_fp16)[name = tensor("op_4314_cast_fp16")]; + tensor var_4315_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_4314_cast_fp16)[name = tensor("op_4315_cast_fp16")]; + tensor mask_33_to_fp16_dtype_0 = const()[name = tensor("mask_33_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_4322_to_fp16 = const()[name = tensor("op_4322_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4315_cast_fp16_to_fp16 = cast(dtype = mask_33_to_fp16_dtype_0, x = var_4315_cast_fp16)[name = tensor("cast_443")]; + tensor var_4323_cast_fp16 = sub(x = var_4315_cast_fp16_to_fp16, y = var_4322_to_fp16)[name = tensor("op_4323_cast_fp16")]; + tensor var_4328 = const()[name = tensor("op_4328"), val = tensor([1, 1, 1, 512])]; + tensor var_4329_cast_fp16 = reshape(shape = var_4328, x = var_4323_cast_fp16)[name = tensor("op_4329_cast_fp16")]; + tensor var_4330_to_fp16 = const()[name = tensor("op_4330_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_4331_cast_fp16 = mul(x = var_4329_cast_fp16, y = var_4330_to_fp16)[name = tensor("op_4331_cast_fp16")]; + tensor scores_67_cast_fp16 = add(x = scores_65_cast_fp16, y = var_4331_cast_fp16)[name = tensor("scores_67_cast_fp16")]; + tensor var_4334 = const()[name = tensor("op_4334"), val = tensor(-1)]; + tensor weights_33_cast_fp16 = softmax(axis = var_4334, x = scores_67_cast_fp16)[name = tensor("weights_33_cast_fp16")]; + tensor attn_33_transpose_x_0 = const()[name = tensor("attn_33_transpose_x_0"), val = tensor(false)]; + tensor attn_33_transpose_y_0 = const()[name = tensor("attn_33_transpose_y_0"), val = tensor(false)]; + tensor values_33_cast_fp16 = transpose(perm = var_4297, x = new_v_cache_33_cast_fp16)[name = tensor("transpose_46")]; + tensor attn_33_cast_fp16 = matmul(transpose_x = attn_33_transpose_x_0, transpose_y = attn_33_transpose_y_0, x = weights_33_cast_fp16, y = values_33_cast_fp16)[name = tensor("attn_33_cast_fp16")]; + tensor var_4342 = const()[name = tensor("op_4342"), val = tensor([0, 2, 1, 3])]; + tensor var_4347 = const()[name = tensor("op_4347"), val = tensor([1, 1, 1024])]; + tensor var_4343_cast_fp16 = transpose(perm = var_4342, x = attn_33_cast_fp16)[name = tensor("transpose_45")]; + tensor input_199_cast_fp16 = reshape(shape = var_4347, x = var_4343_cast_fp16)[name = tensor("input_199_cast_fp16")]; + tensor attn16_out_proj_weight_to_fp16 = const()[name = tensor("attn16_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(409171648)))]; + tensor linear_66_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn16_out_proj_weight_to_fp16, x = input_199_cast_fp16)[name = tensor("linear_66_cast_fp16")]; + tensor var_4353_to_fp16 = const()[name = tensor("op_4353_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4354_cast_fp16 = add(x = position16_to_fp16, y = var_4353_to_fp16)[name = tensor("op_4354_cast_fp16")]; + tensor var_4354_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4354_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_201_cast_fp16 = add(x = input_195_cast_fp16, y = linear_66_cast_fp16)[name = tensor("input_201_cast_fp16")]; + tensor input_203_axes_0 = const()[name = tensor("input_203_axes_0"), val = tensor([-1])]; + tensor norm16_2_weight_to_fp16 = const()[name = tensor("norm16_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411268864)))]; + tensor norm16_2_bias_to_fp16 = const()[name = tensor("norm16_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411270976)))]; + tensor var_4358_to_fp16 = const()[name = tensor("op_4358_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_203_cast_fp16 = layer_norm(axes = input_203_axes_0, beta = norm16_2_bias_to_fp16, epsilon = var_4358_to_fp16, gamma = norm16_2_weight_to_fp16, x = input_201_cast_fp16)[name = tensor("input_203_cast_fp16")]; + tensor linear16_1_weight_to_fp16 = const()[name = tensor("linear16_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411273088)))]; + tensor linear_67_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear16_1_weight_to_fp16, x = input_203_cast_fp16)[name = tensor("linear_67_cast_fp16")]; + tensor input_205_mode_0 = const()[name = tensor("input_205_mode_0"), val = tensor("EXACT")]; + tensor input_205_cast_fp16 = gelu(mode = input_205_mode_0, x = linear_67_cast_fp16)[name = tensor("input_205_cast_fp16")]; + tensor linear16_2_weight_to_fp16 = const()[name = tensor("linear16_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(419661760)))]; + tensor linear_68_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear16_2_weight_to_fp16, x = input_205_cast_fp16)[name = tensor("linear_68_cast_fp16")]; + tensor input_207_cast_fp16 = add(x = input_201_cast_fp16, y = linear_68_cast_fp16)[name = tensor("input_207_cast_fp16")]; + tensor input_209_axes_0 = const()[name = tensor("input_209_axes_0"), val = tensor([-1])]; + tensor norm17_1_weight_to_fp16 = const()[name = tensor("norm17_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(428050432)))]; + tensor norm17_1_bias_to_fp16 = const()[name = tensor("norm17_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(428052544)))]; + tensor var_4375_to_fp16 = const()[name = tensor("op_4375_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_209_cast_fp16 = layer_norm(axes = input_209_axes_0, beta = norm17_1_bias_to_fp16, epsilon = var_4375_to_fp16, gamma = norm17_1_weight_to_fp16, x = input_207_cast_fp16)[name = tensor("input_209_cast_fp16")]; + tensor attn17_in_proj_weight_to_fp16 = const()[name = tensor("attn17_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(428054656)))]; + tensor linear_69_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn17_in_proj_weight_to_fp16, x = input_209_cast_fp16)[name = tensor("linear_69_cast_fp16")]; + tensor var_4388 = const()[name = tensor("op_4388"), val = tensor([1, 3, 16, 64])]; + tensor qkv_35_cast_fp16 = reshape(shape = var_4388, x = linear_69_cast_fp16)[name = tensor("qkv_35_cast_fp16")]; + tensor x_69_begin_0 = const()[name = tensor("x_69_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_69_end_0 = const()[name = tensor("x_69_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_69_end_mask_0 = const()[name = tensor("x_69_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_69_squeeze_mask_0 = const()[name = tensor("x_69_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_69_cast_fp16 = slice_by_index(begin = x_69_begin_0, end = x_69_end_0, end_mask = x_69_end_mask_0, squeeze_mask = x_69_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("x_69_cast_fp16")]; + tensor x_71_begin_0 = const()[name = tensor("x_71_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_71_end_0 = const()[name = tensor("x_71_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_71_end_mask_0 = const()[name = tensor("x_71_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_71_squeeze_mask_0 = const()[name = tensor("x_71_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_71_cast_fp16 = slice_by_index(begin = x_71_begin_0, end = x_71_end_0, end_mask = x_71_end_mask_0, squeeze_mask = x_71_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("x_71_cast_fp16")]; + tensor v_35_begin_0 = const()[name = tensor("v_35_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_35_end_0 = const()[name = tensor("v_35_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_35_end_mask_0 = const()[name = tensor("v_35_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_35_squeeze_mask_0 = const()[name = tensor("v_35_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_35_cast_fp16 = slice_by_index(begin = v_35_begin_0, end = v_35_end_0, end_mask = v_35_end_mask_0, squeeze_mask = v_35_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("v_35_cast_fp16")]; + tensor var_4417 = const()[name = tensor("op_4417"), val = tensor([1, 1, 1])]; + tensor position17_to_fp16_dtype_0 = const()[name = tensor("position17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position17_to_fp16 = cast(dtype = position17_to_fp16_dtype_0, x = position17)[name = tensor("cast_441")]; + tensor var_4418_cast_fp16 = reshape(shape = var_4417, x = position17_to_fp16)[name = tensor("op_4418_cast_fp16")]; + tensor angles_69_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_4418_cast_fp16)[name = tensor("angles_69_cast_fp16")]; + tensor rotr_69_cast_fp16 = cos(x = angles_69_cast_fp16)[name = tensor("rotr_69_cast_fp16")]; + tensor roti_69_cast_fp16 = sin(x = angles_69_cast_fp16)[name = tensor("roti_69_cast_fp16")]; + tensor var_4426 = const()[name = tensor("op_4426"), val = tensor([1, 16, 32, 2])]; + tensor pairs_69_cast_fp16 = reshape(shape = var_4426, x = x_69_cast_fp16)[name = tensor("pairs_69_cast_fp16")]; + tensor xr_69_begin_0 = const()[name = tensor("xr_69_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_69_end_0 = const()[name = tensor("xr_69_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_69_end_mask_0 = const()[name = tensor("xr_69_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_69_squeeze_mask_0 = const()[name = tensor("xr_69_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_69_cast_fp16 = slice_by_index(begin = xr_69_begin_0, end = xr_69_end_0, end_mask = xr_69_end_mask_0, squeeze_mask = xr_69_squeeze_mask_0, x = pairs_69_cast_fp16)[name = tensor("xr_69_cast_fp16")]; + tensor xi_69_begin_0 = const()[name = tensor("xi_69_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_69_end_0 = const()[name = tensor("xi_69_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_69_end_mask_0 = const()[name = tensor("xi_69_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_69_squeeze_mask_0 = const()[name = tensor("xi_69_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_69_cast_fp16 = slice_by_index(begin = xi_69_begin_0, end = xi_69_end_0, end_mask = xi_69_end_mask_0, squeeze_mask = xi_69_squeeze_mask_0, x = pairs_69_cast_fp16)[name = tensor("xi_69_cast_fp16")]; + tensor var_4434_cast_fp16 = mul(x = xr_69_cast_fp16, y = rotr_69_cast_fp16)[name = tensor("op_4434_cast_fp16")]; + tensor var_4435_cast_fp16 = mul(x = xi_69_cast_fp16, y = roti_69_cast_fp16)[name = tensor("op_4435_cast_fp16")]; + tensor xor__69_cast_fp16 = sub(x = var_4434_cast_fp16, y = var_4435_cast_fp16)[name = tensor("xor__69_cast_fp16")]; + tensor var_4438_cast_fp16 = mul(x = xr_69_cast_fp16, y = roti_69_cast_fp16)[name = tensor("op_4438_cast_fp16")]; + tensor var_4439_cast_fp16 = mul(x = xi_69_cast_fp16, y = rotr_69_cast_fp16)[name = tensor("op_4439_cast_fp16")]; + tensor xoi_69_cast_fp16 = add(x = var_4438_cast_fp16, y = var_4439_cast_fp16)[name = tensor("xoi_69_cast_fp16")]; + tensor out_69_axis_0 = const()[name = tensor("out_69_axis_0"), val = tensor(-1)]; + tensor out_69_cast_fp16 = stack(axis = out_69_axis_0, values = (xor__69_cast_fp16, xoi_69_cast_fp16))[name = tensor("out_69_cast_fp16")]; + tensor var_4462 = const()[name = tensor("op_4462"), val = tensor([1, 16, 32, 2])]; + tensor pairs_71_cast_fp16 = reshape(shape = var_4462, x = x_71_cast_fp16)[name = tensor("pairs_71_cast_fp16")]; + tensor xr_71_begin_0 = const()[name = tensor("xr_71_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_71_end_0 = const()[name = tensor("xr_71_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_71_end_mask_0 = const()[name = tensor("xr_71_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_71_squeeze_mask_0 = const()[name = tensor("xr_71_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_71_cast_fp16 = slice_by_index(begin = xr_71_begin_0, end = xr_71_end_0, end_mask = xr_71_end_mask_0, squeeze_mask = xr_71_squeeze_mask_0, x = pairs_71_cast_fp16)[name = tensor("xr_71_cast_fp16")]; + tensor xi_71_begin_0 = const()[name = tensor("xi_71_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_71_end_0 = const()[name = tensor("xi_71_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_71_end_mask_0 = const()[name = tensor("xi_71_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_71_squeeze_mask_0 = const()[name = tensor("xi_71_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_71_cast_fp16 = slice_by_index(begin = xi_71_begin_0, end = xi_71_end_0, end_mask = xi_71_end_mask_0, squeeze_mask = xi_71_squeeze_mask_0, x = pairs_71_cast_fp16)[name = tensor("xi_71_cast_fp16")]; + tensor var_4470_cast_fp16 = mul(x = xr_71_cast_fp16, y = rotr_69_cast_fp16)[name = tensor("op_4470_cast_fp16")]; + tensor var_4471_cast_fp16 = mul(x = xi_71_cast_fp16, y = roti_69_cast_fp16)[name = tensor("op_4471_cast_fp16")]; + tensor xor__71_cast_fp16 = sub(x = var_4470_cast_fp16, y = var_4471_cast_fp16)[name = tensor("xor__71_cast_fp16")]; + tensor var_4474_cast_fp16 = mul(x = xr_71_cast_fp16, y = roti_69_cast_fp16)[name = tensor("op_4474_cast_fp16")]; + tensor var_4475_cast_fp16 = mul(x = xi_71_cast_fp16, y = rotr_69_cast_fp16)[name = tensor("op_4475_cast_fp16")]; + tensor xoi_71_cast_fp16 = add(x = var_4474_cast_fp16, y = var_4475_cast_fp16)[name = tensor("xoi_71_cast_fp16")]; + tensor out_71_axis_0 = const()[name = tensor("out_71_axis_0"), val = tensor(-1)]; + tensor out_71_cast_fp16 = stack(axis = out_71_axis_0, values = (xor__71_cast_fp16, xoi_71_cast_fp16))[name = tensor("out_71_cast_fp16")]; + tensor _inversed_4487_y_0_to_fp16 = const()[name = tensor("_inversed_4487_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_4487_cast_fp16 = mul(x = position17_to_fp16, y = _inversed_4487_y_0_to_fp16)[name = tensor("_inversed_4487_cast_fp16")]; + tensor var_4488_cast_fp16 = floor(x = _inversed_4487_cast_fp16)[name = tensor("op_4488_cast_fp16")]; + tensor var_4489_to_fp16 = const()[name = tensor("op_4489_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4490_cast_fp16 = mul(x = var_4488_cast_fp16, y = var_4489_to_fp16)[name = tensor("op_4490_cast_fp16")]; + tensor write_idx_35_cast_fp16 = sub(x = position17_to_fp16, y = var_4490_cast_fp16)[name = tensor("write_idx_35_cast_fp16")]; + tensor var_4493_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_35_cast_fp16)[name = tensor("op_4493_cast_fp16")]; + tensor var_4503 = const()[name = tensor("op_4503"), val = tensor([1, 512, 1, 1])]; + tensor onehot_69_to_fp16_dtype_0 = const()[name = tensor("onehot_69_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_4493_cast_fp16_to_fp16 = cast(dtype = onehot_69_to_fp16_dtype_0, x = var_4493_cast_fp16)[name = tensor("cast_440")]; + tensor onehot_71_cast_fp16 = reshape(shape = var_4503, x = var_4493_cast_fp16_to_fp16)[name = tensor("onehot_71_cast_fp16")]; + tensor var_4505_to_fp16 = const()[name = tensor("op_4505_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4507_cast_fp16 = sub(x = var_4505_to_fp16, y = onehot_71_cast_fp16)[name = tensor("op_4507_cast_fp16")]; + tensor k_cache17_to_fp16_dtype_0 = const()[name = tensor("k_cache17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache17_to_fp16 = cast(dtype = k_cache17_to_fp16_dtype_0, x = k_cache17)[name = tensor("cast_439")]; + tensor var_4508_cast_fp16 = mul(x = k_cache17_to_fp16, y = var_4507_cast_fp16)[name = tensor("op_4508_cast_fp16")]; + tensor var_4513 = const()[name = tensor("op_4513"), val = tensor([1, 1, 16, 64])]; + tensor var_4514_cast_fp16 = reshape(shape = var_4513, x = out_71_cast_fp16)[name = tensor("op_4514_cast_fp16")]; + tensor var_4515_cast_fp16 = mul(x = var_4514_cast_fp16, y = onehot_71_cast_fp16)[name = tensor("op_4515_cast_fp16")]; + tensor new_k_cache_35_cast_fp16 = add(x = var_4508_cast_fp16, y = var_4515_cast_fp16)[name = tensor("new_k_cache_35_cast_fp16")]; + tensor new_k_cache_35_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_35_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache17_to_fp16_dtype_0 = const()[name = tensor("v_cache17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache17_to_fp16 = cast(dtype = v_cache17_to_fp16_dtype_0, x = v_cache17)[name = tensor("cast_437")]; + tensor var_4521_cast_fp16 = mul(x = v_cache17_to_fp16, y = var_4507_cast_fp16)[name = tensor("op_4521_cast_fp16")]; + tensor var_4526 = const()[name = tensor("op_4526"), val = tensor([1, 1, 16, 64])]; + tensor var_4527_cast_fp16 = reshape(shape = var_4526, x = v_35_cast_fp16)[name = tensor("op_4527_cast_fp16")]; + tensor var_4528_cast_fp16 = mul(x = var_4527_cast_fp16, y = onehot_71_cast_fp16)[name = tensor("op_4528_cast_fp16")]; + tensor new_v_cache_35_cast_fp16 = add(x = var_4521_cast_fp16, y = var_4528_cast_fp16)[name = tensor("new_v_cache_35_cast_fp16")]; + tensor new_v_cache_35_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_35_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_17_perm_0 = const()[name = tensor("transpose_17_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_4541 = const()[name = tensor("op_4541"), val = tensor([0, 2, 1, 3])]; + tensor var_4547 = const()[name = tensor("op_4547"), val = tensor([1, 16, 1, 64])]; + tensor q4_35_cast_fp16 = reshape(shape = var_4547, x = out_69_cast_fp16)[name = tensor("q4_35_cast_fp16")]; + tensor var_4552_transpose_x_0 = const()[name = tensor("op_4552_transpose_x_0"), val = tensor(false)]; + tensor var_4552_transpose_y_0 = const()[name = tensor("op_4552_transpose_y_0"), val = tensor(false)]; + tensor transpose_17_cast_fp16 = transpose(perm = transpose_17_perm_0, x = new_k_cache_35_cast_fp16)[name = tensor("transpose_44")]; + tensor var_4552_cast_fp16 = matmul(transpose_x = var_4552_transpose_x_0, transpose_y = var_4552_transpose_y_0, x = q4_35_cast_fp16, y = transpose_17_cast_fp16)[name = tensor("op_4552_cast_fp16")]; + tensor var_4553_to_fp16 = const()[name = tensor("op_4553_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_69_cast_fp16 = mul(x = var_4552_cast_fp16, y = var_4553_to_fp16)[name = tensor("scores_69_cast_fp16")]; + tensor var_4557 = const()[name = tensor("op_4557"), val = tensor([1, 1])]; + tensor var_4558_cast_fp16 = reshape(shape = var_4557, x = position17_to_fp16)[name = tensor("op_4558_cast_fp16")]; + tensor var_4559_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_4558_cast_fp16)[name = tensor("op_4559_cast_fp16")]; + tensor mask_35_to_fp16_dtype_0 = const()[name = tensor("mask_35_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_4566_to_fp16 = const()[name = tensor("op_4566_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4559_cast_fp16_to_fp16 = cast(dtype = mask_35_to_fp16_dtype_0, x = var_4559_cast_fp16)[name = tensor("cast_435")]; + tensor var_4567_cast_fp16 = sub(x = var_4559_cast_fp16_to_fp16, y = var_4566_to_fp16)[name = tensor("op_4567_cast_fp16")]; + tensor var_4572 = const()[name = tensor("op_4572"), val = tensor([1, 1, 1, 512])]; + tensor var_4573_cast_fp16 = reshape(shape = var_4572, x = var_4567_cast_fp16)[name = tensor("op_4573_cast_fp16")]; + tensor var_4574_to_fp16 = const()[name = tensor("op_4574_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_4575_cast_fp16 = mul(x = var_4573_cast_fp16, y = var_4574_to_fp16)[name = tensor("op_4575_cast_fp16")]; + tensor scores_71_cast_fp16 = add(x = scores_69_cast_fp16, y = var_4575_cast_fp16)[name = tensor("scores_71_cast_fp16")]; + tensor var_4578 = const()[name = tensor("op_4578"), val = tensor(-1)]; + tensor weights_35_cast_fp16 = softmax(axis = var_4578, x = scores_71_cast_fp16)[name = tensor("weights_35_cast_fp16")]; + tensor attn_35_transpose_x_0 = const()[name = tensor("attn_35_transpose_x_0"), val = tensor(false)]; + tensor attn_35_transpose_y_0 = const()[name = tensor("attn_35_transpose_y_0"), val = tensor(false)]; + tensor values_35_cast_fp16 = transpose(perm = var_4541, x = new_v_cache_35_cast_fp16)[name = tensor("transpose_43")]; + tensor attn_35_cast_fp16 = matmul(transpose_x = attn_35_transpose_x_0, transpose_y = attn_35_transpose_y_0, x = weights_35_cast_fp16, y = values_35_cast_fp16)[name = tensor("attn_35_cast_fp16")]; + tensor var_4586 = const()[name = tensor("op_4586"), val = tensor([0, 2, 1, 3])]; + tensor var_4591 = const()[name = tensor("op_4591"), val = tensor([1, 1, 1024])]; + tensor var_4587_cast_fp16 = transpose(perm = var_4586, x = attn_35_cast_fp16)[name = tensor("transpose_42")]; + tensor input_211_cast_fp16 = reshape(shape = var_4591, x = var_4587_cast_fp16)[name = tensor("input_211_cast_fp16")]; + tensor attn17_out_proj_weight_to_fp16 = const()[name = tensor("attn17_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(434346176)))]; + tensor linear_70_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn17_out_proj_weight_to_fp16, x = input_211_cast_fp16)[name = tensor("linear_70_cast_fp16")]; + tensor var_4597_to_fp16 = const()[name = tensor("op_4597_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4598_cast_fp16 = add(x = position17_to_fp16, y = var_4597_to_fp16)[name = tensor("op_4598_cast_fp16")]; + tensor var_4598_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4598_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_213_cast_fp16 = add(x = input_207_cast_fp16, y = linear_70_cast_fp16)[name = tensor("input_213_cast_fp16")]; + tensor input_215_axes_0 = const()[name = tensor("input_215_axes_0"), val = tensor([-1])]; + tensor norm17_2_weight_to_fp16 = const()[name = tensor("norm17_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(436443392)))]; + tensor norm17_2_bias_to_fp16 = const()[name = tensor("norm17_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(436445504)))]; + tensor var_4602_to_fp16 = const()[name = tensor("op_4602_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_215_cast_fp16 = layer_norm(axes = input_215_axes_0, beta = norm17_2_bias_to_fp16, epsilon = var_4602_to_fp16, gamma = norm17_2_weight_to_fp16, x = input_213_cast_fp16)[name = tensor("input_215_cast_fp16")]; + tensor linear17_1_weight_to_fp16 = const()[name = tensor("linear17_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(436447616)))]; + tensor linear_71_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear17_1_weight_to_fp16, x = input_215_cast_fp16)[name = tensor("linear_71_cast_fp16")]; + tensor input_217_mode_0 = const()[name = tensor("input_217_mode_0"), val = tensor("EXACT")]; + tensor input_217_cast_fp16 = gelu(mode = input_217_mode_0, x = linear_71_cast_fp16)[name = tensor("input_217_cast_fp16")]; + tensor linear17_2_weight_to_fp16 = const()[name = tensor("linear17_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(444836288)))]; + tensor linear_72_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear17_2_weight_to_fp16, x = input_217_cast_fp16)[name = tensor("linear_72_cast_fp16")]; + tensor input_219_cast_fp16 = add(x = input_213_cast_fp16, y = linear_72_cast_fp16)[name = tensor("input_219_cast_fp16")]; + tensor input_221_axes_0 = const()[name = tensor("input_221_axes_0"), val = tensor([-1])]; + tensor norm18_1_weight_to_fp16 = const()[name = tensor("norm18_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(453224960)))]; + tensor norm18_1_bias_to_fp16 = const()[name = tensor("norm18_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(453227072)))]; + tensor var_4619_to_fp16 = const()[name = tensor("op_4619_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_221_cast_fp16 = layer_norm(axes = input_221_axes_0, beta = norm18_1_bias_to_fp16, epsilon = var_4619_to_fp16, gamma = norm18_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("input_221_cast_fp16")]; + tensor attn18_in_proj_weight_to_fp16 = const()[name = tensor("attn18_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(453229184)))]; + tensor linear_73_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn18_in_proj_weight_to_fp16, x = input_221_cast_fp16)[name = tensor("linear_73_cast_fp16")]; + tensor var_4632 = const()[name = tensor("op_4632"), val = tensor([1, 3, 16, 64])]; + tensor qkv_37_cast_fp16 = reshape(shape = var_4632, x = linear_73_cast_fp16)[name = tensor("qkv_37_cast_fp16")]; + tensor x_73_begin_0 = const()[name = tensor("x_73_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_73_end_0 = const()[name = tensor("x_73_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_73_end_mask_0 = const()[name = tensor("x_73_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_73_squeeze_mask_0 = const()[name = tensor("x_73_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_73_cast_fp16 = slice_by_index(begin = x_73_begin_0, end = x_73_end_0, end_mask = x_73_end_mask_0, squeeze_mask = x_73_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("x_73_cast_fp16")]; + tensor x_75_begin_0 = const()[name = tensor("x_75_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_75_end_0 = const()[name = tensor("x_75_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_75_end_mask_0 = const()[name = tensor("x_75_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_75_squeeze_mask_0 = const()[name = tensor("x_75_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_75_cast_fp16 = slice_by_index(begin = x_75_begin_0, end = x_75_end_0, end_mask = x_75_end_mask_0, squeeze_mask = x_75_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("x_75_cast_fp16")]; + tensor v_37_begin_0 = const()[name = tensor("v_37_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_37_end_0 = const()[name = tensor("v_37_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_37_end_mask_0 = const()[name = tensor("v_37_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_37_squeeze_mask_0 = const()[name = tensor("v_37_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_37_cast_fp16 = slice_by_index(begin = v_37_begin_0, end = v_37_end_0, end_mask = v_37_end_mask_0, squeeze_mask = v_37_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("v_37_cast_fp16")]; + tensor var_4661 = const()[name = tensor("op_4661"), val = tensor([1, 1, 1])]; + tensor position18_to_fp16_dtype_0 = const()[name = tensor("position18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position18_to_fp16 = cast(dtype = position18_to_fp16_dtype_0, x = position18)[name = tensor("cast_433")]; + tensor var_4662_cast_fp16 = reshape(shape = var_4661, x = position18_to_fp16)[name = tensor("op_4662_cast_fp16")]; + tensor angles_73_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_4662_cast_fp16)[name = tensor("angles_73_cast_fp16")]; + tensor rotr_73_cast_fp16 = cos(x = angles_73_cast_fp16)[name = tensor("rotr_73_cast_fp16")]; + tensor roti_73_cast_fp16 = sin(x = angles_73_cast_fp16)[name = tensor("roti_73_cast_fp16")]; + tensor var_4670 = const()[name = tensor("op_4670"), val = tensor([1, 16, 32, 2])]; + tensor pairs_73_cast_fp16 = reshape(shape = var_4670, x = x_73_cast_fp16)[name = tensor("pairs_73_cast_fp16")]; + tensor xr_73_begin_0 = const()[name = tensor("xr_73_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_73_end_0 = const()[name = tensor("xr_73_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_73_end_mask_0 = const()[name = tensor("xr_73_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_73_squeeze_mask_0 = const()[name = tensor("xr_73_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_73_cast_fp16 = slice_by_index(begin = xr_73_begin_0, end = xr_73_end_0, end_mask = xr_73_end_mask_0, squeeze_mask = xr_73_squeeze_mask_0, x = pairs_73_cast_fp16)[name = tensor("xr_73_cast_fp16")]; + tensor xi_73_begin_0 = const()[name = tensor("xi_73_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_73_end_0 = const()[name = tensor("xi_73_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_73_end_mask_0 = const()[name = tensor("xi_73_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_73_squeeze_mask_0 = const()[name = tensor("xi_73_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_73_cast_fp16 = slice_by_index(begin = xi_73_begin_0, end = xi_73_end_0, end_mask = xi_73_end_mask_0, squeeze_mask = xi_73_squeeze_mask_0, x = pairs_73_cast_fp16)[name = tensor("xi_73_cast_fp16")]; + tensor var_4678_cast_fp16 = mul(x = xr_73_cast_fp16, y = rotr_73_cast_fp16)[name = tensor("op_4678_cast_fp16")]; + tensor var_4679_cast_fp16 = mul(x = xi_73_cast_fp16, y = roti_73_cast_fp16)[name = tensor("op_4679_cast_fp16")]; + tensor xor__73_cast_fp16 = sub(x = var_4678_cast_fp16, y = var_4679_cast_fp16)[name = tensor("xor__73_cast_fp16")]; + tensor var_4682_cast_fp16 = mul(x = xr_73_cast_fp16, y = roti_73_cast_fp16)[name = tensor("op_4682_cast_fp16")]; + tensor var_4683_cast_fp16 = mul(x = xi_73_cast_fp16, y = rotr_73_cast_fp16)[name = tensor("op_4683_cast_fp16")]; + tensor xoi_73_cast_fp16 = add(x = var_4682_cast_fp16, y = var_4683_cast_fp16)[name = tensor("xoi_73_cast_fp16")]; + tensor out_73_axis_0 = const()[name = tensor("out_73_axis_0"), val = tensor(-1)]; + tensor out_73_cast_fp16 = stack(axis = out_73_axis_0, values = (xor__73_cast_fp16, xoi_73_cast_fp16))[name = tensor("out_73_cast_fp16")]; + tensor var_4706 = const()[name = tensor("op_4706"), val = tensor([1, 16, 32, 2])]; + tensor pairs_75_cast_fp16 = reshape(shape = var_4706, x = x_75_cast_fp16)[name = tensor("pairs_75_cast_fp16")]; + tensor xr_75_begin_0 = const()[name = tensor("xr_75_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_75_end_0 = const()[name = tensor("xr_75_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_75_end_mask_0 = const()[name = tensor("xr_75_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_75_squeeze_mask_0 = const()[name = tensor("xr_75_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_75_cast_fp16 = slice_by_index(begin = xr_75_begin_0, end = xr_75_end_0, end_mask = xr_75_end_mask_0, squeeze_mask = xr_75_squeeze_mask_0, x = pairs_75_cast_fp16)[name = tensor("xr_75_cast_fp16")]; + tensor xi_75_begin_0 = const()[name = tensor("xi_75_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_75_end_0 = const()[name = tensor("xi_75_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_75_end_mask_0 = const()[name = tensor("xi_75_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_75_squeeze_mask_0 = const()[name = tensor("xi_75_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_75_cast_fp16 = slice_by_index(begin = xi_75_begin_0, end = xi_75_end_0, end_mask = xi_75_end_mask_0, squeeze_mask = xi_75_squeeze_mask_0, x = pairs_75_cast_fp16)[name = tensor("xi_75_cast_fp16")]; + tensor var_4714_cast_fp16 = mul(x = xr_75_cast_fp16, y = rotr_73_cast_fp16)[name = tensor("op_4714_cast_fp16")]; + tensor var_4715_cast_fp16 = mul(x = xi_75_cast_fp16, y = roti_73_cast_fp16)[name = tensor("op_4715_cast_fp16")]; + tensor xor__75_cast_fp16 = sub(x = var_4714_cast_fp16, y = var_4715_cast_fp16)[name = tensor("xor__75_cast_fp16")]; + tensor var_4718_cast_fp16 = mul(x = xr_75_cast_fp16, y = roti_73_cast_fp16)[name = tensor("op_4718_cast_fp16")]; + tensor var_4719_cast_fp16 = mul(x = xi_75_cast_fp16, y = rotr_73_cast_fp16)[name = tensor("op_4719_cast_fp16")]; + tensor xoi_75_cast_fp16 = add(x = var_4718_cast_fp16, y = var_4719_cast_fp16)[name = tensor("xoi_75_cast_fp16")]; + tensor out_75_axis_0 = const()[name = tensor("out_75_axis_0"), val = tensor(-1)]; + tensor out_75_cast_fp16 = stack(axis = out_75_axis_0, values = (xor__75_cast_fp16, xoi_75_cast_fp16))[name = tensor("out_75_cast_fp16")]; + tensor _inversed_4731_y_0_to_fp16 = const()[name = tensor("_inversed_4731_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_4731_cast_fp16 = mul(x = position18_to_fp16, y = _inversed_4731_y_0_to_fp16)[name = tensor("_inversed_4731_cast_fp16")]; + tensor var_4732_cast_fp16 = floor(x = _inversed_4731_cast_fp16)[name = tensor("op_4732_cast_fp16")]; + tensor var_4733_to_fp16 = const()[name = tensor("op_4733_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4734_cast_fp16 = mul(x = var_4732_cast_fp16, y = var_4733_to_fp16)[name = tensor("op_4734_cast_fp16")]; + tensor write_idx_37_cast_fp16 = sub(x = position18_to_fp16, y = var_4734_cast_fp16)[name = tensor("write_idx_37_cast_fp16")]; + tensor var_4737_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_37_cast_fp16)[name = tensor("op_4737_cast_fp16")]; + tensor var_4747 = const()[name = tensor("op_4747"), val = tensor([1, 512, 1, 1])]; + tensor onehot_73_to_fp16_dtype_0 = const()[name = tensor("onehot_73_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_4737_cast_fp16_to_fp16 = cast(dtype = onehot_73_to_fp16_dtype_0, x = var_4737_cast_fp16)[name = tensor("cast_432")]; + tensor onehot_75_cast_fp16 = reshape(shape = var_4747, x = var_4737_cast_fp16_to_fp16)[name = tensor("onehot_75_cast_fp16")]; + tensor var_4749_to_fp16 = const()[name = tensor("op_4749_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4751_cast_fp16 = sub(x = var_4749_to_fp16, y = onehot_75_cast_fp16)[name = tensor("op_4751_cast_fp16")]; + tensor k_cache18_to_fp16_dtype_0 = const()[name = tensor("k_cache18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache18_to_fp16 = cast(dtype = k_cache18_to_fp16_dtype_0, x = k_cache18)[name = tensor("cast_431")]; + tensor var_4752_cast_fp16 = mul(x = k_cache18_to_fp16, y = var_4751_cast_fp16)[name = tensor("op_4752_cast_fp16")]; + tensor var_4757 = const()[name = tensor("op_4757"), val = tensor([1, 1, 16, 64])]; + tensor var_4758_cast_fp16 = reshape(shape = var_4757, x = out_75_cast_fp16)[name = tensor("op_4758_cast_fp16")]; + tensor var_4759_cast_fp16 = mul(x = var_4758_cast_fp16, y = onehot_75_cast_fp16)[name = tensor("op_4759_cast_fp16")]; + tensor new_k_cache_37_cast_fp16 = add(x = var_4752_cast_fp16, y = var_4759_cast_fp16)[name = tensor("new_k_cache_37_cast_fp16")]; + tensor new_k_cache_37_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_37_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache18_to_fp16_dtype_0 = const()[name = tensor("v_cache18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache18_to_fp16 = cast(dtype = v_cache18_to_fp16_dtype_0, x = v_cache18)[name = tensor("cast_429")]; + tensor var_4765_cast_fp16 = mul(x = v_cache18_to_fp16, y = var_4751_cast_fp16)[name = tensor("op_4765_cast_fp16")]; + tensor var_4770 = const()[name = tensor("op_4770"), val = tensor([1, 1, 16, 64])]; + tensor var_4771_cast_fp16 = reshape(shape = var_4770, x = v_37_cast_fp16)[name = tensor("op_4771_cast_fp16")]; + tensor var_4772_cast_fp16 = mul(x = var_4771_cast_fp16, y = onehot_75_cast_fp16)[name = tensor("op_4772_cast_fp16")]; + tensor new_v_cache_37_cast_fp16 = add(x = var_4765_cast_fp16, y = var_4772_cast_fp16)[name = tensor("new_v_cache_37_cast_fp16")]; + tensor new_v_cache_37_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_37_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_18_perm_0 = const()[name = tensor("transpose_18_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_4785 = const()[name = tensor("op_4785"), val = tensor([0, 2, 1, 3])]; + tensor var_4791 = const()[name = tensor("op_4791"), val = tensor([1, 16, 1, 64])]; + tensor q4_37_cast_fp16 = reshape(shape = var_4791, x = out_73_cast_fp16)[name = tensor("q4_37_cast_fp16")]; + tensor var_4796_transpose_x_0 = const()[name = tensor("op_4796_transpose_x_0"), val = tensor(false)]; + tensor var_4796_transpose_y_0 = const()[name = tensor("op_4796_transpose_y_0"), val = tensor(false)]; + tensor transpose_18_cast_fp16 = transpose(perm = transpose_18_perm_0, x = new_k_cache_37_cast_fp16)[name = tensor("transpose_41")]; + tensor var_4796_cast_fp16 = matmul(transpose_x = var_4796_transpose_x_0, transpose_y = var_4796_transpose_y_0, x = q4_37_cast_fp16, y = transpose_18_cast_fp16)[name = tensor("op_4796_cast_fp16")]; + tensor var_4797_to_fp16 = const()[name = tensor("op_4797_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_73_cast_fp16 = mul(x = var_4796_cast_fp16, y = var_4797_to_fp16)[name = tensor("scores_73_cast_fp16")]; + tensor var_4801 = const()[name = tensor("op_4801"), val = tensor([1, 1])]; + tensor var_4802_cast_fp16 = reshape(shape = var_4801, x = position18_to_fp16)[name = tensor("op_4802_cast_fp16")]; + tensor var_4803_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_4802_cast_fp16)[name = tensor("op_4803_cast_fp16")]; + tensor mask_37_to_fp16_dtype_0 = const()[name = tensor("mask_37_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_4810_to_fp16 = const()[name = tensor("op_4810_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4803_cast_fp16_to_fp16 = cast(dtype = mask_37_to_fp16_dtype_0, x = var_4803_cast_fp16)[name = tensor("cast_427")]; + tensor var_4811_cast_fp16 = sub(x = var_4803_cast_fp16_to_fp16, y = var_4810_to_fp16)[name = tensor("op_4811_cast_fp16")]; + tensor var_4816 = const()[name = tensor("op_4816"), val = tensor([1, 1, 1, 512])]; + tensor var_4817_cast_fp16 = reshape(shape = var_4816, x = var_4811_cast_fp16)[name = tensor("op_4817_cast_fp16")]; + tensor var_4818_to_fp16 = const()[name = tensor("op_4818_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_4819_cast_fp16 = mul(x = var_4817_cast_fp16, y = var_4818_to_fp16)[name = tensor("op_4819_cast_fp16")]; + tensor scores_75_cast_fp16 = add(x = scores_73_cast_fp16, y = var_4819_cast_fp16)[name = tensor("scores_75_cast_fp16")]; + tensor var_4822 = const()[name = tensor("op_4822"), val = tensor(-1)]; + tensor weights_37_cast_fp16 = softmax(axis = var_4822, x = scores_75_cast_fp16)[name = tensor("weights_37_cast_fp16")]; + tensor attn_37_transpose_x_0 = const()[name = tensor("attn_37_transpose_x_0"), val = tensor(false)]; + tensor attn_37_transpose_y_0 = const()[name = tensor("attn_37_transpose_y_0"), val = tensor(false)]; + tensor values_37_cast_fp16 = transpose(perm = var_4785, x = new_v_cache_37_cast_fp16)[name = tensor("transpose_40")]; + tensor attn_37_cast_fp16 = matmul(transpose_x = attn_37_transpose_x_0, transpose_y = attn_37_transpose_y_0, x = weights_37_cast_fp16, y = values_37_cast_fp16)[name = tensor("attn_37_cast_fp16")]; + tensor var_4830 = const()[name = tensor("op_4830"), val = tensor([0, 2, 1, 3])]; + tensor var_4835 = const()[name = tensor("op_4835"), val = tensor([1, 1, 1024])]; + tensor var_4831_cast_fp16 = transpose(perm = var_4830, x = attn_37_cast_fp16)[name = tensor("transpose_39")]; + tensor input_223_cast_fp16 = reshape(shape = var_4835, x = var_4831_cast_fp16)[name = tensor("input_223_cast_fp16")]; + tensor attn18_out_proj_weight_to_fp16 = const()[name = tensor("attn18_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(459520704)))]; + tensor linear_74_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn18_out_proj_weight_to_fp16, x = input_223_cast_fp16)[name = tensor("linear_74_cast_fp16")]; + tensor var_4841_to_fp16 = const()[name = tensor("op_4841_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4842_cast_fp16 = add(x = position18_to_fp16, y = var_4841_to_fp16)[name = tensor("op_4842_cast_fp16")]; + tensor var_4842_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4842_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_225_cast_fp16 = add(x = input_219_cast_fp16, y = linear_74_cast_fp16)[name = tensor("input_225_cast_fp16")]; + tensor input_227_axes_0 = const()[name = tensor("input_227_axes_0"), val = tensor([-1])]; + tensor norm18_2_weight_to_fp16 = const()[name = tensor("norm18_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(461617920)))]; + tensor norm18_2_bias_to_fp16 = const()[name = tensor("norm18_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(461620032)))]; + tensor var_4846_to_fp16 = const()[name = tensor("op_4846_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_227_cast_fp16 = layer_norm(axes = input_227_axes_0, beta = norm18_2_bias_to_fp16, epsilon = var_4846_to_fp16, gamma = norm18_2_weight_to_fp16, x = input_225_cast_fp16)[name = tensor("input_227_cast_fp16")]; + tensor linear18_1_weight_to_fp16 = const()[name = tensor("linear18_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(461622144)))]; + tensor linear_75_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear18_1_weight_to_fp16, x = input_227_cast_fp16)[name = tensor("linear_75_cast_fp16")]; + tensor input_229_mode_0 = const()[name = tensor("input_229_mode_0"), val = tensor("EXACT")]; + tensor input_229_cast_fp16 = gelu(mode = input_229_mode_0, x = linear_75_cast_fp16)[name = tensor("input_229_cast_fp16")]; + tensor linear18_2_weight_to_fp16 = const()[name = tensor("linear18_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(470010816)))]; + tensor linear_76_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear18_2_weight_to_fp16, x = input_229_cast_fp16)[name = tensor("linear_76_cast_fp16")]; + tensor input_231_cast_fp16 = add(x = input_225_cast_fp16, y = linear_76_cast_fp16)[name = tensor("input_231_cast_fp16")]; + tensor input_233_axes_0 = const()[name = tensor("input_233_axes_0"), val = tensor([-1])]; + tensor norm19_1_weight_to_fp16 = const()[name = tensor("norm19_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(478399488)))]; + tensor norm19_1_bias_to_fp16 = const()[name = tensor("norm19_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(478401600)))]; + tensor var_4863_to_fp16 = const()[name = tensor("op_4863_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_233_cast_fp16 = layer_norm(axes = input_233_axes_0, beta = norm19_1_bias_to_fp16, epsilon = var_4863_to_fp16, gamma = norm19_1_weight_to_fp16, x = input_231_cast_fp16)[name = tensor("input_233_cast_fp16")]; + tensor attn19_in_proj_weight_to_fp16 = const()[name = tensor("attn19_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(478403712)))]; + tensor linear_77_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn19_in_proj_weight_to_fp16, x = input_233_cast_fp16)[name = tensor("linear_77_cast_fp16")]; + tensor var_4876 = const()[name = tensor("op_4876"), val = tensor([1, 3, 16, 64])]; + tensor qkv_39_cast_fp16 = reshape(shape = var_4876, x = linear_77_cast_fp16)[name = tensor("qkv_39_cast_fp16")]; + tensor x_77_begin_0 = const()[name = tensor("x_77_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_77_end_0 = const()[name = tensor("x_77_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_77_end_mask_0 = const()[name = tensor("x_77_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_77_squeeze_mask_0 = const()[name = tensor("x_77_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_77_cast_fp16 = slice_by_index(begin = x_77_begin_0, end = x_77_end_0, end_mask = x_77_end_mask_0, squeeze_mask = x_77_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("x_77_cast_fp16")]; + tensor x_79_begin_0 = const()[name = tensor("x_79_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_79_end_0 = const()[name = tensor("x_79_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_79_end_mask_0 = const()[name = tensor("x_79_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_79_squeeze_mask_0 = const()[name = tensor("x_79_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_79_cast_fp16 = slice_by_index(begin = x_79_begin_0, end = x_79_end_0, end_mask = x_79_end_mask_0, squeeze_mask = x_79_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("x_79_cast_fp16")]; + tensor v_39_begin_0 = const()[name = tensor("v_39_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_39_end_0 = const()[name = tensor("v_39_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_39_end_mask_0 = const()[name = tensor("v_39_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_39_squeeze_mask_0 = const()[name = tensor("v_39_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_39_cast_fp16 = slice_by_index(begin = v_39_begin_0, end = v_39_end_0, end_mask = v_39_end_mask_0, squeeze_mask = v_39_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("v_39_cast_fp16")]; + tensor var_4905 = const()[name = tensor("op_4905"), val = tensor([1, 1, 1])]; + tensor position19_to_fp16_dtype_0 = const()[name = tensor("position19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position19_to_fp16 = cast(dtype = position19_to_fp16_dtype_0, x = position19)[name = tensor("cast_425")]; + tensor var_4906_cast_fp16 = reshape(shape = var_4905, x = position19_to_fp16)[name = tensor("op_4906_cast_fp16")]; + tensor angles_77_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_4906_cast_fp16)[name = tensor("angles_77_cast_fp16")]; + tensor rotr_77_cast_fp16 = cos(x = angles_77_cast_fp16)[name = tensor("rotr_77_cast_fp16")]; + tensor roti_77_cast_fp16 = sin(x = angles_77_cast_fp16)[name = tensor("roti_77_cast_fp16")]; + tensor var_4914 = const()[name = tensor("op_4914"), val = tensor([1, 16, 32, 2])]; + tensor pairs_77_cast_fp16 = reshape(shape = var_4914, x = x_77_cast_fp16)[name = tensor("pairs_77_cast_fp16")]; + tensor xr_77_begin_0 = const()[name = tensor("xr_77_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_77_end_0 = const()[name = tensor("xr_77_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_77_end_mask_0 = const()[name = tensor("xr_77_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_77_squeeze_mask_0 = const()[name = tensor("xr_77_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_77_cast_fp16 = slice_by_index(begin = xr_77_begin_0, end = xr_77_end_0, end_mask = xr_77_end_mask_0, squeeze_mask = xr_77_squeeze_mask_0, x = pairs_77_cast_fp16)[name = tensor("xr_77_cast_fp16")]; + tensor xi_77_begin_0 = const()[name = tensor("xi_77_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_77_end_0 = const()[name = tensor("xi_77_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_77_end_mask_0 = const()[name = tensor("xi_77_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_77_squeeze_mask_0 = const()[name = tensor("xi_77_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_77_cast_fp16 = slice_by_index(begin = xi_77_begin_0, end = xi_77_end_0, end_mask = xi_77_end_mask_0, squeeze_mask = xi_77_squeeze_mask_0, x = pairs_77_cast_fp16)[name = tensor("xi_77_cast_fp16")]; + tensor var_4922_cast_fp16 = mul(x = xr_77_cast_fp16, y = rotr_77_cast_fp16)[name = tensor("op_4922_cast_fp16")]; + tensor var_4923_cast_fp16 = mul(x = xi_77_cast_fp16, y = roti_77_cast_fp16)[name = tensor("op_4923_cast_fp16")]; + tensor xor__77_cast_fp16 = sub(x = var_4922_cast_fp16, y = var_4923_cast_fp16)[name = tensor("xor__77_cast_fp16")]; + tensor var_4926_cast_fp16 = mul(x = xr_77_cast_fp16, y = roti_77_cast_fp16)[name = tensor("op_4926_cast_fp16")]; + tensor var_4927_cast_fp16 = mul(x = xi_77_cast_fp16, y = rotr_77_cast_fp16)[name = tensor("op_4927_cast_fp16")]; + tensor xoi_77_cast_fp16 = add(x = var_4926_cast_fp16, y = var_4927_cast_fp16)[name = tensor("xoi_77_cast_fp16")]; + tensor out_77_axis_0 = const()[name = tensor("out_77_axis_0"), val = tensor(-1)]; + tensor out_77_cast_fp16 = stack(axis = out_77_axis_0, values = (xor__77_cast_fp16, xoi_77_cast_fp16))[name = tensor("out_77_cast_fp16")]; + tensor var_4950 = const()[name = tensor("op_4950"), val = tensor([1, 16, 32, 2])]; + tensor pairs_79_cast_fp16 = reshape(shape = var_4950, x = x_79_cast_fp16)[name = tensor("pairs_79_cast_fp16")]; + tensor xr_79_begin_0 = const()[name = tensor("xr_79_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_79_end_0 = const()[name = tensor("xr_79_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_79_end_mask_0 = const()[name = tensor("xr_79_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_79_squeeze_mask_0 = const()[name = tensor("xr_79_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_79_cast_fp16 = slice_by_index(begin = xr_79_begin_0, end = xr_79_end_0, end_mask = xr_79_end_mask_0, squeeze_mask = xr_79_squeeze_mask_0, x = pairs_79_cast_fp16)[name = tensor("xr_79_cast_fp16")]; + tensor xi_79_begin_0 = const()[name = tensor("xi_79_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_79_end_0 = const()[name = tensor("xi_79_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_79_end_mask_0 = const()[name = tensor("xi_79_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_79_squeeze_mask_0 = const()[name = tensor("xi_79_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_79_cast_fp16 = slice_by_index(begin = xi_79_begin_0, end = xi_79_end_0, end_mask = xi_79_end_mask_0, squeeze_mask = xi_79_squeeze_mask_0, x = pairs_79_cast_fp16)[name = tensor("xi_79_cast_fp16")]; + tensor var_4958_cast_fp16 = mul(x = xr_79_cast_fp16, y = rotr_77_cast_fp16)[name = tensor("op_4958_cast_fp16")]; + tensor var_4959_cast_fp16 = mul(x = xi_79_cast_fp16, y = roti_77_cast_fp16)[name = tensor("op_4959_cast_fp16")]; + tensor xor__79_cast_fp16 = sub(x = var_4958_cast_fp16, y = var_4959_cast_fp16)[name = tensor("xor__79_cast_fp16")]; + tensor var_4962_cast_fp16 = mul(x = xr_79_cast_fp16, y = roti_77_cast_fp16)[name = tensor("op_4962_cast_fp16")]; + tensor var_4963_cast_fp16 = mul(x = xi_79_cast_fp16, y = rotr_77_cast_fp16)[name = tensor("op_4963_cast_fp16")]; + tensor xoi_79_cast_fp16 = add(x = var_4962_cast_fp16, y = var_4963_cast_fp16)[name = tensor("xoi_79_cast_fp16")]; + tensor out_79_axis_0 = const()[name = tensor("out_79_axis_0"), val = tensor(-1)]; + tensor out_79_cast_fp16 = stack(axis = out_79_axis_0, values = (xor__79_cast_fp16, xoi_79_cast_fp16))[name = tensor("out_79_cast_fp16")]; + tensor _inversed_4975_y_0_to_fp16 = const()[name = tensor("_inversed_4975_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_4975_cast_fp16 = mul(x = position19_to_fp16, y = _inversed_4975_y_0_to_fp16)[name = tensor("_inversed_4975_cast_fp16")]; + tensor var_4976_cast_fp16 = floor(x = _inversed_4975_cast_fp16)[name = tensor("op_4976_cast_fp16")]; + tensor var_4977_to_fp16 = const()[name = tensor("op_4977_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4978_cast_fp16 = mul(x = var_4976_cast_fp16, y = var_4977_to_fp16)[name = tensor("op_4978_cast_fp16")]; + tensor write_idx_39_cast_fp16 = sub(x = position19_to_fp16, y = var_4978_cast_fp16)[name = tensor("write_idx_39_cast_fp16")]; + tensor var_4981_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_39_cast_fp16)[name = tensor("op_4981_cast_fp16")]; + tensor var_4991 = const()[name = tensor("op_4991"), val = tensor([1, 512, 1, 1])]; + tensor onehot_77_to_fp16_dtype_0 = const()[name = tensor("onehot_77_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_4981_cast_fp16_to_fp16 = cast(dtype = onehot_77_to_fp16_dtype_0, x = var_4981_cast_fp16)[name = tensor("cast_424")]; + tensor onehot_79_cast_fp16 = reshape(shape = var_4991, x = var_4981_cast_fp16_to_fp16)[name = tensor("onehot_79_cast_fp16")]; + tensor var_4993_to_fp16 = const()[name = tensor("op_4993_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4995_cast_fp16 = sub(x = var_4993_to_fp16, y = onehot_79_cast_fp16)[name = tensor("op_4995_cast_fp16")]; + tensor k_cache19_to_fp16_dtype_0 = const()[name = tensor("k_cache19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache19_to_fp16 = cast(dtype = k_cache19_to_fp16_dtype_0, x = k_cache19)[name = tensor("cast_423")]; + tensor var_4996_cast_fp16 = mul(x = k_cache19_to_fp16, y = var_4995_cast_fp16)[name = tensor("op_4996_cast_fp16")]; + tensor var_5001 = const()[name = tensor("op_5001"), val = tensor([1, 1, 16, 64])]; + tensor var_5002_cast_fp16 = reshape(shape = var_5001, x = out_79_cast_fp16)[name = tensor("op_5002_cast_fp16")]; + tensor var_5003_cast_fp16 = mul(x = var_5002_cast_fp16, y = onehot_79_cast_fp16)[name = tensor("op_5003_cast_fp16")]; + tensor new_k_cache_39_cast_fp16 = add(x = var_4996_cast_fp16, y = var_5003_cast_fp16)[name = tensor("new_k_cache_39_cast_fp16")]; + tensor new_k_cache_39_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_39_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache19_to_fp16_dtype_0 = const()[name = tensor("v_cache19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache19_to_fp16 = cast(dtype = v_cache19_to_fp16_dtype_0, x = v_cache19)[name = tensor("cast_421")]; + tensor var_5009_cast_fp16 = mul(x = v_cache19_to_fp16, y = var_4995_cast_fp16)[name = tensor("op_5009_cast_fp16")]; + tensor var_5014 = const()[name = tensor("op_5014"), val = tensor([1, 1, 16, 64])]; + tensor var_5015_cast_fp16 = reshape(shape = var_5014, x = v_39_cast_fp16)[name = tensor("op_5015_cast_fp16")]; + tensor var_5016_cast_fp16 = mul(x = var_5015_cast_fp16, y = onehot_79_cast_fp16)[name = tensor("op_5016_cast_fp16")]; + tensor new_v_cache_39_cast_fp16 = add(x = var_5009_cast_fp16, y = var_5016_cast_fp16)[name = tensor("new_v_cache_39_cast_fp16")]; + tensor new_v_cache_39_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_39_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_19_perm_0 = const()[name = tensor("transpose_19_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_5029 = const()[name = tensor("op_5029"), val = tensor([0, 2, 1, 3])]; + tensor var_5035 = const()[name = tensor("op_5035"), val = tensor([1, 16, 1, 64])]; + tensor q4_39_cast_fp16 = reshape(shape = var_5035, x = out_77_cast_fp16)[name = tensor("q4_39_cast_fp16")]; + tensor var_5040_transpose_x_0 = const()[name = tensor("op_5040_transpose_x_0"), val = tensor(false)]; + tensor var_5040_transpose_y_0 = const()[name = tensor("op_5040_transpose_y_0"), val = tensor(false)]; + tensor transpose_19_cast_fp16 = transpose(perm = transpose_19_perm_0, x = new_k_cache_39_cast_fp16)[name = tensor("transpose_38")]; + tensor var_5040_cast_fp16 = matmul(transpose_x = var_5040_transpose_x_0, transpose_y = var_5040_transpose_y_0, x = q4_39_cast_fp16, y = transpose_19_cast_fp16)[name = tensor("op_5040_cast_fp16")]; + tensor var_5041_to_fp16 = const()[name = tensor("op_5041_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_77_cast_fp16 = mul(x = var_5040_cast_fp16, y = var_5041_to_fp16)[name = tensor("scores_77_cast_fp16")]; + tensor var_5045 = const()[name = tensor("op_5045"), val = tensor([1, 1])]; + tensor var_5046_cast_fp16 = reshape(shape = var_5045, x = position19_to_fp16)[name = tensor("op_5046_cast_fp16")]; + tensor var_5047_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_5046_cast_fp16)[name = tensor("op_5047_cast_fp16")]; + tensor mask_39_to_fp16_dtype_0 = const()[name = tensor("mask_39_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_5054_to_fp16 = const()[name = tensor("op_5054_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5047_cast_fp16_to_fp16 = cast(dtype = mask_39_to_fp16_dtype_0, x = var_5047_cast_fp16)[name = tensor("cast_419")]; + tensor var_5055_cast_fp16 = sub(x = var_5047_cast_fp16_to_fp16, y = var_5054_to_fp16)[name = tensor("op_5055_cast_fp16")]; + tensor var_5060 = const()[name = tensor("op_5060"), val = tensor([1, 1, 1, 512])]; + tensor var_5061_cast_fp16 = reshape(shape = var_5060, x = var_5055_cast_fp16)[name = tensor("op_5061_cast_fp16")]; + tensor var_5062_to_fp16 = const()[name = tensor("op_5062_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_5063_cast_fp16 = mul(x = var_5061_cast_fp16, y = var_5062_to_fp16)[name = tensor("op_5063_cast_fp16")]; + tensor scores_79_cast_fp16 = add(x = scores_77_cast_fp16, y = var_5063_cast_fp16)[name = tensor("scores_79_cast_fp16")]; + tensor var_5066 = const()[name = tensor("op_5066"), val = tensor(-1)]; + tensor weights_39_cast_fp16 = softmax(axis = var_5066, x = scores_79_cast_fp16)[name = tensor("weights_39_cast_fp16")]; + tensor attn_39_transpose_x_0 = const()[name = tensor("attn_39_transpose_x_0"), val = tensor(false)]; + tensor attn_39_transpose_y_0 = const()[name = tensor("attn_39_transpose_y_0"), val = tensor(false)]; + tensor values_39_cast_fp16 = transpose(perm = var_5029, x = new_v_cache_39_cast_fp16)[name = tensor("transpose_37")]; + tensor attn_39_cast_fp16 = matmul(transpose_x = attn_39_transpose_x_0, transpose_y = attn_39_transpose_y_0, x = weights_39_cast_fp16, y = values_39_cast_fp16)[name = tensor("attn_39_cast_fp16")]; + tensor var_5074 = const()[name = tensor("op_5074"), val = tensor([0, 2, 1, 3])]; + tensor var_5079 = const()[name = tensor("op_5079"), val = tensor([1, 1, 1024])]; + tensor var_5075_cast_fp16 = transpose(perm = var_5074, x = attn_39_cast_fp16)[name = tensor("transpose_36")]; + tensor input_235_cast_fp16 = reshape(shape = var_5079, x = var_5075_cast_fp16)[name = tensor("input_235_cast_fp16")]; + tensor attn19_out_proj_weight_to_fp16 = const()[name = tensor("attn19_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(484695232)))]; + tensor linear_78_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn19_out_proj_weight_to_fp16, x = input_235_cast_fp16)[name = tensor("linear_78_cast_fp16")]; + tensor var_5085_to_fp16 = const()[name = tensor("op_5085_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5086_cast_fp16 = add(x = position19_to_fp16, y = var_5085_to_fp16)[name = tensor("op_5086_cast_fp16")]; + tensor var_5086_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5086_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_237_cast_fp16 = add(x = input_231_cast_fp16, y = linear_78_cast_fp16)[name = tensor("input_237_cast_fp16")]; + tensor input_239_axes_0 = const()[name = tensor("input_239_axes_0"), val = tensor([-1])]; + tensor norm19_2_weight_to_fp16 = const()[name = tensor("norm19_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486792448)))]; + tensor norm19_2_bias_to_fp16 = const()[name = tensor("norm19_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486794560)))]; + tensor var_5090_to_fp16 = const()[name = tensor("op_5090_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_239_cast_fp16 = layer_norm(axes = input_239_axes_0, beta = norm19_2_bias_to_fp16, epsilon = var_5090_to_fp16, gamma = norm19_2_weight_to_fp16, x = input_237_cast_fp16)[name = tensor("input_239_cast_fp16")]; + tensor linear19_1_weight_to_fp16 = const()[name = tensor("linear19_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486796672)))]; + tensor linear_79_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear19_1_weight_to_fp16, x = input_239_cast_fp16)[name = tensor("linear_79_cast_fp16")]; + tensor input_241_mode_0 = const()[name = tensor("input_241_mode_0"), val = tensor("EXACT")]; + tensor input_241_cast_fp16 = gelu(mode = input_241_mode_0, x = linear_79_cast_fp16)[name = tensor("input_241_cast_fp16")]; + tensor linear19_2_weight_to_fp16 = const()[name = tensor("linear19_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(495185344)))]; + tensor linear_80_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear19_2_weight_to_fp16, x = input_241_cast_fp16)[name = tensor("linear_80_cast_fp16")]; + tensor input_243_cast_fp16 = add(x = input_237_cast_fp16, y = linear_80_cast_fp16)[name = tensor("input_243_cast_fp16")]; + tensor input_245_axes_0 = const()[name = tensor("input_245_axes_0"), val = tensor([-1])]; + tensor norm20_1_weight_to_fp16 = const()[name = tensor("norm20_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(503574016)))]; + tensor norm20_1_bias_to_fp16 = const()[name = tensor("norm20_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(503576128)))]; + tensor var_5107_to_fp16 = const()[name = tensor("op_5107_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_245_cast_fp16 = layer_norm(axes = input_245_axes_0, beta = norm20_1_bias_to_fp16, epsilon = var_5107_to_fp16, gamma = norm20_1_weight_to_fp16, x = input_243_cast_fp16)[name = tensor("input_245_cast_fp16")]; + tensor attn20_in_proj_weight_to_fp16 = const()[name = tensor("attn20_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(503578240)))]; + tensor linear_81_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn20_in_proj_weight_to_fp16, x = input_245_cast_fp16)[name = tensor("linear_81_cast_fp16")]; + tensor var_5120 = const()[name = tensor("op_5120"), val = tensor([1, 3, 16, 64])]; + tensor qkv_41_cast_fp16 = reshape(shape = var_5120, x = linear_81_cast_fp16)[name = tensor("qkv_41_cast_fp16")]; + tensor x_81_begin_0 = const()[name = tensor("x_81_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_81_end_0 = const()[name = tensor("x_81_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_81_end_mask_0 = const()[name = tensor("x_81_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_81_squeeze_mask_0 = const()[name = tensor("x_81_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_81_cast_fp16 = slice_by_index(begin = x_81_begin_0, end = x_81_end_0, end_mask = x_81_end_mask_0, squeeze_mask = x_81_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("x_81_cast_fp16")]; + tensor x_83_begin_0 = const()[name = tensor("x_83_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_83_end_0 = const()[name = tensor("x_83_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_83_end_mask_0 = const()[name = tensor("x_83_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_83_squeeze_mask_0 = const()[name = tensor("x_83_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_83_cast_fp16 = slice_by_index(begin = x_83_begin_0, end = x_83_end_0, end_mask = x_83_end_mask_0, squeeze_mask = x_83_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("x_83_cast_fp16")]; + tensor v_41_begin_0 = const()[name = tensor("v_41_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_41_end_0 = const()[name = tensor("v_41_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_41_end_mask_0 = const()[name = tensor("v_41_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_41_squeeze_mask_0 = const()[name = tensor("v_41_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_41_cast_fp16 = slice_by_index(begin = v_41_begin_0, end = v_41_end_0, end_mask = v_41_end_mask_0, squeeze_mask = v_41_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("v_41_cast_fp16")]; + tensor var_5149 = const()[name = tensor("op_5149"), val = tensor([1, 1, 1])]; + tensor position20_to_fp16_dtype_0 = const()[name = tensor("position20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position20_to_fp16 = cast(dtype = position20_to_fp16_dtype_0, x = position20)[name = tensor("cast_417")]; + tensor var_5150_cast_fp16 = reshape(shape = var_5149, x = position20_to_fp16)[name = tensor("op_5150_cast_fp16")]; + tensor angles_81_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_5150_cast_fp16)[name = tensor("angles_81_cast_fp16")]; + tensor rotr_81_cast_fp16 = cos(x = angles_81_cast_fp16)[name = tensor("rotr_81_cast_fp16")]; + tensor roti_81_cast_fp16 = sin(x = angles_81_cast_fp16)[name = tensor("roti_81_cast_fp16")]; + tensor var_5158 = const()[name = tensor("op_5158"), val = tensor([1, 16, 32, 2])]; + tensor pairs_81_cast_fp16 = reshape(shape = var_5158, x = x_81_cast_fp16)[name = tensor("pairs_81_cast_fp16")]; + tensor xr_81_begin_0 = const()[name = tensor("xr_81_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_81_end_0 = const()[name = tensor("xr_81_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_81_end_mask_0 = const()[name = tensor("xr_81_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_81_squeeze_mask_0 = const()[name = tensor("xr_81_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_81_cast_fp16 = slice_by_index(begin = xr_81_begin_0, end = xr_81_end_0, end_mask = xr_81_end_mask_0, squeeze_mask = xr_81_squeeze_mask_0, x = pairs_81_cast_fp16)[name = tensor("xr_81_cast_fp16")]; + tensor xi_81_begin_0 = const()[name = tensor("xi_81_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_81_end_0 = const()[name = tensor("xi_81_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_81_end_mask_0 = const()[name = tensor("xi_81_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_81_squeeze_mask_0 = const()[name = tensor("xi_81_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_81_cast_fp16 = slice_by_index(begin = xi_81_begin_0, end = xi_81_end_0, end_mask = xi_81_end_mask_0, squeeze_mask = xi_81_squeeze_mask_0, x = pairs_81_cast_fp16)[name = tensor("xi_81_cast_fp16")]; + tensor var_5166_cast_fp16 = mul(x = xr_81_cast_fp16, y = rotr_81_cast_fp16)[name = tensor("op_5166_cast_fp16")]; + tensor var_5167_cast_fp16 = mul(x = xi_81_cast_fp16, y = roti_81_cast_fp16)[name = tensor("op_5167_cast_fp16")]; + tensor xor__81_cast_fp16 = sub(x = var_5166_cast_fp16, y = var_5167_cast_fp16)[name = tensor("xor__81_cast_fp16")]; + tensor var_5170_cast_fp16 = mul(x = xr_81_cast_fp16, y = roti_81_cast_fp16)[name = tensor("op_5170_cast_fp16")]; + tensor var_5171_cast_fp16 = mul(x = xi_81_cast_fp16, y = rotr_81_cast_fp16)[name = tensor("op_5171_cast_fp16")]; + tensor xoi_81_cast_fp16 = add(x = var_5170_cast_fp16, y = var_5171_cast_fp16)[name = tensor("xoi_81_cast_fp16")]; + tensor out_81_axis_0 = const()[name = tensor("out_81_axis_0"), val = tensor(-1)]; + tensor out_81_cast_fp16 = stack(axis = out_81_axis_0, values = (xor__81_cast_fp16, xoi_81_cast_fp16))[name = tensor("out_81_cast_fp16")]; + tensor var_5194 = const()[name = tensor("op_5194"), val = tensor([1, 16, 32, 2])]; + tensor pairs_83_cast_fp16 = reshape(shape = var_5194, x = x_83_cast_fp16)[name = tensor("pairs_83_cast_fp16")]; + tensor xr_83_begin_0 = const()[name = tensor("xr_83_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_83_end_0 = const()[name = tensor("xr_83_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_83_end_mask_0 = const()[name = tensor("xr_83_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_83_squeeze_mask_0 = const()[name = tensor("xr_83_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_83_cast_fp16 = slice_by_index(begin = xr_83_begin_0, end = xr_83_end_0, end_mask = xr_83_end_mask_0, squeeze_mask = xr_83_squeeze_mask_0, x = pairs_83_cast_fp16)[name = tensor("xr_83_cast_fp16")]; + tensor xi_83_begin_0 = const()[name = tensor("xi_83_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_83_end_0 = const()[name = tensor("xi_83_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_83_end_mask_0 = const()[name = tensor("xi_83_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_83_squeeze_mask_0 = const()[name = tensor("xi_83_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_83_cast_fp16 = slice_by_index(begin = xi_83_begin_0, end = xi_83_end_0, end_mask = xi_83_end_mask_0, squeeze_mask = xi_83_squeeze_mask_0, x = pairs_83_cast_fp16)[name = tensor("xi_83_cast_fp16")]; + tensor var_5202_cast_fp16 = mul(x = xr_83_cast_fp16, y = rotr_81_cast_fp16)[name = tensor("op_5202_cast_fp16")]; + tensor var_5203_cast_fp16 = mul(x = xi_83_cast_fp16, y = roti_81_cast_fp16)[name = tensor("op_5203_cast_fp16")]; + tensor xor__83_cast_fp16 = sub(x = var_5202_cast_fp16, y = var_5203_cast_fp16)[name = tensor("xor__83_cast_fp16")]; + tensor var_5206_cast_fp16 = mul(x = xr_83_cast_fp16, y = roti_81_cast_fp16)[name = tensor("op_5206_cast_fp16")]; + tensor var_5207_cast_fp16 = mul(x = xi_83_cast_fp16, y = rotr_81_cast_fp16)[name = tensor("op_5207_cast_fp16")]; + tensor xoi_83_cast_fp16 = add(x = var_5206_cast_fp16, y = var_5207_cast_fp16)[name = tensor("xoi_83_cast_fp16")]; + tensor out_83_axis_0 = const()[name = tensor("out_83_axis_0"), val = tensor(-1)]; + tensor out_83_cast_fp16 = stack(axis = out_83_axis_0, values = (xor__83_cast_fp16, xoi_83_cast_fp16))[name = tensor("out_83_cast_fp16")]; + tensor _inversed_5219_y_0_to_fp16 = const()[name = tensor("_inversed_5219_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5219_cast_fp16 = mul(x = position20_to_fp16, y = _inversed_5219_y_0_to_fp16)[name = tensor("_inversed_5219_cast_fp16")]; + tensor var_5220_cast_fp16 = floor(x = _inversed_5219_cast_fp16)[name = tensor("op_5220_cast_fp16")]; + tensor var_5221_to_fp16 = const()[name = tensor("op_5221_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5222_cast_fp16 = mul(x = var_5220_cast_fp16, y = var_5221_to_fp16)[name = tensor("op_5222_cast_fp16")]; + tensor write_idx_41_cast_fp16 = sub(x = position20_to_fp16, y = var_5222_cast_fp16)[name = tensor("write_idx_41_cast_fp16")]; + tensor var_5225_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_41_cast_fp16)[name = tensor("op_5225_cast_fp16")]; + tensor var_5235 = const()[name = tensor("op_5235"), val = tensor([1, 512, 1, 1])]; + tensor onehot_81_to_fp16_dtype_0 = const()[name = tensor("onehot_81_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_5225_cast_fp16_to_fp16 = cast(dtype = onehot_81_to_fp16_dtype_0, x = var_5225_cast_fp16)[name = tensor("cast_416")]; + tensor onehot_83_cast_fp16 = reshape(shape = var_5235, x = var_5225_cast_fp16_to_fp16)[name = tensor("onehot_83_cast_fp16")]; + tensor var_5237_to_fp16 = const()[name = tensor("op_5237_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5239_cast_fp16 = sub(x = var_5237_to_fp16, y = onehot_83_cast_fp16)[name = tensor("op_5239_cast_fp16")]; + tensor k_cache20_to_fp16_dtype_0 = const()[name = tensor("k_cache20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache20_to_fp16 = cast(dtype = k_cache20_to_fp16_dtype_0, x = k_cache20)[name = tensor("cast_415")]; + tensor var_5240_cast_fp16 = mul(x = k_cache20_to_fp16, y = var_5239_cast_fp16)[name = tensor("op_5240_cast_fp16")]; + tensor var_5245 = const()[name = tensor("op_5245"), val = tensor([1, 1, 16, 64])]; + tensor var_5246_cast_fp16 = reshape(shape = var_5245, x = out_83_cast_fp16)[name = tensor("op_5246_cast_fp16")]; + tensor var_5247_cast_fp16 = mul(x = var_5246_cast_fp16, y = onehot_83_cast_fp16)[name = tensor("op_5247_cast_fp16")]; + tensor new_k_cache_41_cast_fp16 = add(x = var_5240_cast_fp16, y = var_5247_cast_fp16)[name = tensor("new_k_cache_41_cast_fp16")]; + tensor new_k_cache_41_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_41_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache20_to_fp16_dtype_0 = const()[name = tensor("v_cache20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache20_to_fp16 = cast(dtype = v_cache20_to_fp16_dtype_0, x = v_cache20)[name = tensor("cast_413")]; + tensor var_5253_cast_fp16 = mul(x = v_cache20_to_fp16, y = var_5239_cast_fp16)[name = tensor("op_5253_cast_fp16")]; + tensor var_5258 = const()[name = tensor("op_5258"), val = tensor([1, 1, 16, 64])]; + tensor var_5259_cast_fp16 = reshape(shape = var_5258, x = v_41_cast_fp16)[name = tensor("op_5259_cast_fp16")]; + tensor var_5260_cast_fp16 = mul(x = var_5259_cast_fp16, y = onehot_83_cast_fp16)[name = tensor("op_5260_cast_fp16")]; + tensor new_v_cache_41_cast_fp16 = add(x = var_5253_cast_fp16, y = var_5260_cast_fp16)[name = tensor("new_v_cache_41_cast_fp16")]; + tensor new_v_cache_41_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_41_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_20_perm_0 = const()[name = tensor("transpose_20_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_5273 = const()[name = tensor("op_5273"), val = tensor([0, 2, 1, 3])]; + tensor var_5279 = const()[name = tensor("op_5279"), val = tensor([1, 16, 1, 64])]; + tensor q4_41_cast_fp16 = reshape(shape = var_5279, x = out_81_cast_fp16)[name = tensor("q4_41_cast_fp16")]; + tensor var_5284_transpose_x_0 = const()[name = tensor("op_5284_transpose_x_0"), val = tensor(false)]; + tensor var_5284_transpose_y_0 = const()[name = tensor("op_5284_transpose_y_0"), val = tensor(false)]; + tensor transpose_20_cast_fp16 = transpose(perm = transpose_20_perm_0, x = new_k_cache_41_cast_fp16)[name = tensor("transpose_35")]; + tensor var_5284_cast_fp16 = matmul(transpose_x = var_5284_transpose_x_0, transpose_y = var_5284_transpose_y_0, x = q4_41_cast_fp16, y = transpose_20_cast_fp16)[name = tensor("op_5284_cast_fp16")]; + tensor var_5285_to_fp16 = const()[name = tensor("op_5285_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_81_cast_fp16 = mul(x = var_5284_cast_fp16, y = var_5285_to_fp16)[name = tensor("scores_81_cast_fp16")]; + tensor var_5289 = const()[name = tensor("op_5289"), val = tensor([1, 1])]; + tensor var_5290_cast_fp16 = reshape(shape = var_5289, x = position20_to_fp16)[name = tensor("op_5290_cast_fp16")]; + tensor var_5291_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_5290_cast_fp16)[name = tensor("op_5291_cast_fp16")]; + tensor mask_41_to_fp16_dtype_0 = const()[name = tensor("mask_41_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_5298_to_fp16 = const()[name = tensor("op_5298_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5291_cast_fp16_to_fp16 = cast(dtype = mask_41_to_fp16_dtype_0, x = var_5291_cast_fp16)[name = tensor("cast_411")]; + tensor var_5299_cast_fp16 = sub(x = var_5291_cast_fp16_to_fp16, y = var_5298_to_fp16)[name = tensor("op_5299_cast_fp16")]; + tensor var_5304 = const()[name = tensor("op_5304"), val = tensor([1, 1, 1, 512])]; + tensor var_5305_cast_fp16 = reshape(shape = var_5304, x = var_5299_cast_fp16)[name = tensor("op_5305_cast_fp16")]; + tensor var_5306_to_fp16 = const()[name = tensor("op_5306_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_5307_cast_fp16 = mul(x = var_5305_cast_fp16, y = var_5306_to_fp16)[name = tensor("op_5307_cast_fp16")]; + tensor scores_83_cast_fp16 = add(x = scores_81_cast_fp16, y = var_5307_cast_fp16)[name = tensor("scores_83_cast_fp16")]; + tensor var_5310 = const()[name = tensor("op_5310"), val = tensor(-1)]; + tensor weights_41_cast_fp16 = softmax(axis = var_5310, x = scores_83_cast_fp16)[name = tensor("weights_41_cast_fp16")]; + tensor attn_41_transpose_x_0 = const()[name = tensor("attn_41_transpose_x_0"), val = tensor(false)]; + tensor attn_41_transpose_y_0 = const()[name = tensor("attn_41_transpose_y_0"), val = tensor(false)]; + tensor values_41_cast_fp16 = transpose(perm = var_5273, x = new_v_cache_41_cast_fp16)[name = tensor("transpose_34")]; + tensor attn_41_cast_fp16 = matmul(transpose_x = attn_41_transpose_x_0, transpose_y = attn_41_transpose_y_0, x = weights_41_cast_fp16, y = values_41_cast_fp16)[name = tensor("attn_41_cast_fp16")]; + tensor var_5318 = const()[name = tensor("op_5318"), val = tensor([0, 2, 1, 3])]; + tensor var_5323 = const()[name = tensor("op_5323"), val = tensor([1, 1, 1024])]; + tensor var_5319_cast_fp16 = transpose(perm = var_5318, x = attn_41_cast_fp16)[name = tensor("transpose_33")]; + tensor input_247_cast_fp16 = reshape(shape = var_5323, x = var_5319_cast_fp16)[name = tensor("input_247_cast_fp16")]; + tensor attn20_out_proj_weight_to_fp16 = const()[name = tensor("attn20_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(509869760)))]; + tensor linear_82_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn20_out_proj_weight_to_fp16, x = input_247_cast_fp16)[name = tensor("linear_82_cast_fp16")]; + tensor var_5329_to_fp16 = const()[name = tensor("op_5329_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5330_cast_fp16 = add(x = position20_to_fp16, y = var_5329_to_fp16)[name = tensor("op_5330_cast_fp16")]; + tensor var_5330_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5330_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_249_cast_fp16 = add(x = input_243_cast_fp16, y = linear_82_cast_fp16)[name = tensor("input_249_cast_fp16")]; + tensor input_251_axes_0 = const()[name = tensor("input_251_axes_0"), val = tensor([-1])]; + tensor norm20_2_weight_to_fp16 = const()[name = tensor("norm20_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(511966976)))]; + tensor norm20_2_bias_to_fp16 = const()[name = tensor("norm20_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(511969088)))]; + tensor var_5334_to_fp16 = const()[name = tensor("op_5334_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_251_cast_fp16 = layer_norm(axes = input_251_axes_0, beta = norm20_2_bias_to_fp16, epsilon = var_5334_to_fp16, gamma = norm20_2_weight_to_fp16, x = input_249_cast_fp16)[name = tensor("input_251_cast_fp16")]; + tensor linear20_1_weight_to_fp16 = const()[name = tensor("linear20_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(511971200)))]; + tensor linear_83_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear20_1_weight_to_fp16, x = input_251_cast_fp16)[name = tensor("linear_83_cast_fp16")]; + tensor input_253_mode_0 = const()[name = tensor("input_253_mode_0"), val = tensor("EXACT")]; + tensor input_253_cast_fp16 = gelu(mode = input_253_mode_0, x = linear_83_cast_fp16)[name = tensor("input_253_cast_fp16")]; + tensor linear20_2_weight_to_fp16 = const()[name = tensor("linear20_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(520359872)))]; + tensor linear_84_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear20_2_weight_to_fp16, x = input_253_cast_fp16)[name = tensor("linear_84_cast_fp16")]; + tensor input_255_cast_fp16 = add(x = input_249_cast_fp16, y = linear_84_cast_fp16)[name = tensor("input_255_cast_fp16")]; + tensor input_257_axes_0 = const()[name = tensor("input_257_axes_0"), val = tensor([-1])]; + tensor norm21_1_weight_to_fp16 = const()[name = tensor("norm21_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(528748544)))]; + tensor norm21_1_bias_to_fp16 = const()[name = tensor("norm21_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(528750656)))]; + tensor var_5351_to_fp16 = const()[name = tensor("op_5351_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_257_cast_fp16 = layer_norm(axes = input_257_axes_0, beta = norm21_1_bias_to_fp16, epsilon = var_5351_to_fp16, gamma = norm21_1_weight_to_fp16, x = input_255_cast_fp16)[name = tensor("input_257_cast_fp16")]; + tensor attn21_in_proj_weight_to_fp16 = const()[name = tensor("attn21_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(528752768)))]; + tensor linear_85_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn21_in_proj_weight_to_fp16, x = input_257_cast_fp16)[name = tensor("linear_85_cast_fp16")]; + tensor var_5364 = const()[name = tensor("op_5364"), val = tensor([1, 3, 16, 64])]; + tensor qkv_43_cast_fp16 = reshape(shape = var_5364, x = linear_85_cast_fp16)[name = tensor("qkv_43_cast_fp16")]; + tensor x_85_begin_0 = const()[name = tensor("x_85_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_85_end_0 = const()[name = tensor("x_85_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_85_end_mask_0 = const()[name = tensor("x_85_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_85_squeeze_mask_0 = const()[name = tensor("x_85_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_85_cast_fp16 = slice_by_index(begin = x_85_begin_0, end = x_85_end_0, end_mask = x_85_end_mask_0, squeeze_mask = x_85_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("x_85_cast_fp16")]; + tensor x_87_begin_0 = const()[name = tensor("x_87_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_87_end_0 = const()[name = tensor("x_87_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_87_end_mask_0 = const()[name = tensor("x_87_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_87_squeeze_mask_0 = const()[name = tensor("x_87_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_87_cast_fp16 = slice_by_index(begin = x_87_begin_0, end = x_87_end_0, end_mask = x_87_end_mask_0, squeeze_mask = x_87_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("x_87_cast_fp16")]; + tensor v_43_begin_0 = const()[name = tensor("v_43_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_43_end_0 = const()[name = tensor("v_43_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_43_end_mask_0 = const()[name = tensor("v_43_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_43_squeeze_mask_0 = const()[name = tensor("v_43_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_43_cast_fp16 = slice_by_index(begin = v_43_begin_0, end = v_43_end_0, end_mask = v_43_end_mask_0, squeeze_mask = v_43_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("v_43_cast_fp16")]; + tensor var_5393 = const()[name = tensor("op_5393"), val = tensor([1, 1, 1])]; + tensor position21_to_fp16_dtype_0 = const()[name = tensor("position21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position21_to_fp16 = cast(dtype = position21_to_fp16_dtype_0, x = position21)[name = tensor("cast_409")]; + tensor var_5394_cast_fp16 = reshape(shape = var_5393, x = position21_to_fp16)[name = tensor("op_5394_cast_fp16")]; + tensor angles_85_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_5394_cast_fp16)[name = tensor("angles_85_cast_fp16")]; + tensor rotr_85_cast_fp16 = cos(x = angles_85_cast_fp16)[name = tensor("rotr_85_cast_fp16")]; + tensor roti_85_cast_fp16 = sin(x = angles_85_cast_fp16)[name = tensor("roti_85_cast_fp16")]; + tensor var_5402 = const()[name = tensor("op_5402"), val = tensor([1, 16, 32, 2])]; + tensor pairs_85_cast_fp16 = reshape(shape = var_5402, x = x_85_cast_fp16)[name = tensor("pairs_85_cast_fp16")]; + tensor xr_85_begin_0 = const()[name = tensor("xr_85_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_85_end_0 = const()[name = tensor("xr_85_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_85_end_mask_0 = const()[name = tensor("xr_85_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_85_squeeze_mask_0 = const()[name = tensor("xr_85_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_85_cast_fp16 = slice_by_index(begin = xr_85_begin_0, end = xr_85_end_0, end_mask = xr_85_end_mask_0, squeeze_mask = xr_85_squeeze_mask_0, x = pairs_85_cast_fp16)[name = tensor("xr_85_cast_fp16")]; + tensor xi_85_begin_0 = const()[name = tensor("xi_85_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_85_end_0 = const()[name = tensor("xi_85_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_85_end_mask_0 = const()[name = tensor("xi_85_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_85_squeeze_mask_0 = const()[name = tensor("xi_85_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_85_cast_fp16 = slice_by_index(begin = xi_85_begin_0, end = xi_85_end_0, end_mask = xi_85_end_mask_0, squeeze_mask = xi_85_squeeze_mask_0, x = pairs_85_cast_fp16)[name = tensor("xi_85_cast_fp16")]; + tensor var_5410_cast_fp16 = mul(x = xr_85_cast_fp16, y = rotr_85_cast_fp16)[name = tensor("op_5410_cast_fp16")]; + tensor var_5411_cast_fp16 = mul(x = xi_85_cast_fp16, y = roti_85_cast_fp16)[name = tensor("op_5411_cast_fp16")]; + tensor xor__85_cast_fp16 = sub(x = var_5410_cast_fp16, y = var_5411_cast_fp16)[name = tensor("xor__85_cast_fp16")]; + tensor var_5414_cast_fp16 = mul(x = xr_85_cast_fp16, y = roti_85_cast_fp16)[name = tensor("op_5414_cast_fp16")]; + tensor var_5415_cast_fp16 = mul(x = xi_85_cast_fp16, y = rotr_85_cast_fp16)[name = tensor("op_5415_cast_fp16")]; + tensor xoi_85_cast_fp16 = add(x = var_5414_cast_fp16, y = var_5415_cast_fp16)[name = tensor("xoi_85_cast_fp16")]; + tensor out_85_axis_0 = const()[name = tensor("out_85_axis_0"), val = tensor(-1)]; + tensor out_85_cast_fp16 = stack(axis = out_85_axis_0, values = (xor__85_cast_fp16, xoi_85_cast_fp16))[name = tensor("out_85_cast_fp16")]; + tensor var_5438 = const()[name = tensor("op_5438"), val = tensor([1, 16, 32, 2])]; + tensor pairs_87_cast_fp16 = reshape(shape = var_5438, x = x_87_cast_fp16)[name = tensor("pairs_87_cast_fp16")]; + tensor xr_87_begin_0 = const()[name = tensor("xr_87_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_87_end_0 = const()[name = tensor("xr_87_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_87_end_mask_0 = const()[name = tensor("xr_87_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_87_squeeze_mask_0 = const()[name = tensor("xr_87_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_87_cast_fp16 = slice_by_index(begin = xr_87_begin_0, end = xr_87_end_0, end_mask = xr_87_end_mask_0, squeeze_mask = xr_87_squeeze_mask_0, x = pairs_87_cast_fp16)[name = tensor("xr_87_cast_fp16")]; + tensor xi_87_begin_0 = const()[name = tensor("xi_87_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_87_end_0 = const()[name = tensor("xi_87_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_87_end_mask_0 = const()[name = tensor("xi_87_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_87_squeeze_mask_0 = const()[name = tensor("xi_87_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_87_cast_fp16 = slice_by_index(begin = xi_87_begin_0, end = xi_87_end_0, end_mask = xi_87_end_mask_0, squeeze_mask = xi_87_squeeze_mask_0, x = pairs_87_cast_fp16)[name = tensor("xi_87_cast_fp16")]; + tensor var_5446_cast_fp16 = mul(x = xr_87_cast_fp16, y = rotr_85_cast_fp16)[name = tensor("op_5446_cast_fp16")]; + tensor var_5447_cast_fp16 = mul(x = xi_87_cast_fp16, y = roti_85_cast_fp16)[name = tensor("op_5447_cast_fp16")]; + tensor xor__87_cast_fp16 = sub(x = var_5446_cast_fp16, y = var_5447_cast_fp16)[name = tensor("xor__87_cast_fp16")]; + tensor var_5450_cast_fp16 = mul(x = xr_87_cast_fp16, y = roti_85_cast_fp16)[name = tensor("op_5450_cast_fp16")]; + tensor var_5451_cast_fp16 = mul(x = xi_87_cast_fp16, y = rotr_85_cast_fp16)[name = tensor("op_5451_cast_fp16")]; + tensor xoi_87_cast_fp16 = add(x = var_5450_cast_fp16, y = var_5451_cast_fp16)[name = tensor("xoi_87_cast_fp16")]; + tensor out_87_axis_0 = const()[name = tensor("out_87_axis_0"), val = tensor(-1)]; + tensor out_87_cast_fp16 = stack(axis = out_87_axis_0, values = (xor__87_cast_fp16, xoi_87_cast_fp16))[name = tensor("out_87_cast_fp16")]; + tensor _inversed_5463_y_0_to_fp16 = const()[name = tensor("_inversed_5463_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5463_cast_fp16 = mul(x = position21_to_fp16, y = _inversed_5463_y_0_to_fp16)[name = tensor("_inversed_5463_cast_fp16")]; + tensor var_5464_cast_fp16 = floor(x = _inversed_5463_cast_fp16)[name = tensor("op_5464_cast_fp16")]; + tensor var_5465_to_fp16 = const()[name = tensor("op_5465_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5466_cast_fp16 = mul(x = var_5464_cast_fp16, y = var_5465_to_fp16)[name = tensor("op_5466_cast_fp16")]; + tensor write_idx_43_cast_fp16 = sub(x = position21_to_fp16, y = var_5466_cast_fp16)[name = tensor("write_idx_43_cast_fp16")]; + tensor var_5469_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_43_cast_fp16)[name = tensor("op_5469_cast_fp16")]; + tensor var_5479 = const()[name = tensor("op_5479"), val = tensor([1, 512, 1, 1])]; + tensor onehot_85_to_fp16_dtype_0 = const()[name = tensor("onehot_85_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_5469_cast_fp16_to_fp16 = cast(dtype = onehot_85_to_fp16_dtype_0, x = var_5469_cast_fp16)[name = tensor("cast_408")]; + tensor onehot_87_cast_fp16 = reshape(shape = var_5479, x = var_5469_cast_fp16_to_fp16)[name = tensor("onehot_87_cast_fp16")]; + tensor var_5481_to_fp16 = const()[name = tensor("op_5481_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5483_cast_fp16 = sub(x = var_5481_to_fp16, y = onehot_87_cast_fp16)[name = tensor("op_5483_cast_fp16")]; + tensor k_cache21_to_fp16_dtype_0 = const()[name = tensor("k_cache21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache21_to_fp16 = cast(dtype = k_cache21_to_fp16_dtype_0, x = k_cache21)[name = tensor("cast_407")]; + tensor var_5484_cast_fp16 = mul(x = k_cache21_to_fp16, y = var_5483_cast_fp16)[name = tensor("op_5484_cast_fp16")]; + tensor var_5489 = const()[name = tensor("op_5489"), val = tensor([1, 1, 16, 64])]; + tensor var_5490_cast_fp16 = reshape(shape = var_5489, x = out_87_cast_fp16)[name = tensor("op_5490_cast_fp16")]; + tensor var_5491_cast_fp16 = mul(x = var_5490_cast_fp16, y = onehot_87_cast_fp16)[name = tensor("op_5491_cast_fp16")]; + tensor new_k_cache_43_cast_fp16 = add(x = var_5484_cast_fp16, y = var_5491_cast_fp16)[name = tensor("new_k_cache_43_cast_fp16")]; + tensor new_k_cache_43_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_43_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache21_to_fp16_dtype_0 = const()[name = tensor("v_cache21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache21_to_fp16 = cast(dtype = v_cache21_to_fp16_dtype_0, x = v_cache21)[name = tensor("cast_405")]; + tensor var_5497_cast_fp16 = mul(x = v_cache21_to_fp16, y = var_5483_cast_fp16)[name = tensor("op_5497_cast_fp16")]; + tensor var_5502 = const()[name = tensor("op_5502"), val = tensor([1, 1, 16, 64])]; + tensor var_5503_cast_fp16 = reshape(shape = var_5502, x = v_43_cast_fp16)[name = tensor("op_5503_cast_fp16")]; + tensor var_5504_cast_fp16 = mul(x = var_5503_cast_fp16, y = onehot_87_cast_fp16)[name = tensor("op_5504_cast_fp16")]; + tensor new_v_cache_43_cast_fp16 = add(x = var_5497_cast_fp16, y = var_5504_cast_fp16)[name = tensor("new_v_cache_43_cast_fp16")]; + tensor new_v_cache_43_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_43_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_21_perm_0 = const()[name = tensor("transpose_21_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_5517 = const()[name = tensor("op_5517"), val = tensor([0, 2, 1, 3])]; + tensor var_5523 = const()[name = tensor("op_5523"), val = tensor([1, 16, 1, 64])]; + tensor q4_43_cast_fp16 = reshape(shape = var_5523, x = out_85_cast_fp16)[name = tensor("q4_43_cast_fp16")]; + tensor var_5528_transpose_x_0 = const()[name = tensor("op_5528_transpose_x_0"), val = tensor(false)]; + tensor var_5528_transpose_y_0 = const()[name = tensor("op_5528_transpose_y_0"), val = tensor(false)]; + tensor transpose_21_cast_fp16 = transpose(perm = transpose_21_perm_0, x = new_k_cache_43_cast_fp16)[name = tensor("transpose_32")]; + tensor var_5528_cast_fp16 = matmul(transpose_x = var_5528_transpose_x_0, transpose_y = var_5528_transpose_y_0, x = q4_43_cast_fp16, y = transpose_21_cast_fp16)[name = tensor("op_5528_cast_fp16")]; + tensor var_5529_to_fp16 = const()[name = tensor("op_5529_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_85_cast_fp16 = mul(x = var_5528_cast_fp16, y = var_5529_to_fp16)[name = tensor("scores_85_cast_fp16")]; + tensor var_5533 = const()[name = tensor("op_5533"), val = tensor([1, 1])]; + tensor var_5534_cast_fp16 = reshape(shape = var_5533, x = position21_to_fp16)[name = tensor("op_5534_cast_fp16")]; + tensor var_5535_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_5534_cast_fp16)[name = tensor("op_5535_cast_fp16")]; + tensor mask_43_to_fp16_dtype_0 = const()[name = tensor("mask_43_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_5542_to_fp16 = const()[name = tensor("op_5542_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5535_cast_fp16_to_fp16 = cast(dtype = mask_43_to_fp16_dtype_0, x = var_5535_cast_fp16)[name = tensor("cast_403")]; + tensor var_5543_cast_fp16 = sub(x = var_5535_cast_fp16_to_fp16, y = var_5542_to_fp16)[name = tensor("op_5543_cast_fp16")]; + tensor var_5548 = const()[name = tensor("op_5548"), val = tensor([1, 1, 1, 512])]; + tensor var_5549_cast_fp16 = reshape(shape = var_5548, x = var_5543_cast_fp16)[name = tensor("op_5549_cast_fp16")]; + tensor var_5550_to_fp16 = const()[name = tensor("op_5550_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_5551_cast_fp16 = mul(x = var_5549_cast_fp16, y = var_5550_to_fp16)[name = tensor("op_5551_cast_fp16")]; + tensor scores_87_cast_fp16 = add(x = scores_85_cast_fp16, y = var_5551_cast_fp16)[name = tensor("scores_87_cast_fp16")]; + tensor var_5554 = const()[name = tensor("op_5554"), val = tensor(-1)]; + tensor weights_43_cast_fp16 = softmax(axis = var_5554, x = scores_87_cast_fp16)[name = tensor("weights_43_cast_fp16")]; + tensor attn_43_transpose_x_0 = const()[name = tensor("attn_43_transpose_x_0"), val = tensor(false)]; + tensor attn_43_transpose_y_0 = const()[name = tensor("attn_43_transpose_y_0"), val = tensor(false)]; + tensor values_43_cast_fp16 = transpose(perm = var_5517, x = new_v_cache_43_cast_fp16)[name = tensor("transpose_31")]; + tensor attn_43_cast_fp16 = matmul(transpose_x = attn_43_transpose_x_0, transpose_y = attn_43_transpose_y_0, x = weights_43_cast_fp16, y = values_43_cast_fp16)[name = tensor("attn_43_cast_fp16")]; + tensor var_5562 = const()[name = tensor("op_5562"), val = tensor([0, 2, 1, 3])]; + tensor var_5567 = const()[name = tensor("op_5567"), val = tensor([1, 1, 1024])]; + tensor var_5563_cast_fp16 = transpose(perm = var_5562, x = attn_43_cast_fp16)[name = tensor("transpose_30")]; + tensor input_259_cast_fp16 = reshape(shape = var_5567, x = var_5563_cast_fp16)[name = tensor("input_259_cast_fp16")]; + tensor attn21_out_proj_weight_to_fp16 = const()[name = tensor("attn21_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(535044288)))]; + tensor linear_86_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn21_out_proj_weight_to_fp16, x = input_259_cast_fp16)[name = tensor("linear_86_cast_fp16")]; + tensor var_5573_to_fp16 = const()[name = tensor("op_5573_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5574_cast_fp16 = add(x = position21_to_fp16, y = var_5573_to_fp16)[name = tensor("op_5574_cast_fp16")]; + tensor var_5574_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5574_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_261_cast_fp16 = add(x = input_255_cast_fp16, y = linear_86_cast_fp16)[name = tensor("input_261_cast_fp16")]; + tensor input_263_axes_0 = const()[name = tensor("input_263_axes_0"), val = tensor([-1])]; + tensor norm21_2_weight_to_fp16 = const()[name = tensor("norm21_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(537141504)))]; + tensor norm21_2_bias_to_fp16 = const()[name = tensor("norm21_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(537143616)))]; + tensor var_5578_to_fp16 = const()[name = tensor("op_5578_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_263_cast_fp16 = layer_norm(axes = input_263_axes_0, beta = norm21_2_bias_to_fp16, epsilon = var_5578_to_fp16, gamma = norm21_2_weight_to_fp16, x = input_261_cast_fp16)[name = tensor("input_263_cast_fp16")]; + tensor linear21_1_weight_to_fp16 = const()[name = tensor("linear21_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(537145728)))]; + tensor linear_87_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear21_1_weight_to_fp16, x = input_263_cast_fp16)[name = tensor("linear_87_cast_fp16")]; + tensor input_265_mode_0 = const()[name = tensor("input_265_mode_0"), val = tensor("EXACT")]; + tensor input_265_cast_fp16 = gelu(mode = input_265_mode_0, x = linear_87_cast_fp16)[name = tensor("input_265_cast_fp16")]; + tensor linear21_2_weight_to_fp16 = const()[name = tensor("linear21_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(545534400)))]; + tensor linear_88_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear21_2_weight_to_fp16, x = input_265_cast_fp16)[name = tensor("linear_88_cast_fp16")]; + tensor input_267_cast_fp16 = add(x = input_261_cast_fp16, y = linear_88_cast_fp16)[name = tensor("input_267_cast_fp16")]; + tensor input_269_axes_0 = const()[name = tensor("input_269_axes_0"), val = tensor([-1])]; + tensor norm22_1_weight_to_fp16 = const()[name = tensor("norm22_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(553923072)))]; + tensor norm22_1_bias_to_fp16 = const()[name = tensor("norm22_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(553925184)))]; + tensor var_5595_to_fp16 = const()[name = tensor("op_5595_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_269_cast_fp16 = layer_norm(axes = input_269_axes_0, beta = norm22_1_bias_to_fp16, epsilon = var_5595_to_fp16, gamma = norm22_1_weight_to_fp16, x = input_267_cast_fp16)[name = tensor("input_269_cast_fp16")]; + tensor attn22_in_proj_weight_to_fp16 = const()[name = tensor("attn22_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(553927296)))]; + tensor linear_89_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn22_in_proj_weight_to_fp16, x = input_269_cast_fp16)[name = tensor("linear_89_cast_fp16")]; + tensor var_5608 = const()[name = tensor("op_5608"), val = tensor([1, 3, 16, 64])]; + tensor qkv_45_cast_fp16 = reshape(shape = var_5608, x = linear_89_cast_fp16)[name = tensor("qkv_45_cast_fp16")]; + tensor x_89_begin_0 = const()[name = tensor("x_89_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_89_end_0 = const()[name = tensor("x_89_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_89_end_mask_0 = const()[name = tensor("x_89_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_89_squeeze_mask_0 = const()[name = tensor("x_89_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_89_cast_fp16 = slice_by_index(begin = x_89_begin_0, end = x_89_end_0, end_mask = x_89_end_mask_0, squeeze_mask = x_89_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("x_89_cast_fp16")]; + tensor x_91_begin_0 = const()[name = tensor("x_91_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_91_end_0 = const()[name = tensor("x_91_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_91_end_mask_0 = const()[name = tensor("x_91_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_91_squeeze_mask_0 = const()[name = tensor("x_91_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_91_cast_fp16 = slice_by_index(begin = x_91_begin_0, end = x_91_end_0, end_mask = x_91_end_mask_0, squeeze_mask = x_91_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("x_91_cast_fp16")]; + tensor v_45_begin_0 = const()[name = tensor("v_45_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_45_end_0 = const()[name = tensor("v_45_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_45_end_mask_0 = const()[name = tensor("v_45_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_45_squeeze_mask_0 = const()[name = tensor("v_45_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_45_cast_fp16 = slice_by_index(begin = v_45_begin_0, end = v_45_end_0, end_mask = v_45_end_mask_0, squeeze_mask = v_45_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("v_45_cast_fp16")]; + tensor var_5637 = const()[name = tensor("op_5637"), val = tensor([1, 1, 1])]; + tensor position22_to_fp16_dtype_0 = const()[name = tensor("position22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position22_to_fp16 = cast(dtype = position22_to_fp16_dtype_0, x = position22)[name = tensor("cast_401")]; + tensor var_5638_cast_fp16 = reshape(shape = var_5637, x = position22_to_fp16)[name = tensor("op_5638_cast_fp16")]; + tensor angles_89_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_5638_cast_fp16)[name = tensor("angles_89_cast_fp16")]; + tensor rotr_89_cast_fp16 = cos(x = angles_89_cast_fp16)[name = tensor("rotr_89_cast_fp16")]; + tensor roti_89_cast_fp16 = sin(x = angles_89_cast_fp16)[name = tensor("roti_89_cast_fp16")]; + tensor var_5646 = const()[name = tensor("op_5646"), val = tensor([1, 16, 32, 2])]; + tensor pairs_89_cast_fp16 = reshape(shape = var_5646, x = x_89_cast_fp16)[name = tensor("pairs_89_cast_fp16")]; + tensor xr_89_begin_0 = const()[name = tensor("xr_89_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_89_end_0 = const()[name = tensor("xr_89_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_89_end_mask_0 = const()[name = tensor("xr_89_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_89_squeeze_mask_0 = const()[name = tensor("xr_89_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_89_cast_fp16 = slice_by_index(begin = xr_89_begin_0, end = xr_89_end_0, end_mask = xr_89_end_mask_0, squeeze_mask = xr_89_squeeze_mask_0, x = pairs_89_cast_fp16)[name = tensor("xr_89_cast_fp16")]; + tensor xi_89_begin_0 = const()[name = tensor("xi_89_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_89_end_0 = const()[name = tensor("xi_89_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_89_end_mask_0 = const()[name = tensor("xi_89_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_89_squeeze_mask_0 = const()[name = tensor("xi_89_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_89_cast_fp16 = slice_by_index(begin = xi_89_begin_0, end = xi_89_end_0, end_mask = xi_89_end_mask_0, squeeze_mask = xi_89_squeeze_mask_0, x = pairs_89_cast_fp16)[name = tensor("xi_89_cast_fp16")]; + tensor var_5654_cast_fp16 = mul(x = xr_89_cast_fp16, y = rotr_89_cast_fp16)[name = tensor("op_5654_cast_fp16")]; + tensor var_5655_cast_fp16 = mul(x = xi_89_cast_fp16, y = roti_89_cast_fp16)[name = tensor("op_5655_cast_fp16")]; + tensor xor__89_cast_fp16 = sub(x = var_5654_cast_fp16, y = var_5655_cast_fp16)[name = tensor("xor__89_cast_fp16")]; + tensor var_5658_cast_fp16 = mul(x = xr_89_cast_fp16, y = roti_89_cast_fp16)[name = tensor("op_5658_cast_fp16")]; + tensor var_5659_cast_fp16 = mul(x = xi_89_cast_fp16, y = rotr_89_cast_fp16)[name = tensor("op_5659_cast_fp16")]; + tensor xoi_89_cast_fp16 = add(x = var_5658_cast_fp16, y = var_5659_cast_fp16)[name = tensor("xoi_89_cast_fp16")]; + tensor out_89_axis_0 = const()[name = tensor("out_89_axis_0"), val = tensor(-1)]; + tensor out_89_cast_fp16 = stack(axis = out_89_axis_0, values = (xor__89_cast_fp16, xoi_89_cast_fp16))[name = tensor("out_89_cast_fp16")]; + tensor var_5682 = const()[name = tensor("op_5682"), val = tensor([1, 16, 32, 2])]; + tensor pairs_91_cast_fp16 = reshape(shape = var_5682, x = x_91_cast_fp16)[name = tensor("pairs_91_cast_fp16")]; + tensor xr_91_begin_0 = const()[name = tensor("xr_91_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_91_end_0 = const()[name = tensor("xr_91_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_91_end_mask_0 = const()[name = tensor("xr_91_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_91_squeeze_mask_0 = const()[name = tensor("xr_91_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_91_cast_fp16 = slice_by_index(begin = xr_91_begin_0, end = xr_91_end_0, end_mask = xr_91_end_mask_0, squeeze_mask = xr_91_squeeze_mask_0, x = pairs_91_cast_fp16)[name = tensor("xr_91_cast_fp16")]; + tensor xi_91_begin_0 = const()[name = tensor("xi_91_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_91_end_0 = const()[name = tensor("xi_91_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_91_end_mask_0 = const()[name = tensor("xi_91_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_91_squeeze_mask_0 = const()[name = tensor("xi_91_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_91_cast_fp16 = slice_by_index(begin = xi_91_begin_0, end = xi_91_end_0, end_mask = xi_91_end_mask_0, squeeze_mask = xi_91_squeeze_mask_0, x = pairs_91_cast_fp16)[name = tensor("xi_91_cast_fp16")]; + tensor var_5690_cast_fp16 = mul(x = xr_91_cast_fp16, y = rotr_89_cast_fp16)[name = tensor("op_5690_cast_fp16")]; + tensor var_5691_cast_fp16 = mul(x = xi_91_cast_fp16, y = roti_89_cast_fp16)[name = tensor("op_5691_cast_fp16")]; + tensor xor__91_cast_fp16 = sub(x = var_5690_cast_fp16, y = var_5691_cast_fp16)[name = tensor("xor__91_cast_fp16")]; + tensor var_5694_cast_fp16 = mul(x = xr_91_cast_fp16, y = roti_89_cast_fp16)[name = tensor("op_5694_cast_fp16")]; + tensor var_5695_cast_fp16 = mul(x = xi_91_cast_fp16, y = rotr_89_cast_fp16)[name = tensor("op_5695_cast_fp16")]; + tensor xoi_91_cast_fp16 = add(x = var_5694_cast_fp16, y = var_5695_cast_fp16)[name = tensor("xoi_91_cast_fp16")]; + tensor out_91_axis_0 = const()[name = tensor("out_91_axis_0"), val = tensor(-1)]; + tensor out_91_cast_fp16 = stack(axis = out_91_axis_0, values = (xor__91_cast_fp16, xoi_91_cast_fp16))[name = tensor("out_91_cast_fp16")]; + tensor _inversed_5707_y_0_to_fp16 = const()[name = tensor("_inversed_5707_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5707_cast_fp16 = mul(x = position22_to_fp16, y = _inversed_5707_y_0_to_fp16)[name = tensor("_inversed_5707_cast_fp16")]; + tensor var_5708_cast_fp16 = floor(x = _inversed_5707_cast_fp16)[name = tensor("op_5708_cast_fp16")]; + tensor var_5709_to_fp16 = const()[name = tensor("op_5709_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5710_cast_fp16 = mul(x = var_5708_cast_fp16, y = var_5709_to_fp16)[name = tensor("op_5710_cast_fp16")]; + tensor write_idx_45_cast_fp16 = sub(x = position22_to_fp16, y = var_5710_cast_fp16)[name = tensor("write_idx_45_cast_fp16")]; + tensor var_5713_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_45_cast_fp16)[name = tensor("op_5713_cast_fp16")]; + tensor var_5723 = const()[name = tensor("op_5723"), val = tensor([1, 512, 1, 1])]; + tensor onehot_89_to_fp16_dtype_0 = const()[name = tensor("onehot_89_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_5713_cast_fp16_to_fp16 = cast(dtype = onehot_89_to_fp16_dtype_0, x = var_5713_cast_fp16)[name = tensor("cast_400")]; + tensor onehot_91_cast_fp16 = reshape(shape = var_5723, x = var_5713_cast_fp16_to_fp16)[name = tensor("onehot_91_cast_fp16")]; + tensor var_5725_to_fp16 = const()[name = tensor("op_5725_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5727_cast_fp16 = sub(x = var_5725_to_fp16, y = onehot_91_cast_fp16)[name = tensor("op_5727_cast_fp16")]; + tensor k_cache22_to_fp16_dtype_0 = const()[name = tensor("k_cache22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache22_to_fp16 = cast(dtype = k_cache22_to_fp16_dtype_0, x = k_cache22)[name = tensor("cast_399")]; + tensor var_5728_cast_fp16 = mul(x = k_cache22_to_fp16, y = var_5727_cast_fp16)[name = tensor("op_5728_cast_fp16")]; + tensor var_5733 = const()[name = tensor("op_5733"), val = tensor([1, 1, 16, 64])]; + tensor var_5734_cast_fp16 = reshape(shape = var_5733, x = out_91_cast_fp16)[name = tensor("op_5734_cast_fp16")]; + tensor var_5735_cast_fp16 = mul(x = var_5734_cast_fp16, y = onehot_91_cast_fp16)[name = tensor("op_5735_cast_fp16")]; + tensor new_k_cache_45_cast_fp16 = add(x = var_5728_cast_fp16, y = var_5735_cast_fp16)[name = tensor("new_k_cache_45_cast_fp16")]; + tensor new_k_cache_45_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_45_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache22_to_fp16_dtype_0 = const()[name = tensor("v_cache22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache22_to_fp16 = cast(dtype = v_cache22_to_fp16_dtype_0, x = v_cache22)[name = tensor("cast_397")]; + tensor var_5741_cast_fp16 = mul(x = v_cache22_to_fp16, y = var_5727_cast_fp16)[name = tensor("op_5741_cast_fp16")]; + tensor var_5746 = const()[name = tensor("op_5746"), val = tensor([1, 1, 16, 64])]; + tensor var_5747_cast_fp16 = reshape(shape = var_5746, x = v_45_cast_fp16)[name = tensor("op_5747_cast_fp16")]; + tensor var_5748_cast_fp16 = mul(x = var_5747_cast_fp16, y = onehot_91_cast_fp16)[name = tensor("op_5748_cast_fp16")]; + tensor new_v_cache_45_cast_fp16 = add(x = var_5741_cast_fp16, y = var_5748_cast_fp16)[name = tensor("new_v_cache_45_cast_fp16")]; + tensor new_v_cache_45_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_45_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_22_perm_0 = const()[name = tensor("transpose_22_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_5761 = const()[name = tensor("op_5761"), val = tensor([0, 2, 1, 3])]; + tensor var_5767 = const()[name = tensor("op_5767"), val = tensor([1, 16, 1, 64])]; + tensor q4_45_cast_fp16 = reshape(shape = var_5767, x = out_89_cast_fp16)[name = tensor("q4_45_cast_fp16")]; + tensor var_5772_transpose_x_0 = const()[name = tensor("op_5772_transpose_x_0"), val = tensor(false)]; + tensor var_5772_transpose_y_0 = const()[name = tensor("op_5772_transpose_y_0"), val = tensor(false)]; + tensor transpose_22_cast_fp16 = transpose(perm = transpose_22_perm_0, x = new_k_cache_45_cast_fp16)[name = tensor("transpose_29")]; + tensor var_5772_cast_fp16 = matmul(transpose_x = var_5772_transpose_x_0, transpose_y = var_5772_transpose_y_0, x = q4_45_cast_fp16, y = transpose_22_cast_fp16)[name = tensor("op_5772_cast_fp16")]; + tensor var_5773_to_fp16 = const()[name = tensor("op_5773_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_89_cast_fp16 = mul(x = var_5772_cast_fp16, y = var_5773_to_fp16)[name = tensor("scores_89_cast_fp16")]; + tensor var_5777 = const()[name = tensor("op_5777"), val = tensor([1, 1])]; + tensor var_5778_cast_fp16 = reshape(shape = var_5777, x = position22_to_fp16)[name = tensor("op_5778_cast_fp16")]; + tensor var_5779_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_5778_cast_fp16)[name = tensor("op_5779_cast_fp16")]; + tensor mask_45_to_fp16_dtype_0 = const()[name = tensor("mask_45_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_5786_to_fp16 = const()[name = tensor("op_5786_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5779_cast_fp16_to_fp16 = cast(dtype = mask_45_to_fp16_dtype_0, x = var_5779_cast_fp16)[name = tensor("cast_395")]; + tensor var_5787_cast_fp16 = sub(x = var_5779_cast_fp16_to_fp16, y = var_5786_to_fp16)[name = tensor("op_5787_cast_fp16")]; + tensor var_5792 = const()[name = tensor("op_5792"), val = tensor([1, 1, 1, 512])]; + tensor var_5793_cast_fp16 = reshape(shape = var_5792, x = var_5787_cast_fp16)[name = tensor("op_5793_cast_fp16")]; + tensor var_5794_to_fp16 = const()[name = tensor("op_5794_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_5795_cast_fp16 = mul(x = var_5793_cast_fp16, y = var_5794_to_fp16)[name = tensor("op_5795_cast_fp16")]; + tensor scores_91_cast_fp16 = add(x = scores_89_cast_fp16, y = var_5795_cast_fp16)[name = tensor("scores_91_cast_fp16")]; + tensor var_5798 = const()[name = tensor("op_5798"), val = tensor(-1)]; + tensor weights_45_cast_fp16 = softmax(axis = var_5798, x = scores_91_cast_fp16)[name = tensor("weights_45_cast_fp16")]; + tensor attn_45_transpose_x_0 = const()[name = tensor("attn_45_transpose_x_0"), val = tensor(false)]; + tensor attn_45_transpose_y_0 = const()[name = tensor("attn_45_transpose_y_0"), val = tensor(false)]; + tensor values_45_cast_fp16 = transpose(perm = var_5761, x = new_v_cache_45_cast_fp16)[name = tensor("transpose_28")]; + tensor attn_45_cast_fp16 = matmul(transpose_x = attn_45_transpose_x_0, transpose_y = attn_45_transpose_y_0, x = weights_45_cast_fp16, y = values_45_cast_fp16)[name = tensor("attn_45_cast_fp16")]; + tensor var_5806 = const()[name = tensor("op_5806"), val = tensor([0, 2, 1, 3])]; + tensor var_5811 = const()[name = tensor("op_5811"), val = tensor([1, 1, 1024])]; + tensor var_5807_cast_fp16 = transpose(perm = var_5806, x = attn_45_cast_fp16)[name = tensor("transpose_27")]; + tensor input_271_cast_fp16 = reshape(shape = var_5811, x = var_5807_cast_fp16)[name = tensor("input_271_cast_fp16")]; + tensor attn22_out_proj_weight_to_fp16 = const()[name = tensor("attn22_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(560218816)))]; + tensor linear_90_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn22_out_proj_weight_to_fp16, x = input_271_cast_fp16)[name = tensor("linear_90_cast_fp16")]; + tensor var_5817_to_fp16 = const()[name = tensor("op_5817_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5818_cast_fp16 = add(x = position22_to_fp16, y = var_5817_to_fp16)[name = tensor("op_5818_cast_fp16")]; + tensor var_5818_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5818_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_273_cast_fp16 = add(x = input_267_cast_fp16, y = linear_90_cast_fp16)[name = tensor("input_273_cast_fp16")]; + tensor input_275_axes_0 = const()[name = tensor("input_275_axes_0"), val = tensor([-1])]; + tensor norm22_2_weight_to_fp16 = const()[name = tensor("norm22_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(562316032)))]; + tensor norm22_2_bias_to_fp16 = const()[name = tensor("norm22_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(562318144)))]; + tensor var_5822_to_fp16 = const()[name = tensor("op_5822_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_275_cast_fp16 = layer_norm(axes = input_275_axes_0, beta = norm22_2_bias_to_fp16, epsilon = var_5822_to_fp16, gamma = norm22_2_weight_to_fp16, x = input_273_cast_fp16)[name = tensor("input_275_cast_fp16")]; + tensor linear22_1_weight_to_fp16 = const()[name = tensor("linear22_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(562320256)))]; + tensor linear_91_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear22_1_weight_to_fp16, x = input_275_cast_fp16)[name = tensor("linear_91_cast_fp16")]; + tensor input_277_mode_0 = const()[name = tensor("input_277_mode_0"), val = tensor("EXACT")]; + tensor input_277_cast_fp16 = gelu(mode = input_277_mode_0, x = linear_91_cast_fp16)[name = tensor("input_277_cast_fp16")]; + tensor linear22_2_weight_to_fp16 = const()[name = tensor("linear22_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(570708928)))]; + tensor linear_92_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear22_2_weight_to_fp16, x = input_277_cast_fp16)[name = tensor("linear_92_cast_fp16")]; + tensor input_279_cast_fp16 = add(x = input_273_cast_fp16, y = linear_92_cast_fp16)[name = tensor("input_279_cast_fp16")]; + tensor input_281_axes_0 = const()[name = tensor("input_281_axes_0"), val = tensor([-1])]; + tensor norm23_1_weight_to_fp16 = const()[name = tensor("norm23_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(579097600)))]; + tensor norm23_1_bias_to_fp16 = const()[name = tensor("norm23_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(579099712)))]; + tensor var_5839_to_fp16 = const()[name = tensor("op_5839_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_281_cast_fp16 = layer_norm(axes = input_281_axes_0, beta = norm23_1_bias_to_fp16, epsilon = var_5839_to_fp16, gamma = norm23_1_weight_to_fp16, x = input_279_cast_fp16)[name = tensor("input_281_cast_fp16")]; + tensor attn23_in_proj_weight_to_fp16 = const()[name = tensor("attn23_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(579101824)))]; + tensor linear_93_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn23_in_proj_weight_to_fp16, x = input_281_cast_fp16)[name = tensor("linear_93_cast_fp16")]; + tensor var_5852 = const()[name = tensor("op_5852"), val = tensor([1, 3, 16, 64])]; + tensor qkv_cast_fp16 = reshape(shape = var_5852, x = linear_93_cast_fp16)[name = tensor("qkv_cast_fp16")]; + tensor x_93_begin_0 = const()[name = tensor("x_93_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_93_end_0 = const()[name = tensor("x_93_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_93_end_mask_0 = const()[name = tensor("x_93_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_93_squeeze_mask_0 = const()[name = tensor("x_93_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_93_cast_fp16 = slice_by_index(begin = x_93_begin_0, end = x_93_end_0, end_mask = x_93_end_mask_0, squeeze_mask = x_93_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("x_93_cast_fp16")]; + tensor x_begin_0 = const()[name = tensor("x_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_end_0 = const()[name = tensor("x_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_end_mask_0 = const()[name = tensor("x_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_squeeze_mask_0 = const()[name = tensor("x_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_cast_fp16 = slice_by_index(begin = x_begin_0, end = x_end_0, end_mask = x_end_mask_0, squeeze_mask = x_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("x_cast_fp16")]; + tensor v_begin_0 = const()[name = tensor("v_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_end_0 = const()[name = tensor("v_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_end_mask_0 = const()[name = tensor("v_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_squeeze_mask_0 = const()[name = tensor("v_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_cast_fp16 = slice_by_index(begin = v_begin_0, end = v_end_0, end_mask = v_end_mask_0, squeeze_mask = v_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("v_cast_fp16")]; + tensor var_5881 = const()[name = tensor("op_5881"), val = tensor([1, 1, 1])]; + tensor position23_to_fp16_dtype_0 = const()[name = tensor("position23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position23_to_fp16 = cast(dtype = position23_to_fp16_dtype_0, x = position23)[name = tensor("cast_393")]; + tensor var_5882_cast_fp16 = reshape(shape = var_5881, x = position23_to_fp16)[name = tensor("op_5882_cast_fp16")]; + tensor angles_93_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_5882_cast_fp16)[name = tensor("angles_93_cast_fp16")]; + tensor rotr_93_cast_fp16 = cos(x = angles_93_cast_fp16)[name = tensor("rotr_93_cast_fp16")]; + tensor roti_93_cast_fp16 = sin(x = angles_93_cast_fp16)[name = tensor("roti_93_cast_fp16")]; + tensor var_5890 = const()[name = tensor("op_5890"), val = tensor([1, 16, 32, 2])]; + tensor pairs_93_cast_fp16 = reshape(shape = var_5890, x = x_93_cast_fp16)[name = tensor("pairs_93_cast_fp16")]; + tensor xr_93_begin_0 = const()[name = tensor("xr_93_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_93_end_0 = const()[name = tensor("xr_93_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_93_end_mask_0 = const()[name = tensor("xr_93_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_93_squeeze_mask_0 = const()[name = tensor("xr_93_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_93_cast_fp16 = slice_by_index(begin = xr_93_begin_0, end = xr_93_end_0, end_mask = xr_93_end_mask_0, squeeze_mask = xr_93_squeeze_mask_0, x = pairs_93_cast_fp16)[name = tensor("xr_93_cast_fp16")]; + tensor xi_93_begin_0 = const()[name = tensor("xi_93_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_93_end_0 = const()[name = tensor("xi_93_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_93_end_mask_0 = const()[name = tensor("xi_93_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_93_squeeze_mask_0 = const()[name = tensor("xi_93_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_93_cast_fp16 = slice_by_index(begin = xi_93_begin_0, end = xi_93_end_0, end_mask = xi_93_end_mask_0, squeeze_mask = xi_93_squeeze_mask_0, x = pairs_93_cast_fp16)[name = tensor("xi_93_cast_fp16")]; + tensor var_5898_cast_fp16 = mul(x = xr_93_cast_fp16, y = rotr_93_cast_fp16)[name = tensor("op_5898_cast_fp16")]; + tensor var_5899_cast_fp16 = mul(x = xi_93_cast_fp16, y = roti_93_cast_fp16)[name = tensor("op_5899_cast_fp16")]; + tensor xor__93_cast_fp16 = sub(x = var_5898_cast_fp16, y = var_5899_cast_fp16)[name = tensor("xor__93_cast_fp16")]; + tensor var_5902_cast_fp16 = mul(x = xr_93_cast_fp16, y = roti_93_cast_fp16)[name = tensor("op_5902_cast_fp16")]; + tensor var_5903_cast_fp16 = mul(x = xi_93_cast_fp16, y = rotr_93_cast_fp16)[name = tensor("op_5903_cast_fp16")]; + tensor xoi_93_cast_fp16 = add(x = var_5902_cast_fp16, y = var_5903_cast_fp16)[name = tensor("xoi_93_cast_fp16")]; + tensor out_93_axis_0 = const()[name = tensor("out_93_axis_0"), val = tensor(-1)]; + tensor out_93_cast_fp16 = stack(axis = out_93_axis_0, values = (xor__93_cast_fp16, xoi_93_cast_fp16))[name = tensor("out_93_cast_fp16")]; + tensor var_5926 = const()[name = tensor("op_5926"), val = tensor([1, 16, 32, 2])]; + tensor pairs_cast_fp16 = reshape(shape = var_5926, x = x_cast_fp16)[name = tensor("pairs_cast_fp16")]; + tensor xr_begin_0 = const()[name = tensor("xr_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_end_0 = const()[name = tensor("xr_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_end_mask_0 = const()[name = tensor("xr_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_squeeze_mask_0 = const()[name = tensor("xr_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_cast_fp16 = slice_by_index(begin = xr_begin_0, end = xr_end_0, end_mask = xr_end_mask_0, squeeze_mask = xr_squeeze_mask_0, x = pairs_cast_fp16)[name = tensor("xr_cast_fp16")]; + tensor xi_begin_0 = const()[name = tensor("xi_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_end_0 = const()[name = tensor("xi_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_end_mask_0 = const()[name = tensor("xi_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_squeeze_mask_0 = const()[name = tensor("xi_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_cast_fp16 = slice_by_index(begin = xi_begin_0, end = xi_end_0, end_mask = xi_end_mask_0, squeeze_mask = xi_squeeze_mask_0, x = pairs_cast_fp16)[name = tensor("xi_cast_fp16")]; + tensor var_5934_cast_fp16 = mul(x = xr_cast_fp16, y = rotr_93_cast_fp16)[name = tensor("op_5934_cast_fp16")]; + tensor var_5935_cast_fp16 = mul(x = xi_cast_fp16, y = roti_93_cast_fp16)[name = tensor("op_5935_cast_fp16")]; + tensor xor__cast_fp16 = sub(x = var_5934_cast_fp16, y = var_5935_cast_fp16)[name = tensor("xor__cast_fp16")]; + tensor var_5938_cast_fp16 = mul(x = xr_cast_fp16, y = roti_93_cast_fp16)[name = tensor("op_5938_cast_fp16")]; + tensor var_5939_cast_fp16 = mul(x = xi_cast_fp16, y = rotr_93_cast_fp16)[name = tensor("op_5939_cast_fp16")]; + tensor xoi_cast_fp16 = add(x = var_5938_cast_fp16, y = var_5939_cast_fp16)[name = tensor("xoi_cast_fp16")]; + tensor out_axis_0 = const()[name = tensor("out_axis_0"), val = tensor(-1)]; + tensor out_cast_fp16 = stack(axis = out_axis_0, values = (xor__cast_fp16, xoi_cast_fp16))[name = tensor("out_cast_fp16")]; + tensor _inversed_5951_y_0_to_fp16 = const()[name = tensor("_inversed_5951_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5951_cast_fp16 = mul(x = position23_to_fp16, y = _inversed_5951_y_0_to_fp16)[name = tensor("_inversed_5951_cast_fp16")]; + tensor var_5952_cast_fp16 = floor(x = _inversed_5951_cast_fp16)[name = tensor("op_5952_cast_fp16")]; + tensor var_5953_to_fp16 = const()[name = tensor("op_5953_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5954_cast_fp16 = mul(x = var_5952_cast_fp16, y = var_5953_to_fp16)[name = tensor("op_5954_cast_fp16")]; + tensor write_idx_cast_fp16 = sub(x = position23_to_fp16, y = var_5954_cast_fp16)[name = tensor("write_idx_cast_fp16")]; + tensor var_5957_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_cast_fp16)[name = tensor("op_5957_cast_fp16")]; + tensor var_5967 = const()[name = tensor("op_5967"), val = tensor([1, 512, 1, 1])]; + tensor onehot_93_to_fp16_dtype_0 = const()[name = tensor("onehot_93_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_5957_cast_fp16_to_fp16 = cast(dtype = onehot_93_to_fp16_dtype_0, x = var_5957_cast_fp16)[name = tensor("cast_392")]; + tensor onehot_cast_fp16 = reshape(shape = var_5967, x = var_5957_cast_fp16_to_fp16)[name = tensor("onehot_cast_fp16")]; + tensor var_5969_to_fp16 = const()[name = tensor("op_5969_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5971_cast_fp16 = sub(x = var_5969_to_fp16, y = onehot_cast_fp16)[name = tensor("op_5971_cast_fp16")]; + tensor k_cache23_to_fp16_dtype_0 = const()[name = tensor("k_cache23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache23_to_fp16 = cast(dtype = k_cache23_to_fp16_dtype_0, x = k_cache23)[name = tensor("cast_391")]; + tensor var_5972_cast_fp16 = mul(x = k_cache23_to_fp16, y = var_5971_cast_fp16)[name = tensor("op_5972_cast_fp16")]; + tensor var_5977 = const()[name = tensor("op_5977"), val = tensor([1, 1, 16, 64])]; + tensor var_5978_cast_fp16 = reshape(shape = var_5977, x = out_cast_fp16)[name = tensor("op_5978_cast_fp16")]; + tensor var_5979_cast_fp16 = mul(x = var_5978_cast_fp16, y = onehot_cast_fp16)[name = tensor("op_5979_cast_fp16")]; + tensor new_k_cache_cast_fp16 = add(x = var_5972_cast_fp16, y = var_5979_cast_fp16)[name = tensor("new_k_cache_cast_fp16")]; + tensor new_k_cache_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache23_to_fp16_dtype_0 = const()[name = tensor("v_cache23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache23_to_fp16 = cast(dtype = v_cache23_to_fp16_dtype_0, x = v_cache23)[name = tensor("cast_389")]; + tensor var_5985_cast_fp16 = mul(x = v_cache23_to_fp16, y = var_5971_cast_fp16)[name = tensor("op_5985_cast_fp16")]; + tensor var_5990 = const()[name = tensor("op_5990"), val = tensor([1, 1, 16, 64])]; + tensor var_5991_cast_fp16 = reshape(shape = var_5990, x = v_cast_fp16)[name = tensor("op_5991_cast_fp16")]; + tensor var_5992_cast_fp16 = mul(x = var_5991_cast_fp16, y = onehot_cast_fp16)[name = tensor("op_5992_cast_fp16")]; + tensor new_v_cache_cast_fp16 = add(x = var_5985_cast_fp16, y = var_5992_cast_fp16)[name = tensor("new_v_cache_cast_fp16")]; + tensor new_v_cache_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_23_perm_0 = const()[name = tensor("transpose_23_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_6005 = const()[name = tensor("op_6005"), val = tensor([0, 2, 1, 3])]; + tensor var_6011 = const()[name = tensor("op_6011"), val = tensor([1, 16, 1, 64])]; + tensor q4_cast_fp16 = reshape(shape = var_6011, x = out_93_cast_fp16)[name = tensor("q4_cast_fp16")]; + tensor var_6016_transpose_x_0 = const()[name = tensor("op_6016_transpose_x_0"), val = tensor(false)]; + tensor var_6016_transpose_y_0 = const()[name = tensor("op_6016_transpose_y_0"), val = tensor(false)]; + tensor transpose_23_cast_fp16 = transpose(perm = transpose_23_perm_0, x = new_k_cache_cast_fp16)[name = tensor("transpose_26")]; + tensor var_6016_cast_fp16 = matmul(transpose_x = var_6016_transpose_x_0, transpose_y = var_6016_transpose_y_0, x = q4_cast_fp16, y = transpose_23_cast_fp16)[name = tensor("op_6016_cast_fp16")]; + tensor var_6017_to_fp16 = const()[name = tensor("op_6017_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_93_cast_fp16 = mul(x = var_6016_cast_fp16, y = var_6017_to_fp16)[name = tensor("scores_93_cast_fp16")]; + tensor var_6021 = const()[name = tensor("op_6021"), val = tensor([1, 1])]; + tensor var_6022_cast_fp16 = reshape(shape = var_6021, x = position23_to_fp16)[name = tensor("op_6022_cast_fp16")]; + tensor var_6023_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_6022_cast_fp16)[name = tensor("op_6023_cast_fp16")]; + tensor mask_to_fp16_dtype_0 = const()[name = tensor("mask_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_6030_to_fp16 = const()[name = tensor("op_6030_to_fp16"), val = tensor(0x1p+0)]; + tensor var_6023_cast_fp16_to_fp16 = cast(dtype = mask_to_fp16_dtype_0, x = var_6023_cast_fp16)[name = tensor("cast_387")]; + tensor var_6031_cast_fp16 = sub(x = var_6023_cast_fp16_to_fp16, y = var_6030_to_fp16)[name = tensor("op_6031_cast_fp16")]; + tensor var_6036 = const()[name = tensor("op_6036"), val = tensor([1, 1, 1, 512])]; + tensor var_6037_cast_fp16 = reshape(shape = var_6036, x = var_6031_cast_fp16)[name = tensor("op_6037_cast_fp16")]; + tensor var_6038_to_fp16 = const()[name = tensor("op_6038_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_6039_cast_fp16 = mul(x = var_6037_cast_fp16, y = var_6038_to_fp16)[name = tensor("op_6039_cast_fp16")]; + tensor scores_cast_fp16 = add(x = scores_93_cast_fp16, y = var_6039_cast_fp16)[name = tensor("scores_cast_fp16")]; + tensor var_6042 = const()[name = tensor("op_6042"), val = tensor(-1)]; + tensor weights_cast_fp16 = softmax(axis = var_6042, x = scores_cast_fp16)[name = tensor("weights_cast_fp16")]; + tensor attn_transpose_x_0 = const()[name = tensor("attn_transpose_x_0"), val = tensor(false)]; + tensor attn_transpose_y_0 = const()[name = tensor("attn_transpose_y_0"), val = tensor(false)]; + tensor values_cast_fp16 = transpose(perm = var_6005, x = new_v_cache_cast_fp16)[name = tensor("transpose_25")]; + tensor attn_cast_fp16 = matmul(transpose_x = attn_transpose_x_0, transpose_y = attn_transpose_y_0, x = weights_cast_fp16, y = values_cast_fp16)[name = tensor("attn_cast_fp16")]; + tensor var_6050 = const()[name = tensor("op_6050"), val = tensor([0, 2, 1, 3])]; + tensor var_6055 = const()[name = tensor("op_6055"), val = tensor([1, 1, 1024])]; + tensor var_6051_cast_fp16 = transpose(perm = var_6050, x = attn_cast_fp16)[name = tensor("transpose_24")]; + tensor input_283_cast_fp16 = reshape(shape = var_6055, x = var_6051_cast_fp16)[name = tensor("input_283_cast_fp16")]; + tensor attn23_out_proj_weight_to_fp16 = const()[name = tensor("attn23_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(585393344)))]; + tensor linear_94_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn23_out_proj_weight_to_fp16, x = input_283_cast_fp16)[name = tensor("linear_94_cast_fp16")]; + tensor var_6061_to_fp16 = const()[name = tensor("op_6061_to_fp16"), val = tensor(0x1p+0)]; + tensor var_6062_cast_fp16 = add(x = position23_to_fp16, y = var_6061_to_fp16)[name = tensor("op_6062_cast_fp16")]; + tensor var_6062_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_6062_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_285_cast_fp16 = add(x = input_279_cast_fp16, y = linear_94_cast_fp16)[name = tensor("input_285_cast_fp16")]; + tensor input_287_axes_0 = const()[name = tensor("input_287_axes_0"), val = tensor([-1])]; + tensor norm23_2_weight_to_fp16 = const()[name = tensor("norm23_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(587490560)))]; + tensor norm23_2_bias_to_fp16 = const()[name = tensor("norm23_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(587492672)))]; + tensor var_6066_to_fp16 = const()[name = tensor("op_6066_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_287_cast_fp16 = layer_norm(axes = input_287_axes_0, beta = norm23_2_bias_to_fp16, epsilon = var_6066_to_fp16, gamma = norm23_2_weight_to_fp16, x = input_285_cast_fp16)[name = tensor("input_287_cast_fp16")]; + tensor linear23_1_weight_to_fp16 = const()[name = tensor("linear23_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(587494784)))]; + tensor linear_95_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear23_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_95_cast_fp16")]; + tensor input_289_mode_0 = const()[name = tensor("input_289_mode_0"), val = tensor("EXACT")]; + tensor input_289_cast_fp16 = gelu(mode = input_289_mode_0, x = linear_95_cast_fp16)[name = tensor("input_289_cast_fp16")]; + tensor linear23_2_weight_to_fp16 = const()[name = tensor("linear23_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(595883456)))]; + tensor linear_96_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear23_2_weight_to_fp16, x = input_289_cast_fp16)[name = tensor("linear_96_cast_fp16")]; + tensor input_291_cast_fp16 = add(x = input_285_cast_fp16, y = linear_96_cast_fp16)[name = tensor("input_291_cast_fp16")]; + tensor input_axes_0 = const()[name = tensor("input_axes_0"), val = tensor([-1])]; + tensor out_norm_weight_to_fp16 = const()[name = tensor("out_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(604272128)))]; + tensor out_norm_bias_to_fp16 = const()[name = tensor("out_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(604274240)))]; + tensor var_6083_to_fp16 = const()[name = tensor("op_6083_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_cast_fp16 = layer_norm(axes = input_axes_0, beta = out_norm_bias_to_fp16, epsilon = var_6083_to_fp16, gamma = out_norm_weight_to_fp16, x = input_291_cast_fp16)[name = tensor("input_cast_fp16")]; + tensor input_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("input_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor out_eos_weight_to_fp16 = const()[name = tensor("out_eos_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(604276352)))]; + tensor out_eos_bias_to_fp16 = const()[name = tensor("out_eos_bias_to_fp16"), val = tensor([-0x1.5ap-3])]; + tensor linear_97_cast_fp16 = linear(bias = out_eos_bias_to_fp16, weight = out_eos_weight_to_fp16, x = input_cast_fp16)[name = tensor("linear_97_cast_fp16")]; + tensor linear_97_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("linear_97_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor is_eos = cast(dtype = linear_97_cast_fp16_to_fp32_dtype_0, x = linear_97_cast_fp16)[name = tensor("cast_384")]; + tensor transformer_out = cast(dtype = input_cast_fp16_to_fp32_dtype_0, x = input_cast_fp16)[name = tensor("cast_385")]; + tensor new_position23 = cast(dtype = var_6062_cast_fp16_to_fp32_dtype_0, x = var_6062_cast_fp16)[name = tensor("cast_386")]; + tensor new_v_cache23 = cast(dtype = new_v_cache_cast_fp16_to_fp32_dtype_0, x = new_v_cache_cast_fp16)[name = tensor("cast_388")]; + tensor new_k_cache23 = cast(dtype = new_k_cache_cast_fp16_to_fp32_dtype_0, x = new_k_cache_cast_fp16)[name = tensor("cast_390")]; + tensor new_position22 = cast(dtype = var_5818_cast_fp16_to_fp32_dtype_0, x = var_5818_cast_fp16)[name = tensor("cast_394")]; + tensor new_v_cache22 = cast(dtype = new_v_cache_45_cast_fp16_to_fp32_dtype_0, x = new_v_cache_45_cast_fp16)[name = tensor("cast_396")]; + tensor new_k_cache22 = cast(dtype = new_k_cache_45_cast_fp16_to_fp32_dtype_0, x = new_k_cache_45_cast_fp16)[name = tensor("cast_398")]; + tensor new_position21 = cast(dtype = var_5574_cast_fp16_to_fp32_dtype_0, x = var_5574_cast_fp16)[name = tensor("cast_402")]; + tensor new_v_cache21 = cast(dtype = new_v_cache_43_cast_fp16_to_fp32_dtype_0, x = new_v_cache_43_cast_fp16)[name = tensor("cast_404")]; + tensor new_k_cache21 = cast(dtype = new_k_cache_43_cast_fp16_to_fp32_dtype_0, x = new_k_cache_43_cast_fp16)[name = tensor("cast_406")]; + tensor new_position20 = cast(dtype = var_5330_cast_fp16_to_fp32_dtype_0, x = var_5330_cast_fp16)[name = tensor("cast_410")]; + tensor new_v_cache20 = cast(dtype = new_v_cache_41_cast_fp16_to_fp32_dtype_0, x = new_v_cache_41_cast_fp16)[name = tensor("cast_412")]; + tensor new_k_cache20 = cast(dtype = new_k_cache_41_cast_fp16_to_fp32_dtype_0, x = new_k_cache_41_cast_fp16)[name = tensor("cast_414")]; + tensor new_position19 = cast(dtype = var_5086_cast_fp16_to_fp32_dtype_0, x = var_5086_cast_fp16)[name = tensor("cast_418")]; + tensor new_v_cache19 = cast(dtype = new_v_cache_39_cast_fp16_to_fp32_dtype_0, x = new_v_cache_39_cast_fp16)[name = tensor("cast_420")]; + tensor new_k_cache19 = cast(dtype = new_k_cache_39_cast_fp16_to_fp32_dtype_0, x = new_k_cache_39_cast_fp16)[name = tensor("cast_422")]; + tensor new_position18 = cast(dtype = var_4842_cast_fp16_to_fp32_dtype_0, x = var_4842_cast_fp16)[name = tensor("cast_426")]; + tensor new_v_cache18 = cast(dtype = new_v_cache_37_cast_fp16_to_fp32_dtype_0, x = new_v_cache_37_cast_fp16)[name = tensor("cast_428")]; + tensor new_k_cache18 = cast(dtype = new_k_cache_37_cast_fp16_to_fp32_dtype_0, x = new_k_cache_37_cast_fp16)[name = tensor("cast_430")]; + tensor new_position17 = cast(dtype = var_4598_cast_fp16_to_fp32_dtype_0, x = var_4598_cast_fp16)[name = tensor("cast_434")]; + tensor new_v_cache17 = cast(dtype = new_v_cache_35_cast_fp16_to_fp32_dtype_0, x = new_v_cache_35_cast_fp16)[name = tensor("cast_436")]; + tensor new_k_cache17 = cast(dtype = new_k_cache_35_cast_fp16_to_fp32_dtype_0, x = new_k_cache_35_cast_fp16)[name = tensor("cast_438")]; + tensor new_position16 = cast(dtype = var_4354_cast_fp16_to_fp32_dtype_0, x = var_4354_cast_fp16)[name = tensor("cast_442")]; + tensor new_v_cache16 = cast(dtype = new_v_cache_33_cast_fp16_to_fp32_dtype_0, x = new_v_cache_33_cast_fp16)[name = tensor("cast_444")]; + tensor new_k_cache16 = cast(dtype = new_k_cache_33_cast_fp16_to_fp32_dtype_0, x = new_k_cache_33_cast_fp16)[name = tensor("cast_446")]; + tensor new_position15 = cast(dtype = var_4110_cast_fp16_to_fp32_dtype_0, x = var_4110_cast_fp16)[name = tensor("cast_450")]; + tensor new_v_cache15 = cast(dtype = new_v_cache_31_cast_fp16_to_fp32_dtype_0, x = new_v_cache_31_cast_fp16)[name = tensor("cast_452")]; + tensor new_k_cache15 = cast(dtype = new_k_cache_31_cast_fp16_to_fp32_dtype_0, x = new_k_cache_31_cast_fp16)[name = tensor("cast_454")]; + tensor new_position14 = cast(dtype = var_3866_cast_fp16_to_fp32_dtype_0, x = var_3866_cast_fp16)[name = tensor("cast_458")]; + tensor new_v_cache14 = cast(dtype = new_v_cache_29_cast_fp16_to_fp32_dtype_0, x = new_v_cache_29_cast_fp16)[name = tensor("cast_460")]; + tensor new_k_cache14 = cast(dtype = new_k_cache_29_cast_fp16_to_fp32_dtype_0, x = new_k_cache_29_cast_fp16)[name = tensor("cast_462")]; + tensor new_position13 = cast(dtype = var_3622_cast_fp16_to_fp32_dtype_0, x = var_3622_cast_fp16)[name = tensor("cast_466")]; + tensor new_v_cache13 = cast(dtype = new_v_cache_27_cast_fp16_to_fp32_dtype_0, x = new_v_cache_27_cast_fp16)[name = tensor("cast_468")]; + tensor new_k_cache13 = cast(dtype = new_k_cache_27_cast_fp16_to_fp32_dtype_0, x = new_k_cache_27_cast_fp16)[name = tensor("cast_470")]; + tensor new_position12 = cast(dtype = var_3378_cast_fp16_to_fp32_dtype_0, x = var_3378_cast_fp16)[name = tensor("cast_474")]; + tensor new_v_cache12 = cast(dtype = new_v_cache_25_cast_fp16_to_fp32_dtype_0, x = new_v_cache_25_cast_fp16)[name = tensor("cast_476")]; + tensor new_k_cache12 = cast(dtype = new_k_cache_25_cast_fp16_to_fp32_dtype_0, x = new_k_cache_25_cast_fp16)[name = tensor("cast_478")]; + tensor new_position11 = cast(dtype = var_3134_cast_fp16_to_fp32_dtype_0, x = var_3134_cast_fp16)[name = tensor("cast_482")]; + tensor new_v_cache11 = cast(dtype = new_v_cache_23_cast_fp16_to_fp32_dtype_0, x = new_v_cache_23_cast_fp16)[name = tensor("cast_484")]; + tensor new_k_cache11 = cast(dtype = new_k_cache_23_cast_fp16_to_fp32_dtype_0, x = new_k_cache_23_cast_fp16)[name = tensor("cast_486")]; + tensor new_position10 = cast(dtype = var_2890_cast_fp16_to_fp32_dtype_0, x = var_2890_cast_fp16)[name = tensor("cast_490")]; + tensor new_v_cache10 = cast(dtype = new_v_cache_21_cast_fp16_to_fp32_dtype_0, x = new_v_cache_21_cast_fp16)[name = tensor("cast_492")]; + tensor new_k_cache10 = cast(dtype = new_k_cache_21_cast_fp16_to_fp32_dtype_0, x = new_k_cache_21_cast_fp16)[name = tensor("cast_494")]; + tensor new_position9 = cast(dtype = var_2646_cast_fp16_to_fp32_dtype_0, x = var_2646_cast_fp16)[name = tensor("cast_498")]; + tensor new_v_cache9 = cast(dtype = new_v_cache_19_cast_fp16_to_fp32_dtype_0, x = new_v_cache_19_cast_fp16)[name = tensor("cast_500")]; + tensor new_k_cache9 = cast(dtype = new_k_cache_19_cast_fp16_to_fp32_dtype_0, x = new_k_cache_19_cast_fp16)[name = tensor("cast_502")]; + tensor new_position8 = cast(dtype = var_2402_cast_fp16_to_fp32_dtype_0, x = var_2402_cast_fp16)[name = tensor("cast_506")]; + tensor new_v_cache8 = cast(dtype = new_v_cache_17_cast_fp16_to_fp32_dtype_0, x = new_v_cache_17_cast_fp16)[name = tensor("cast_508")]; + tensor new_k_cache8 = cast(dtype = new_k_cache_17_cast_fp16_to_fp32_dtype_0, x = new_k_cache_17_cast_fp16)[name = tensor("cast_510")]; + tensor new_position7 = cast(dtype = var_2158_cast_fp16_to_fp32_dtype_0, x = var_2158_cast_fp16)[name = tensor("cast_514")]; + tensor new_v_cache7 = cast(dtype = new_v_cache_15_cast_fp16_to_fp32_dtype_0, x = new_v_cache_15_cast_fp16)[name = tensor("cast_516")]; + tensor new_k_cache7 = cast(dtype = new_k_cache_15_cast_fp16_to_fp32_dtype_0, x = new_k_cache_15_cast_fp16)[name = tensor("cast_518")]; + tensor new_position6 = cast(dtype = var_1914_cast_fp16_to_fp32_dtype_0, x = var_1914_cast_fp16)[name = tensor("cast_522")]; + tensor new_v_cache6 = cast(dtype = new_v_cache_13_cast_fp16_to_fp32_dtype_0, x = new_v_cache_13_cast_fp16)[name = tensor("cast_524")]; + tensor new_k_cache6 = cast(dtype = new_k_cache_13_cast_fp16_to_fp32_dtype_0, x = new_k_cache_13_cast_fp16)[name = tensor("cast_526")]; + tensor new_position5 = cast(dtype = var_1670_cast_fp16_to_fp32_dtype_0, x = var_1670_cast_fp16)[name = tensor("cast_530")]; + tensor new_v_cache5 = cast(dtype = new_v_cache_11_cast_fp16_to_fp32_dtype_0, x = new_v_cache_11_cast_fp16)[name = tensor("cast_532")]; + tensor new_k_cache5 = cast(dtype = new_k_cache_11_cast_fp16_to_fp32_dtype_0, x = new_k_cache_11_cast_fp16)[name = tensor("cast_534")]; + tensor new_position4 = cast(dtype = var_1426_cast_fp16_to_fp32_dtype_0, x = var_1426_cast_fp16)[name = tensor("cast_538")]; + tensor new_v_cache4 = cast(dtype = new_v_cache_9_cast_fp16_to_fp32_dtype_0, x = new_v_cache_9_cast_fp16)[name = tensor("cast_540")]; + tensor new_k_cache4 = cast(dtype = new_k_cache_9_cast_fp16_to_fp32_dtype_0, x = new_k_cache_9_cast_fp16)[name = tensor("cast_542")]; + tensor new_position3 = cast(dtype = var_1182_cast_fp16_to_fp32_dtype_0, x = var_1182_cast_fp16)[name = tensor("cast_546")]; + tensor new_v_cache3 = cast(dtype = new_v_cache_7_cast_fp16_to_fp32_dtype_0, x = new_v_cache_7_cast_fp16)[name = tensor("cast_548")]; + tensor new_k_cache3 = cast(dtype = new_k_cache_7_cast_fp16_to_fp32_dtype_0, x = new_k_cache_7_cast_fp16)[name = tensor("cast_550")]; + tensor new_position2 = cast(dtype = var_938_cast_fp16_to_fp32_dtype_0, x = var_938_cast_fp16)[name = tensor("cast_554")]; + tensor new_v_cache2 = cast(dtype = new_v_cache_5_cast_fp16_to_fp32_dtype_0, x = new_v_cache_5_cast_fp16)[name = tensor("cast_556")]; + tensor new_k_cache2 = cast(dtype = new_k_cache_5_cast_fp16_to_fp32_dtype_0, x = new_k_cache_5_cast_fp16)[name = tensor("cast_558")]; + tensor new_position1 = cast(dtype = var_694_cast_fp16_to_fp32_dtype_0, x = var_694_cast_fp16)[name = tensor("cast_562")]; + tensor new_v_cache1 = cast(dtype = new_v_cache_3_cast_fp16_to_fp32_dtype_0, x = new_v_cache_3_cast_fp16)[name = tensor("cast_564")]; + tensor new_k_cache1 = cast(dtype = new_k_cache_3_cast_fp16_to_fp32_dtype_0, x = new_k_cache_3_cast_fp16)[name = tensor("cast_566")]; + tensor new_position0 = cast(dtype = var_450_cast_fp16_to_fp32_dtype_0, x = var_450_cast_fp16)[name = tensor("cast_570")]; + tensor new_v_cache0 = cast(dtype = new_v_cache_1_cast_fp16_to_fp32_dtype_0, x = new_v_cache_1_cast_fp16)[name = tensor("cast_572")]; + tensor new_k_cache0 = cast(dtype = new_k_cache_1_cast_fp16_to_fp32_dtype_0, x = new_k_cache_1_cast_fp16)[name = tensor("cast_574")]; + } -> (transformer_out, is_eos, new_k_cache0, new_v_cache0, new_position0, new_k_cache1, new_v_cache1, new_position1, new_k_cache2, new_v_cache2, new_position2, new_k_cache3, new_v_cache3, new_position3, new_k_cache4, new_v_cache4, new_position4, new_k_cache5, new_v_cache5, new_position5, new_k_cache6, new_v_cache6, new_position6, new_k_cache7, new_v_cache7, new_position7, new_k_cache8, new_v_cache8, new_position8, new_k_cache9, new_v_cache9, new_position9, new_k_cache10, new_v_cache10, new_position10, new_k_cache11, new_v_cache11, new_position11, new_k_cache12, new_v_cache12, new_position12, new_k_cache13, new_v_cache13, new_position13, new_k_cache14, new_v_cache14, new_position14, new_k_cache15, new_v_cache15, new_position15, new_k_cache16, new_v_cache16, new_position16, new_k_cache17, new_v_cache17, new_position17, new_k_cache18, new_v_cache18, new_position18, new_k_cache19, new_v_cache19, new_position19, new_k_cache20, new_v_cache20, new_position20, new_k_cache21, new_v_cache21, new_position21, new_k_cache22, new_v_cache22, new_position22, new_k_cache23, new_v_cache23, new_position23); +} \ No newline at end of file diff --git a/v2.1/portuguese_24l/flowlm_step_ane.mlmodelc/weights/weight.bin b/v2.1/portuguese_24l/flowlm_step_ane.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..ae9e25656238b56b2a47625f68bf646451e6677e --- /dev/null +++ b/v2.1/portuguese_24l/flowlm_step_ane.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f6d244e9ded32fdfe5e22316d798b56ece4e93b0b74b3d40e14c8a2b659227e2 +size 604278464 diff --git a/v2.1/portuguese_24l/flowlm_step_ane.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/portuguese_24l/flowlm_step_ane.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..1ae33479b3fb46e1ad324a8e953a295a74c2bc59 --- /dev/null +++ b/v2.1/portuguese_24l/flowlm_step_ane.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c3241f3b9cb3303e8a3fc17810661d044bc35bcda29a497ddfe29442c1062a55 +size 545925 diff --git a/v2.1/portuguese_24l/flowlm_step_ane.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/portuguese_24l/flowlm_step_ane.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..ae9e25656238b56b2a47625f68bf646451e6677e --- /dev/null +++ b/v2.1/portuguese_24l/flowlm_step_ane.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f6d244e9ded32fdfe5e22316d798b56ece4e93b0b74b3d40e14c8a2b659227e2 +size 604278464 diff --git a/v2.1/portuguese_24l/flowlm_step_ane.mlpackage/Manifest.json b/v2.1/portuguese_24l/flowlm_step_ane.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..d8479fd77d763e96951fb770ad15e292cbd32d90 --- /dev/null +++ b/v2.1/portuguese_24l/flowlm_step_ane.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "28A89CCA-9AA8-4484-84CF-354AB04F5BA2": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "A9D63875-A9E6-48A4-AA72-A64BE035E2AD": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "28A89CCA-9AA8-4484-84CF-354AB04F5BA2" +} diff --git a/v2.1/spanish/cond_prefill_ane.mlmodelc/analytics/coremldata.bin b/v2.1/spanish/cond_prefill_ane.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..c49ad2794b681cbf3277312569d5b0bf92707e91 --- /dev/null +++ b/v2.1/spanish/cond_prefill_ane.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5f90794da9a6d5752523ffa0073239a144634cfddff3cb750e0f9769fc739cdd +size 243 diff --git a/v2.1/spanish/cond_prefill_ane.mlmodelc/coremldata.bin b/v2.1/spanish/cond_prefill_ane.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..4b2358f26261a81bf828929f6acd10517b5f807f --- /dev/null +++ b/v2.1/spanish/cond_prefill_ane.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:36b8209e6d9cc441b4892358ead92750a111bdd6ca546ee23fd04d7a923e94d1 +size 1388 diff --git a/v2.1/spanish/cond_prefill_ane.mlmodelc/model.mil b/v2.1/spanish/cond_prefill_ane.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..2aef44ccd7c895876ef314e46a742909df52fe28 --- /dev/null +++ b/v2.1/spanish/cond_prefill_ane.mlmodelc/model.mil @@ -0,0 +1,804 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.9.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor conditioning, tensor k_cache0, tensor k_cache1, tensor k_cache2, tensor k_cache3, tensor k_cache4, tensor k_cache5, tensor position0, tensor position1, tensor position2, tensor position3, tensor position4, tensor position5, tensor v_cache0, tensor v_cache1, tensor v_cache2, tensor v_cache3, tensor v_cache4, tensor v_cache5, tensor valid_len) { + tensor token_indices_to_fp16 = const()[name = tensor("token_indices_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor position0_to_fp16_dtype_0 = const()[name = tensor("position0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position0_to_fp16 = cast(dtype = position0_to_fp16_dtype_0, x = position0)[name = tensor("cast_128")]; + tensor ts_cast_fp16 = add(x = token_indices_to_fp16, y = position0_to_fp16)[name = tensor("ts_cast_fp16")]; + tensor var_69 = const()[name = tensor("op_69"), val = tensor([256, 1, 1])]; + tensor var_70_cast_fp16 = reshape(shape = var_69, x = ts_cast_fp16)[name = tensor("op_70_cast_fp16")]; + tensor var_65_to_fp16 = const()[name = tensor("op_65_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(640)))]; + tensor angles_cast_fp16 = mul(x = var_65_to_fp16, y = var_70_cast_fp16)[name = tensor("angles_cast_fp16")]; + tensor rotr_cast_fp16 = cos(x = angles_cast_fp16)[name = tensor("rotr_cast_fp16")]; + tensor roti_cast_fp16 = sin(x = angles_cast_fp16)[name = tensor("roti_cast_fp16")]; + tensor _inversed_75_y_0_to_fp16 = const()[name = tensor("_inversed_75_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_75_cast_fp16 = mul(x = ts_cast_fp16, y = _inversed_75_y_0_to_fp16)[name = tensor("_inversed_75_cast_fp16")]; + tensor var_76_cast_fp16 = floor(x = _inversed_75_cast_fp16)[name = tensor("op_76_cast_fp16")]; + tensor var_77_to_fp16 = const()[name = tensor("op_77_to_fp16"), val = tensor(0x1p+9)]; + tensor var_78_cast_fp16 = mul(x = var_76_cast_fp16, y = var_77_to_fp16)[name = tensor("op_78_cast_fp16")]; + tensor write_idx_1_cast_fp16 = sub(x = ts_cast_fp16, y = var_78_cast_fp16)[name = tensor("write_idx_1_cast_fp16")]; + tensor valid_len_to_fp16_dtype_0 = const()[name = tensor("valid_len_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor valid_len_to_fp16 = cast(dtype = valid_len_to_fp16_dtype_0, x = valid_len)[name = tensor("cast_127")]; + tensor var_84_cast_fp16 = less(x = token_indices_to_fp16, y = valid_len_to_fp16)[name = tensor("op_84_cast_fp16")]; + tensor token_valid_to_fp16_dtype_0 = const()[name = tensor("token_valid_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_84_cast_fp16_to_fp16 = cast(dtype = token_valid_to_fp16_dtype_0, x = var_84_cast_fp16)[name = tensor("cast_126")]; + tensor var_90_cast_fp16 = mul(x = write_idx_1_cast_fp16, y = var_84_cast_fp16_to_fp16)[name = tensor("op_90_cast_fp16")]; + tensor var_91_to_fp16 = const()[name = tensor("op_91_to_fp16"), val = tensor(0x1p+0)]; + tensor var_93_cast_fp16 = sub(x = var_91_to_fp16, y = var_84_cast_fp16_to_fp16)[name = tensor("op_93_cast_fp16")]; + tensor var_94_to_fp16 = const()[name = tensor("op_94_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_95_cast_fp16 = mul(x = var_93_cast_fp16, y = var_94_to_fp16)[name = tensor("op_95_cast_fp16")]; + tensor write_idx_cast_fp16 = add(x = var_90_cast_fp16, y = var_95_cast_fp16)[name = tensor("write_idx_cast_fp16")]; + tensor var_100 = const()[name = tensor("op_100"), val = tensor([256, 1])]; + tensor var_101_cast_fp16 = reshape(shape = var_100, x = write_idx_cast_fp16)[name = tensor("op_101_cast_fp16")]; + tensor var_105_to_fp16 = const()[name = tensor("op_105_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(768)))]; + tensor var_106_cast_fp16 = equal(x = var_101_cast_fp16, y = var_105_to_fp16)[name = tensor("op_106_cast_fp16")]; + tensor var_116_axes_0 = const()[name = tensor("op_116_axes_0"), val = tensor([0])]; + tensor var_116_keep_dims_0 = const()[name = tensor("op_116_keep_dims_0"), val = tensor(false)]; + tensor assign_to_fp16_dtype_0 = const()[name = tensor("assign_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_106_cast_fp16_to_fp16 = cast(dtype = assign_to_fp16_dtype_0, x = var_106_cast_fp16)[name = tensor("cast_125")]; + tensor var_116_cast_fp16 = reduce_sum(axes = var_116_axes_0, keep_dims = var_116_keep_dims_0, x = var_106_cast_fp16_to_fp16)[name = tensor("op_116_cast_fp16")]; + tensor const_0_to_fp16 = const()[name = tensor("const_0_to_fp16"), val = tensor(-inf)]; + tensor var_118_to_fp16 = const()[name = tensor("op_118_to_fp16"), val = tensor(0x1p+0)]; + tensor clip_0_cast_fp16 = clip(alpha = const_0_to_fp16, beta = var_118_to_fp16, x = var_116_cast_fp16)[name = tensor("clip_0_cast_fp16")]; + tensor var_122 = const()[name = tensor("op_122"), val = tensor([1, 1])]; + tensor var_123_cast_fp16 = reshape(shape = var_122, x = position0_to_fp16)[name = tensor("op_123_cast_fp16")]; + tensor var_127_to_fp16 = const()[name = tensor("op_127_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1856)))]; + tensor q_positions_cast_fp16 = add(x = var_123_cast_fp16, y = var_127_to_fp16)[name = tensor("q_positions_cast_fp16")]; + tensor var_135_cast_fp16 = add(x = position0_to_fp16, y = valid_len_to_fp16)[name = tensor("op_135_cast_fp16")]; + tensor var_138 = const()[name = tensor("op_138"), val = tensor([1, 1])]; + tensor valid_end_cast_fp16 = reshape(shape = var_138, x = var_135_cast_fp16)[name = tensor("valid_end_cast_fp16")]; + tensor var_140_cast_fp16 = less(x = var_105_to_fp16, y = valid_end_cast_fp16)[name = tensor("op_140_cast_fp16")]; + tensor var_141_cast_fp16 = less_equal(x = var_105_to_fp16, y = q_positions_cast_fp16)[name = tensor("op_141_cast_fp16")]; + tensor var_142 = logical_and(x = var_140_cast_fp16, y = var_141_cast_fp16)[name = tensor("op_142")]; + tensor mask_to_fp16_dtype_0 = const()[name = tensor("mask_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_149_to_fp16 = const()[name = tensor("op_149_to_fp16"), val = tensor(0x1p+0)]; + tensor var_142_to_fp16 = cast(dtype = mask_to_fp16_dtype_0, x = var_142)[name = tensor("cast_124")]; + tensor var_150_cast_fp16 = sub(x = var_142_to_fp16, y = var_149_to_fp16)[name = tensor("op_150_cast_fp16")]; + tensor var_151_to_fp16 = const()[name = tensor("op_151_to_fp16"), val = tensor(0x1.388p+13)]; + tensor mask_bias_cast_fp16 = mul(x = var_150_cast_fp16, y = var_151_to_fp16)[name = tensor("mask_bias_cast_fp16")]; + tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; + tensor conditioning_to_fp16_dtype_0 = const()[name = tensor("conditioning_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor norm0_1_weight_to_fp16 = const()[name = tensor("norm0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2432)))]; + tensor norm0_1_bias_to_fp16 = const()[name = tensor("norm0_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4544)))]; + tensor var_154_to_fp16 = const()[name = tensor("op_154_to_fp16"), val = tensor(0x1.5p-17)]; + tensor conditioning_to_fp16 = cast(dtype = conditioning_to_fp16_dtype_0, x = conditioning)[name = tensor("cast_123")]; + tensor input_3_cast_fp16 = layer_norm(axes = input_3_axes_0, beta = norm0_1_bias_to_fp16, epsilon = var_154_to_fp16, gamma = norm0_1_weight_to_fp16, x = conditioning_to_fp16)[name = tensor("input_3_cast_fp16")]; + tensor attn0_in_proj_weight_to_fp16 = const()[name = tensor("attn0_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6656)))]; + tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6298176)))]; + tensor linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn0_in_proj_weight_to_fp16, x = input_3_cast_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor var_168 = const()[name = tensor("op_168"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_1_cast_fp16 = reshape(shape = var_168, x = linear_0_cast_fp16)[name = tensor("qkv_1_cast_fp16")]; + tensor x_1_begin_0 = const()[name = tensor("x_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_1_end_0 = const()[name = tensor("x_1_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_1_end_mask_0 = const()[name = tensor("x_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_1_squeeze_mask_0 = const()[name = tensor("x_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_1_cast_fp16 = slice_by_index(begin = x_1_begin_0, end = x_1_end_0, end_mask = x_1_end_mask_0, squeeze_mask = x_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("x_1_cast_fp16")]; + tensor x_3_begin_0 = const()[name = tensor("x_3_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_3_end_0 = const()[name = tensor("x_3_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_3_end_mask_0 = const()[name = tensor("x_3_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_3_squeeze_mask_0 = const()[name = tensor("x_3_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_3_cast_fp16 = slice_by_index(begin = x_3_begin_0, end = x_3_end_0, end_mask = x_3_end_mask_0, squeeze_mask = x_3_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("x_3_cast_fp16")]; + tensor v_1_begin_0 = const()[name = tensor("v_1_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_1_end_0 = const()[name = tensor("v_1_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_1_end_mask_0 = const()[name = tensor("v_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_1_squeeze_mask_0 = const()[name = tensor("v_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_1_cast_fp16 = slice_by_index(begin = v_1_begin_0, end = v_1_end_0, end_mask = v_1_end_mask_0, squeeze_mask = v_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("v_1_cast_fp16")]; + tensor var_213 = const()[name = tensor("op_213"), val = tensor([256, 16, 32, 2])]; + tensor pairs_1_cast_fp16 = reshape(shape = var_213, x = x_1_cast_fp16)[name = tensor("pairs_1_cast_fp16")]; + tensor xr_1_begin_0 = const()[name = tensor("xr_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_1_end_0 = const()[name = tensor("xr_1_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_1_end_mask_0 = const()[name = tensor("xr_1_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_1_squeeze_mask_0 = const()[name = tensor("xr_1_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_1_cast_fp16 = slice_by_index(begin = xr_1_begin_0, end = xr_1_end_0, end_mask = xr_1_end_mask_0, squeeze_mask = xr_1_squeeze_mask_0, x = pairs_1_cast_fp16)[name = tensor("xr_1_cast_fp16")]; + tensor xi_1_begin_0 = const()[name = tensor("xi_1_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_1_end_0 = const()[name = tensor("xi_1_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_1_end_mask_0 = const()[name = tensor("xi_1_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_1_squeeze_mask_0 = const()[name = tensor("xi_1_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_1_cast_fp16 = slice_by_index(begin = xi_1_begin_0, end = xi_1_end_0, end_mask = xi_1_end_mask_0, squeeze_mask = xi_1_squeeze_mask_0, x = pairs_1_cast_fp16)[name = tensor("xi_1_cast_fp16")]; + tensor var_221_cast_fp16 = mul(x = xr_1_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_221_cast_fp16")]; + tensor var_222_cast_fp16 = mul(x = xi_1_cast_fp16, y = roti_cast_fp16)[name = tensor("op_222_cast_fp16")]; + tensor xor__1_cast_fp16 = sub(x = var_221_cast_fp16, y = var_222_cast_fp16)[name = tensor("xor__1_cast_fp16")]; + tensor var_225_cast_fp16 = mul(x = xr_1_cast_fp16, y = roti_cast_fp16)[name = tensor("op_225_cast_fp16")]; + tensor var_226_cast_fp16 = mul(x = xi_1_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_226_cast_fp16")]; + tensor xoi_1_cast_fp16 = add(x = var_225_cast_fp16, y = var_226_cast_fp16)[name = tensor("xoi_1_cast_fp16")]; + tensor out_1_axis_0 = const()[name = tensor("out_1_axis_0"), val = tensor(-1)]; + tensor out_1_cast_fp16 = stack(axis = out_1_axis_0, values = (xor__1_cast_fp16, xoi_1_cast_fp16))[name = tensor("out_1_cast_fp16")]; + tensor var_236 = const()[name = tensor("op_236"), val = tensor([1, 256, 16, 64])]; + tensor q_1_cast_fp16 = reshape(shape = var_236, x = out_1_cast_fp16)[name = tensor("q_1_cast_fp16")]; + tensor var_242 = const()[name = tensor("op_242"), val = tensor([256, 16, 32, 2])]; + tensor pairs_3_cast_fp16 = reshape(shape = var_242, x = x_3_cast_fp16)[name = tensor("pairs_3_cast_fp16")]; + tensor xr_3_begin_0 = const()[name = tensor("xr_3_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_3_end_0 = const()[name = tensor("xr_3_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_3_end_mask_0 = const()[name = tensor("xr_3_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_3_squeeze_mask_0 = const()[name = tensor("xr_3_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_3_cast_fp16 = slice_by_index(begin = xr_3_begin_0, end = xr_3_end_0, end_mask = xr_3_end_mask_0, squeeze_mask = xr_3_squeeze_mask_0, x = pairs_3_cast_fp16)[name = tensor("xr_3_cast_fp16")]; + tensor xi_3_begin_0 = const()[name = tensor("xi_3_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_3_end_0 = const()[name = tensor("xi_3_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_3_end_mask_0 = const()[name = tensor("xi_3_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_3_squeeze_mask_0 = const()[name = tensor("xi_3_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_3_cast_fp16 = slice_by_index(begin = xi_3_begin_0, end = xi_3_end_0, end_mask = xi_3_end_mask_0, squeeze_mask = xi_3_squeeze_mask_0, x = pairs_3_cast_fp16)[name = tensor("xi_3_cast_fp16")]; + tensor var_250_cast_fp16 = mul(x = xr_3_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_250_cast_fp16")]; + tensor var_251_cast_fp16 = mul(x = xi_3_cast_fp16, y = roti_cast_fp16)[name = tensor("op_251_cast_fp16")]; + tensor xor__3_cast_fp16 = sub(x = var_250_cast_fp16, y = var_251_cast_fp16)[name = tensor("xor__3_cast_fp16")]; + tensor var_254_cast_fp16 = mul(x = xr_3_cast_fp16, y = roti_cast_fp16)[name = tensor("op_254_cast_fp16")]; + tensor var_255_cast_fp16 = mul(x = xi_3_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_255_cast_fp16")]; + tensor xoi_3_cast_fp16 = add(x = var_254_cast_fp16, y = var_255_cast_fp16)[name = tensor("xoi_3_cast_fp16")]; + tensor out_3_axis_0 = const()[name = tensor("out_3_axis_0"), val = tensor(-1)]; + tensor out_3_cast_fp16 = stack(axis = out_3_axis_0, values = (xor__3_cast_fp16, xoi_3_cast_fp16))[name = tensor("out_3_cast_fp16")]; + tensor var_272 = const()[name = tensor("op_272"), val = tensor([256, 1024])]; + tensor var_273_cast_fp16 = reshape(shape = var_272, x = out_3_cast_fp16)[name = tensor("op_273_cast_fp16")]; + tensor written_k_1_transpose_x_1 = const()[name = tensor("written_k_1_transpose_x_1"), val = tensor(true)]; + tensor written_k_1_transpose_y_1 = const()[name = tensor("written_k_1_transpose_y_1"), val = tensor(false)]; + tensor written_k_1_cast_fp16 = matmul(transpose_x = written_k_1_transpose_x_1, transpose_y = written_k_1_transpose_y_1, x = var_106_cast_fp16_to_fp16, y = var_273_cast_fp16)[name = tensor("written_k_1_cast_fp16")]; + tensor var_280 = const()[name = tensor("op_280"), val = tensor([256, 1024])]; + tensor var_281_cast_fp16 = reshape(shape = var_280, x = v_1_cast_fp16)[name = tensor("op_281_cast_fp16")]; + tensor written_v_1_transpose_x_1 = const()[name = tensor("written_v_1_transpose_x_1"), val = tensor(true)]; + tensor written_v_1_transpose_y_1 = const()[name = tensor("written_v_1_transpose_y_1"), val = tensor(false)]; + tensor written_v_1_cast_fp16 = matmul(transpose_x = written_v_1_transpose_x_1, transpose_y = written_v_1_transpose_y_1, x = var_106_cast_fp16_to_fp16, y = var_281_cast_fp16)[name = tensor("written_v_1_cast_fp16")]; + tensor var_283_to_fp16 = const()[name = tensor("op_283_to_fp16"), val = tensor(0x1p+0)]; + tensor var_285_cast_fp16 = sub(x = var_283_to_fp16, y = clip_0_cast_fp16)[name = tensor("op_285_cast_fp16")]; + tensor var_290 = const()[name = tensor("op_290"), val = tensor([1, 512, 1, 1])]; + tensor keep_1_cast_fp16 = reshape(shape = var_290, x = var_285_cast_fp16)[name = tensor("keep_1_cast_fp16")]; + tensor k_cache0_to_fp16_dtype_0 = const()[name = tensor("k_cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache0_to_fp16 = cast(dtype = k_cache0_to_fp16_dtype_0, x = k_cache0)[name = tensor("cast_122")]; + tensor var_292_cast_fp16 = mul(x = k_cache0_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_292_cast_fp16")]; + tensor var_297 = const()[name = tensor("op_297"), val = tensor([1, 512, 16, 64])]; + tensor var_298_cast_fp16 = reshape(shape = var_297, x = written_k_1_cast_fp16)[name = tensor("op_298_cast_fp16")]; + tensor new_k_cache_1_cast_fp16 = add(x = var_292_cast_fp16, y = var_298_cast_fp16)[name = tensor("new_k_cache_1_cast_fp16")]; + tensor new_k_cache_1_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_1_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache0_to_fp16_dtype_0 = const()[name = tensor("v_cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache0_to_fp16 = cast(dtype = v_cache0_to_fp16_dtype_0, x = v_cache0)[name = tensor("cast_120")]; + tensor var_301_cast_fp16 = mul(x = v_cache0_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_301_cast_fp16")]; + tensor var_306 = const()[name = tensor("op_306"), val = tensor([1, 512, 16, 64])]; + tensor var_307_cast_fp16 = reshape(shape = var_306, x = written_v_1_cast_fp16)[name = tensor("op_307_cast_fp16")]; + tensor new_v_cache_1_cast_fp16 = add(x = var_301_cast_fp16, y = var_307_cast_fp16)[name = tensor("new_v_cache_1_cast_fp16")]; + tensor new_v_cache_1_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_1_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_320 = const()[name = tensor("op_320"), val = tensor([0, 2, 1, 3])]; + tensor var_331_transpose_x_0 = const()[name = tensor("op_331_transpose_x_0"), val = tensor(false)]; + tensor var_331_transpose_y_0 = const()[name = tensor("op_331_transpose_y_0"), val = tensor(false)]; + tensor transpose_15_perm_0 = const()[name = tensor("transpose_15_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_16_perm_0 = const()[name = tensor("transpose_16_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_16 = transpose(perm = transpose_16_perm_0, x = new_k_cache_1_cast_fp16)[name = tensor("transpose_42")]; + tensor transpose_15 = transpose(perm = transpose_15_perm_0, x = q_1_cast_fp16)[name = tensor("transpose_43")]; + tensor var_331_cast_fp16 = matmul(transpose_x = var_331_transpose_x_0, transpose_y = var_331_transpose_y_0, x = transpose_15, y = transpose_16)[name = tensor("op_331_cast_fp16")]; + tensor var_332_to_fp16 = const()[name = tensor("op_332_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_1_cast_fp16 = mul(x = var_331_cast_fp16, y = var_332_to_fp16)[name = tensor("scores_1_cast_fp16")]; + tensor var_338 = const()[name = tensor("op_338"), val = tensor([1, 1, 256, 512])]; + tensor var_339_cast_fp16 = reshape(shape = var_338, x = mask_bias_cast_fp16)[name = tensor("op_339_cast_fp16")]; + tensor scores_3_cast_fp16 = add(x = scores_1_cast_fp16, y = var_339_cast_fp16)[name = tensor("scores_3_cast_fp16")]; + tensor var_342 = const()[name = tensor("op_342"), val = tensor(-1)]; + tensor weights_1_cast_fp16 = softmax(axis = var_342, x = scores_3_cast_fp16)[name = tensor("weights_1_cast_fp16")]; + tensor attn_1_transpose_x_0 = const()[name = tensor("attn_1_transpose_x_0"), val = tensor(false)]; + tensor attn_1_transpose_y_0 = const()[name = tensor("attn_1_transpose_y_0"), val = tensor(false)]; + tensor values_1_cast_fp16 = transpose(perm = var_320, x = new_v_cache_1_cast_fp16)[name = tensor("transpose_44")]; + tensor attn_1_cast_fp16 = matmul(transpose_x = attn_1_transpose_x_0, transpose_y = attn_1_transpose_y_0, x = weights_1_cast_fp16, y = values_1_cast_fp16)[name = tensor("attn_1_cast_fp16")]; + tensor var_350 = const()[name = tensor("op_350"), val = tensor([0, 2, 1, 3])]; + tensor var_355 = const()[name = tensor("op_355"), val = tensor([1, 256, 1024])]; + tensor var_351_cast_fp16 = transpose(perm = var_350, x = attn_1_cast_fp16)[name = tensor("transpose_41")]; + tensor input_5_cast_fp16 = reshape(shape = var_355, x = var_351_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor attn0_out_proj_weight_to_fp16 = const()[name = tensor("attn0_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6304384)))]; + tensor linear_1_bias_0_to_fp16 = const()[name = tensor("linear_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401600)))]; + tensor linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn0_out_proj_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("linear_1_cast_fp16")]; + tensor var_364_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_364_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_7_cast_fp16 = add(x = conditioning_to_fp16, y = linear_1_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor input_9_axes_0 = const()[name = tensor("input_9_axes_0"), val = tensor([-1])]; + tensor norm0_2_weight_to_fp16 = const()[name = tensor("norm0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8403712)))]; + tensor norm0_2_bias_to_fp16 = const()[name = tensor("norm0_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8405824)))]; + tensor var_368_to_fp16 = const()[name = tensor("op_368_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_9_cast_fp16 = layer_norm(axes = input_9_axes_0, beta = norm0_2_bias_to_fp16, epsilon = var_368_to_fp16, gamma = norm0_2_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor linear0_1_weight_to_fp16 = const()[name = tensor("linear0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8407936)))]; + tensor linear_2_bias_0_to_fp16 = const()[name = tensor("linear_2_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16796608)))]; + tensor linear_2_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear0_1_weight_to_fp16, x = input_9_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor input_11_mode_0 = const()[name = tensor("input_11_mode_0"), val = tensor("EXACT")]; + tensor input_11_cast_fp16 = gelu(mode = input_11_mode_0, x = linear_2_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor linear0_2_weight_to_fp16 = const()[name = tensor("linear0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16804864)))]; + tensor linear_3_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear0_2_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("linear_3_cast_fp16")]; + tensor input_13_cast_fp16 = add(x = input_7_cast_fp16, y = linear_3_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor input_15_axes_0 = const()[name = tensor("input_15_axes_0"), val = tensor([-1])]; + tensor norm1_1_weight_to_fp16 = const()[name = tensor("norm1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25193536)))]; + tensor norm1_1_bias_to_fp16 = const()[name = tensor("norm1_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25195648)))]; + tensor var_385_to_fp16 = const()[name = tensor("op_385_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_15_cast_fp16 = layer_norm(axes = input_15_axes_0, beta = norm1_1_bias_to_fp16, epsilon = var_385_to_fp16, gamma = norm1_1_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor attn1_in_proj_weight_to_fp16 = const()[name = tensor("attn1_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25197760)))]; + tensor linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn1_in_proj_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor var_399 = const()[name = tensor("op_399"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_3_cast_fp16 = reshape(shape = var_399, x = linear_4_cast_fp16)[name = tensor("qkv_3_cast_fp16")]; + tensor x_5_begin_0 = const()[name = tensor("x_5_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_5_end_0 = const()[name = tensor("x_5_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_5_end_mask_0 = const()[name = tensor("x_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_5_squeeze_mask_0 = const()[name = tensor("x_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_5_cast_fp16 = slice_by_index(begin = x_5_begin_0, end = x_5_end_0, end_mask = x_5_end_mask_0, squeeze_mask = x_5_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("x_5_cast_fp16")]; + tensor x_7_begin_0 = const()[name = tensor("x_7_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_7_end_0 = const()[name = tensor("x_7_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_7_end_mask_0 = const()[name = tensor("x_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_7_squeeze_mask_0 = const()[name = tensor("x_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_7_cast_fp16 = slice_by_index(begin = x_7_begin_0, end = x_7_end_0, end_mask = x_7_end_mask_0, squeeze_mask = x_7_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("x_7_cast_fp16")]; + tensor v_3_begin_0 = const()[name = tensor("v_3_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_3_end_0 = const()[name = tensor("v_3_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_3_end_mask_0 = const()[name = tensor("v_3_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_3_squeeze_mask_0 = const()[name = tensor("v_3_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_3_cast_fp16 = slice_by_index(begin = v_3_begin_0, end = v_3_end_0, end_mask = v_3_end_mask_0, squeeze_mask = v_3_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("v_3_cast_fp16")]; + tensor var_444 = const()[name = tensor("op_444"), val = tensor([256, 16, 32, 2])]; + tensor pairs_5_cast_fp16 = reshape(shape = var_444, x = x_5_cast_fp16)[name = tensor("pairs_5_cast_fp16")]; + tensor xr_5_begin_0 = const()[name = tensor("xr_5_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_5_end_0 = const()[name = tensor("xr_5_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_5_end_mask_0 = const()[name = tensor("xr_5_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_5_squeeze_mask_0 = const()[name = tensor("xr_5_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_5_cast_fp16 = slice_by_index(begin = xr_5_begin_0, end = xr_5_end_0, end_mask = xr_5_end_mask_0, squeeze_mask = xr_5_squeeze_mask_0, x = pairs_5_cast_fp16)[name = tensor("xr_5_cast_fp16")]; + tensor xi_5_begin_0 = const()[name = tensor("xi_5_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_5_end_0 = const()[name = tensor("xi_5_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_5_end_mask_0 = const()[name = tensor("xi_5_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_5_squeeze_mask_0 = const()[name = tensor("xi_5_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_5_cast_fp16 = slice_by_index(begin = xi_5_begin_0, end = xi_5_end_0, end_mask = xi_5_end_mask_0, squeeze_mask = xi_5_squeeze_mask_0, x = pairs_5_cast_fp16)[name = tensor("xi_5_cast_fp16")]; + tensor var_452_cast_fp16 = mul(x = xr_5_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_452_cast_fp16")]; + tensor var_453_cast_fp16 = mul(x = xi_5_cast_fp16, y = roti_cast_fp16)[name = tensor("op_453_cast_fp16")]; + tensor xor__5_cast_fp16 = sub(x = var_452_cast_fp16, y = var_453_cast_fp16)[name = tensor("xor__5_cast_fp16")]; + tensor var_456_cast_fp16 = mul(x = xr_5_cast_fp16, y = roti_cast_fp16)[name = tensor("op_456_cast_fp16")]; + tensor var_457_cast_fp16 = mul(x = xi_5_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_457_cast_fp16")]; + tensor xoi_5_cast_fp16 = add(x = var_456_cast_fp16, y = var_457_cast_fp16)[name = tensor("xoi_5_cast_fp16")]; + tensor out_5_axis_0 = const()[name = tensor("out_5_axis_0"), val = tensor(-1)]; + tensor out_5_cast_fp16 = stack(axis = out_5_axis_0, values = (xor__5_cast_fp16, xoi_5_cast_fp16))[name = tensor("out_5_cast_fp16")]; + tensor var_467 = const()[name = tensor("op_467"), val = tensor([1, 256, 16, 64])]; + tensor q_3_cast_fp16 = reshape(shape = var_467, x = out_5_cast_fp16)[name = tensor("q_3_cast_fp16")]; + tensor var_473 = const()[name = tensor("op_473"), val = tensor([256, 16, 32, 2])]; + tensor pairs_7_cast_fp16 = reshape(shape = var_473, x = x_7_cast_fp16)[name = tensor("pairs_7_cast_fp16")]; + tensor xr_7_begin_0 = const()[name = tensor("xr_7_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_7_end_0 = const()[name = tensor("xr_7_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_7_end_mask_0 = const()[name = tensor("xr_7_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_7_squeeze_mask_0 = const()[name = tensor("xr_7_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_7_cast_fp16 = slice_by_index(begin = xr_7_begin_0, end = xr_7_end_0, end_mask = xr_7_end_mask_0, squeeze_mask = xr_7_squeeze_mask_0, x = pairs_7_cast_fp16)[name = tensor("xr_7_cast_fp16")]; + tensor xi_7_begin_0 = const()[name = tensor("xi_7_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_7_end_0 = const()[name = tensor("xi_7_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_7_end_mask_0 = const()[name = tensor("xi_7_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_7_squeeze_mask_0 = const()[name = tensor("xi_7_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_7_cast_fp16 = slice_by_index(begin = xi_7_begin_0, end = xi_7_end_0, end_mask = xi_7_end_mask_0, squeeze_mask = xi_7_squeeze_mask_0, x = pairs_7_cast_fp16)[name = tensor("xi_7_cast_fp16")]; + tensor var_481_cast_fp16 = mul(x = xr_7_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_481_cast_fp16")]; + tensor var_482_cast_fp16 = mul(x = xi_7_cast_fp16, y = roti_cast_fp16)[name = tensor("op_482_cast_fp16")]; + tensor xor__7_cast_fp16 = sub(x = var_481_cast_fp16, y = var_482_cast_fp16)[name = tensor("xor__7_cast_fp16")]; + tensor var_485_cast_fp16 = mul(x = xr_7_cast_fp16, y = roti_cast_fp16)[name = tensor("op_485_cast_fp16")]; + tensor var_486_cast_fp16 = mul(x = xi_7_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_486_cast_fp16")]; + tensor xoi_7_cast_fp16 = add(x = var_485_cast_fp16, y = var_486_cast_fp16)[name = tensor("xoi_7_cast_fp16")]; + tensor out_7_axis_0 = const()[name = tensor("out_7_axis_0"), val = tensor(-1)]; + tensor out_7_cast_fp16 = stack(axis = out_7_axis_0, values = (xor__7_cast_fp16, xoi_7_cast_fp16))[name = tensor("out_7_cast_fp16")]; + tensor var_503 = const()[name = tensor("op_503"), val = tensor([256, 1024])]; + tensor var_504_cast_fp16 = reshape(shape = var_503, x = out_7_cast_fp16)[name = tensor("op_504_cast_fp16")]; + tensor written_k_3_transpose_x_1 = const()[name = tensor("written_k_3_transpose_x_1"), val = tensor(true)]; + tensor written_k_3_transpose_y_1 = const()[name = tensor("written_k_3_transpose_y_1"), val = tensor(false)]; + tensor written_k_3_cast_fp16 = matmul(transpose_x = written_k_3_transpose_x_1, transpose_y = written_k_3_transpose_y_1, x = var_106_cast_fp16_to_fp16, y = var_504_cast_fp16)[name = tensor("written_k_3_cast_fp16")]; + tensor var_511 = const()[name = tensor("op_511"), val = tensor([256, 1024])]; + tensor var_512_cast_fp16 = reshape(shape = var_511, x = v_3_cast_fp16)[name = tensor("op_512_cast_fp16")]; + tensor written_v_3_transpose_x_1 = const()[name = tensor("written_v_3_transpose_x_1"), val = tensor(true)]; + tensor written_v_3_transpose_y_1 = const()[name = tensor("written_v_3_transpose_y_1"), val = tensor(false)]; + tensor written_v_3_cast_fp16 = matmul(transpose_x = written_v_3_transpose_x_1, transpose_y = written_v_3_transpose_y_1, x = var_106_cast_fp16_to_fp16, y = var_512_cast_fp16)[name = tensor("written_v_3_cast_fp16")]; + tensor k_cache1_to_fp16_dtype_0 = const()[name = tensor("k_cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache1_to_fp16 = cast(dtype = k_cache1_to_fp16_dtype_0, x = k_cache1)[name = tensor("cast_117")]; + tensor var_523_cast_fp16 = mul(x = k_cache1_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_523_cast_fp16")]; + tensor var_528 = const()[name = tensor("op_528"), val = tensor([1, 512, 16, 64])]; + tensor var_529_cast_fp16 = reshape(shape = var_528, x = written_k_3_cast_fp16)[name = tensor("op_529_cast_fp16")]; + tensor new_k_cache_3_cast_fp16 = add(x = var_523_cast_fp16, y = var_529_cast_fp16)[name = tensor("new_k_cache_3_cast_fp16")]; + tensor new_k_cache_3_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_3_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache1_to_fp16_dtype_0 = const()[name = tensor("v_cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache1_to_fp16 = cast(dtype = v_cache1_to_fp16_dtype_0, x = v_cache1)[name = tensor("cast_115")]; + tensor var_532_cast_fp16 = mul(x = v_cache1_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_532_cast_fp16")]; + tensor var_537 = const()[name = tensor("op_537"), val = tensor([1, 512, 16, 64])]; + tensor var_538_cast_fp16 = reshape(shape = var_537, x = written_v_3_cast_fp16)[name = tensor("op_538_cast_fp16")]; + tensor new_v_cache_3_cast_fp16 = add(x = var_532_cast_fp16, y = var_538_cast_fp16)[name = tensor("new_v_cache_3_cast_fp16")]; + tensor new_v_cache_3_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_3_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_551 = const()[name = tensor("op_551"), val = tensor([0, 2, 1, 3])]; + tensor var_562_transpose_x_0 = const()[name = tensor("op_562_transpose_x_0"), val = tensor(false)]; + tensor var_562_transpose_y_0 = const()[name = tensor("op_562_transpose_y_0"), val = tensor(false)]; + tensor transpose_17_perm_0 = const()[name = tensor("transpose_17_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_18_perm_0 = const()[name = tensor("transpose_18_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_18 = transpose(perm = transpose_18_perm_0, x = new_k_cache_3_cast_fp16)[name = tensor("transpose_38")]; + tensor transpose_17 = transpose(perm = transpose_17_perm_0, x = q_3_cast_fp16)[name = tensor("transpose_39")]; + tensor var_562_cast_fp16 = matmul(transpose_x = var_562_transpose_x_0, transpose_y = var_562_transpose_y_0, x = transpose_17, y = transpose_18)[name = tensor("op_562_cast_fp16")]; + tensor var_563_to_fp16 = const()[name = tensor("op_563_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_5_cast_fp16 = mul(x = var_562_cast_fp16, y = var_563_to_fp16)[name = tensor("scores_5_cast_fp16")]; + tensor scores_7_cast_fp16 = add(x = scores_5_cast_fp16, y = var_339_cast_fp16)[name = tensor("scores_7_cast_fp16")]; + tensor var_573 = const()[name = tensor("op_573"), val = tensor(-1)]; + tensor weights_3_cast_fp16 = softmax(axis = var_573, x = scores_7_cast_fp16)[name = tensor("weights_3_cast_fp16")]; + tensor attn_3_transpose_x_0 = const()[name = tensor("attn_3_transpose_x_0"), val = tensor(false)]; + tensor attn_3_transpose_y_0 = const()[name = tensor("attn_3_transpose_y_0"), val = tensor(false)]; + tensor values_3_cast_fp16 = transpose(perm = var_551, x = new_v_cache_3_cast_fp16)[name = tensor("transpose_40")]; + tensor attn_3_cast_fp16 = matmul(transpose_x = attn_3_transpose_x_0, transpose_y = attn_3_transpose_y_0, x = weights_3_cast_fp16, y = values_3_cast_fp16)[name = tensor("attn_3_cast_fp16")]; + tensor var_581 = const()[name = tensor("op_581"), val = tensor([0, 2, 1, 3])]; + tensor var_586 = const()[name = tensor("op_586"), val = tensor([1, 256, 1024])]; + tensor var_582_cast_fp16 = transpose(perm = var_581, x = attn_3_cast_fp16)[name = tensor("transpose_37")]; + tensor input_17_cast_fp16 = reshape(shape = var_586, x = var_582_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor attn1_out_proj_weight_to_fp16 = const()[name = tensor("attn1_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31489280)))]; + tensor linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn1_out_proj_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor position1_to_fp16_dtype_0 = const()[name = tensor("position1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position1_to_fp16 = cast(dtype = position1_to_fp16_dtype_0, x = position1)[name = tensor("cast_113")]; + tensor var_595_cast_fp16 = add(x = position1_to_fp16, y = valid_len_to_fp16)[name = tensor("op_595_cast_fp16")]; + tensor var_595_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_595_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_19_cast_fp16 = add(x = input_13_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor input_21_axes_0 = const()[name = tensor("input_21_axes_0"), val = tensor([-1])]; + tensor norm1_2_weight_to_fp16 = const()[name = tensor("norm1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33586496)))]; + tensor norm1_2_bias_to_fp16 = const()[name = tensor("norm1_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33588608)))]; + tensor var_599_to_fp16 = const()[name = tensor("op_599_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_21_cast_fp16 = layer_norm(axes = input_21_axes_0, beta = norm1_2_bias_to_fp16, epsilon = var_599_to_fp16, gamma = norm1_2_weight_to_fp16, x = input_19_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor linear1_1_weight_to_fp16 = const()[name = tensor("linear1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33590720)))]; + tensor linear_6_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear1_1_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor input_23_mode_0 = const()[name = tensor("input_23_mode_0"), val = tensor("EXACT")]; + tensor input_23_cast_fp16 = gelu(mode = input_23_mode_0, x = linear_6_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor linear1_2_weight_to_fp16 = const()[name = tensor("linear1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41979392)))]; + tensor linear_7_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear1_2_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_25_cast_fp16 = add(x = input_19_cast_fp16, y = linear_7_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor input_27_axes_0 = const()[name = tensor("input_27_axes_0"), val = tensor([-1])]; + tensor norm2_1_weight_to_fp16 = const()[name = tensor("norm2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50368064)))]; + tensor norm2_1_bias_to_fp16 = const()[name = tensor("norm2_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50370176)))]; + tensor var_616_to_fp16 = const()[name = tensor("op_616_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = norm2_1_bias_to_fp16, epsilon = var_616_to_fp16, gamma = norm2_1_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor attn2_in_proj_weight_to_fp16 = const()[name = tensor("attn2_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50372288)))]; + tensor linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn2_in_proj_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor var_630 = const()[name = tensor("op_630"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_5_cast_fp16 = reshape(shape = var_630, x = linear_8_cast_fp16)[name = tensor("qkv_5_cast_fp16")]; + tensor x_9_begin_0 = const()[name = tensor("x_9_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_9_end_0 = const()[name = tensor("x_9_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_9_end_mask_0 = const()[name = tensor("x_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_9_squeeze_mask_0 = const()[name = tensor("x_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_9_cast_fp16 = slice_by_index(begin = x_9_begin_0, end = x_9_end_0, end_mask = x_9_end_mask_0, squeeze_mask = x_9_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("x_9_cast_fp16")]; + tensor x_11_begin_0 = const()[name = tensor("x_11_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_11_end_0 = const()[name = tensor("x_11_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_11_end_mask_0 = const()[name = tensor("x_11_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_11_squeeze_mask_0 = const()[name = tensor("x_11_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_11_cast_fp16 = slice_by_index(begin = x_11_begin_0, end = x_11_end_0, end_mask = x_11_end_mask_0, squeeze_mask = x_11_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("x_11_cast_fp16")]; + tensor v_5_begin_0 = const()[name = tensor("v_5_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_5_end_0 = const()[name = tensor("v_5_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_5_end_mask_0 = const()[name = tensor("v_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_5_squeeze_mask_0 = const()[name = tensor("v_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_5_cast_fp16 = slice_by_index(begin = v_5_begin_0, end = v_5_end_0, end_mask = v_5_end_mask_0, squeeze_mask = v_5_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("v_5_cast_fp16")]; + tensor var_675 = const()[name = tensor("op_675"), val = tensor([256, 16, 32, 2])]; + tensor pairs_9_cast_fp16 = reshape(shape = var_675, x = x_9_cast_fp16)[name = tensor("pairs_9_cast_fp16")]; + tensor xr_9_begin_0 = const()[name = tensor("xr_9_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_9_end_0 = const()[name = tensor("xr_9_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_9_end_mask_0 = const()[name = tensor("xr_9_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_9_squeeze_mask_0 = const()[name = tensor("xr_9_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_9_cast_fp16 = slice_by_index(begin = xr_9_begin_0, end = xr_9_end_0, end_mask = xr_9_end_mask_0, squeeze_mask = xr_9_squeeze_mask_0, x = pairs_9_cast_fp16)[name = tensor("xr_9_cast_fp16")]; + tensor xi_9_begin_0 = const()[name = tensor("xi_9_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_9_end_0 = const()[name = tensor("xi_9_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_9_end_mask_0 = const()[name = tensor("xi_9_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_9_squeeze_mask_0 = const()[name = tensor("xi_9_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_9_cast_fp16 = slice_by_index(begin = xi_9_begin_0, end = xi_9_end_0, end_mask = xi_9_end_mask_0, squeeze_mask = xi_9_squeeze_mask_0, x = pairs_9_cast_fp16)[name = tensor("xi_9_cast_fp16")]; + tensor var_683_cast_fp16 = mul(x = xr_9_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_683_cast_fp16")]; + tensor var_684_cast_fp16 = mul(x = xi_9_cast_fp16, y = roti_cast_fp16)[name = tensor("op_684_cast_fp16")]; + tensor xor__9_cast_fp16 = sub(x = var_683_cast_fp16, y = var_684_cast_fp16)[name = tensor("xor__9_cast_fp16")]; + tensor var_687_cast_fp16 = mul(x = xr_9_cast_fp16, y = roti_cast_fp16)[name = tensor("op_687_cast_fp16")]; + tensor var_688_cast_fp16 = mul(x = xi_9_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_688_cast_fp16")]; + tensor xoi_9_cast_fp16 = add(x = var_687_cast_fp16, y = var_688_cast_fp16)[name = tensor("xoi_9_cast_fp16")]; + tensor out_9_axis_0 = const()[name = tensor("out_9_axis_0"), val = tensor(-1)]; + tensor out_9_cast_fp16 = stack(axis = out_9_axis_0, values = (xor__9_cast_fp16, xoi_9_cast_fp16))[name = tensor("out_9_cast_fp16")]; + tensor var_698 = const()[name = tensor("op_698"), val = tensor([1, 256, 16, 64])]; + tensor q_5_cast_fp16 = reshape(shape = var_698, x = out_9_cast_fp16)[name = tensor("q_5_cast_fp16")]; + tensor var_704 = const()[name = tensor("op_704"), val = tensor([256, 16, 32, 2])]; + tensor pairs_11_cast_fp16 = reshape(shape = var_704, x = x_11_cast_fp16)[name = tensor("pairs_11_cast_fp16")]; + tensor xr_11_begin_0 = const()[name = tensor("xr_11_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_11_end_0 = const()[name = tensor("xr_11_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_11_end_mask_0 = const()[name = tensor("xr_11_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_11_squeeze_mask_0 = const()[name = tensor("xr_11_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_11_cast_fp16 = slice_by_index(begin = xr_11_begin_0, end = xr_11_end_0, end_mask = xr_11_end_mask_0, squeeze_mask = xr_11_squeeze_mask_0, x = pairs_11_cast_fp16)[name = tensor("xr_11_cast_fp16")]; + tensor xi_11_begin_0 = const()[name = tensor("xi_11_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_11_end_0 = const()[name = tensor("xi_11_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_11_end_mask_0 = const()[name = tensor("xi_11_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_11_squeeze_mask_0 = const()[name = tensor("xi_11_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_11_cast_fp16 = slice_by_index(begin = xi_11_begin_0, end = xi_11_end_0, end_mask = xi_11_end_mask_0, squeeze_mask = xi_11_squeeze_mask_0, x = pairs_11_cast_fp16)[name = tensor("xi_11_cast_fp16")]; + tensor var_712_cast_fp16 = mul(x = xr_11_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_712_cast_fp16")]; + tensor var_713_cast_fp16 = mul(x = xi_11_cast_fp16, y = roti_cast_fp16)[name = tensor("op_713_cast_fp16")]; + tensor xor__11_cast_fp16 = sub(x = var_712_cast_fp16, y = var_713_cast_fp16)[name = tensor("xor__11_cast_fp16")]; + tensor var_716_cast_fp16 = mul(x = xr_11_cast_fp16, y = roti_cast_fp16)[name = tensor("op_716_cast_fp16")]; + tensor var_717_cast_fp16 = mul(x = xi_11_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_717_cast_fp16")]; + tensor xoi_11_cast_fp16 = add(x = var_716_cast_fp16, y = var_717_cast_fp16)[name = tensor("xoi_11_cast_fp16")]; + tensor out_11_axis_0 = const()[name = tensor("out_11_axis_0"), val = tensor(-1)]; + tensor out_11_cast_fp16 = stack(axis = out_11_axis_0, values = (xor__11_cast_fp16, xoi_11_cast_fp16))[name = tensor("out_11_cast_fp16")]; + tensor var_734 = const()[name = tensor("op_734"), val = tensor([256, 1024])]; + tensor var_735_cast_fp16 = reshape(shape = var_734, x = out_11_cast_fp16)[name = tensor("op_735_cast_fp16")]; + tensor written_k_5_transpose_x_1 = const()[name = tensor("written_k_5_transpose_x_1"), val = tensor(true)]; + tensor written_k_5_transpose_y_1 = const()[name = tensor("written_k_5_transpose_y_1"), val = tensor(false)]; + tensor written_k_5_cast_fp16 = matmul(transpose_x = written_k_5_transpose_x_1, transpose_y = written_k_5_transpose_y_1, x = var_106_cast_fp16_to_fp16, y = var_735_cast_fp16)[name = tensor("written_k_5_cast_fp16")]; + tensor var_742 = const()[name = tensor("op_742"), val = tensor([256, 1024])]; + tensor var_743_cast_fp16 = reshape(shape = var_742, x = v_5_cast_fp16)[name = tensor("op_743_cast_fp16")]; + tensor written_v_5_transpose_x_1 = const()[name = tensor("written_v_5_transpose_x_1"), val = tensor(true)]; + tensor written_v_5_transpose_y_1 = const()[name = tensor("written_v_5_transpose_y_1"), val = tensor(false)]; + tensor written_v_5_cast_fp16 = matmul(transpose_x = written_v_5_transpose_x_1, transpose_y = written_v_5_transpose_y_1, x = var_106_cast_fp16_to_fp16, y = var_743_cast_fp16)[name = tensor("written_v_5_cast_fp16")]; + tensor k_cache2_to_fp16_dtype_0 = const()[name = tensor("k_cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache2_to_fp16 = cast(dtype = k_cache2_to_fp16_dtype_0, x = k_cache2)[name = tensor("cast_111")]; + tensor var_754_cast_fp16 = mul(x = k_cache2_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_754_cast_fp16")]; + tensor var_759 = const()[name = tensor("op_759"), val = tensor([1, 512, 16, 64])]; + tensor var_760_cast_fp16 = reshape(shape = var_759, x = written_k_5_cast_fp16)[name = tensor("op_760_cast_fp16")]; + tensor new_k_cache_5_cast_fp16 = add(x = var_754_cast_fp16, y = var_760_cast_fp16)[name = tensor("new_k_cache_5_cast_fp16")]; + tensor new_k_cache_5_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_5_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache2_to_fp16_dtype_0 = const()[name = tensor("v_cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache2_to_fp16 = cast(dtype = v_cache2_to_fp16_dtype_0, x = v_cache2)[name = tensor("cast_109")]; + tensor var_763_cast_fp16 = mul(x = v_cache2_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_763_cast_fp16")]; + tensor var_768 = const()[name = tensor("op_768"), val = tensor([1, 512, 16, 64])]; + tensor var_769_cast_fp16 = reshape(shape = var_768, x = written_v_5_cast_fp16)[name = tensor("op_769_cast_fp16")]; + tensor new_v_cache_5_cast_fp16 = add(x = var_763_cast_fp16, y = var_769_cast_fp16)[name = tensor("new_v_cache_5_cast_fp16")]; + tensor new_v_cache_5_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_5_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_782 = const()[name = tensor("op_782"), val = tensor([0, 2, 1, 3])]; + tensor var_793_transpose_x_0 = const()[name = tensor("op_793_transpose_x_0"), val = tensor(false)]; + tensor var_793_transpose_y_0 = const()[name = tensor("op_793_transpose_y_0"), val = tensor(false)]; + tensor transpose_19_perm_0 = const()[name = tensor("transpose_19_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_20_perm_0 = const()[name = tensor("transpose_20_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_20 = transpose(perm = transpose_20_perm_0, x = new_k_cache_5_cast_fp16)[name = tensor("transpose_34")]; + tensor transpose_19 = transpose(perm = transpose_19_perm_0, x = q_5_cast_fp16)[name = tensor("transpose_35")]; + tensor var_793_cast_fp16 = matmul(transpose_x = var_793_transpose_x_0, transpose_y = var_793_transpose_y_0, x = transpose_19, y = transpose_20)[name = tensor("op_793_cast_fp16")]; + tensor var_794_to_fp16 = const()[name = tensor("op_794_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_9_cast_fp16 = mul(x = var_793_cast_fp16, y = var_794_to_fp16)[name = tensor("scores_9_cast_fp16")]; + tensor scores_11_cast_fp16 = add(x = scores_9_cast_fp16, y = var_339_cast_fp16)[name = tensor("scores_11_cast_fp16")]; + tensor var_804 = const()[name = tensor("op_804"), val = tensor(-1)]; + tensor weights_5_cast_fp16 = softmax(axis = var_804, x = scores_11_cast_fp16)[name = tensor("weights_5_cast_fp16")]; + tensor attn_5_transpose_x_0 = const()[name = tensor("attn_5_transpose_x_0"), val = tensor(false)]; + tensor attn_5_transpose_y_0 = const()[name = tensor("attn_5_transpose_y_0"), val = tensor(false)]; + tensor values_5_cast_fp16 = transpose(perm = var_782, x = new_v_cache_5_cast_fp16)[name = tensor("transpose_36")]; + tensor attn_5_cast_fp16 = matmul(transpose_x = attn_5_transpose_x_0, transpose_y = attn_5_transpose_y_0, x = weights_5_cast_fp16, y = values_5_cast_fp16)[name = tensor("attn_5_cast_fp16")]; + tensor var_812 = const()[name = tensor("op_812"), val = tensor([0, 2, 1, 3])]; + tensor var_817 = const()[name = tensor("op_817"), val = tensor([1, 256, 1024])]; + tensor var_813_cast_fp16 = transpose(perm = var_812, x = attn_5_cast_fp16)[name = tensor("transpose_33")]; + tensor input_29_cast_fp16 = reshape(shape = var_817, x = var_813_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor attn2_out_proj_weight_to_fp16 = const()[name = tensor("attn2_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(56663808)))]; + tensor linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn2_out_proj_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor position2_to_fp16_dtype_0 = const()[name = tensor("position2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position2_to_fp16 = cast(dtype = position2_to_fp16_dtype_0, x = position2)[name = tensor("cast_107")]; + tensor var_826_cast_fp16 = add(x = position2_to_fp16, y = valid_len_to_fp16)[name = tensor("op_826_cast_fp16")]; + tensor var_826_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_826_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_31_cast_fp16 = add(x = input_25_cast_fp16, y = linear_9_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; + tensor norm2_2_weight_to_fp16 = const()[name = tensor("norm2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58761024)))]; + tensor norm2_2_bias_to_fp16 = const()[name = tensor("norm2_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58763136)))]; + tensor var_830_to_fp16 = const()[name = tensor("op_830_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_33_cast_fp16 = layer_norm(axes = input_33_axes_0, beta = norm2_2_bias_to_fp16, epsilon = var_830_to_fp16, gamma = norm2_2_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor linear2_1_weight_to_fp16 = const()[name = tensor("linear2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58765248)))]; + tensor linear_10_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear2_1_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor input_35_mode_0 = const()[name = tensor("input_35_mode_0"), val = tensor("EXACT")]; + tensor input_35_cast_fp16 = gelu(mode = input_35_mode_0, x = linear_10_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor linear2_2_weight_to_fp16 = const()[name = tensor("linear2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67153920)))]; + tensor linear_11_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear2_2_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor input_37_cast_fp16 = add(x = input_31_cast_fp16, y = linear_11_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor input_39_axes_0 = const()[name = tensor("input_39_axes_0"), val = tensor([-1])]; + tensor norm3_1_weight_to_fp16 = const()[name = tensor("norm3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75542592)))]; + tensor norm3_1_bias_to_fp16 = const()[name = tensor("norm3_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75544704)))]; + tensor var_847_to_fp16 = const()[name = tensor("op_847_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_39_cast_fp16 = layer_norm(axes = input_39_axes_0, beta = norm3_1_bias_to_fp16, epsilon = var_847_to_fp16, gamma = norm3_1_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor attn3_in_proj_weight_to_fp16 = const()[name = tensor("attn3_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75546816)))]; + tensor linear_12_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn3_in_proj_weight_to_fp16, x = input_39_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor var_861 = const()[name = tensor("op_861"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_7_cast_fp16 = reshape(shape = var_861, x = linear_12_cast_fp16)[name = tensor("qkv_7_cast_fp16")]; + tensor x_13_begin_0 = const()[name = tensor("x_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_13_end_0 = const()[name = tensor("x_13_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_13_end_mask_0 = const()[name = tensor("x_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_13_squeeze_mask_0 = const()[name = tensor("x_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_13_cast_fp16 = slice_by_index(begin = x_13_begin_0, end = x_13_end_0, end_mask = x_13_end_mask_0, squeeze_mask = x_13_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("x_13_cast_fp16")]; + tensor x_15_begin_0 = const()[name = tensor("x_15_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_15_end_0 = const()[name = tensor("x_15_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_15_end_mask_0 = const()[name = tensor("x_15_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_15_squeeze_mask_0 = const()[name = tensor("x_15_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_15_cast_fp16 = slice_by_index(begin = x_15_begin_0, end = x_15_end_0, end_mask = x_15_end_mask_0, squeeze_mask = x_15_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("x_15_cast_fp16")]; + tensor v_7_begin_0 = const()[name = tensor("v_7_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_7_end_0 = const()[name = tensor("v_7_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_7_end_mask_0 = const()[name = tensor("v_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_7_squeeze_mask_0 = const()[name = tensor("v_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_7_cast_fp16 = slice_by_index(begin = v_7_begin_0, end = v_7_end_0, end_mask = v_7_end_mask_0, squeeze_mask = v_7_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("v_7_cast_fp16")]; + tensor var_906 = const()[name = tensor("op_906"), val = tensor([256, 16, 32, 2])]; + tensor pairs_13_cast_fp16 = reshape(shape = var_906, x = x_13_cast_fp16)[name = tensor("pairs_13_cast_fp16")]; + tensor xr_13_begin_0 = const()[name = tensor("xr_13_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_13_end_0 = const()[name = tensor("xr_13_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_13_end_mask_0 = const()[name = tensor("xr_13_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_13_squeeze_mask_0 = const()[name = tensor("xr_13_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_13_cast_fp16 = slice_by_index(begin = xr_13_begin_0, end = xr_13_end_0, end_mask = xr_13_end_mask_0, squeeze_mask = xr_13_squeeze_mask_0, x = pairs_13_cast_fp16)[name = tensor("xr_13_cast_fp16")]; + tensor xi_13_begin_0 = const()[name = tensor("xi_13_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_13_end_0 = const()[name = tensor("xi_13_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_13_end_mask_0 = const()[name = tensor("xi_13_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_13_squeeze_mask_0 = const()[name = tensor("xi_13_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_13_cast_fp16 = slice_by_index(begin = xi_13_begin_0, end = xi_13_end_0, end_mask = xi_13_end_mask_0, squeeze_mask = xi_13_squeeze_mask_0, x = pairs_13_cast_fp16)[name = tensor("xi_13_cast_fp16")]; + tensor var_914_cast_fp16 = mul(x = xr_13_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_914_cast_fp16")]; + tensor var_915_cast_fp16 = mul(x = xi_13_cast_fp16, y = roti_cast_fp16)[name = tensor("op_915_cast_fp16")]; + tensor xor__13_cast_fp16 = sub(x = var_914_cast_fp16, y = var_915_cast_fp16)[name = tensor("xor__13_cast_fp16")]; + tensor var_918_cast_fp16 = mul(x = xr_13_cast_fp16, y = roti_cast_fp16)[name = tensor("op_918_cast_fp16")]; + tensor var_919_cast_fp16 = mul(x = xi_13_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_919_cast_fp16")]; + tensor xoi_13_cast_fp16 = add(x = var_918_cast_fp16, y = var_919_cast_fp16)[name = tensor("xoi_13_cast_fp16")]; + tensor out_13_axis_0 = const()[name = tensor("out_13_axis_0"), val = tensor(-1)]; + tensor out_13_cast_fp16 = stack(axis = out_13_axis_0, values = (xor__13_cast_fp16, xoi_13_cast_fp16))[name = tensor("out_13_cast_fp16")]; + tensor var_929 = const()[name = tensor("op_929"), val = tensor([1, 256, 16, 64])]; + tensor q_7_cast_fp16 = reshape(shape = var_929, x = out_13_cast_fp16)[name = tensor("q_7_cast_fp16")]; + tensor var_935 = const()[name = tensor("op_935"), val = tensor([256, 16, 32, 2])]; + tensor pairs_15_cast_fp16 = reshape(shape = var_935, x = x_15_cast_fp16)[name = tensor("pairs_15_cast_fp16")]; + tensor xr_15_begin_0 = const()[name = tensor("xr_15_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_15_end_0 = const()[name = tensor("xr_15_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_15_end_mask_0 = const()[name = tensor("xr_15_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_15_squeeze_mask_0 = const()[name = tensor("xr_15_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_15_cast_fp16 = slice_by_index(begin = xr_15_begin_0, end = xr_15_end_0, end_mask = xr_15_end_mask_0, squeeze_mask = xr_15_squeeze_mask_0, x = pairs_15_cast_fp16)[name = tensor("xr_15_cast_fp16")]; + tensor xi_15_begin_0 = const()[name = tensor("xi_15_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_15_end_0 = const()[name = tensor("xi_15_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_15_end_mask_0 = const()[name = tensor("xi_15_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_15_squeeze_mask_0 = const()[name = tensor("xi_15_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_15_cast_fp16 = slice_by_index(begin = xi_15_begin_0, end = xi_15_end_0, end_mask = xi_15_end_mask_0, squeeze_mask = xi_15_squeeze_mask_0, x = pairs_15_cast_fp16)[name = tensor("xi_15_cast_fp16")]; + tensor var_943_cast_fp16 = mul(x = xr_15_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_943_cast_fp16")]; + tensor var_944_cast_fp16 = mul(x = xi_15_cast_fp16, y = roti_cast_fp16)[name = tensor("op_944_cast_fp16")]; + tensor xor__15_cast_fp16 = sub(x = var_943_cast_fp16, y = var_944_cast_fp16)[name = tensor("xor__15_cast_fp16")]; + tensor var_947_cast_fp16 = mul(x = xr_15_cast_fp16, y = roti_cast_fp16)[name = tensor("op_947_cast_fp16")]; + tensor var_948_cast_fp16 = mul(x = xi_15_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_948_cast_fp16")]; + tensor xoi_15_cast_fp16 = add(x = var_947_cast_fp16, y = var_948_cast_fp16)[name = tensor("xoi_15_cast_fp16")]; + tensor out_15_axis_0 = const()[name = tensor("out_15_axis_0"), val = tensor(-1)]; + tensor out_15_cast_fp16 = stack(axis = out_15_axis_0, values = (xor__15_cast_fp16, xoi_15_cast_fp16))[name = tensor("out_15_cast_fp16")]; + tensor var_965 = const()[name = tensor("op_965"), val = tensor([256, 1024])]; + tensor var_966_cast_fp16 = reshape(shape = var_965, x = out_15_cast_fp16)[name = tensor("op_966_cast_fp16")]; + tensor written_k_7_transpose_x_1 = const()[name = tensor("written_k_7_transpose_x_1"), val = tensor(true)]; + tensor written_k_7_transpose_y_1 = const()[name = tensor("written_k_7_transpose_y_1"), val = tensor(false)]; + tensor written_k_7_cast_fp16 = matmul(transpose_x = written_k_7_transpose_x_1, transpose_y = written_k_7_transpose_y_1, x = var_106_cast_fp16_to_fp16, y = var_966_cast_fp16)[name = tensor("written_k_7_cast_fp16")]; + tensor var_973 = const()[name = tensor("op_973"), val = tensor([256, 1024])]; + tensor var_974_cast_fp16 = reshape(shape = var_973, x = v_7_cast_fp16)[name = tensor("op_974_cast_fp16")]; + tensor written_v_7_transpose_x_1 = const()[name = tensor("written_v_7_transpose_x_1"), val = tensor(true)]; + tensor written_v_7_transpose_y_1 = const()[name = tensor("written_v_7_transpose_y_1"), val = tensor(false)]; + tensor written_v_7_cast_fp16 = matmul(transpose_x = written_v_7_transpose_x_1, transpose_y = written_v_7_transpose_y_1, x = var_106_cast_fp16_to_fp16, y = var_974_cast_fp16)[name = tensor("written_v_7_cast_fp16")]; + tensor k_cache3_to_fp16_dtype_0 = const()[name = tensor("k_cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache3_to_fp16 = cast(dtype = k_cache3_to_fp16_dtype_0, x = k_cache3)[name = tensor("cast_105")]; + tensor var_985_cast_fp16 = mul(x = k_cache3_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_985_cast_fp16")]; + tensor var_990 = const()[name = tensor("op_990"), val = tensor([1, 512, 16, 64])]; + tensor var_991_cast_fp16 = reshape(shape = var_990, x = written_k_7_cast_fp16)[name = tensor("op_991_cast_fp16")]; + tensor new_k_cache_7_cast_fp16 = add(x = var_985_cast_fp16, y = var_991_cast_fp16)[name = tensor("new_k_cache_7_cast_fp16")]; + tensor new_k_cache_7_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_7_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache3_to_fp16_dtype_0 = const()[name = tensor("v_cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache3_to_fp16 = cast(dtype = v_cache3_to_fp16_dtype_0, x = v_cache3)[name = tensor("cast_103")]; + tensor var_994_cast_fp16 = mul(x = v_cache3_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_994_cast_fp16")]; + tensor var_999 = const()[name = tensor("op_999"), val = tensor([1, 512, 16, 64])]; + tensor var_1000_cast_fp16 = reshape(shape = var_999, x = written_v_7_cast_fp16)[name = tensor("op_1000_cast_fp16")]; + tensor new_v_cache_7_cast_fp16 = add(x = var_994_cast_fp16, y = var_1000_cast_fp16)[name = tensor("new_v_cache_7_cast_fp16")]; + tensor new_v_cache_7_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_7_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_1013 = const()[name = tensor("op_1013"), val = tensor([0, 2, 1, 3])]; + tensor var_1024_transpose_x_0 = const()[name = tensor("op_1024_transpose_x_0"), val = tensor(false)]; + tensor var_1024_transpose_y_0 = const()[name = tensor("op_1024_transpose_y_0"), val = tensor(false)]; + tensor transpose_21_perm_0 = const()[name = tensor("transpose_21_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_22_perm_0 = const()[name = tensor("transpose_22_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_22 = transpose(perm = transpose_22_perm_0, x = new_k_cache_7_cast_fp16)[name = tensor("transpose_30")]; + tensor transpose_21 = transpose(perm = transpose_21_perm_0, x = q_7_cast_fp16)[name = tensor("transpose_31")]; + tensor var_1024_cast_fp16 = matmul(transpose_x = var_1024_transpose_x_0, transpose_y = var_1024_transpose_y_0, x = transpose_21, y = transpose_22)[name = tensor("op_1024_cast_fp16")]; + tensor var_1025_to_fp16 = const()[name = tensor("op_1025_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_13_cast_fp16 = mul(x = var_1024_cast_fp16, y = var_1025_to_fp16)[name = tensor("scores_13_cast_fp16")]; + tensor scores_15_cast_fp16 = add(x = scores_13_cast_fp16, y = var_339_cast_fp16)[name = tensor("scores_15_cast_fp16")]; + tensor var_1035 = const()[name = tensor("op_1035"), val = tensor(-1)]; + tensor weights_7_cast_fp16 = softmax(axis = var_1035, x = scores_15_cast_fp16)[name = tensor("weights_7_cast_fp16")]; + tensor attn_7_transpose_x_0 = const()[name = tensor("attn_7_transpose_x_0"), val = tensor(false)]; + tensor attn_7_transpose_y_0 = const()[name = tensor("attn_7_transpose_y_0"), val = tensor(false)]; + tensor values_7_cast_fp16 = transpose(perm = var_1013, x = new_v_cache_7_cast_fp16)[name = tensor("transpose_32")]; + tensor attn_7_cast_fp16 = matmul(transpose_x = attn_7_transpose_x_0, transpose_y = attn_7_transpose_y_0, x = weights_7_cast_fp16, y = values_7_cast_fp16)[name = tensor("attn_7_cast_fp16")]; + tensor var_1043 = const()[name = tensor("op_1043"), val = tensor([0, 2, 1, 3])]; + tensor var_1048 = const()[name = tensor("op_1048"), val = tensor([1, 256, 1024])]; + tensor var_1044_cast_fp16 = transpose(perm = var_1043, x = attn_7_cast_fp16)[name = tensor("transpose_29")]; + tensor input_41_cast_fp16 = reshape(shape = var_1048, x = var_1044_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor attn3_out_proj_weight_to_fp16 = const()[name = tensor("attn3_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(81838336)))]; + tensor linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn3_out_proj_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor position3_to_fp16_dtype_0 = const()[name = tensor("position3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position3_to_fp16 = cast(dtype = position3_to_fp16_dtype_0, x = position3)[name = tensor("cast_101")]; + tensor var_1057_cast_fp16 = add(x = position3_to_fp16, y = valid_len_to_fp16)[name = tensor("op_1057_cast_fp16")]; + tensor var_1057_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1057_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_43_cast_fp16 = add(x = input_37_cast_fp16, y = linear_13_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; + tensor norm3_2_weight_to_fp16 = const()[name = tensor("norm3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(83935552)))]; + tensor norm3_2_bias_to_fp16 = const()[name = tensor("norm3_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(83937664)))]; + tensor var_1061_to_fp16 = const()[name = tensor("op_1061_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_45_cast_fp16 = layer_norm(axes = input_45_axes_0, beta = norm3_2_bias_to_fp16, epsilon = var_1061_to_fp16, gamma = norm3_2_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor linear3_1_weight_to_fp16 = const()[name = tensor("linear3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(83939776)))]; + tensor linear_14_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear3_1_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor input_47_mode_0 = const()[name = tensor("input_47_mode_0"), val = tensor("EXACT")]; + tensor input_47_cast_fp16 = gelu(mode = input_47_mode_0, x = linear_14_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor linear3_2_weight_to_fp16 = const()[name = tensor("linear3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(92328448)))]; + tensor linear_15_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear3_2_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor input_49_cast_fp16 = add(x = input_43_cast_fp16, y = linear_15_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor input_51_axes_0 = const()[name = tensor("input_51_axes_0"), val = tensor([-1])]; + tensor norm4_1_weight_to_fp16 = const()[name = tensor("norm4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100717120)))]; + tensor norm4_1_bias_to_fp16 = const()[name = tensor("norm4_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100719232)))]; + tensor var_1078_to_fp16 = const()[name = tensor("op_1078_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_51_cast_fp16 = layer_norm(axes = input_51_axes_0, beta = norm4_1_bias_to_fp16, epsilon = var_1078_to_fp16, gamma = norm4_1_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor attn4_in_proj_weight_to_fp16 = const()[name = tensor("attn4_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100721344)))]; + tensor linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn4_in_proj_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor var_1092 = const()[name = tensor("op_1092"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_9_cast_fp16 = reshape(shape = var_1092, x = linear_16_cast_fp16)[name = tensor("qkv_9_cast_fp16")]; + tensor x_17_begin_0 = const()[name = tensor("x_17_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_17_end_0 = const()[name = tensor("x_17_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_17_end_mask_0 = const()[name = tensor("x_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_17_squeeze_mask_0 = const()[name = tensor("x_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_17_cast_fp16 = slice_by_index(begin = x_17_begin_0, end = x_17_end_0, end_mask = x_17_end_mask_0, squeeze_mask = x_17_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("x_17_cast_fp16")]; + tensor x_19_begin_0 = const()[name = tensor("x_19_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_19_end_0 = const()[name = tensor("x_19_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_19_end_mask_0 = const()[name = tensor("x_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_19_squeeze_mask_0 = const()[name = tensor("x_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_19_cast_fp16 = slice_by_index(begin = x_19_begin_0, end = x_19_end_0, end_mask = x_19_end_mask_0, squeeze_mask = x_19_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("x_19_cast_fp16")]; + tensor v_9_begin_0 = const()[name = tensor("v_9_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_9_end_0 = const()[name = tensor("v_9_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_9_end_mask_0 = const()[name = tensor("v_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_9_squeeze_mask_0 = const()[name = tensor("v_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_9_cast_fp16 = slice_by_index(begin = v_9_begin_0, end = v_9_end_0, end_mask = v_9_end_mask_0, squeeze_mask = v_9_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("v_9_cast_fp16")]; + tensor var_1137 = const()[name = tensor("op_1137"), val = tensor([256, 16, 32, 2])]; + tensor pairs_17_cast_fp16 = reshape(shape = var_1137, x = x_17_cast_fp16)[name = tensor("pairs_17_cast_fp16")]; + tensor xr_17_begin_0 = const()[name = tensor("xr_17_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_17_end_0 = const()[name = tensor("xr_17_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_17_end_mask_0 = const()[name = tensor("xr_17_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_17_squeeze_mask_0 = const()[name = tensor("xr_17_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_17_cast_fp16 = slice_by_index(begin = xr_17_begin_0, end = xr_17_end_0, end_mask = xr_17_end_mask_0, squeeze_mask = xr_17_squeeze_mask_0, x = pairs_17_cast_fp16)[name = tensor("xr_17_cast_fp16")]; + tensor xi_17_begin_0 = const()[name = tensor("xi_17_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_17_end_0 = const()[name = tensor("xi_17_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_17_end_mask_0 = const()[name = tensor("xi_17_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_17_squeeze_mask_0 = const()[name = tensor("xi_17_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_17_cast_fp16 = slice_by_index(begin = xi_17_begin_0, end = xi_17_end_0, end_mask = xi_17_end_mask_0, squeeze_mask = xi_17_squeeze_mask_0, x = pairs_17_cast_fp16)[name = tensor("xi_17_cast_fp16")]; + tensor var_1145_cast_fp16 = mul(x = xr_17_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1145_cast_fp16")]; + tensor var_1146_cast_fp16 = mul(x = xi_17_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1146_cast_fp16")]; + tensor xor__17_cast_fp16 = sub(x = var_1145_cast_fp16, y = var_1146_cast_fp16)[name = tensor("xor__17_cast_fp16")]; + tensor var_1149_cast_fp16 = mul(x = xr_17_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1149_cast_fp16")]; + tensor var_1150_cast_fp16 = mul(x = xi_17_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1150_cast_fp16")]; + tensor xoi_17_cast_fp16 = add(x = var_1149_cast_fp16, y = var_1150_cast_fp16)[name = tensor("xoi_17_cast_fp16")]; + tensor out_17_axis_0 = const()[name = tensor("out_17_axis_0"), val = tensor(-1)]; + tensor out_17_cast_fp16 = stack(axis = out_17_axis_0, values = (xor__17_cast_fp16, xoi_17_cast_fp16))[name = tensor("out_17_cast_fp16")]; + tensor var_1160 = const()[name = tensor("op_1160"), val = tensor([1, 256, 16, 64])]; + tensor q_9_cast_fp16 = reshape(shape = var_1160, x = out_17_cast_fp16)[name = tensor("q_9_cast_fp16")]; + tensor var_1166 = const()[name = tensor("op_1166"), val = tensor([256, 16, 32, 2])]; + tensor pairs_19_cast_fp16 = reshape(shape = var_1166, x = x_19_cast_fp16)[name = tensor("pairs_19_cast_fp16")]; + tensor xr_19_begin_0 = const()[name = tensor("xr_19_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_19_end_0 = const()[name = tensor("xr_19_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_19_end_mask_0 = const()[name = tensor("xr_19_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_19_squeeze_mask_0 = const()[name = tensor("xr_19_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_19_cast_fp16 = slice_by_index(begin = xr_19_begin_0, end = xr_19_end_0, end_mask = xr_19_end_mask_0, squeeze_mask = xr_19_squeeze_mask_0, x = pairs_19_cast_fp16)[name = tensor("xr_19_cast_fp16")]; + tensor xi_19_begin_0 = const()[name = tensor("xi_19_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_19_end_0 = const()[name = tensor("xi_19_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_19_end_mask_0 = const()[name = tensor("xi_19_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_19_squeeze_mask_0 = const()[name = tensor("xi_19_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_19_cast_fp16 = slice_by_index(begin = xi_19_begin_0, end = xi_19_end_0, end_mask = xi_19_end_mask_0, squeeze_mask = xi_19_squeeze_mask_0, x = pairs_19_cast_fp16)[name = tensor("xi_19_cast_fp16")]; + tensor var_1174_cast_fp16 = mul(x = xr_19_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1174_cast_fp16")]; + tensor var_1175_cast_fp16 = mul(x = xi_19_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1175_cast_fp16")]; + tensor xor__19_cast_fp16 = sub(x = var_1174_cast_fp16, y = var_1175_cast_fp16)[name = tensor("xor__19_cast_fp16")]; + tensor var_1178_cast_fp16 = mul(x = xr_19_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1178_cast_fp16")]; + tensor var_1179_cast_fp16 = mul(x = xi_19_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1179_cast_fp16")]; + tensor xoi_19_cast_fp16 = add(x = var_1178_cast_fp16, y = var_1179_cast_fp16)[name = tensor("xoi_19_cast_fp16")]; + tensor out_19_axis_0 = const()[name = tensor("out_19_axis_0"), val = tensor(-1)]; + tensor out_19_cast_fp16 = stack(axis = out_19_axis_0, values = (xor__19_cast_fp16, xoi_19_cast_fp16))[name = tensor("out_19_cast_fp16")]; + tensor var_1196 = const()[name = tensor("op_1196"), val = tensor([256, 1024])]; + tensor var_1197_cast_fp16 = reshape(shape = var_1196, x = out_19_cast_fp16)[name = tensor("op_1197_cast_fp16")]; + tensor written_k_9_transpose_x_1 = const()[name = tensor("written_k_9_transpose_x_1"), val = tensor(true)]; + tensor written_k_9_transpose_y_1 = const()[name = tensor("written_k_9_transpose_y_1"), val = tensor(false)]; + tensor written_k_9_cast_fp16 = matmul(transpose_x = written_k_9_transpose_x_1, transpose_y = written_k_9_transpose_y_1, x = var_106_cast_fp16_to_fp16, y = var_1197_cast_fp16)[name = tensor("written_k_9_cast_fp16")]; + tensor var_1204 = const()[name = tensor("op_1204"), val = tensor([256, 1024])]; + tensor var_1205_cast_fp16 = reshape(shape = var_1204, x = v_9_cast_fp16)[name = tensor("op_1205_cast_fp16")]; + tensor written_v_9_transpose_x_1 = const()[name = tensor("written_v_9_transpose_x_1"), val = tensor(true)]; + tensor written_v_9_transpose_y_1 = const()[name = tensor("written_v_9_transpose_y_1"), val = tensor(false)]; + tensor written_v_9_cast_fp16 = matmul(transpose_x = written_v_9_transpose_x_1, transpose_y = written_v_9_transpose_y_1, x = var_106_cast_fp16_to_fp16, y = var_1205_cast_fp16)[name = tensor("written_v_9_cast_fp16")]; + tensor k_cache4_to_fp16_dtype_0 = const()[name = tensor("k_cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache4_to_fp16 = cast(dtype = k_cache4_to_fp16_dtype_0, x = k_cache4)[name = tensor("cast_99")]; + tensor var_1216_cast_fp16 = mul(x = k_cache4_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_1216_cast_fp16")]; + tensor var_1221 = const()[name = tensor("op_1221"), val = tensor([1, 512, 16, 64])]; + tensor var_1222_cast_fp16 = reshape(shape = var_1221, x = written_k_9_cast_fp16)[name = tensor("op_1222_cast_fp16")]; + tensor new_k_cache_9_cast_fp16 = add(x = var_1216_cast_fp16, y = var_1222_cast_fp16)[name = tensor("new_k_cache_9_cast_fp16")]; + tensor new_k_cache_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache4_to_fp16_dtype_0 = const()[name = tensor("v_cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache4_to_fp16 = cast(dtype = v_cache4_to_fp16_dtype_0, x = v_cache4)[name = tensor("cast_97")]; + tensor var_1225_cast_fp16 = mul(x = v_cache4_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_1225_cast_fp16")]; + tensor var_1230 = const()[name = tensor("op_1230"), val = tensor([1, 512, 16, 64])]; + tensor var_1231_cast_fp16 = reshape(shape = var_1230, x = written_v_9_cast_fp16)[name = tensor("op_1231_cast_fp16")]; + tensor new_v_cache_9_cast_fp16 = add(x = var_1225_cast_fp16, y = var_1231_cast_fp16)[name = tensor("new_v_cache_9_cast_fp16")]; + tensor new_v_cache_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_1244 = const()[name = tensor("op_1244"), val = tensor([0, 2, 1, 3])]; + tensor var_1255_transpose_x_0 = const()[name = tensor("op_1255_transpose_x_0"), val = tensor(false)]; + tensor var_1255_transpose_y_0 = const()[name = tensor("op_1255_transpose_y_0"), val = tensor(false)]; + tensor transpose_23_perm_0 = const()[name = tensor("transpose_23_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_24_perm_0 = const()[name = tensor("transpose_24_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_24 = transpose(perm = transpose_24_perm_0, x = new_k_cache_9_cast_fp16)[name = tensor("transpose_26")]; + tensor transpose_23 = transpose(perm = transpose_23_perm_0, x = q_9_cast_fp16)[name = tensor("transpose_27")]; + tensor var_1255_cast_fp16 = matmul(transpose_x = var_1255_transpose_x_0, transpose_y = var_1255_transpose_y_0, x = transpose_23, y = transpose_24)[name = tensor("op_1255_cast_fp16")]; + tensor var_1256_to_fp16 = const()[name = tensor("op_1256_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_17_cast_fp16 = mul(x = var_1255_cast_fp16, y = var_1256_to_fp16)[name = tensor("scores_17_cast_fp16")]; + tensor scores_19_cast_fp16 = add(x = scores_17_cast_fp16, y = var_339_cast_fp16)[name = tensor("scores_19_cast_fp16")]; + tensor var_1266 = const()[name = tensor("op_1266"), val = tensor(-1)]; + tensor weights_9_cast_fp16 = softmax(axis = var_1266, x = scores_19_cast_fp16)[name = tensor("weights_9_cast_fp16")]; + tensor attn_9_transpose_x_0 = const()[name = tensor("attn_9_transpose_x_0"), val = tensor(false)]; + tensor attn_9_transpose_y_0 = const()[name = tensor("attn_9_transpose_y_0"), val = tensor(false)]; + tensor values_9_cast_fp16 = transpose(perm = var_1244, x = new_v_cache_9_cast_fp16)[name = tensor("transpose_28")]; + tensor attn_9_cast_fp16 = matmul(transpose_x = attn_9_transpose_x_0, transpose_y = attn_9_transpose_y_0, x = weights_9_cast_fp16, y = values_9_cast_fp16)[name = tensor("attn_9_cast_fp16")]; + tensor var_1274 = const()[name = tensor("op_1274"), val = tensor([0, 2, 1, 3])]; + tensor var_1279 = const()[name = tensor("op_1279"), val = tensor([1, 256, 1024])]; + tensor var_1275_cast_fp16 = transpose(perm = var_1274, x = attn_9_cast_fp16)[name = tensor("transpose_25")]; + tensor input_53_cast_fp16 = reshape(shape = var_1279, x = var_1275_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor attn4_out_proj_weight_to_fp16 = const()[name = tensor("attn4_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(107012864)))]; + tensor linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn4_out_proj_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor position4_to_fp16_dtype_0 = const()[name = tensor("position4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position4_to_fp16 = cast(dtype = position4_to_fp16_dtype_0, x = position4)[name = tensor("cast_95")]; + tensor var_1288_cast_fp16 = add(x = position4_to_fp16, y = valid_len_to_fp16)[name = tensor("op_1288_cast_fp16")]; + tensor var_1288_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1288_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_55_cast_fp16 = add(x = input_49_cast_fp16, y = linear_17_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor input_57_axes_0 = const()[name = tensor("input_57_axes_0"), val = tensor([-1])]; + tensor norm4_2_weight_to_fp16 = const()[name = tensor("norm4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109110080)))]; + tensor norm4_2_bias_to_fp16 = const()[name = tensor("norm4_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109112192)))]; + tensor var_1292_to_fp16 = const()[name = tensor("op_1292_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_57_cast_fp16 = layer_norm(axes = input_57_axes_0, beta = norm4_2_bias_to_fp16, epsilon = var_1292_to_fp16, gamma = norm4_2_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor linear4_1_weight_to_fp16 = const()[name = tensor("linear4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109114304)))]; + tensor linear_18_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear4_1_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor input_59_mode_0 = const()[name = tensor("input_59_mode_0"), val = tensor("EXACT")]; + tensor input_59_cast_fp16 = gelu(mode = input_59_mode_0, x = linear_18_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor linear4_2_weight_to_fp16 = const()[name = tensor("linear4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(117502976)))]; + tensor linear_19_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear4_2_weight_to_fp16, x = input_59_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_61_cast_fp16 = add(x = input_55_cast_fp16, y = linear_19_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor input_63_axes_0 = const()[name = tensor("input_63_axes_0"), val = tensor([-1])]; + tensor norm5_1_weight_to_fp16 = const()[name = tensor("norm5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125891648)))]; + tensor norm5_1_bias_to_fp16 = const()[name = tensor("norm5_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125893760)))]; + tensor var_1309_to_fp16 = const()[name = tensor("op_1309_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_63_cast_fp16 = layer_norm(axes = input_63_axes_0, beta = norm5_1_bias_to_fp16, epsilon = var_1309_to_fp16, gamma = norm5_1_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor attn5_in_proj_weight_to_fp16 = const()[name = tensor("attn5_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125895872)))]; + tensor linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn5_in_proj_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor var_1323 = const()[name = tensor("op_1323"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_cast_fp16 = reshape(shape = var_1323, x = linear_20_cast_fp16)[name = tensor("qkv_cast_fp16")]; + tensor x_begin_0 = const()[name = tensor("x_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_end_0 = const()[name = tensor("x_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_end_mask_0 = const()[name = tensor("x_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_squeeze_mask_0 = const()[name = tensor("x_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_cast_fp16 = slice_by_index(begin = x_begin_0, end = x_end_0, end_mask = x_end_mask_0, squeeze_mask = x_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("x_cast_fp16")]; + tensor v_begin_0 = const()[name = tensor("v_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_end_0 = const()[name = tensor("v_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_end_mask_0 = const()[name = tensor("v_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_squeeze_mask_0 = const()[name = tensor("v_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_cast_fp16 = slice_by_index(begin = v_begin_0, end = v_end_0, end_mask = v_end_mask_0, squeeze_mask = v_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("v_cast_fp16")]; + tensor var_1355 = const()[name = tensor("op_1355"), val = tensor([256, 16, 32, 2])]; + tensor pairs_cast_fp16 = reshape(shape = var_1355, x = x_cast_fp16)[name = tensor("pairs_cast_fp16")]; + tensor xr_begin_0 = const()[name = tensor("xr_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_end_0 = const()[name = tensor("xr_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_end_mask_0 = const()[name = tensor("xr_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_squeeze_mask_0 = const()[name = tensor("xr_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_cast_fp16 = slice_by_index(begin = xr_begin_0, end = xr_end_0, end_mask = xr_end_mask_0, squeeze_mask = xr_squeeze_mask_0, x = pairs_cast_fp16)[name = tensor("xr_cast_fp16")]; + tensor xi_begin_0 = const()[name = tensor("xi_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_end_0 = const()[name = tensor("xi_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_end_mask_0 = const()[name = tensor("xi_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_squeeze_mask_0 = const()[name = tensor("xi_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_cast_fp16 = slice_by_index(begin = xi_begin_0, end = xi_end_0, end_mask = xi_end_mask_0, squeeze_mask = xi_squeeze_mask_0, x = pairs_cast_fp16)[name = tensor("xi_cast_fp16")]; + tensor var_1363_cast_fp16 = mul(x = xr_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1363_cast_fp16")]; + tensor var_1364_cast_fp16 = mul(x = xi_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1364_cast_fp16")]; + tensor xor__cast_fp16 = sub(x = var_1363_cast_fp16, y = var_1364_cast_fp16)[name = tensor("xor__cast_fp16")]; + tensor var_1367_cast_fp16 = mul(x = xr_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1367_cast_fp16")]; + tensor var_1368_cast_fp16 = mul(x = xi_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1368_cast_fp16")]; + tensor xoi_cast_fp16 = add(x = var_1367_cast_fp16, y = var_1368_cast_fp16)[name = tensor("xoi_cast_fp16")]; + tensor out_axis_0 = const()[name = tensor("out_axis_0"), val = tensor(-1)]; + tensor out_cast_fp16 = stack(axis = out_axis_0, values = (xor__cast_fp16, xoi_cast_fp16))[name = tensor("out_cast_fp16")]; + tensor var_1385 = const()[name = tensor("op_1385"), val = tensor([256, 1024])]; + tensor var_1386_cast_fp16 = reshape(shape = var_1385, x = out_cast_fp16)[name = tensor("op_1386_cast_fp16")]; + tensor written_k_transpose_x_1 = const()[name = tensor("written_k_transpose_x_1"), val = tensor(true)]; + tensor written_k_transpose_y_1 = const()[name = tensor("written_k_transpose_y_1"), val = tensor(false)]; + tensor written_k_cast_fp16 = matmul(transpose_x = written_k_transpose_x_1, transpose_y = written_k_transpose_y_1, x = var_106_cast_fp16_to_fp16, y = var_1386_cast_fp16)[name = tensor("written_k_cast_fp16")]; + tensor var_1393 = const()[name = tensor("op_1393"), val = tensor([256, 1024])]; + tensor var_1394_cast_fp16 = reshape(shape = var_1393, x = v_cast_fp16)[name = tensor("op_1394_cast_fp16")]; + tensor written_v_transpose_x_1 = const()[name = tensor("written_v_transpose_x_1"), val = tensor(true)]; + tensor written_v_transpose_y_1 = const()[name = tensor("written_v_transpose_y_1"), val = tensor(false)]; + tensor written_v_cast_fp16 = matmul(transpose_x = written_v_transpose_x_1, transpose_y = written_v_transpose_y_1, x = var_106_cast_fp16_to_fp16, y = var_1394_cast_fp16)[name = tensor("written_v_cast_fp16")]; + tensor k_cache5_to_fp16_dtype_0 = const()[name = tensor("k_cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache5_to_fp16 = cast(dtype = k_cache5_to_fp16_dtype_0, x = k_cache5)[name = tensor("cast_93")]; + tensor var_1405_cast_fp16 = mul(x = k_cache5_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_1405_cast_fp16")]; + tensor var_1410 = const()[name = tensor("op_1410"), val = tensor([1, 512, 16, 64])]; + tensor var_1411_cast_fp16 = reshape(shape = var_1410, x = written_k_cast_fp16)[name = tensor("op_1411_cast_fp16")]; + tensor new_k_cache_cast_fp16 = add(x = var_1405_cast_fp16, y = var_1411_cast_fp16)[name = tensor("new_k_cache_cast_fp16")]; + tensor new_k_cache_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache5_to_fp16_dtype_0 = const()[name = tensor("v_cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache5_to_fp16 = cast(dtype = v_cache5_to_fp16_dtype_0, x = v_cache5)[name = tensor("cast_91")]; + tensor var_1414_cast_fp16 = mul(x = v_cache5_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_1414_cast_fp16")]; + tensor var_1419 = const()[name = tensor("op_1419"), val = tensor([1, 512, 16, 64])]; + tensor var_1420_cast_fp16 = reshape(shape = var_1419, x = written_v_cast_fp16)[name = tensor("op_1420_cast_fp16")]; + tensor new_v_cache_cast_fp16 = add(x = var_1414_cast_fp16, y = var_1420_cast_fp16)[name = tensor("new_v_cache_cast_fp16")]; + tensor new_v_cache_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor position5_to_fp16_dtype_0 = const()[name = tensor("position5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position5_to_fp16 = cast(dtype = position5_to_fp16_dtype_0, x = position5)[name = tensor("cast_89")]; + tensor var_1427_cast_fp16 = add(x = position5_to_fp16, y = valid_len_to_fp16)[name = tensor("op_1427_cast_fp16")]; + tensor var_1427_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1427_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor new_position5 = cast(dtype = var_1427_cast_fp16_to_fp32_dtype_0, x = var_1427_cast_fp16)[name = tensor("cast_88")]; + tensor new_v_cache5 = cast(dtype = new_v_cache_cast_fp16_to_fp32_dtype_0, x = new_v_cache_cast_fp16)[name = tensor("cast_90")]; + tensor new_k_cache5 = cast(dtype = new_k_cache_cast_fp16_to_fp32_dtype_0, x = new_k_cache_cast_fp16)[name = tensor("cast_92")]; + tensor new_position4 = cast(dtype = var_1288_cast_fp16_to_fp32_dtype_0, x = var_1288_cast_fp16)[name = tensor("cast_94")]; + tensor new_v_cache4 = cast(dtype = new_v_cache_9_cast_fp16_to_fp32_dtype_0, x = new_v_cache_9_cast_fp16)[name = tensor("cast_96")]; + tensor new_k_cache4 = cast(dtype = new_k_cache_9_cast_fp16_to_fp32_dtype_0, x = new_k_cache_9_cast_fp16)[name = tensor("cast_98")]; + tensor new_position3 = cast(dtype = var_1057_cast_fp16_to_fp32_dtype_0, x = var_1057_cast_fp16)[name = tensor("cast_100")]; + tensor new_v_cache3 = cast(dtype = new_v_cache_7_cast_fp16_to_fp32_dtype_0, x = new_v_cache_7_cast_fp16)[name = tensor("cast_102")]; + tensor new_k_cache3 = cast(dtype = new_k_cache_7_cast_fp16_to_fp32_dtype_0, x = new_k_cache_7_cast_fp16)[name = tensor("cast_104")]; + tensor new_position2 = cast(dtype = var_826_cast_fp16_to_fp32_dtype_0, x = var_826_cast_fp16)[name = tensor("cast_106")]; + tensor new_v_cache2 = cast(dtype = new_v_cache_5_cast_fp16_to_fp32_dtype_0, x = new_v_cache_5_cast_fp16)[name = tensor("cast_108")]; + tensor new_k_cache2 = cast(dtype = new_k_cache_5_cast_fp16_to_fp32_dtype_0, x = new_k_cache_5_cast_fp16)[name = tensor("cast_110")]; + tensor new_position1 = cast(dtype = var_595_cast_fp16_to_fp32_dtype_0, x = var_595_cast_fp16)[name = tensor("cast_112")]; + tensor new_v_cache1 = cast(dtype = new_v_cache_3_cast_fp16_to_fp32_dtype_0, x = new_v_cache_3_cast_fp16)[name = tensor("cast_114")]; + tensor new_k_cache1 = cast(dtype = new_k_cache_3_cast_fp16_to_fp32_dtype_0, x = new_k_cache_3_cast_fp16)[name = tensor("cast_116")]; + tensor new_position0 = cast(dtype = var_364_cast_fp16_to_fp32_dtype_0, x = var_135_cast_fp16)[name = tensor("cast_118")]; + tensor new_v_cache0 = cast(dtype = new_v_cache_1_cast_fp16_to_fp32_dtype_0, x = new_v_cache_1_cast_fp16)[name = tensor("cast_119")]; + tensor new_k_cache0 = cast(dtype = new_k_cache_1_cast_fp16_to_fp32_dtype_0, x = new_k_cache_1_cast_fp16)[name = tensor("cast_121")]; + } -> (new_k_cache0, new_v_cache0, new_position0, new_k_cache1, new_v_cache1, new_position1, new_k_cache2, new_v_cache2, new_position2, new_k_cache3, new_v_cache3, new_position3, new_k_cache4, new_v_cache4, new_position4, new_k_cache5, new_v_cache5, new_position5); +} \ No newline at end of file diff --git a/v2.1/spanish/cond_prefill_ane.mlmodelc/weights/weight.bin b/v2.1/spanish/cond_prefill_ane.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..fbee5efb97fe41dfb227c45041a1295e23c00da5 --- /dev/null +++ b/v2.1/spanish/cond_prefill_ane.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e1bbdd868c072b8896acf2913d73bf97c5de86aed18a105ae03ccc863684f542 +size 132187392 diff --git a/v2.1/spanish/cond_prefill_ane.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/spanish/cond_prefill_ane.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..9a719798a131cb2e125dc5b5c6be7881a3cc7436 --- /dev/null +++ b/v2.1/spanish/cond_prefill_ane.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:429ba499715207d86da0041189f6d3e364f99f05b0d83a3fa0c76116fe18c329 +size 119086 diff --git a/v2.1/spanish/cond_prefill_ane.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/spanish/cond_prefill_ane.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..fbee5efb97fe41dfb227c45041a1295e23c00da5 --- /dev/null +++ b/v2.1/spanish/cond_prefill_ane.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e1bbdd868c072b8896acf2913d73bf97c5de86aed18a105ae03ccc863684f542 +size 132187392 diff --git a/v2.1/spanish/cond_prefill_ane.mlpackage/Manifest.json b/v2.1/spanish/cond_prefill_ane.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..6ef62ef0808dd28b3b45786c1161323025a22abd --- /dev/null +++ b/v2.1/spanish/cond_prefill_ane.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "467F81E6-05FE-4029-BBBC-ECEFFBFB87B0": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "966BAFCF-9A4B-4252-A821-F7BF6DA4CAF5": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "966BAFCF-9A4B-4252-A821-F7BF6DA4CAF5" +} diff --git a/v2.1/spanish/flowlm_step_ane.mlmodelc/analytics/coremldata.bin b/v2.1/spanish/flowlm_step_ane.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..75da37e1484a1fb3c3e34ba0f9caef3f46321cd8 --- /dev/null +++ b/v2.1/spanish/flowlm_step_ane.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4a256145410acf49659dc151044127d55a93d8d49728b314009c2b4860a023c1 +size 243 diff --git a/v2.1/spanish/flowlm_step_ane.mlmodelc/coremldata.bin b/v2.1/spanish/flowlm_step_ane.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..4cfd4cf5eda852f5bca64bb3ef6760a5b0247618 --- /dev/null +++ b/v2.1/spanish/flowlm_step_ane.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9a3dfca03ae0b61c961b8b25e6f9fd8dc8c7aa5eb759698a1b43bd6cf2f2b9b3 +size 1414 diff --git a/v2.1/spanish/flowlm_step_ane.mlmodelc/model.mil b/v2.1/spanish/flowlm_step_ane.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..47440c00f93629c0591f4e79607fb505c0e94160 --- /dev/null +++ b/v2.1/spanish/flowlm_step_ane.mlmodelc/model.mil @@ -0,0 +1,945 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.9.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor k_cache0, tensor k_cache1, tensor k_cache2, tensor k_cache3, tensor k_cache4, tensor k_cache5, tensor position0, tensor position1, tensor position2, tensor position3, tensor position4, tensor position5, tensor sequence, tensor v_cache0, tensor v_cache1, tensor v_cache2, tensor v_cache3, tensor v_cache4, tensor v_cache5) { + tensor sequence_to_fp16_dtype_0 = const()[name = tensor("sequence_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor input_linear_weight_to_fp16 = const()[name = tensor("input_linear_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(65664)))]; + tensor sequence_to_fp16 = cast(dtype = sequence_to_fp16_dtype_0, x = sequence)[name = tensor("cast_146")]; + tensor linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = input_linear_weight_to_fp16, x = sequence_to_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; + tensor norm0_1_weight_to_fp16 = const()[name = tensor("norm0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67776)))]; + tensor norm0_1_bias_to_fp16 = const()[name = tensor("norm0_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(69888)))]; + tensor var_65_to_fp16 = const()[name = tensor("op_65_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_5_cast_fp16 = layer_norm(axes = input_5_axes_0, beta = norm0_1_bias_to_fp16, epsilon = var_65_to_fp16, gamma = norm0_1_weight_to_fp16, x = linear_0_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor attn0_in_proj_weight_to_fp16 = const()[name = tensor("attn0_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(72000)))]; + tensor linear_1_bias_0_to_fp16 = const()[name = tensor("linear_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6363520)))]; + tensor linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn0_in_proj_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("linear_1_cast_fp16")]; + tensor var_78 = const()[name = tensor("op_78"), val = tensor([1, 3, 16, 64])]; + tensor qkv_1_cast_fp16 = reshape(shape = var_78, x = linear_1_cast_fp16)[name = tensor("qkv_1_cast_fp16")]; + tensor x_1_begin_0 = const()[name = tensor("x_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_1_end_0 = const()[name = tensor("x_1_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_1_end_mask_0 = const()[name = tensor("x_1_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_1_squeeze_mask_0 = const()[name = tensor("x_1_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_1_cast_fp16 = slice_by_index(begin = x_1_begin_0, end = x_1_end_0, end_mask = x_1_end_mask_0, squeeze_mask = x_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("x_1_cast_fp16")]; + tensor x_3_begin_0 = const()[name = tensor("x_3_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_3_end_0 = const()[name = tensor("x_3_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_3_end_mask_0 = const()[name = tensor("x_3_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_3_squeeze_mask_0 = const()[name = tensor("x_3_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_3_cast_fp16 = slice_by_index(begin = x_3_begin_0, end = x_3_end_0, end_mask = x_3_end_mask_0, squeeze_mask = x_3_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("x_3_cast_fp16")]; + tensor v_1_begin_0 = const()[name = tensor("v_1_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_1_end_0 = const()[name = tensor("v_1_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_1_end_mask_0 = const()[name = tensor("v_1_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_1_squeeze_mask_0 = const()[name = tensor("v_1_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_1_cast_fp16 = slice_by_index(begin = v_1_begin_0, end = v_1_end_0, end_mask = v_1_end_mask_0, squeeze_mask = v_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("v_1_cast_fp16")]; + tensor var_107 = const()[name = tensor("op_107"), val = tensor([1, 1, 1])]; + tensor position0_to_fp16_dtype_0 = const()[name = tensor("position0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position0_to_fp16 = cast(dtype = position0_to_fp16_dtype_0, x = position0)[name = tensor("cast_145")]; + tensor var_108_cast_fp16 = reshape(shape = var_107, x = position0_to_fp16)[name = tensor("op_108_cast_fp16")]; + tensor rope_freqs_to_fp16 = const()[name = tensor("rope_freqs_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6369728)))]; + tensor angles_1_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_108_cast_fp16)[name = tensor("angles_1_cast_fp16")]; + tensor rotr_1_cast_fp16 = cos(x = angles_1_cast_fp16)[name = tensor("rotr_1_cast_fp16")]; + tensor roti_1_cast_fp16 = sin(x = angles_1_cast_fp16)[name = tensor("roti_1_cast_fp16")]; + tensor var_116 = const()[name = tensor("op_116"), val = tensor([1, 16, 32, 2])]; + tensor pairs_1_cast_fp16 = reshape(shape = var_116, x = x_1_cast_fp16)[name = tensor("pairs_1_cast_fp16")]; + tensor xr_1_begin_0 = const()[name = tensor("xr_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_1_end_0 = const()[name = tensor("xr_1_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_1_end_mask_0 = const()[name = tensor("xr_1_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_1_squeeze_mask_0 = const()[name = tensor("xr_1_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_1_cast_fp16 = slice_by_index(begin = xr_1_begin_0, end = xr_1_end_0, end_mask = xr_1_end_mask_0, squeeze_mask = xr_1_squeeze_mask_0, x = pairs_1_cast_fp16)[name = tensor("xr_1_cast_fp16")]; + tensor xi_1_begin_0 = const()[name = tensor("xi_1_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_1_end_0 = const()[name = tensor("xi_1_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_1_end_mask_0 = const()[name = tensor("xi_1_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_1_squeeze_mask_0 = const()[name = tensor("xi_1_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_1_cast_fp16 = slice_by_index(begin = xi_1_begin_0, end = xi_1_end_0, end_mask = xi_1_end_mask_0, squeeze_mask = xi_1_squeeze_mask_0, x = pairs_1_cast_fp16)[name = tensor("xi_1_cast_fp16")]; + tensor var_124_cast_fp16 = mul(x = xr_1_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_124_cast_fp16")]; + tensor var_125_cast_fp16 = mul(x = xi_1_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_125_cast_fp16")]; + tensor xor__1_cast_fp16 = sub(x = var_124_cast_fp16, y = var_125_cast_fp16)[name = tensor("xor__1_cast_fp16")]; + tensor var_128_cast_fp16 = mul(x = xr_1_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_128_cast_fp16")]; + tensor var_129_cast_fp16 = mul(x = xi_1_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_129_cast_fp16")]; + tensor xoi_1_cast_fp16 = add(x = var_128_cast_fp16, y = var_129_cast_fp16)[name = tensor("xoi_1_cast_fp16")]; + tensor out_1_axis_0 = const()[name = tensor("out_1_axis_0"), val = tensor(-1)]; + tensor out_1_cast_fp16 = stack(axis = out_1_axis_0, values = (xor__1_cast_fp16, xoi_1_cast_fp16))[name = tensor("out_1_cast_fp16")]; + tensor var_152 = const()[name = tensor("op_152"), val = tensor([1, 16, 32, 2])]; + tensor pairs_3_cast_fp16 = reshape(shape = var_152, x = x_3_cast_fp16)[name = tensor("pairs_3_cast_fp16")]; + tensor xr_3_begin_0 = const()[name = tensor("xr_3_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_3_end_0 = const()[name = tensor("xr_3_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_3_end_mask_0 = const()[name = tensor("xr_3_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_3_squeeze_mask_0 = const()[name = tensor("xr_3_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_3_cast_fp16 = slice_by_index(begin = xr_3_begin_0, end = xr_3_end_0, end_mask = xr_3_end_mask_0, squeeze_mask = xr_3_squeeze_mask_0, x = pairs_3_cast_fp16)[name = tensor("xr_3_cast_fp16")]; + tensor xi_3_begin_0 = const()[name = tensor("xi_3_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_3_end_0 = const()[name = tensor("xi_3_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_3_end_mask_0 = const()[name = tensor("xi_3_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_3_squeeze_mask_0 = const()[name = tensor("xi_3_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_3_cast_fp16 = slice_by_index(begin = xi_3_begin_0, end = xi_3_end_0, end_mask = xi_3_end_mask_0, squeeze_mask = xi_3_squeeze_mask_0, x = pairs_3_cast_fp16)[name = tensor("xi_3_cast_fp16")]; + tensor var_160_cast_fp16 = mul(x = xr_3_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_160_cast_fp16")]; + tensor var_161_cast_fp16 = mul(x = xi_3_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_161_cast_fp16")]; + tensor xor__3_cast_fp16 = sub(x = var_160_cast_fp16, y = var_161_cast_fp16)[name = tensor("xor__3_cast_fp16")]; + tensor var_164_cast_fp16 = mul(x = xr_3_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_164_cast_fp16")]; + tensor var_165_cast_fp16 = mul(x = xi_3_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_165_cast_fp16")]; + tensor xoi_3_cast_fp16 = add(x = var_164_cast_fp16, y = var_165_cast_fp16)[name = tensor("xoi_3_cast_fp16")]; + tensor out_3_axis_0 = const()[name = tensor("out_3_axis_0"), val = tensor(-1)]; + tensor out_3_cast_fp16 = stack(axis = out_3_axis_0, values = (xor__3_cast_fp16, xoi_3_cast_fp16))[name = tensor("out_3_cast_fp16")]; + tensor _inversed_177_y_0_to_fp16 = const()[name = tensor("_inversed_177_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_177_cast_fp16 = mul(x = position0_to_fp16, y = _inversed_177_y_0_to_fp16)[name = tensor("_inversed_177_cast_fp16")]; + tensor var_178_cast_fp16 = floor(x = _inversed_177_cast_fp16)[name = tensor("op_178_cast_fp16")]; + tensor var_179_to_fp16 = const()[name = tensor("op_179_to_fp16"), val = tensor(0x1p+9)]; + tensor var_180_cast_fp16 = mul(x = var_178_cast_fp16, y = var_179_to_fp16)[name = tensor("op_180_cast_fp16")]; + tensor write_idx_1_cast_fp16 = sub(x = position0_to_fp16, y = var_180_cast_fp16)[name = tensor("write_idx_1_cast_fp16")]; + tensor slot_indices_to_fp16 = const()[name = tensor("slot_indices_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6369856)))]; + tensor var_183_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_1_cast_fp16)[name = tensor("op_183_cast_fp16")]; + tensor var_193 = const()[name = tensor("op_193"), val = tensor([1, 512, 1, 1])]; + tensor onehot_1_to_fp16_dtype_0 = const()[name = tensor("onehot_1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_183_cast_fp16_to_fp16 = cast(dtype = onehot_1_to_fp16_dtype_0, x = var_183_cast_fp16)[name = tensor("cast_144")]; + tensor onehot_3_cast_fp16 = reshape(shape = var_193, x = var_183_cast_fp16_to_fp16)[name = tensor("onehot_3_cast_fp16")]; + tensor var_195_to_fp16 = const()[name = tensor("op_195_to_fp16"), val = tensor(0x1p+0)]; + tensor var_197_cast_fp16 = sub(x = var_195_to_fp16, y = onehot_3_cast_fp16)[name = tensor("op_197_cast_fp16")]; + tensor k_cache0_to_fp16_dtype_0 = const()[name = tensor("k_cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache0_to_fp16 = cast(dtype = k_cache0_to_fp16_dtype_0, x = k_cache0)[name = tensor("cast_143")]; + tensor var_198_cast_fp16 = mul(x = k_cache0_to_fp16, y = var_197_cast_fp16)[name = tensor("op_198_cast_fp16")]; + tensor var_203 = const()[name = tensor("op_203"), val = tensor([1, 1, 16, 64])]; + tensor var_204_cast_fp16 = reshape(shape = var_203, x = out_3_cast_fp16)[name = tensor("op_204_cast_fp16")]; + tensor var_205_cast_fp16 = mul(x = var_204_cast_fp16, y = onehot_3_cast_fp16)[name = tensor("op_205_cast_fp16")]; + tensor new_k_cache_1_cast_fp16 = add(x = var_198_cast_fp16, y = var_205_cast_fp16)[name = tensor("new_k_cache_1_cast_fp16")]; + tensor new_k_cache_1_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_1_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache0_to_fp16_dtype_0 = const()[name = tensor("v_cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache0_to_fp16 = cast(dtype = v_cache0_to_fp16_dtype_0, x = v_cache0)[name = tensor("cast_141")]; + tensor var_211_cast_fp16 = mul(x = v_cache0_to_fp16, y = var_197_cast_fp16)[name = tensor("op_211_cast_fp16")]; + tensor var_216 = const()[name = tensor("op_216"), val = tensor([1, 1, 16, 64])]; + tensor var_217_cast_fp16 = reshape(shape = var_216, x = v_1_cast_fp16)[name = tensor("op_217_cast_fp16")]; + tensor var_218_cast_fp16 = mul(x = var_217_cast_fp16, y = onehot_3_cast_fp16)[name = tensor("op_218_cast_fp16")]; + tensor new_v_cache_1_cast_fp16 = add(x = var_211_cast_fp16, y = var_218_cast_fp16)[name = tensor("new_v_cache_1_cast_fp16")]; + tensor new_v_cache_1_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_1_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_0_perm_0 = const()[name = tensor("transpose_0_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_231 = const()[name = tensor("op_231"), val = tensor([0, 2, 1, 3])]; + tensor var_237 = const()[name = tensor("op_237"), val = tensor([1, 16, 1, 64])]; + tensor q4_1_cast_fp16 = reshape(shape = var_237, x = out_1_cast_fp16)[name = tensor("q4_1_cast_fp16")]; + tensor var_242_transpose_x_0 = const()[name = tensor("op_242_transpose_x_0"), val = tensor(false)]; + tensor var_242_transpose_y_0 = const()[name = tensor("op_242_transpose_y_0"), val = tensor(false)]; + tensor transpose_0_cast_fp16 = transpose(perm = transpose_0_perm_0, x = new_k_cache_1_cast_fp16)[name = tensor("transpose_23")]; + tensor var_242_cast_fp16 = matmul(transpose_x = var_242_transpose_x_0, transpose_y = var_242_transpose_y_0, x = q4_1_cast_fp16, y = transpose_0_cast_fp16)[name = tensor("op_242_cast_fp16")]; + tensor var_243_to_fp16 = const()[name = tensor("op_243_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_1_cast_fp16 = mul(x = var_242_cast_fp16, y = var_243_to_fp16)[name = tensor("scores_1_cast_fp16")]; + tensor var_247 = const()[name = tensor("op_247"), val = tensor([1, 1])]; + tensor var_248_cast_fp16 = reshape(shape = var_247, x = position0_to_fp16)[name = tensor("op_248_cast_fp16")]; + tensor var_249_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_248_cast_fp16)[name = tensor("op_249_cast_fp16")]; + tensor mask_1_to_fp16_dtype_0 = const()[name = tensor("mask_1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_256_to_fp16 = const()[name = tensor("op_256_to_fp16"), val = tensor(0x1p+0)]; + tensor var_249_cast_fp16_to_fp16 = cast(dtype = mask_1_to_fp16_dtype_0, x = var_249_cast_fp16)[name = tensor("cast_139")]; + tensor var_257_cast_fp16 = sub(x = var_249_cast_fp16_to_fp16, y = var_256_to_fp16)[name = tensor("op_257_cast_fp16")]; + tensor var_262 = const()[name = tensor("op_262"), val = tensor([1, 1, 1, 512])]; + tensor var_263_cast_fp16 = reshape(shape = var_262, x = var_257_cast_fp16)[name = tensor("op_263_cast_fp16")]; + tensor var_264_to_fp16 = const()[name = tensor("op_264_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_265_cast_fp16 = mul(x = var_263_cast_fp16, y = var_264_to_fp16)[name = tensor("op_265_cast_fp16")]; + tensor scores_3_cast_fp16 = add(x = scores_1_cast_fp16, y = var_265_cast_fp16)[name = tensor("scores_3_cast_fp16")]; + tensor var_268 = const()[name = tensor("op_268"), val = tensor(-1)]; + tensor weights_1_cast_fp16 = softmax(axis = var_268, x = scores_3_cast_fp16)[name = tensor("weights_1_cast_fp16")]; + tensor attn_1_transpose_x_0 = const()[name = tensor("attn_1_transpose_x_0"), val = tensor(false)]; + tensor attn_1_transpose_y_0 = const()[name = tensor("attn_1_transpose_y_0"), val = tensor(false)]; + tensor values_1_cast_fp16 = transpose(perm = var_231, x = new_v_cache_1_cast_fp16)[name = tensor("transpose_22")]; + tensor attn_1_cast_fp16 = matmul(transpose_x = attn_1_transpose_x_0, transpose_y = attn_1_transpose_y_0, x = weights_1_cast_fp16, y = values_1_cast_fp16)[name = tensor("attn_1_cast_fp16")]; + tensor var_276 = const()[name = tensor("op_276"), val = tensor([0, 2, 1, 3])]; + tensor var_281 = const()[name = tensor("op_281"), val = tensor([1, 1, 1024])]; + tensor var_277_cast_fp16 = transpose(perm = var_276, x = attn_1_cast_fp16)[name = tensor("transpose_21")]; + tensor input_7_cast_fp16 = reshape(shape = var_281, x = var_277_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor attn0_out_proj_weight_to_fp16 = const()[name = tensor("attn0_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6370944)))]; + tensor linear_2_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn0_out_proj_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor var_287_to_fp16 = const()[name = tensor("op_287_to_fp16"), val = tensor(0x1p+0)]; + tensor var_288_cast_fp16 = add(x = position0_to_fp16, y = var_287_to_fp16)[name = tensor("op_288_cast_fp16")]; + tensor var_288_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_288_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_9_cast_fp16 = add(x = linear_0_cast_fp16, y = linear_2_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor input_11_axes_0 = const()[name = tensor("input_11_axes_0"), val = tensor([-1])]; + tensor norm0_2_weight_to_fp16 = const()[name = tensor("norm0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8468160)))]; + tensor norm0_2_bias_to_fp16 = const()[name = tensor("norm0_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8470272)))]; + tensor var_292_to_fp16 = const()[name = tensor("op_292_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_11_cast_fp16 = layer_norm(axes = input_11_axes_0, beta = norm0_2_bias_to_fp16, epsilon = var_292_to_fp16, gamma = norm0_2_weight_to_fp16, x = input_9_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor linear0_1_weight_to_fp16 = const()[name = tensor("linear0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8472384)))]; + tensor linear_3_bias_0_to_fp16 = const()[name = tensor("linear_3_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16861056)))]; + tensor linear_3_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear0_1_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("linear_3_cast_fp16")]; + tensor input_13_mode_0 = const()[name = tensor("input_13_mode_0"), val = tensor("EXACT")]; + tensor input_13_cast_fp16 = gelu(mode = input_13_mode_0, x = linear_3_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor linear0_2_weight_to_fp16 = const()[name = tensor("linear0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16869312)))]; + tensor linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear0_2_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor input_15_cast_fp16 = add(x = input_9_cast_fp16, y = linear_4_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor input_17_axes_0 = const()[name = tensor("input_17_axes_0"), val = tensor([-1])]; + tensor norm1_1_weight_to_fp16 = const()[name = tensor("norm1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25257984)))]; + tensor norm1_1_bias_to_fp16 = const()[name = tensor("norm1_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25260096)))]; + tensor var_309_to_fp16 = const()[name = tensor("op_309_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_17_cast_fp16 = layer_norm(axes = input_17_axes_0, beta = norm1_1_bias_to_fp16, epsilon = var_309_to_fp16, gamma = norm1_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor attn1_in_proj_weight_to_fp16 = const()[name = tensor("attn1_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25262208)))]; + tensor linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn1_in_proj_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor var_322 = const()[name = tensor("op_322"), val = tensor([1, 3, 16, 64])]; + tensor qkv_3_cast_fp16 = reshape(shape = var_322, x = linear_5_cast_fp16)[name = tensor("qkv_3_cast_fp16")]; + tensor x_5_begin_0 = const()[name = tensor("x_5_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_5_end_0 = const()[name = tensor("x_5_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_5_end_mask_0 = const()[name = tensor("x_5_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_5_squeeze_mask_0 = const()[name = tensor("x_5_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_5_cast_fp16 = slice_by_index(begin = x_5_begin_0, end = x_5_end_0, end_mask = x_5_end_mask_0, squeeze_mask = x_5_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("x_5_cast_fp16")]; + tensor x_7_begin_0 = const()[name = tensor("x_7_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_7_end_0 = const()[name = tensor("x_7_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_7_end_mask_0 = const()[name = tensor("x_7_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_7_squeeze_mask_0 = const()[name = tensor("x_7_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_7_cast_fp16 = slice_by_index(begin = x_7_begin_0, end = x_7_end_0, end_mask = x_7_end_mask_0, squeeze_mask = x_7_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("x_7_cast_fp16")]; + tensor v_3_begin_0 = const()[name = tensor("v_3_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_3_end_0 = const()[name = tensor("v_3_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_3_end_mask_0 = const()[name = tensor("v_3_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_3_squeeze_mask_0 = const()[name = tensor("v_3_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_3_cast_fp16 = slice_by_index(begin = v_3_begin_0, end = v_3_end_0, end_mask = v_3_end_mask_0, squeeze_mask = v_3_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("v_3_cast_fp16")]; + tensor var_351 = const()[name = tensor("op_351"), val = tensor([1, 1, 1])]; + tensor position1_to_fp16_dtype_0 = const()[name = tensor("position1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position1_to_fp16 = cast(dtype = position1_to_fp16_dtype_0, x = position1)[name = tensor("cast_137")]; + tensor var_352_cast_fp16 = reshape(shape = var_351, x = position1_to_fp16)[name = tensor("op_352_cast_fp16")]; + tensor angles_5_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_352_cast_fp16)[name = tensor("angles_5_cast_fp16")]; + tensor rotr_5_cast_fp16 = cos(x = angles_5_cast_fp16)[name = tensor("rotr_5_cast_fp16")]; + tensor roti_5_cast_fp16 = sin(x = angles_5_cast_fp16)[name = tensor("roti_5_cast_fp16")]; + tensor var_360 = const()[name = tensor("op_360"), val = tensor([1, 16, 32, 2])]; + tensor pairs_5_cast_fp16 = reshape(shape = var_360, x = x_5_cast_fp16)[name = tensor("pairs_5_cast_fp16")]; + tensor xr_5_begin_0 = const()[name = tensor("xr_5_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_5_end_0 = const()[name = tensor("xr_5_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_5_end_mask_0 = const()[name = tensor("xr_5_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_5_squeeze_mask_0 = const()[name = tensor("xr_5_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_5_cast_fp16 = slice_by_index(begin = xr_5_begin_0, end = xr_5_end_0, end_mask = xr_5_end_mask_0, squeeze_mask = xr_5_squeeze_mask_0, x = pairs_5_cast_fp16)[name = tensor("xr_5_cast_fp16")]; + tensor xi_5_begin_0 = const()[name = tensor("xi_5_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_5_end_0 = const()[name = tensor("xi_5_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_5_end_mask_0 = const()[name = tensor("xi_5_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_5_squeeze_mask_0 = const()[name = tensor("xi_5_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_5_cast_fp16 = slice_by_index(begin = xi_5_begin_0, end = xi_5_end_0, end_mask = xi_5_end_mask_0, squeeze_mask = xi_5_squeeze_mask_0, x = pairs_5_cast_fp16)[name = tensor("xi_5_cast_fp16")]; + tensor var_368_cast_fp16 = mul(x = xr_5_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_368_cast_fp16")]; + tensor var_369_cast_fp16 = mul(x = xi_5_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_369_cast_fp16")]; + tensor xor__5_cast_fp16 = sub(x = var_368_cast_fp16, y = var_369_cast_fp16)[name = tensor("xor__5_cast_fp16")]; + tensor var_372_cast_fp16 = mul(x = xr_5_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_372_cast_fp16")]; + tensor var_373_cast_fp16 = mul(x = xi_5_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_373_cast_fp16")]; + tensor xoi_5_cast_fp16 = add(x = var_372_cast_fp16, y = var_373_cast_fp16)[name = tensor("xoi_5_cast_fp16")]; + tensor out_5_axis_0 = const()[name = tensor("out_5_axis_0"), val = tensor(-1)]; + tensor out_5_cast_fp16 = stack(axis = out_5_axis_0, values = (xor__5_cast_fp16, xoi_5_cast_fp16))[name = tensor("out_5_cast_fp16")]; + tensor var_396 = const()[name = tensor("op_396"), val = tensor([1, 16, 32, 2])]; + tensor pairs_7_cast_fp16 = reshape(shape = var_396, x = x_7_cast_fp16)[name = tensor("pairs_7_cast_fp16")]; + tensor xr_7_begin_0 = const()[name = tensor("xr_7_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_7_end_0 = const()[name = tensor("xr_7_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_7_end_mask_0 = const()[name = tensor("xr_7_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_7_squeeze_mask_0 = const()[name = tensor("xr_7_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_7_cast_fp16 = slice_by_index(begin = xr_7_begin_0, end = xr_7_end_0, end_mask = xr_7_end_mask_0, squeeze_mask = xr_7_squeeze_mask_0, x = pairs_7_cast_fp16)[name = tensor("xr_7_cast_fp16")]; + tensor xi_7_begin_0 = const()[name = tensor("xi_7_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_7_end_0 = const()[name = tensor("xi_7_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_7_end_mask_0 = const()[name = tensor("xi_7_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_7_squeeze_mask_0 = const()[name = tensor("xi_7_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_7_cast_fp16 = slice_by_index(begin = xi_7_begin_0, end = xi_7_end_0, end_mask = xi_7_end_mask_0, squeeze_mask = xi_7_squeeze_mask_0, x = pairs_7_cast_fp16)[name = tensor("xi_7_cast_fp16")]; + tensor var_404_cast_fp16 = mul(x = xr_7_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_404_cast_fp16")]; + tensor var_405_cast_fp16 = mul(x = xi_7_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_405_cast_fp16")]; + tensor xor__7_cast_fp16 = sub(x = var_404_cast_fp16, y = var_405_cast_fp16)[name = tensor("xor__7_cast_fp16")]; + tensor var_408_cast_fp16 = mul(x = xr_7_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_408_cast_fp16")]; + tensor var_409_cast_fp16 = mul(x = xi_7_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_409_cast_fp16")]; + tensor xoi_7_cast_fp16 = add(x = var_408_cast_fp16, y = var_409_cast_fp16)[name = tensor("xoi_7_cast_fp16")]; + tensor out_7_axis_0 = const()[name = tensor("out_7_axis_0"), val = tensor(-1)]; + tensor out_7_cast_fp16 = stack(axis = out_7_axis_0, values = (xor__7_cast_fp16, xoi_7_cast_fp16))[name = tensor("out_7_cast_fp16")]; + tensor _inversed_421_y_0_to_fp16 = const()[name = tensor("_inversed_421_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_421_cast_fp16 = mul(x = position1_to_fp16, y = _inversed_421_y_0_to_fp16)[name = tensor("_inversed_421_cast_fp16")]; + tensor var_422_cast_fp16 = floor(x = _inversed_421_cast_fp16)[name = tensor("op_422_cast_fp16")]; + tensor var_423_to_fp16 = const()[name = tensor("op_423_to_fp16"), val = tensor(0x1p+9)]; + tensor var_424_cast_fp16 = mul(x = var_422_cast_fp16, y = var_423_to_fp16)[name = tensor("op_424_cast_fp16")]; + tensor write_idx_3_cast_fp16 = sub(x = position1_to_fp16, y = var_424_cast_fp16)[name = tensor("write_idx_3_cast_fp16")]; + tensor var_427_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_3_cast_fp16)[name = tensor("op_427_cast_fp16")]; + tensor var_437 = const()[name = tensor("op_437"), val = tensor([1, 512, 1, 1])]; + tensor onehot_5_to_fp16_dtype_0 = const()[name = tensor("onehot_5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_427_cast_fp16_to_fp16 = cast(dtype = onehot_5_to_fp16_dtype_0, x = var_427_cast_fp16)[name = tensor("cast_136")]; + tensor onehot_7_cast_fp16 = reshape(shape = var_437, x = var_427_cast_fp16_to_fp16)[name = tensor("onehot_7_cast_fp16")]; + tensor var_439_to_fp16 = const()[name = tensor("op_439_to_fp16"), val = tensor(0x1p+0)]; + tensor var_441_cast_fp16 = sub(x = var_439_to_fp16, y = onehot_7_cast_fp16)[name = tensor("op_441_cast_fp16")]; + tensor k_cache1_to_fp16_dtype_0 = const()[name = tensor("k_cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache1_to_fp16 = cast(dtype = k_cache1_to_fp16_dtype_0, x = k_cache1)[name = tensor("cast_135")]; + tensor var_442_cast_fp16 = mul(x = k_cache1_to_fp16, y = var_441_cast_fp16)[name = tensor("op_442_cast_fp16")]; + tensor var_447 = const()[name = tensor("op_447"), val = tensor([1, 1, 16, 64])]; + tensor var_448_cast_fp16 = reshape(shape = var_447, x = out_7_cast_fp16)[name = tensor("op_448_cast_fp16")]; + tensor var_449_cast_fp16 = mul(x = var_448_cast_fp16, y = onehot_7_cast_fp16)[name = tensor("op_449_cast_fp16")]; + tensor new_k_cache_3_cast_fp16 = add(x = var_442_cast_fp16, y = var_449_cast_fp16)[name = tensor("new_k_cache_3_cast_fp16")]; + tensor new_k_cache_3_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_3_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache1_to_fp16_dtype_0 = const()[name = tensor("v_cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache1_to_fp16 = cast(dtype = v_cache1_to_fp16_dtype_0, x = v_cache1)[name = tensor("cast_133")]; + tensor var_455_cast_fp16 = mul(x = v_cache1_to_fp16, y = var_441_cast_fp16)[name = tensor("op_455_cast_fp16")]; + tensor var_460 = const()[name = tensor("op_460"), val = tensor([1, 1, 16, 64])]; + tensor var_461_cast_fp16 = reshape(shape = var_460, x = v_3_cast_fp16)[name = tensor("op_461_cast_fp16")]; + tensor var_462_cast_fp16 = mul(x = var_461_cast_fp16, y = onehot_7_cast_fp16)[name = tensor("op_462_cast_fp16")]; + tensor new_v_cache_3_cast_fp16 = add(x = var_455_cast_fp16, y = var_462_cast_fp16)[name = tensor("new_v_cache_3_cast_fp16")]; + tensor new_v_cache_3_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_3_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_1_perm_0 = const()[name = tensor("transpose_1_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_475 = const()[name = tensor("op_475"), val = tensor([0, 2, 1, 3])]; + tensor var_481 = const()[name = tensor("op_481"), val = tensor([1, 16, 1, 64])]; + tensor q4_3_cast_fp16 = reshape(shape = var_481, x = out_5_cast_fp16)[name = tensor("q4_3_cast_fp16")]; + tensor var_486_transpose_x_0 = const()[name = tensor("op_486_transpose_x_0"), val = tensor(false)]; + tensor var_486_transpose_y_0 = const()[name = tensor("op_486_transpose_y_0"), val = tensor(false)]; + tensor transpose_1_cast_fp16 = transpose(perm = transpose_1_perm_0, x = new_k_cache_3_cast_fp16)[name = tensor("transpose_20")]; + tensor var_486_cast_fp16 = matmul(transpose_x = var_486_transpose_x_0, transpose_y = var_486_transpose_y_0, x = q4_3_cast_fp16, y = transpose_1_cast_fp16)[name = tensor("op_486_cast_fp16")]; + tensor var_487_to_fp16 = const()[name = tensor("op_487_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_5_cast_fp16 = mul(x = var_486_cast_fp16, y = var_487_to_fp16)[name = tensor("scores_5_cast_fp16")]; + tensor var_491 = const()[name = tensor("op_491"), val = tensor([1, 1])]; + tensor var_492_cast_fp16 = reshape(shape = var_491, x = position1_to_fp16)[name = tensor("op_492_cast_fp16")]; + tensor var_493_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_492_cast_fp16)[name = tensor("op_493_cast_fp16")]; + tensor mask_3_to_fp16_dtype_0 = const()[name = tensor("mask_3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_500_to_fp16 = const()[name = tensor("op_500_to_fp16"), val = tensor(0x1p+0)]; + tensor var_493_cast_fp16_to_fp16 = cast(dtype = mask_3_to_fp16_dtype_0, x = var_493_cast_fp16)[name = tensor("cast_131")]; + tensor var_501_cast_fp16 = sub(x = var_493_cast_fp16_to_fp16, y = var_500_to_fp16)[name = tensor("op_501_cast_fp16")]; + tensor var_506 = const()[name = tensor("op_506"), val = tensor([1, 1, 1, 512])]; + tensor var_507_cast_fp16 = reshape(shape = var_506, x = var_501_cast_fp16)[name = tensor("op_507_cast_fp16")]; + tensor var_508_to_fp16 = const()[name = tensor("op_508_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_509_cast_fp16 = mul(x = var_507_cast_fp16, y = var_508_to_fp16)[name = tensor("op_509_cast_fp16")]; + tensor scores_7_cast_fp16 = add(x = scores_5_cast_fp16, y = var_509_cast_fp16)[name = tensor("scores_7_cast_fp16")]; + tensor var_512 = const()[name = tensor("op_512"), val = tensor(-1)]; + tensor weights_3_cast_fp16 = softmax(axis = var_512, x = scores_7_cast_fp16)[name = tensor("weights_3_cast_fp16")]; + tensor attn_3_transpose_x_0 = const()[name = tensor("attn_3_transpose_x_0"), val = tensor(false)]; + tensor attn_3_transpose_y_0 = const()[name = tensor("attn_3_transpose_y_0"), val = tensor(false)]; + tensor values_3_cast_fp16 = transpose(perm = var_475, x = new_v_cache_3_cast_fp16)[name = tensor("transpose_19")]; + tensor attn_3_cast_fp16 = matmul(transpose_x = attn_3_transpose_x_0, transpose_y = attn_3_transpose_y_0, x = weights_3_cast_fp16, y = values_3_cast_fp16)[name = tensor("attn_3_cast_fp16")]; + tensor var_520 = const()[name = tensor("op_520"), val = tensor([0, 2, 1, 3])]; + tensor var_525 = const()[name = tensor("op_525"), val = tensor([1, 1, 1024])]; + tensor var_521_cast_fp16 = transpose(perm = var_520, x = attn_3_cast_fp16)[name = tensor("transpose_18")]; + tensor input_19_cast_fp16 = reshape(shape = var_525, x = var_521_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor attn1_out_proj_weight_to_fp16 = const()[name = tensor("attn1_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31553728)))]; + tensor linear_6_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn1_out_proj_weight_to_fp16, x = input_19_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor var_531_to_fp16 = const()[name = tensor("op_531_to_fp16"), val = tensor(0x1p+0)]; + tensor var_532_cast_fp16 = add(x = position1_to_fp16, y = var_531_to_fp16)[name = tensor("op_532_cast_fp16")]; + tensor var_532_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_532_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_21_cast_fp16 = add(x = input_15_cast_fp16, y = linear_6_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor input_23_axes_0 = const()[name = tensor("input_23_axes_0"), val = tensor([-1])]; + tensor norm1_2_weight_to_fp16 = const()[name = tensor("norm1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33650944)))]; + tensor norm1_2_bias_to_fp16 = const()[name = tensor("norm1_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33653056)))]; + tensor var_536_to_fp16 = const()[name = tensor("op_536_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_23_cast_fp16 = layer_norm(axes = input_23_axes_0, beta = norm1_2_bias_to_fp16, epsilon = var_536_to_fp16, gamma = norm1_2_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor linear1_1_weight_to_fp16 = const()[name = tensor("linear1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33655168)))]; + tensor linear_7_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear1_1_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_25_mode_0 = const()[name = tensor("input_25_mode_0"), val = tensor("EXACT")]; + tensor input_25_cast_fp16 = gelu(mode = input_25_mode_0, x = linear_7_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor linear1_2_weight_to_fp16 = const()[name = tensor("linear1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42043840)))]; + tensor linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear1_2_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor input_27_cast_fp16 = add(x = input_21_cast_fp16, y = linear_8_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor input_29_axes_0 = const()[name = tensor("input_29_axes_0"), val = tensor([-1])]; + tensor norm2_1_weight_to_fp16 = const()[name = tensor("norm2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50432512)))]; + tensor norm2_1_bias_to_fp16 = const()[name = tensor("norm2_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50434624)))]; + tensor var_553_to_fp16 = const()[name = tensor("op_553_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_29_cast_fp16 = layer_norm(axes = input_29_axes_0, beta = norm2_1_bias_to_fp16, epsilon = var_553_to_fp16, gamma = norm2_1_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor attn2_in_proj_weight_to_fp16 = const()[name = tensor("attn2_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50436736)))]; + tensor linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn2_in_proj_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor var_566 = const()[name = tensor("op_566"), val = tensor([1, 3, 16, 64])]; + tensor qkv_5_cast_fp16 = reshape(shape = var_566, x = linear_9_cast_fp16)[name = tensor("qkv_5_cast_fp16")]; + tensor x_9_begin_0 = const()[name = tensor("x_9_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_9_end_0 = const()[name = tensor("x_9_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_9_end_mask_0 = const()[name = tensor("x_9_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_9_squeeze_mask_0 = const()[name = tensor("x_9_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_9_cast_fp16 = slice_by_index(begin = x_9_begin_0, end = x_9_end_0, end_mask = x_9_end_mask_0, squeeze_mask = x_9_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("x_9_cast_fp16")]; + tensor x_11_begin_0 = const()[name = tensor("x_11_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_11_end_0 = const()[name = tensor("x_11_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_11_end_mask_0 = const()[name = tensor("x_11_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_11_squeeze_mask_0 = const()[name = tensor("x_11_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_11_cast_fp16 = slice_by_index(begin = x_11_begin_0, end = x_11_end_0, end_mask = x_11_end_mask_0, squeeze_mask = x_11_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("x_11_cast_fp16")]; + tensor v_5_begin_0 = const()[name = tensor("v_5_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_5_end_0 = const()[name = tensor("v_5_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_5_end_mask_0 = const()[name = tensor("v_5_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_5_squeeze_mask_0 = const()[name = tensor("v_5_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_5_cast_fp16 = slice_by_index(begin = v_5_begin_0, end = v_5_end_0, end_mask = v_5_end_mask_0, squeeze_mask = v_5_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("v_5_cast_fp16")]; + tensor var_595 = const()[name = tensor("op_595"), val = tensor([1, 1, 1])]; + tensor position2_to_fp16_dtype_0 = const()[name = tensor("position2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position2_to_fp16 = cast(dtype = position2_to_fp16_dtype_0, x = position2)[name = tensor("cast_129")]; + tensor var_596_cast_fp16 = reshape(shape = var_595, x = position2_to_fp16)[name = tensor("op_596_cast_fp16")]; + tensor angles_9_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_596_cast_fp16)[name = tensor("angles_9_cast_fp16")]; + tensor rotr_9_cast_fp16 = cos(x = angles_9_cast_fp16)[name = tensor("rotr_9_cast_fp16")]; + tensor roti_9_cast_fp16 = sin(x = angles_9_cast_fp16)[name = tensor("roti_9_cast_fp16")]; + tensor var_604 = const()[name = tensor("op_604"), val = tensor([1, 16, 32, 2])]; + tensor pairs_9_cast_fp16 = reshape(shape = var_604, x = x_9_cast_fp16)[name = tensor("pairs_9_cast_fp16")]; + tensor xr_9_begin_0 = const()[name = tensor("xr_9_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_9_end_0 = const()[name = tensor("xr_9_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_9_end_mask_0 = const()[name = tensor("xr_9_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_9_squeeze_mask_0 = const()[name = tensor("xr_9_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_9_cast_fp16 = slice_by_index(begin = xr_9_begin_0, end = xr_9_end_0, end_mask = xr_9_end_mask_0, squeeze_mask = xr_9_squeeze_mask_0, x = pairs_9_cast_fp16)[name = tensor("xr_9_cast_fp16")]; + tensor xi_9_begin_0 = const()[name = tensor("xi_9_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_9_end_0 = const()[name = tensor("xi_9_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_9_end_mask_0 = const()[name = tensor("xi_9_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_9_squeeze_mask_0 = const()[name = tensor("xi_9_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_9_cast_fp16 = slice_by_index(begin = xi_9_begin_0, end = xi_9_end_0, end_mask = xi_9_end_mask_0, squeeze_mask = xi_9_squeeze_mask_0, x = pairs_9_cast_fp16)[name = tensor("xi_9_cast_fp16")]; + tensor var_612_cast_fp16 = mul(x = xr_9_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_612_cast_fp16")]; + tensor var_613_cast_fp16 = mul(x = xi_9_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_613_cast_fp16")]; + tensor xor__9_cast_fp16 = sub(x = var_612_cast_fp16, y = var_613_cast_fp16)[name = tensor("xor__9_cast_fp16")]; + tensor var_616_cast_fp16 = mul(x = xr_9_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_616_cast_fp16")]; + tensor var_617_cast_fp16 = mul(x = xi_9_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_617_cast_fp16")]; + tensor xoi_9_cast_fp16 = add(x = var_616_cast_fp16, y = var_617_cast_fp16)[name = tensor("xoi_9_cast_fp16")]; + tensor out_9_axis_0 = const()[name = tensor("out_9_axis_0"), val = tensor(-1)]; + tensor out_9_cast_fp16 = stack(axis = out_9_axis_0, values = (xor__9_cast_fp16, xoi_9_cast_fp16))[name = tensor("out_9_cast_fp16")]; + tensor var_640 = const()[name = tensor("op_640"), val = tensor([1, 16, 32, 2])]; + tensor pairs_11_cast_fp16 = reshape(shape = var_640, x = x_11_cast_fp16)[name = tensor("pairs_11_cast_fp16")]; + tensor xr_11_begin_0 = const()[name = tensor("xr_11_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_11_end_0 = const()[name = tensor("xr_11_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_11_end_mask_0 = const()[name = tensor("xr_11_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_11_squeeze_mask_0 = const()[name = tensor("xr_11_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_11_cast_fp16 = slice_by_index(begin = xr_11_begin_0, end = xr_11_end_0, end_mask = xr_11_end_mask_0, squeeze_mask = xr_11_squeeze_mask_0, x = pairs_11_cast_fp16)[name = tensor("xr_11_cast_fp16")]; + tensor xi_11_begin_0 = const()[name = tensor("xi_11_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_11_end_0 = const()[name = tensor("xi_11_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_11_end_mask_0 = const()[name = tensor("xi_11_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_11_squeeze_mask_0 = const()[name = tensor("xi_11_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_11_cast_fp16 = slice_by_index(begin = xi_11_begin_0, end = xi_11_end_0, end_mask = xi_11_end_mask_0, squeeze_mask = xi_11_squeeze_mask_0, x = pairs_11_cast_fp16)[name = tensor("xi_11_cast_fp16")]; + tensor var_648_cast_fp16 = mul(x = xr_11_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_648_cast_fp16")]; + tensor var_649_cast_fp16 = mul(x = xi_11_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_649_cast_fp16")]; + tensor xor__11_cast_fp16 = sub(x = var_648_cast_fp16, y = var_649_cast_fp16)[name = tensor("xor__11_cast_fp16")]; + tensor var_652_cast_fp16 = mul(x = xr_11_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_652_cast_fp16")]; + tensor var_653_cast_fp16 = mul(x = xi_11_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_653_cast_fp16")]; + tensor xoi_11_cast_fp16 = add(x = var_652_cast_fp16, y = var_653_cast_fp16)[name = tensor("xoi_11_cast_fp16")]; + tensor out_11_axis_0 = const()[name = tensor("out_11_axis_0"), val = tensor(-1)]; + tensor out_11_cast_fp16 = stack(axis = out_11_axis_0, values = (xor__11_cast_fp16, xoi_11_cast_fp16))[name = tensor("out_11_cast_fp16")]; + tensor _inversed_665_y_0_to_fp16 = const()[name = tensor("_inversed_665_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_665_cast_fp16 = mul(x = position2_to_fp16, y = _inversed_665_y_0_to_fp16)[name = tensor("_inversed_665_cast_fp16")]; + tensor var_666_cast_fp16 = floor(x = _inversed_665_cast_fp16)[name = tensor("op_666_cast_fp16")]; + tensor var_667_to_fp16 = const()[name = tensor("op_667_to_fp16"), val = tensor(0x1p+9)]; + tensor var_668_cast_fp16 = mul(x = var_666_cast_fp16, y = var_667_to_fp16)[name = tensor("op_668_cast_fp16")]; + tensor write_idx_5_cast_fp16 = sub(x = position2_to_fp16, y = var_668_cast_fp16)[name = tensor("write_idx_5_cast_fp16")]; + tensor var_671_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_5_cast_fp16)[name = tensor("op_671_cast_fp16")]; + tensor var_681 = const()[name = tensor("op_681"), val = tensor([1, 512, 1, 1])]; + tensor onehot_9_to_fp16_dtype_0 = const()[name = tensor("onehot_9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_671_cast_fp16_to_fp16 = cast(dtype = onehot_9_to_fp16_dtype_0, x = var_671_cast_fp16)[name = tensor("cast_128")]; + tensor onehot_11_cast_fp16 = reshape(shape = var_681, x = var_671_cast_fp16_to_fp16)[name = tensor("onehot_11_cast_fp16")]; + tensor var_683_to_fp16 = const()[name = tensor("op_683_to_fp16"), val = tensor(0x1p+0)]; + tensor var_685_cast_fp16 = sub(x = var_683_to_fp16, y = onehot_11_cast_fp16)[name = tensor("op_685_cast_fp16")]; + tensor k_cache2_to_fp16_dtype_0 = const()[name = tensor("k_cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache2_to_fp16 = cast(dtype = k_cache2_to_fp16_dtype_0, x = k_cache2)[name = tensor("cast_127")]; + tensor var_686_cast_fp16 = mul(x = k_cache2_to_fp16, y = var_685_cast_fp16)[name = tensor("op_686_cast_fp16")]; + tensor var_691 = const()[name = tensor("op_691"), val = tensor([1, 1, 16, 64])]; + tensor var_692_cast_fp16 = reshape(shape = var_691, x = out_11_cast_fp16)[name = tensor("op_692_cast_fp16")]; + tensor var_693_cast_fp16 = mul(x = var_692_cast_fp16, y = onehot_11_cast_fp16)[name = tensor("op_693_cast_fp16")]; + tensor new_k_cache_5_cast_fp16 = add(x = var_686_cast_fp16, y = var_693_cast_fp16)[name = tensor("new_k_cache_5_cast_fp16")]; + tensor new_k_cache_5_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_5_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache2_to_fp16_dtype_0 = const()[name = tensor("v_cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache2_to_fp16 = cast(dtype = v_cache2_to_fp16_dtype_0, x = v_cache2)[name = tensor("cast_125")]; + tensor var_699_cast_fp16 = mul(x = v_cache2_to_fp16, y = var_685_cast_fp16)[name = tensor("op_699_cast_fp16")]; + tensor var_704 = const()[name = tensor("op_704"), val = tensor([1, 1, 16, 64])]; + tensor var_705_cast_fp16 = reshape(shape = var_704, x = v_5_cast_fp16)[name = tensor("op_705_cast_fp16")]; + tensor var_706_cast_fp16 = mul(x = var_705_cast_fp16, y = onehot_11_cast_fp16)[name = tensor("op_706_cast_fp16")]; + tensor new_v_cache_5_cast_fp16 = add(x = var_699_cast_fp16, y = var_706_cast_fp16)[name = tensor("new_v_cache_5_cast_fp16")]; + tensor new_v_cache_5_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_5_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_2_perm_0 = const()[name = tensor("transpose_2_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_719 = const()[name = tensor("op_719"), val = tensor([0, 2, 1, 3])]; + tensor var_725 = const()[name = tensor("op_725"), val = tensor([1, 16, 1, 64])]; + tensor q4_5_cast_fp16 = reshape(shape = var_725, x = out_9_cast_fp16)[name = tensor("q4_5_cast_fp16")]; + tensor var_730_transpose_x_0 = const()[name = tensor("op_730_transpose_x_0"), val = tensor(false)]; + tensor var_730_transpose_y_0 = const()[name = tensor("op_730_transpose_y_0"), val = tensor(false)]; + tensor transpose_2_cast_fp16 = transpose(perm = transpose_2_perm_0, x = new_k_cache_5_cast_fp16)[name = tensor("transpose_17")]; + tensor var_730_cast_fp16 = matmul(transpose_x = var_730_transpose_x_0, transpose_y = var_730_transpose_y_0, x = q4_5_cast_fp16, y = transpose_2_cast_fp16)[name = tensor("op_730_cast_fp16")]; + tensor var_731_to_fp16 = const()[name = tensor("op_731_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_9_cast_fp16 = mul(x = var_730_cast_fp16, y = var_731_to_fp16)[name = tensor("scores_9_cast_fp16")]; + tensor var_735 = const()[name = tensor("op_735"), val = tensor([1, 1])]; + tensor var_736_cast_fp16 = reshape(shape = var_735, x = position2_to_fp16)[name = tensor("op_736_cast_fp16")]; + tensor var_737_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_736_cast_fp16)[name = tensor("op_737_cast_fp16")]; + tensor mask_5_to_fp16_dtype_0 = const()[name = tensor("mask_5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_744_to_fp16 = const()[name = tensor("op_744_to_fp16"), val = tensor(0x1p+0)]; + tensor var_737_cast_fp16_to_fp16 = cast(dtype = mask_5_to_fp16_dtype_0, x = var_737_cast_fp16)[name = tensor("cast_123")]; + tensor var_745_cast_fp16 = sub(x = var_737_cast_fp16_to_fp16, y = var_744_to_fp16)[name = tensor("op_745_cast_fp16")]; + tensor var_750 = const()[name = tensor("op_750"), val = tensor([1, 1, 1, 512])]; + tensor var_751_cast_fp16 = reshape(shape = var_750, x = var_745_cast_fp16)[name = tensor("op_751_cast_fp16")]; + tensor var_752_to_fp16 = const()[name = tensor("op_752_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_753_cast_fp16 = mul(x = var_751_cast_fp16, y = var_752_to_fp16)[name = tensor("op_753_cast_fp16")]; + tensor scores_11_cast_fp16 = add(x = scores_9_cast_fp16, y = var_753_cast_fp16)[name = tensor("scores_11_cast_fp16")]; + tensor var_756 = const()[name = tensor("op_756"), val = tensor(-1)]; + tensor weights_5_cast_fp16 = softmax(axis = var_756, x = scores_11_cast_fp16)[name = tensor("weights_5_cast_fp16")]; + tensor attn_5_transpose_x_0 = const()[name = tensor("attn_5_transpose_x_0"), val = tensor(false)]; + tensor attn_5_transpose_y_0 = const()[name = tensor("attn_5_transpose_y_0"), val = tensor(false)]; + tensor values_5_cast_fp16 = transpose(perm = var_719, x = new_v_cache_5_cast_fp16)[name = tensor("transpose_16")]; + tensor attn_5_cast_fp16 = matmul(transpose_x = attn_5_transpose_x_0, transpose_y = attn_5_transpose_y_0, x = weights_5_cast_fp16, y = values_5_cast_fp16)[name = tensor("attn_5_cast_fp16")]; + tensor var_764 = const()[name = tensor("op_764"), val = tensor([0, 2, 1, 3])]; + tensor var_769 = const()[name = tensor("op_769"), val = tensor([1, 1, 1024])]; + tensor var_765_cast_fp16 = transpose(perm = var_764, x = attn_5_cast_fp16)[name = tensor("transpose_15")]; + tensor input_31_cast_fp16 = reshape(shape = var_769, x = var_765_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor attn2_out_proj_weight_to_fp16 = const()[name = tensor("attn2_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(56728256)))]; + tensor linear_10_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn2_out_proj_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor var_775_to_fp16 = const()[name = tensor("op_775_to_fp16"), val = tensor(0x1p+0)]; + tensor var_776_cast_fp16 = add(x = position2_to_fp16, y = var_775_to_fp16)[name = tensor("op_776_cast_fp16")]; + tensor var_776_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_776_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_33_cast_fp16 = add(x = input_27_cast_fp16, y = linear_10_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor input_35_axes_0 = const()[name = tensor("input_35_axes_0"), val = tensor([-1])]; + tensor norm2_2_weight_to_fp16 = const()[name = tensor("norm2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58825472)))]; + tensor norm2_2_bias_to_fp16 = const()[name = tensor("norm2_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58827584)))]; + tensor var_780_to_fp16 = const()[name = tensor("op_780_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_35_cast_fp16 = layer_norm(axes = input_35_axes_0, beta = norm2_2_bias_to_fp16, epsilon = var_780_to_fp16, gamma = norm2_2_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor linear2_1_weight_to_fp16 = const()[name = tensor("linear2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58829696)))]; + tensor linear_11_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear2_1_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor input_37_mode_0 = const()[name = tensor("input_37_mode_0"), val = tensor("EXACT")]; + tensor input_37_cast_fp16 = gelu(mode = input_37_mode_0, x = linear_11_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor linear2_2_weight_to_fp16 = const()[name = tensor("linear2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67218368)))]; + tensor linear_12_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear2_2_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor input_39_cast_fp16 = add(x = input_33_cast_fp16, y = linear_12_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor input_41_axes_0 = const()[name = tensor("input_41_axes_0"), val = tensor([-1])]; + tensor norm3_1_weight_to_fp16 = const()[name = tensor("norm3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75607040)))]; + tensor norm3_1_bias_to_fp16 = const()[name = tensor("norm3_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75609152)))]; + tensor var_797_to_fp16 = const()[name = tensor("op_797_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_41_cast_fp16 = layer_norm(axes = input_41_axes_0, beta = norm3_1_bias_to_fp16, epsilon = var_797_to_fp16, gamma = norm3_1_weight_to_fp16, x = input_39_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor attn3_in_proj_weight_to_fp16 = const()[name = tensor("attn3_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75611264)))]; + tensor linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn3_in_proj_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor var_810 = const()[name = tensor("op_810"), val = tensor([1, 3, 16, 64])]; + tensor qkv_7_cast_fp16 = reshape(shape = var_810, x = linear_13_cast_fp16)[name = tensor("qkv_7_cast_fp16")]; + tensor x_13_begin_0 = const()[name = tensor("x_13_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_13_end_0 = const()[name = tensor("x_13_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_13_end_mask_0 = const()[name = tensor("x_13_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_13_squeeze_mask_0 = const()[name = tensor("x_13_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_13_cast_fp16 = slice_by_index(begin = x_13_begin_0, end = x_13_end_0, end_mask = x_13_end_mask_0, squeeze_mask = x_13_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("x_13_cast_fp16")]; + tensor x_15_begin_0 = const()[name = tensor("x_15_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_15_end_0 = const()[name = tensor("x_15_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_15_end_mask_0 = const()[name = tensor("x_15_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_15_squeeze_mask_0 = const()[name = tensor("x_15_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_15_cast_fp16 = slice_by_index(begin = x_15_begin_0, end = x_15_end_0, end_mask = x_15_end_mask_0, squeeze_mask = x_15_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("x_15_cast_fp16")]; + tensor v_7_begin_0 = const()[name = tensor("v_7_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_7_end_0 = const()[name = tensor("v_7_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_7_end_mask_0 = const()[name = tensor("v_7_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_7_squeeze_mask_0 = const()[name = tensor("v_7_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_7_cast_fp16 = slice_by_index(begin = v_7_begin_0, end = v_7_end_0, end_mask = v_7_end_mask_0, squeeze_mask = v_7_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("v_7_cast_fp16")]; + tensor var_839 = const()[name = tensor("op_839"), val = tensor([1, 1, 1])]; + tensor position3_to_fp16_dtype_0 = const()[name = tensor("position3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position3_to_fp16 = cast(dtype = position3_to_fp16_dtype_0, x = position3)[name = tensor("cast_121")]; + tensor var_840_cast_fp16 = reshape(shape = var_839, x = position3_to_fp16)[name = tensor("op_840_cast_fp16")]; + tensor angles_13_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_840_cast_fp16)[name = tensor("angles_13_cast_fp16")]; + tensor rotr_13_cast_fp16 = cos(x = angles_13_cast_fp16)[name = tensor("rotr_13_cast_fp16")]; + tensor roti_13_cast_fp16 = sin(x = angles_13_cast_fp16)[name = tensor("roti_13_cast_fp16")]; + tensor var_848 = const()[name = tensor("op_848"), val = tensor([1, 16, 32, 2])]; + tensor pairs_13_cast_fp16 = reshape(shape = var_848, x = x_13_cast_fp16)[name = tensor("pairs_13_cast_fp16")]; + tensor xr_13_begin_0 = const()[name = tensor("xr_13_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_13_end_0 = const()[name = tensor("xr_13_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_13_end_mask_0 = const()[name = tensor("xr_13_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_13_squeeze_mask_0 = const()[name = tensor("xr_13_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_13_cast_fp16 = slice_by_index(begin = xr_13_begin_0, end = xr_13_end_0, end_mask = xr_13_end_mask_0, squeeze_mask = xr_13_squeeze_mask_0, x = pairs_13_cast_fp16)[name = tensor("xr_13_cast_fp16")]; + tensor xi_13_begin_0 = const()[name = tensor("xi_13_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_13_end_0 = const()[name = tensor("xi_13_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_13_end_mask_0 = const()[name = tensor("xi_13_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_13_squeeze_mask_0 = const()[name = tensor("xi_13_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_13_cast_fp16 = slice_by_index(begin = xi_13_begin_0, end = xi_13_end_0, end_mask = xi_13_end_mask_0, squeeze_mask = xi_13_squeeze_mask_0, x = pairs_13_cast_fp16)[name = tensor("xi_13_cast_fp16")]; + tensor var_856_cast_fp16 = mul(x = xr_13_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_856_cast_fp16")]; + tensor var_857_cast_fp16 = mul(x = xi_13_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_857_cast_fp16")]; + tensor xor__13_cast_fp16 = sub(x = var_856_cast_fp16, y = var_857_cast_fp16)[name = tensor("xor__13_cast_fp16")]; + tensor var_860_cast_fp16 = mul(x = xr_13_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_860_cast_fp16")]; + tensor var_861_cast_fp16 = mul(x = xi_13_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_861_cast_fp16")]; + tensor xoi_13_cast_fp16 = add(x = var_860_cast_fp16, y = var_861_cast_fp16)[name = tensor("xoi_13_cast_fp16")]; + tensor out_13_axis_0 = const()[name = tensor("out_13_axis_0"), val = tensor(-1)]; + tensor out_13_cast_fp16 = stack(axis = out_13_axis_0, values = (xor__13_cast_fp16, xoi_13_cast_fp16))[name = tensor("out_13_cast_fp16")]; + tensor var_884 = const()[name = tensor("op_884"), val = tensor([1, 16, 32, 2])]; + tensor pairs_15_cast_fp16 = reshape(shape = var_884, x = x_15_cast_fp16)[name = tensor("pairs_15_cast_fp16")]; + tensor xr_15_begin_0 = const()[name = tensor("xr_15_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_15_end_0 = const()[name = tensor("xr_15_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_15_end_mask_0 = const()[name = tensor("xr_15_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_15_squeeze_mask_0 = const()[name = tensor("xr_15_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_15_cast_fp16 = slice_by_index(begin = xr_15_begin_0, end = xr_15_end_0, end_mask = xr_15_end_mask_0, squeeze_mask = xr_15_squeeze_mask_0, x = pairs_15_cast_fp16)[name = tensor("xr_15_cast_fp16")]; + tensor xi_15_begin_0 = const()[name = tensor("xi_15_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_15_end_0 = const()[name = tensor("xi_15_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_15_end_mask_0 = const()[name = tensor("xi_15_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_15_squeeze_mask_0 = const()[name = tensor("xi_15_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_15_cast_fp16 = slice_by_index(begin = xi_15_begin_0, end = xi_15_end_0, end_mask = xi_15_end_mask_0, squeeze_mask = xi_15_squeeze_mask_0, x = pairs_15_cast_fp16)[name = tensor("xi_15_cast_fp16")]; + tensor var_892_cast_fp16 = mul(x = xr_15_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_892_cast_fp16")]; + tensor var_893_cast_fp16 = mul(x = xi_15_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_893_cast_fp16")]; + tensor xor__15_cast_fp16 = sub(x = var_892_cast_fp16, y = var_893_cast_fp16)[name = tensor("xor__15_cast_fp16")]; + tensor var_896_cast_fp16 = mul(x = xr_15_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_896_cast_fp16")]; + tensor var_897_cast_fp16 = mul(x = xi_15_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_897_cast_fp16")]; + tensor xoi_15_cast_fp16 = add(x = var_896_cast_fp16, y = var_897_cast_fp16)[name = tensor("xoi_15_cast_fp16")]; + tensor out_15_axis_0 = const()[name = tensor("out_15_axis_0"), val = tensor(-1)]; + tensor out_15_cast_fp16 = stack(axis = out_15_axis_0, values = (xor__15_cast_fp16, xoi_15_cast_fp16))[name = tensor("out_15_cast_fp16")]; + tensor _inversed_909_y_0_to_fp16 = const()[name = tensor("_inversed_909_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_909_cast_fp16 = mul(x = position3_to_fp16, y = _inversed_909_y_0_to_fp16)[name = tensor("_inversed_909_cast_fp16")]; + tensor var_910_cast_fp16 = floor(x = _inversed_909_cast_fp16)[name = tensor("op_910_cast_fp16")]; + tensor var_911_to_fp16 = const()[name = tensor("op_911_to_fp16"), val = tensor(0x1p+9)]; + tensor var_912_cast_fp16 = mul(x = var_910_cast_fp16, y = var_911_to_fp16)[name = tensor("op_912_cast_fp16")]; + tensor write_idx_7_cast_fp16 = sub(x = position3_to_fp16, y = var_912_cast_fp16)[name = tensor("write_idx_7_cast_fp16")]; + tensor var_915_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_7_cast_fp16)[name = tensor("op_915_cast_fp16")]; + tensor var_925 = const()[name = tensor("op_925"), val = tensor([1, 512, 1, 1])]; + tensor onehot_13_to_fp16_dtype_0 = const()[name = tensor("onehot_13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_915_cast_fp16_to_fp16 = cast(dtype = onehot_13_to_fp16_dtype_0, x = var_915_cast_fp16)[name = tensor("cast_120")]; + tensor onehot_15_cast_fp16 = reshape(shape = var_925, x = var_915_cast_fp16_to_fp16)[name = tensor("onehot_15_cast_fp16")]; + tensor var_927_to_fp16 = const()[name = tensor("op_927_to_fp16"), val = tensor(0x1p+0)]; + tensor var_929_cast_fp16 = sub(x = var_927_to_fp16, y = onehot_15_cast_fp16)[name = tensor("op_929_cast_fp16")]; + tensor k_cache3_to_fp16_dtype_0 = const()[name = tensor("k_cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache3_to_fp16 = cast(dtype = k_cache3_to_fp16_dtype_0, x = k_cache3)[name = tensor("cast_119")]; + tensor var_930_cast_fp16 = mul(x = k_cache3_to_fp16, y = var_929_cast_fp16)[name = tensor("op_930_cast_fp16")]; + tensor var_935 = const()[name = tensor("op_935"), val = tensor([1, 1, 16, 64])]; + tensor var_936_cast_fp16 = reshape(shape = var_935, x = out_15_cast_fp16)[name = tensor("op_936_cast_fp16")]; + tensor var_937_cast_fp16 = mul(x = var_936_cast_fp16, y = onehot_15_cast_fp16)[name = tensor("op_937_cast_fp16")]; + tensor new_k_cache_7_cast_fp16 = add(x = var_930_cast_fp16, y = var_937_cast_fp16)[name = tensor("new_k_cache_7_cast_fp16")]; + tensor new_k_cache_7_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_7_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache3_to_fp16_dtype_0 = const()[name = tensor("v_cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache3_to_fp16 = cast(dtype = v_cache3_to_fp16_dtype_0, x = v_cache3)[name = tensor("cast_117")]; + tensor var_943_cast_fp16 = mul(x = v_cache3_to_fp16, y = var_929_cast_fp16)[name = tensor("op_943_cast_fp16")]; + tensor var_948 = const()[name = tensor("op_948"), val = tensor([1, 1, 16, 64])]; + tensor var_949_cast_fp16 = reshape(shape = var_948, x = v_7_cast_fp16)[name = tensor("op_949_cast_fp16")]; + tensor var_950_cast_fp16 = mul(x = var_949_cast_fp16, y = onehot_15_cast_fp16)[name = tensor("op_950_cast_fp16")]; + tensor new_v_cache_7_cast_fp16 = add(x = var_943_cast_fp16, y = var_950_cast_fp16)[name = tensor("new_v_cache_7_cast_fp16")]; + tensor new_v_cache_7_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_7_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_3_perm_0 = const()[name = tensor("transpose_3_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_963 = const()[name = tensor("op_963"), val = tensor([0, 2, 1, 3])]; + tensor var_969 = const()[name = tensor("op_969"), val = tensor([1, 16, 1, 64])]; + tensor q4_7_cast_fp16 = reshape(shape = var_969, x = out_13_cast_fp16)[name = tensor("q4_7_cast_fp16")]; + tensor var_974_transpose_x_0 = const()[name = tensor("op_974_transpose_x_0"), val = tensor(false)]; + tensor var_974_transpose_y_0 = const()[name = tensor("op_974_transpose_y_0"), val = tensor(false)]; + tensor transpose_3_cast_fp16 = transpose(perm = transpose_3_perm_0, x = new_k_cache_7_cast_fp16)[name = tensor("transpose_14")]; + tensor var_974_cast_fp16 = matmul(transpose_x = var_974_transpose_x_0, transpose_y = var_974_transpose_y_0, x = q4_7_cast_fp16, y = transpose_3_cast_fp16)[name = tensor("op_974_cast_fp16")]; + tensor var_975_to_fp16 = const()[name = tensor("op_975_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_13_cast_fp16 = mul(x = var_974_cast_fp16, y = var_975_to_fp16)[name = tensor("scores_13_cast_fp16")]; + tensor var_979 = const()[name = tensor("op_979"), val = tensor([1, 1])]; + tensor var_980_cast_fp16 = reshape(shape = var_979, x = position3_to_fp16)[name = tensor("op_980_cast_fp16")]; + tensor var_981_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_980_cast_fp16)[name = tensor("op_981_cast_fp16")]; + tensor mask_7_to_fp16_dtype_0 = const()[name = tensor("mask_7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_988_to_fp16 = const()[name = tensor("op_988_to_fp16"), val = tensor(0x1p+0)]; + tensor var_981_cast_fp16_to_fp16 = cast(dtype = mask_7_to_fp16_dtype_0, x = var_981_cast_fp16)[name = tensor("cast_115")]; + tensor var_989_cast_fp16 = sub(x = var_981_cast_fp16_to_fp16, y = var_988_to_fp16)[name = tensor("op_989_cast_fp16")]; + tensor var_994 = const()[name = tensor("op_994"), val = tensor([1, 1, 1, 512])]; + tensor var_995_cast_fp16 = reshape(shape = var_994, x = var_989_cast_fp16)[name = tensor("op_995_cast_fp16")]; + tensor var_996_to_fp16 = const()[name = tensor("op_996_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_997_cast_fp16 = mul(x = var_995_cast_fp16, y = var_996_to_fp16)[name = tensor("op_997_cast_fp16")]; + tensor scores_15_cast_fp16 = add(x = scores_13_cast_fp16, y = var_997_cast_fp16)[name = tensor("scores_15_cast_fp16")]; + tensor var_1000 = const()[name = tensor("op_1000"), val = tensor(-1)]; + tensor weights_7_cast_fp16 = softmax(axis = var_1000, x = scores_15_cast_fp16)[name = tensor("weights_7_cast_fp16")]; + tensor attn_7_transpose_x_0 = const()[name = tensor("attn_7_transpose_x_0"), val = tensor(false)]; + tensor attn_7_transpose_y_0 = const()[name = tensor("attn_7_transpose_y_0"), val = tensor(false)]; + tensor values_7_cast_fp16 = transpose(perm = var_963, x = new_v_cache_7_cast_fp16)[name = tensor("transpose_13")]; + tensor attn_7_cast_fp16 = matmul(transpose_x = attn_7_transpose_x_0, transpose_y = attn_7_transpose_y_0, x = weights_7_cast_fp16, y = values_7_cast_fp16)[name = tensor("attn_7_cast_fp16")]; + tensor var_1008 = const()[name = tensor("op_1008"), val = tensor([0, 2, 1, 3])]; + tensor var_1013 = const()[name = tensor("op_1013"), val = tensor([1, 1, 1024])]; + tensor var_1009_cast_fp16 = transpose(perm = var_1008, x = attn_7_cast_fp16)[name = tensor("transpose_12")]; + tensor input_43_cast_fp16 = reshape(shape = var_1013, x = var_1009_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor attn3_out_proj_weight_to_fp16 = const()[name = tensor("attn3_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(81902784)))]; + tensor linear_14_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn3_out_proj_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor var_1019_to_fp16 = const()[name = tensor("op_1019_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1020_cast_fp16 = add(x = position3_to_fp16, y = var_1019_to_fp16)[name = tensor("op_1020_cast_fp16")]; + tensor var_1020_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1020_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_45_cast_fp16 = add(x = input_39_cast_fp16, y = linear_14_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor input_47_axes_0 = const()[name = tensor("input_47_axes_0"), val = tensor([-1])]; + tensor norm3_2_weight_to_fp16 = const()[name = tensor("norm3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84000000)))]; + tensor norm3_2_bias_to_fp16 = const()[name = tensor("norm3_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84002112)))]; + tensor var_1024_to_fp16 = const()[name = tensor("op_1024_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_47_cast_fp16 = layer_norm(axes = input_47_axes_0, beta = norm3_2_bias_to_fp16, epsilon = var_1024_to_fp16, gamma = norm3_2_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor linear3_1_weight_to_fp16 = const()[name = tensor("linear3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84004224)))]; + tensor linear_15_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear3_1_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor input_49_mode_0 = const()[name = tensor("input_49_mode_0"), val = tensor("EXACT")]; + tensor input_49_cast_fp16 = gelu(mode = input_49_mode_0, x = linear_15_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor linear3_2_weight_to_fp16 = const()[name = tensor("linear3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(92392896)))]; + tensor linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear3_2_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor input_51_cast_fp16 = add(x = input_45_cast_fp16, y = linear_16_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor input_53_axes_0 = const()[name = tensor("input_53_axes_0"), val = tensor([-1])]; + tensor norm4_1_weight_to_fp16 = const()[name = tensor("norm4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100781568)))]; + tensor norm4_1_bias_to_fp16 = const()[name = tensor("norm4_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100783680)))]; + tensor var_1041_to_fp16 = const()[name = tensor("op_1041_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_53_cast_fp16 = layer_norm(axes = input_53_axes_0, beta = norm4_1_bias_to_fp16, epsilon = var_1041_to_fp16, gamma = norm4_1_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor attn4_in_proj_weight_to_fp16 = const()[name = tensor("attn4_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100785792)))]; + tensor linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn4_in_proj_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor var_1054 = const()[name = tensor("op_1054"), val = tensor([1, 3, 16, 64])]; + tensor qkv_9_cast_fp16 = reshape(shape = var_1054, x = linear_17_cast_fp16)[name = tensor("qkv_9_cast_fp16")]; + tensor x_17_begin_0 = const()[name = tensor("x_17_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_17_end_0 = const()[name = tensor("x_17_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_17_end_mask_0 = const()[name = tensor("x_17_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_17_squeeze_mask_0 = const()[name = tensor("x_17_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_17_cast_fp16 = slice_by_index(begin = x_17_begin_0, end = x_17_end_0, end_mask = x_17_end_mask_0, squeeze_mask = x_17_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("x_17_cast_fp16")]; + tensor x_19_begin_0 = const()[name = tensor("x_19_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_19_end_0 = const()[name = tensor("x_19_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_19_end_mask_0 = const()[name = tensor("x_19_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_19_squeeze_mask_0 = const()[name = tensor("x_19_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_19_cast_fp16 = slice_by_index(begin = x_19_begin_0, end = x_19_end_0, end_mask = x_19_end_mask_0, squeeze_mask = x_19_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("x_19_cast_fp16")]; + tensor v_9_begin_0 = const()[name = tensor("v_9_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_9_end_0 = const()[name = tensor("v_9_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_9_end_mask_0 = const()[name = tensor("v_9_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_9_squeeze_mask_0 = const()[name = tensor("v_9_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_9_cast_fp16 = slice_by_index(begin = v_9_begin_0, end = v_9_end_0, end_mask = v_9_end_mask_0, squeeze_mask = v_9_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("v_9_cast_fp16")]; + tensor var_1083 = const()[name = tensor("op_1083"), val = tensor([1, 1, 1])]; + tensor position4_to_fp16_dtype_0 = const()[name = tensor("position4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position4_to_fp16 = cast(dtype = position4_to_fp16_dtype_0, x = position4)[name = tensor("cast_113")]; + tensor var_1084_cast_fp16 = reshape(shape = var_1083, x = position4_to_fp16)[name = tensor("op_1084_cast_fp16")]; + tensor angles_17_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_1084_cast_fp16)[name = tensor("angles_17_cast_fp16")]; + tensor rotr_17_cast_fp16 = cos(x = angles_17_cast_fp16)[name = tensor("rotr_17_cast_fp16")]; + tensor roti_17_cast_fp16 = sin(x = angles_17_cast_fp16)[name = tensor("roti_17_cast_fp16")]; + tensor var_1092 = const()[name = tensor("op_1092"), val = tensor([1, 16, 32, 2])]; + tensor pairs_17_cast_fp16 = reshape(shape = var_1092, x = x_17_cast_fp16)[name = tensor("pairs_17_cast_fp16")]; + tensor xr_17_begin_0 = const()[name = tensor("xr_17_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_17_end_0 = const()[name = tensor("xr_17_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_17_end_mask_0 = const()[name = tensor("xr_17_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_17_squeeze_mask_0 = const()[name = tensor("xr_17_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_17_cast_fp16 = slice_by_index(begin = xr_17_begin_0, end = xr_17_end_0, end_mask = xr_17_end_mask_0, squeeze_mask = xr_17_squeeze_mask_0, x = pairs_17_cast_fp16)[name = tensor("xr_17_cast_fp16")]; + tensor xi_17_begin_0 = const()[name = tensor("xi_17_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_17_end_0 = const()[name = tensor("xi_17_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_17_end_mask_0 = const()[name = tensor("xi_17_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_17_squeeze_mask_0 = const()[name = tensor("xi_17_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_17_cast_fp16 = slice_by_index(begin = xi_17_begin_0, end = xi_17_end_0, end_mask = xi_17_end_mask_0, squeeze_mask = xi_17_squeeze_mask_0, x = pairs_17_cast_fp16)[name = tensor("xi_17_cast_fp16")]; + tensor var_1100_cast_fp16 = mul(x = xr_17_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_1100_cast_fp16")]; + tensor var_1101_cast_fp16 = mul(x = xi_17_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_1101_cast_fp16")]; + tensor xor__17_cast_fp16 = sub(x = var_1100_cast_fp16, y = var_1101_cast_fp16)[name = tensor("xor__17_cast_fp16")]; + tensor var_1104_cast_fp16 = mul(x = xr_17_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_1104_cast_fp16")]; + tensor var_1105_cast_fp16 = mul(x = xi_17_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_1105_cast_fp16")]; + tensor xoi_17_cast_fp16 = add(x = var_1104_cast_fp16, y = var_1105_cast_fp16)[name = tensor("xoi_17_cast_fp16")]; + tensor out_17_axis_0 = const()[name = tensor("out_17_axis_0"), val = tensor(-1)]; + tensor out_17_cast_fp16 = stack(axis = out_17_axis_0, values = (xor__17_cast_fp16, xoi_17_cast_fp16))[name = tensor("out_17_cast_fp16")]; + tensor var_1128 = const()[name = tensor("op_1128"), val = tensor([1, 16, 32, 2])]; + tensor pairs_19_cast_fp16 = reshape(shape = var_1128, x = x_19_cast_fp16)[name = tensor("pairs_19_cast_fp16")]; + tensor xr_19_begin_0 = const()[name = tensor("xr_19_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_19_end_0 = const()[name = tensor("xr_19_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_19_end_mask_0 = const()[name = tensor("xr_19_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_19_squeeze_mask_0 = const()[name = tensor("xr_19_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_19_cast_fp16 = slice_by_index(begin = xr_19_begin_0, end = xr_19_end_0, end_mask = xr_19_end_mask_0, squeeze_mask = xr_19_squeeze_mask_0, x = pairs_19_cast_fp16)[name = tensor("xr_19_cast_fp16")]; + tensor xi_19_begin_0 = const()[name = tensor("xi_19_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_19_end_0 = const()[name = tensor("xi_19_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_19_end_mask_0 = const()[name = tensor("xi_19_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_19_squeeze_mask_0 = const()[name = tensor("xi_19_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_19_cast_fp16 = slice_by_index(begin = xi_19_begin_0, end = xi_19_end_0, end_mask = xi_19_end_mask_0, squeeze_mask = xi_19_squeeze_mask_0, x = pairs_19_cast_fp16)[name = tensor("xi_19_cast_fp16")]; + tensor var_1136_cast_fp16 = mul(x = xr_19_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_1136_cast_fp16")]; + tensor var_1137_cast_fp16 = mul(x = xi_19_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_1137_cast_fp16")]; + tensor xor__19_cast_fp16 = sub(x = var_1136_cast_fp16, y = var_1137_cast_fp16)[name = tensor("xor__19_cast_fp16")]; + tensor var_1140_cast_fp16 = mul(x = xr_19_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_1140_cast_fp16")]; + tensor var_1141_cast_fp16 = mul(x = xi_19_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_1141_cast_fp16")]; + tensor xoi_19_cast_fp16 = add(x = var_1140_cast_fp16, y = var_1141_cast_fp16)[name = tensor("xoi_19_cast_fp16")]; + tensor out_19_axis_0 = const()[name = tensor("out_19_axis_0"), val = tensor(-1)]; + tensor out_19_cast_fp16 = stack(axis = out_19_axis_0, values = (xor__19_cast_fp16, xoi_19_cast_fp16))[name = tensor("out_19_cast_fp16")]; + tensor _inversed_1153_y_0_to_fp16 = const()[name = tensor("_inversed_1153_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1153_cast_fp16 = mul(x = position4_to_fp16, y = _inversed_1153_y_0_to_fp16)[name = tensor("_inversed_1153_cast_fp16")]; + tensor var_1154_cast_fp16 = floor(x = _inversed_1153_cast_fp16)[name = tensor("op_1154_cast_fp16")]; + tensor var_1155_to_fp16 = const()[name = tensor("op_1155_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1156_cast_fp16 = mul(x = var_1154_cast_fp16, y = var_1155_to_fp16)[name = tensor("op_1156_cast_fp16")]; + tensor write_idx_9_cast_fp16 = sub(x = position4_to_fp16, y = var_1156_cast_fp16)[name = tensor("write_idx_9_cast_fp16")]; + tensor var_1159_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_9_cast_fp16)[name = tensor("op_1159_cast_fp16")]; + tensor var_1169 = const()[name = tensor("op_1169"), val = tensor([1, 512, 1, 1])]; + tensor onehot_17_to_fp16_dtype_0 = const()[name = tensor("onehot_17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_1159_cast_fp16_to_fp16 = cast(dtype = onehot_17_to_fp16_dtype_0, x = var_1159_cast_fp16)[name = tensor("cast_112")]; + tensor onehot_19_cast_fp16 = reshape(shape = var_1169, x = var_1159_cast_fp16_to_fp16)[name = tensor("onehot_19_cast_fp16")]; + tensor var_1171_to_fp16 = const()[name = tensor("op_1171_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1173_cast_fp16 = sub(x = var_1171_to_fp16, y = onehot_19_cast_fp16)[name = tensor("op_1173_cast_fp16")]; + tensor k_cache4_to_fp16_dtype_0 = const()[name = tensor("k_cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache4_to_fp16 = cast(dtype = k_cache4_to_fp16_dtype_0, x = k_cache4)[name = tensor("cast_111")]; + tensor var_1174_cast_fp16 = mul(x = k_cache4_to_fp16, y = var_1173_cast_fp16)[name = tensor("op_1174_cast_fp16")]; + tensor var_1179 = const()[name = tensor("op_1179"), val = tensor([1, 1, 16, 64])]; + tensor var_1180_cast_fp16 = reshape(shape = var_1179, x = out_19_cast_fp16)[name = tensor("op_1180_cast_fp16")]; + tensor var_1181_cast_fp16 = mul(x = var_1180_cast_fp16, y = onehot_19_cast_fp16)[name = tensor("op_1181_cast_fp16")]; + tensor new_k_cache_9_cast_fp16 = add(x = var_1174_cast_fp16, y = var_1181_cast_fp16)[name = tensor("new_k_cache_9_cast_fp16")]; + tensor new_k_cache_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache4_to_fp16_dtype_0 = const()[name = tensor("v_cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache4_to_fp16 = cast(dtype = v_cache4_to_fp16_dtype_0, x = v_cache4)[name = tensor("cast_109")]; + tensor var_1187_cast_fp16 = mul(x = v_cache4_to_fp16, y = var_1173_cast_fp16)[name = tensor("op_1187_cast_fp16")]; + tensor var_1192 = const()[name = tensor("op_1192"), val = tensor([1, 1, 16, 64])]; + tensor var_1193_cast_fp16 = reshape(shape = var_1192, x = v_9_cast_fp16)[name = tensor("op_1193_cast_fp16")]; + tensor var_1194_cast_fp16 = mul(x = var_1193_cast_fp16, y = onehot_19_cast_fp16)[name = tensor("op_1194_cast_fp16")]; + tensor new_v_cache_9_cast_fp16 = add(x = var_1187_cast_fp16, y = var_1194_cast_fp16)[name = tensor("new_v_cache_9_cast_fp16")]; + tensor new_v_cache_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_4_perm_0 = const()[name = tensor("transpose_4_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_1207 = const()[name = tensor("op_1207"), val = tensor([0, 2, 1, 3])]; + tensor var_1213 = const()[name = tensor("op_1213"), val = tensor([1, 16, 1, 64])]; + tensor q4_9_cast_fp16 = reshape(shape = var_1213, x = out_17_cast_fp16)[name = tensor("q4_9_cast_fp16")]; + tensor var_1218_transpose_x_0 = const()[name = tensor("op_1218_transpose_x_0"), val = tensor(false)]; + tensor var_1218_transpose_y_0 = const()[name = tensor("op_1218_transpose_y_0"), val = tensor(false)]; + tensor transpose_4_cast_fp16 = transpose(perm = transpose_4_perm_0, x = new_k_cache_9_cast_fp16)[name = tensor("transpose_11")]; + tensor var_1218_cast_fp16 = matmul(transpose_x = var_1218_transpose_x_0, transpose_y = var_1218_transpose_y_0, x = q4_9_cast_fp16, y = transpose_4_cast_fp16)[name = tensor("op_1218_cast_fp16")]; + tensor var_1219_to_fp16 = const()[name = tensor("op_1219_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_17_cast_fp16 = mul(x = var_1218_cast_fp16, y = var_1219_to_fp16)[name = tensor("scores_17_cast_fp16")]; + tensor var_1223 = const()[name = tensor("op_1223"), val = tensor([1, 1])]; + tensor var_1224_cast_fp16 = reshape(shape = var_1223, x = position4_to_fp16)[name = tensor("op_1224_cast_fp16")]; + tensor var_1225_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_1224_cast_fp16)[name = tensor("op_1225_cast_fp16")]; + tensor mask_9_to_fp16_dtype_0 = const()[name = tensor("mask_9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_1232_to_fp16 = const()[name = tensor("op_1232_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1225_cast_fp16_to_fp16 = cast(dtype = mask_9_to_fp16_dtype_0, x = var_1225_cast_fp16)[name = tensor("cast_107")]; + tensor var_1233_cast_fp16 = sub(x = var_1225_cast_fp16_to_fp16, y = var_1232_to_fp16)[name = tensor("op_1233_cast_fp16")]; + tensor var_1238 = const()[name = tensor("op_1238"), val = tensor([1, 1, 1, 512])]; + tensor var_1239_cast_fp16 = reshape(shape = var_1238, x = var_1233_cast_fp16)[name = tensor("op_1239_cast_fp16")]; + tensor var_1240_to_fp16 = const()[name = tensor("op_1240_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_1241_cast_fp16 = mul(x = var_1239_cast_fp16, y = var_1240_to_fp16)[name = tensor("op_1241_cast_fp16")]; + tensor scores_19_cast_fp16 = add(x = scores_17_cast_fp16, y = var_1241_cast_fp16)[name = tensor("scores_19_cast_fp16")]; + tensor var_1244 = const()[name = tensor("op_1244"), val = tensor(-1)]; + tensor weights_9_cast_fp16 = softmax(axis = var_1244, x = scores_19_cast_fp16)[name = tensor("weights_9_cast_fp16")]; + tensor attn_9_transpose_x_0 = const()[name = tensor("attn_9_transpose_x_0"), val = tensor(false)]; + tensor attn_9_transpose_y_0 = const()[name = tensor("attn_9_transpose_y_0"), val = tensor(false)]; + tensor values_9_cast_fp16 = transpose(perm = var_1207, x = new_v_cache_9_cast_fp16)[name = tensor("transpose_10")]; + tensor attn_9_cast_fp16 = matmul(transpose_x = attn_9_transpose_x_0, transpose_y = attn_9_transpose_y_0, x = weights_9_cast_fp16, y = values_9_cast_fp16)[name = tensor("attn_9_cast_fp16")]; + tensor var_1252 = const()[name = tensor("op_1252"), val = tensor([0, 2, 1, 3])]; + tensor var_1257 = const()[name = tensor("op_1257"), val = tensor([1, 1, 1024])]; + tensor var_1253_cast_fp16 = transpose(perm = var_1252, x = attn_9_cast_fp16)[name = tensor("transpose_9")]; + tensor input_55_cast_fp16 = reshape(shape = var_1257, x = var_1253_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor attn4_out_proj_weight_to_fp16 = const()[name = tensor("attn4_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(107077312)))]; + tensor linear_18_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn4_out_proj_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor var_1263_to_fp16 = const()[name = tensor("op_1263_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1264_cast_fp16 = add(x = position4_to_fp16, y = var_1263_to_fp16)[name = tensor("op_1264_cast_fp16")]; + tensor var_1264_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1264_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_57_cast_fp16 = add(x = input_51_cast_fp16, y = linear_18_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor input_59_axes_0 = const()[name = tensor("input_59_axes_0"), val = tensor([-1])]; + tensor norm4_2_weight_to_fp16 = const()[name = tensor("norm4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109174528)))]; + tensor norm4_2_bias_to_fp16 = const()[name = tensor("norm4_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109176640)))]; + tensor var_1268_to_fp16 = const()[name = tensor("op_1268_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_59_cast_fp16 = layer_norm(axes = input_59_axes_0, beta = norm4_2_bias_to_fp16, epsilon = var_1268_to_fp16, gamma = norm4_2_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor linear4_1_weight_to_fp16 = const()[name = tensor("linear4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109178752)))]; + tensor linear_19_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear4_1_weight_to_fp16, x = input_59_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_61_mode_0 = const()[name = tensor("input_61_mode_0"), val = tensor("EXACT")]; + tensor input_61_cast_fp16 = gelu(mode = input_61_mode_0, x = linear_19_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor linear4_2_weight_to_fp16 = const()[name = tensor("linear4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(117567424)))]; + tensor linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear4_2_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor input_63_cast_fp16 = add(x = input_57_cast_fp16, y = linear_20_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor input_65_axes_0 = const()[name = tensor("input_65_axes_0"), val = tensor([-1])]; + tensor norm5_1_weight_to_fp16 = const()[name = tensor("norm5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125956096)))]; + tensor norm5_1_bias_to_fp16 = const()[name = tensor("norm5_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125958208)))]; + tensor var_1285_to_fp16 = const()[name = tensor("op_1285_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_65_cast_fp16 = layer_norm(axes = input_65_axes_0, beta = norm5_1_bias_to_fp16, epsilon = var_1285_to_fp16, gamma = norm5_1_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("input_65_cast_fp16")]; + tensor attn5_in_proj_weight_to_fp16 = const()[name = tensor("attn5_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125960320)))]; + tensor linear_21_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn5_in_proj_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("linear_21_cast_fp16")]; + tensor var_1298 = const()[name = tensor("op_1298"), val = tensor([1, 3, 16, 64])]; + tensor qkv_cast_fp16 = reshape(shape = var_1298, x = linear_21_cast_fp16)[name = tensor("qkv_cast_fp16")]; + tensor x_21_begin_0 = const()[name = tensor("x_21_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_21_end_0 = const()[name = tensor("x_21_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_21_end_mask_0 = const()[name = tensor("x_21_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_21_squeeze_mask_0 = const()[name = tensor("x_21_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_21_cast_fp16 = slice_by_index(begin = x_21_begin_0, end = x_21_end_0, end_mask = x_21_end_mask_0, squeeze_mask = x_21_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("x_21_cast_fp16")]; + tensor x_begin_0 = const()[name = tensor("x_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_end_0 = const()[name = tensor("x_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_end_mask_0 = const()[name = tensor("x_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_squeeze_mask_0 = const()[name = tensor("x_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_cast_fp16 = slice_by_index(begin = x_begin_0, end = x_end_0, end_mask = x_end_mask_0, squeeze_mask = x_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("x_cast_fp16")]; + tensor v_begin_0 = const()[name = tensor("v_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_end_0 = const()[name = tensor("v_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_end_mask_0 = const()[name = tensor("v_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_squeeze_mask_0 = const()[name = tensor("v_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_cast_fp16 = slice_by_index(begin = v_begin_0, end = v_end_0, end_mask = v_end_mask_0, squeeze_mask = v_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("v_cast_fp16")]; + tensor var_1327 = const()[name = tensor("op_1327"), val = tensor([1, 1, 1])]; + tensor position5_to_fp16_dtype_0 = const()[name = tensor("position5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position5_to_fp16 = cast(dtype = position5_to_fp16_dtype_0, x = position5)[name = tensor("cast_105")]; + tensor var_1328_cast_fp16 = reshape(shape = var_1327, x = position5_to_fp16)[name = tensor("op_1328_cast_fp16")]; + tensor angles_21_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_1328_cast_fp16)[name = tensor("angles_21_cast_fp16")]; + tensor rotr_21_cast_fp16 = cos(x = angles_21_cast_fp16)[name = tensor("rotr_21_cast_fp16")]; + tensor roti_21_cast_fp16 = sin(x = angles_21_cast_fp16)[name = tensor("roti_21_cast_fp16")]; + tensor var_1336 = const()[name = tensor("op_1336"), val = tensor([1, 16, 32, 2])]; + tensor pairs_21_cast_fp16 = reshape(shape = var_1336, x = x_21_cast_fp16)[name = tensor("pairs_21_cast_fp16")]; + tensor xr_21_begin_0 = const()[name = tensor("xr_21_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_21_end_0 = const()[name = tensor("xr_21_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_21_end_mask_0 = const()[name = tensor("xr_21_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_21_squeeze_mask_0 = const()[name = tensor("xr_21_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_21_cast_fp16 = slice_by_index(begin = xr_21_begin_0, end = xr_21_end_0, end_mask = xr_21_end_mask_0, squeeze_mask = xr_21_squeeze_mask_0, x = pairs_21_cast_fp16)[name = tensor("xr_21_cast_fp16")]; + tensor xi_21_begin_0 = const()[name = tensor("xi_21_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_21_end_0 = const()[name = tensor("xi_21_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_21_end_mask_0 = const()[name = tensor("xi_21_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_21_squeeze_mask_0 = const()[name = tensor("xi_21_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_21_cast_fp16 = slice_by_index(begin = xi_21_begin_0, end = xi_21_end_0, end_mask = xi_21_end_mask_0, squeeze_mask = xi_21_squeeze_mask_0, x = pairs_21_cast_fp16)[name = tensor("xi_21_cast_fp16")]; + tensor var_1344_cast_fp16 = mul(x = xr_21_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_1344_cast_fp16")]; + tensor var_1345_cast_fp16 = mul(x = xi_21_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_1345_cast_fp16")]; + tensor xor__21_cast_fp16 = sub(x = var_1344_cast_fp16, y = var_1345_cast_fp16)[name = tensor("xor__21_cast_fp16")]; + tensor var_1348_cast_fp16 = mul(x = xr_21_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_1348_cast_fp16")]; + tensor var_1349_cast_fp16 = mul(x = xi_21_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_1349_cast_fp16")]; + tensor xoi_21_cast_fp16 = add(x = var_1348_cast_fp16, y = var_1349_cast_fp16)[name = tensor("xoi_21_cast_fp16")]; + tensor out_21_axis_0 = const()[name = tensor("out_21_axis_0"), val = tensor(-1)]; + tensor out_21_cast_fp16 = stack(axis = out_21_axis_0, values = (xor__21_cast_fp16, xoi_21_cast_fp16))[name = tensor("out_21_cast_fp16")]; + tensor var_1372 = const()[name = tensor("op_1372"), val = tensor([1, 16, 32, 2])]; + tensor pairs_cast_fp16 = reshape(shape = var_1372, x = x_cast_fp16)[name = tensor("pairs_cast_fp16")]; + tensor xr_begin_0 = const()[name = tensor("xr_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_end_0 = const()[name = tensor("xr_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_end_mask_0 = const()[name = tensor("xr_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_squeeze_mask_0 = const()[name = tensor("xr_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_cast_fp16 = slice_by_index(begin = xr_begin_0, end = xr_end_0, end_mask = xr_end_mask_0, squeeze_mask = xr_squeeze_mask_0, x = pairs_cast_fp16)[name = tensor("xr_cast_fp16")]; + tensor xi_begin_0 = const()[name = tensor("xi_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_end_0 = const()[name = tensor("xi_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_end_mask_0 = const()[name = tensor("xi_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_squeeze_mask_0 = const()[name = tensor("xi_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_cast_fp16 = slice_by_index(begin = xi_begin_0, end = xi_end_0, end_mask = xi_end_mask_0, squeeze_mask = xi_squeeze_mask_0, x = pairs_cast_fp16)[name = tensor("xi_cast_fp16")]; + tensor var_1380_cast_fp16 = mul(x = xr_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_1380_cast_fp16")]; + tensor var_1381_cast_fp16 = mul(x = xi_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_1381_cast_fp16")]; + tensor xor__cast_fp16 = sub(x = var_1380_cast_fp16, y = var_1381_cast_fp16)[name = tensor("xor__cast_fp16")]; + tensor var_1384_cast_fp16 = mul(x = xr_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_1384_cast_fp16")]; + tensor var_1385_cast_fp16 = mul(x = xi_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_1385_cast_fp16")]; + tensor xoi_cast_fp16 = add(x = var_1384_cast_fp16, y = var_1385_cast_fp16)[name = tensor("xoi_cast_fp16")]; + tensor out_axis_0 = const()[name = tensor("out_axis_0"), val = tensor(-1)]; + tensor out_cast_fp16 = stack(axis = out_axis_0, values = (xor__cast_fp16, xoi_cast_fp16))[name = tensor("out_cast_fp16")]; + tensor _inversed_1397_y_0_to_fp16 = const()[name = tensor("_inversed_1397_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1397_cast_fp16 = mul(x = position5_to_fp16, y = _inversed_1397_y_0_to_fp16)[name = tensor("_inversed_1397_cast_fp16")]; + tensor var_1398_cast_fp16 = floor(x = _inversed_1397_cast_fp16)[name = tensor("op_1398_cast_fp16")]; + tensor var_1399_to_fp16 = const()[name = tensor("op_1399_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1400_cast_fp16 = mul(x = var_1398_cast_fp16, y = var_1399_to_fp16)[name = tensor("op_1400_cast_fp16")]; + tensor write_idx_cast_fp16 = sub(x = position5_to_fp16, y = var_1400_cast_fp16)[name = tensor("write_idx_cast_fp16")]; + tensor var_1403_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_cast_fp16)[name = tensor("op_1403_cast_fp16")]; + tensor var_1413 = const()[name = tensor("op_1413"), val = tensor([1, 512, 1, 1])]; + tensor onehot_21_to_fp16_dtype_0 = const()[name = tensor("onehot_21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_1403_cast_fp16_to_fp16 = cast(dtype = onehot_21_to_fp16_dtype_0, x = var_1403_cast_fp16)[name = tensor("cast_104")]; + tensor onehot_cast_fp16 = reshape(shape = var_1413, x = var_1403_cast_fp16_to_fp16)[name = tensor("onehot_cast_fp16")]; + tensor var_1415_to_fp16 = const()[name = tensor("op_1415_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1417_cast_fp16 = sub(x = var_1415_to_fp16, y = onehot_cast_fp16)[name = tensor("op_1417_cast_fp16")]; + tensor k_cache5_to_fp16_dtype_0 = const()[name = tensor("k_cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache5_to_fp16 = cast(dtype = k_cache5_to_fp16_dtype_0, x = k_cache5)[name = tensor("cast_103")]; + tensor var_1418_cast_fp16 = mul(x = k_cache5_to_fp16, y = var_1417_cast_fp16)[name = tensor("op_1418_cast_fp16")]; + tensor var_1423 = const()[name = tensor("op_1423"), val = tensor([1, 1, 16, 64])]; + tensor var_1424_cast_fp16 = reshape(shape = var_1423, x = out_cast_fp16)[name = tensor("op_1424_cast_fp16")]; + tensor var_1425_cast_fp16 = mul(x = var_1424_cast_fp16, y = onehot_cast_fp16)[name = tensor("op_1425_cast_fp16")]; + tensor new_k_cache_cast_fp16 = add(x = var_1418_cast_fp16, y = var_1425_cast_fp16)[name = tensor("new_k_cache_cast_fp16")]; + tensor new_k_cache_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache5_to_fp16_dtype_0 = const()[name = tensor("v_cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache5_to_fp16 = cast(dtype = v_cache5_to_fp16_dtype_0, x = v_cache5)[name = tensor("cast_101")]; + tensor var_1431_cast_fp16 = mul(x = v_cache5_to_fp16, y = var_1417_cast_fp16)[name = tensor("op_1431_cast_fp16")]; + tensor var_1436 = const()[name = tensor("op_1436"), val = tensor([1, 1, 16, 64])]; + tensor var_1437_cast_fp16 = reshape(shape = var_1436, x = v_cast_fp16)[name = tensor("op_1437_cast_fp16")]; + tensor var_1438_cast_fp16 = mul(x = var_1437_cast_fp16, y = onehot_cast_fp16)[name = tensor("op_1438_cast_fp16")]; + tensor new_v_cache_cast_fp16 = add(x = var_1431_cast_fp16, y = var_1438_cast_fp16)[name = tensor("new_v_cache_cast_fp16")]; + tensor new_v_cache_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_5_perm_0 = const()[name = tensor("transpose_5_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_1451 = const()[name = tensor("op_1451"), val = tensor([0, 2, 1, 3])]; + tensor var_1457 = const()[name = tensor("op_1457"), val = tensor([1, 16, 1, 64])]; + tensor q4_cast_fp16 = reshape(shape = var_1457, x = out_21_cast_fp16)[name = tensor("q4_cast_fp16")]; + tensor var_1462_transpose_x_0 = const()[name = tensor("op_1462_transpose_x_0"), val = tensor(false)]; + tensor var_1462_transpose_y_0 = const()[name = tensor("op_1462_transpose_y_0"), val = tensor(false)]; + tensor transpose_5_cast_fp16 = transpose(perm = transpose_5_perm_0, x = new_k_cache_cast_fp16)[name = tensor("transpose_8")]; + tensor var_1462_cast_fp16 = matmul(transpose_x = var_1462_transpose_x_0, transpose_y = var_1462_transpose_y_0, x = q4_cast_fp16, y = transpose_5_cast_fp16)[name = tensor("op_1462_cast_fp16")]; + tensor var_1463_to_fp16 = const()[name = tensor("op_1463_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_21_cast_fp16 = mul(x = var_1462_cast_fp16, y = var_1463_to_fp16)[name = tensor("scores_21_cast_fp16")]; + tensor var_1467 = const()[name = tensor("op_1467"), val = tensor([1, 1])]; + tensor var_1468_cast_fp16 = reshape(shape = var_1467, x = position5_to_fp16)[name = tensor("op_1468_cast_fp16")]; + tensor var_1469_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_1468_cast_fp16)[name = tensor("op_1469_cast_fp16")]; + tensor mask_to_fp16_dtype_0 = const()[name = tensor("mask_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_1476_to_fp16 = const()[name = tensor("op_1476_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1469_cast_fp16_to_fp16 = cast(dtype = mask_to_fp16_dtype_0, x = var_1469_cast_fp16)[name = tensor("cast_99")]; + tensor var_1477_cast_fp16 = sub(x = var_1469_cast_fp16_to_fp16, y = var_1476_to_fp16)[name = tensor("op_1477_cast_fp16")]; + tensor var_1482 = const()[name = tensor("op_1482"), val = tensor([1, 1, 1, 512])]; + tensor var_1483_cast_fp16 = reshape(shape = var_1482, x = var_1477_cast_fp16)[name = tensor("op_1483_cast_fp16")]; + tensor var_1484_to_fp16 = const()[name = tensor("op_1484_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_1485_cast_fp16 = mul(x = var_1483_cast_fp16, y = var_1484_to_fp16)[name = tensor("op_1485_cast_fp16")]; + tensor scores_cast_fp16 = add(x = scores_21_cast_fp16, y = var_1485_cast_fp16)[name = tensor("scores_cast_fp16")]; + tensor var_1488 = const()[name = tensor("op_1488"), val = tensor(-1)]; + tensor weights_cast_fp16 = softmax(axis = var_1488, x = scores_cast_fp16)[name = tensor("weights_cast_fp16")]; + tensor attn_transpose_x_0 = const()[name = tensor("attn_transpose_x_0"), val = tensor(false)]; + tensor attn_transpose_y_0 = const()[name = tensor("attn_transpose_y_0"), val = tensor(false)]; + tensor values_cast_fp16 = transpose(perm = var_1451, x = new_v_cache_cast_fp16)[name = tensor("transpose_7")]; + tensor attn_cast_fp16 = matmul(transpose_x = attn_transpose_x_0, transpose_y = attn_transpose_y_0, x = weights_cast_fp16, y = values_cast_fp16)[name = tensor("attn_cast_fp16")]; + tensor var_1496 = const()[name = tensor("op_1496"), val = tensor([0, 2, 1, 3])]; + tensor var_1501 = const()[name = tensor("op_1501"), val = tensor([1, 1, 1024])]; + tensor var_1497_cast_fp16 = transpose(perm = var_1496, x = attn_cast_fp16)[name = tensor("transpose_6")]; + tensor input_67_cast_fp16 = reshape(shape = var_1501, x = var_1497_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor attn5_out_proj_weight_to_fp16 = const()[name = tensor("attn5_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132251840)))]; + tensor linear_22_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn5_out_proj_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("linear_22_cast_fp16")]; + tensor var_1507_to_fp16 = const()[name = tensor("op_1507_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1508_cast_fp16 = add(x = position5_to_fp16, y = var_1507_to_fp16)[name = tensor("op_1508_cast_fp16")]; + tensor var_1508_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1508_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_69_cast_fp16 = add(x = input_63_cast_fp16, y = linear_22_cast_fp16)[name = tensor("input_69_cast_fp16")]; + tensor input_71_axes_0 = const()[name = tensor("input_71_axes_0"), val = tensor([-1])]; + tensor norm5_2_weight_to_fp16 = const()[name = tensor("norm5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134349056)))]; + tensor norm5_2_bias_to_fp16 = const()[name = tensor("norm5_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134351168)))]; + tensor var_1512_to_fp16 = const()[name = tensor("op_1512_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_71_cast_fp16 = layer_norm(axes = input_71_axes_0, beta = norm5_2_bias_to_fp16, epsilon = var_1512_to_fp16, gamma = norm5_2_weight_to_fp16, x = input_69_cast_fp16)[name = tensor("input_71_cast_fp16")]; + tensor linear5_1_weight_to_fp16 = const()[name = tensor("linear5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134353280)))]; + tensor linear_23_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear5_1_weight_to_fp16, x = input_71_cast_fp16)[name = tensor("linear_23_cast_fp16")]; + tensor input_73_mode_0 = const()[name = tensor("input_73_mode_0"), val = tensor("EXACT")]; + tensor input_73_cast_fp16 = gelu(mode = input_73_mode_0, x = linear_23_cast_fp16)[name = tensor("input_73_cast_fp16")]; + tensor linear5_2_weight_to_fp16 = const()[name = tensor("linear5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142741952)))]; + tensor linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear5_2_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("linear_24_cast_fp16")]; + tensor input_75_cast_fp16 = add(x = input_69_cast_fp16, y = linear_24_cast_fp16)[name = tensor("input_75_cast_fp16")]; + tensor input_axes_0 = const()[name = tensor("input_axes_0"), val = tensor([-1])]; + tensor out_norm_weight_to_fp16 = const()[name = tensor("out_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151130624)))]; + tensor out_norm_bias_to_fp16 = const()[name = tensor("out_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151132736)))]; + tensor var_1529_to_fp16 = const()[name = tensor("op_1529_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_cast_fp16 = layer_norm(axes = input_axes_0, beta = out_norm_bias_to_fp16, epsilon = var_1529_to_fp16, gamma = out_norm_weight_to_fp16, x = input_75_cast_fp16)[name = tensor("input_cast_fp16")]; + tensor input_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("input_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor out_eos_weight_to_fp16 = const()[name = tensor("out_eos_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151134848)))]; + tensor out_eos_bias_to_fp16 = const()[name = tensor("out_eos_bias_to_fp16"), val = tensor([-0x1.5ep-3])]; + tensor linear_25_cast_fp16 = linear(bias = out_eos_bias_to_fp16, weight = out_eos_weight_to_fp16, x = input_cast_fp16)[name = tensor("linear_25_cast_fp16")]; + tensor linear_25_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("linear_25_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor is_eos = cast(dtype = linear_25_cast_fp16_to_fp32_dtype_0, x = linear_25_cast_fp16)[name = tensor("cast_96")]; + tensor transformer_out = cast(dtype = input_cast_fp16_to_fp32_dtype_0, x = input_cast_fp16)[name = tensor("cast_97")]; + tensor new_position5 = cast(dtype = var_1508_cast_fp16_to_fp32_dtype_0, x = var_1508_cast_fp16)[name = tensor("cast_98")]; + tensor new_v_cache5 = cast(dtype = new_v_cache_cast_fp16_to_fp32_dtype_0, x = new_v_cache_cast_fp16)[name = tensor("cast_100")]; + tensor new_k_cache5 = cast(dtype = new_k_cache_cast_fp16_to_fp32_dtype_0, x = new_k_cache_cast_fp16)[name = tensor("cast_102")]; + tensor new_position4 = cast(dtype = var_1264_cast_fp16_to_fp32_dtype_0, x = var_1264_cast_fp16)[name = tensor("cast_106")]; + tensor new_v_cache4 = cast(dtype = new_v_cache_9_cast_fp16_to_fp32_dtype_0, x = new_v_cache_9_cast_fp16)[name = tensor("cast_108")]; + tensor new_k_cache4 = cast(dtype = new_k_cache_9_cast_fp16_to_fp32_dtype_0, x = new_k_cache_9_cast_fp16)[name = tensor("cast_110")]; + tensor new_position3 = cast(dtype = var_1020_cast_fp16_to_fp32_dtype_0, x = var_1020_cast_fp16)[name = tensor("cast_114")]; + tensor new_v_cache3 = cast(dtype = new_v_cache_7_cast_fp16_to_fp32_dtype_0, x = new_v_cache_7_cast_fp16)[name = tensor("cast_116")]; + tensor new_k_cache3 = cast(dtype = new_k_cache_7_cast_fp16_to_fp32_dtype_0, x = new_k_cache_7_cast_fp16)[name = tensor("cast_118")]; + tensor new_position2 = cast(dtype = var_776_cast_fp16_to_fp32_dtype_0, x = var_776_cast_fp16)[name = tensor("cast_122")]; + tensor new_v_cache2 = cast(dtype = new_v_cache_5_cast_fp16_to_fp32_dtype_0, x = new_v_cache_5_cast_fp16)[name = tensor("cast_124")]; + tensor new_k_cache2 = cast(dtype = new_k_cache_5_cast_fp16_to_fp32_dtype_0, x = new_k_cache_5_cast_fp16)[name = tensor("cast_126")]; + tensor new_position1 = cast(dtype = var_532_cast_fp16_to_fp32_dtype_0, x = var_532_cast_fp16)[name = tensor("cast_130")]; + tensor new_v_cache1 = cast(dtype = new_v_cache_3_cast_fp16_to_fp32_dtype_0, x = new_v_cache_3_cast_fp16)[name = tensor("cast_132")]; + tensor new_k_cache1 = cast(dtype = new_k_cache_3_cast_fp16_to_fp32_dtype_0, x = new_k_cache_3_cast_fp16)[name = tensor("cast_134")]; + tensor new_position0 = cast(dtype = var_288_cast_fp16_to_fp32_dtype_0, x = var_288_cast_fp16)[name = tensor("cast_138")]; + tensor new_v_cache0 = cast(dtype = new_v_cache_1_cast_fp16_to_fp32_dtype_0, x = new_v_cache_1_cast_fp16)[name = tensor("cast_140")]; + tensor new_k_cache0 = cast(dtype = new_k_cache_1_cast_fp16_to_fp32_dtype_0, x = new_k_cache_1_cast_fp16)[name = tensor("cast_142")]; + } -> (transformer_out, is_eos, new_k_cache0, new_v_cache0, new_position0, new_k_cache1, new_v_cache1, new_position1, new_k_cache2, new_v_cache2, new_position2, new_k_cache3, new_v_cache3, new_position3, new_k_cache4, new_v_cache4, new_position4, new_k_cache5, new_v_cache5, new_position5); +} \ No newline at end of file diff --git a/v2.1/spanish/flowlm_step_ane.mlmodelc/weights/weight.bin b/v2.1/spanish/flowlm_step_ane.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..f07e8f98a7be36bb99378a9d6d438f1a669a8284 --- /dev/null +++ b/v2.1/spanish/flowlm_step_ane.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8f043a6fda69cf1234532e01d7f0fb1020a78673ae18a6f33094da2f4d283c6d +size 151136960 diff --git a/v2.1/spanish/flowlm_step_ane.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/spanish/flowlm_step_ane.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..cdb1717c468367cfc23062bb2e18890a77480dc1 --- /dev/null +++ b/v2.1/spanish/flowlm_step_ane.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:99485c3c072c06d878dea941fe18f349fed502850b24fff2d4524b0340a9c18b +size 137731 diff --git a/v2.1/spanish/flowlm_step_ane.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/spanish/flowlm_step_ane.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..f07e8f98a7be36bb99378a9d6d438f1a669a8284 --- /dev/null +++ b/v2.1/spanish/flowlm_step_ane.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8f043a6fda69cf1234532e01d7f0fb1020a78673ae18a6f33094da2f4d283c6d +size 151136960 diff --git a/v2.1/spanish/flowlm_step_ane.mlpackage/Manifest.json b/v2.1/spanish/flowlm_step_ane.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..55f28fb9c995da32f379a242fb9ace93514e16ed --- /dev/null +++ b/v2.1/spanish/flowlm_step_ane.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "09BA8E5A-8F98-43AF-BC97-4A0B374557BA": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "4C49894B-B4CF-4F95-9FBA-3F23547079A9": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "09BA8E5A-8F98-43AF-BC97-4A0B374557BA" +} diff --git a/v2.1/spanish_24l/cond_prefill_ane.mlmodelc/analytics/coremldata.bin b/v2.1/spanish_24l/cond_prefill_ane.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..be1f6f77b89744fa10deb2a1677da327f91569f0 --- /dev/null +++ b/v2.1/spanish_24l/cond_prefill_ane.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4acfc96a614cd09e855179801189877f26c4eaefd093f2b58f6a0f2a281cf290 +size 243 diff --git a/v2.1/spanish_24l/cond_prefill_ane.mlmodelc/coremldata.bin b/v2.1/spanish_24l/cond_prefill_ane.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..5c57493adff56b5584b403baed10bddc4fd3a033 --- /dev/null +++ b/v2.1/spanish_24l/cond_prefill_ane.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a6c64f9d45d385d9ddd035022bb685dbbd5d86ed472f71c6d3fd01b0eca2bc6e +size 4496 diff --git a/v2.1/spanish_24l/cond_prefill_ane.mlmodelc/model.mil b/v2.1/spanish_24l/cond_prefill_ane.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..f102b6e147b28d002e028bb659c273f9e9c7f4c9 --- /dev/null +++ b/v2.1/spanish_24l/cond_prefill_ane.mlmodelc/model.mil @@ -0,0 +1,3198 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.9.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor conditioning, tensor k_cache0, tensor k_cache1, tensor k_cache10, tensor k_cache11, tensor k_cache12, tensor k_cache13, tensor k_cache14, tensor k_cache15, tensor k_cache16, tensor k_cache17, tensor k_cache18, tensor k_cache19, tensor k_cache2, tensor k_cache20, tensor k_cache21, tensor k_cache22, tensor k_cache23, tensor k_cache3, tensor k_cache4, tensor k_cache5, tensor k_cache6, tensor k_cache7, tensor k_cache8, tensor k_cache9, tensor position0, tensor position1, tensor position10, tensor position11, tensor position12, tensor position13, tensor position14, tensor position15, tensor position16, tensor position17, tensor position18, tensor position19, tensor position2, tensor position20, tensor position21, tensor position22, tensor position23, tensor position3, tensor position4, tensor position5, tensor position6, tensor position7, tensor position8, tensor position9, tensor v_cache0, tensor v_cache1, tensor v_cache10, tensor v_cache11, tensor v_cache12, tensor v_cache13, tensor v_cache14, tensor v_cache15, tensor v_cache16, tensor v_cache17, tensor v_cache18, tensor v_cache19, tensor v_cache2, tensor v_cache20, tensor v_cache21, tensor v_cache22, tensor v_cache23, tensor v_cache3, tensor v_cache4, tensor v_cache5, tensor v_cache6, tensor v_cache7, tensor v_cache8, tensor v_cache9, tensor valid_len) { + tensor token_indices_to_fp16 = const()[name = tensor("token_indices_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor position0_to_fp16_dtype_0 = const()[name = tensor("position0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position0_to_fp16 = cast(dtype = position0_to_fp16_dtype_0, x = position0)[name = tensor("cast_470")]; + tensor ts_cast_fp16 = add(x = token_indices_to_fp16, y = position0_to_fp16)[name = tensor("ts_cast_fp16")]; + tensor var_231 = const()[name = tensor("op_231"), val = tensor([256, 1, 1])]; + tensor var_232_cast_fp16 = reshape(shape = var_231, x = ts_cast_fp16)[name = tensor("op_232_cast_fp16")]; + tensor var_227_to_fp16 = const()[name = tensor("op_227_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(640)))]; + tensor angles_cast_fp16 = mul(x = var_227_to_fp16, y = var_232_cast_fp16)[name = tensor("angles_cast_fp16")]; + tensor rotr_cast_fp16 = cos(x = angles_cast_fp16)[name = tensor("rotr_cast_fp16")]; + tensor roti_cast_fp16 = sin(x = angles_cast_fp16)[name = tensor("roti_cast_fp16")]; + tensor _inversed_237_y_0_to_fp16 = const()[name = tensor("_inversed_237_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_237_cast_fp16 = mul(x = ts_cast_fp16, y = _inversed_237_y_0_to_fp16)[name = tensor("_inversed_237_cast_fp16")]; + tensor var_238_cast_fp16 = floor(x = _inversed_237_cast_fp16)[name = tensor("op_238_cast_fp16")]; + tensor var_239_to_fp16 = const()[name = tensor("op_239_to_fp16"), val = tensor(0x1p+9)]; + tensor var_240_cast_fp16 = mul(x = var_238_cast_fp16, y = var_239_to_fp16)[name = tensor("op_240_cast_fp16")]; + tensor write_idx_1_cast_fp16 = sub(x = ts_cast_fp16, y = var_240_cast_fp16)[name = tensor("write_idx_1_cast_fp16")]; + tensor valid_len_to_fp16_dtype_0 = const()[name = tensor("valid_len_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor valid_len_to_fp16 = cast(dtype = valid_len_to_fp16_dtype_0, x = valid_len)[name = tensor("cast_469")]; + tensor var_246_cast_fp16 = less(x = token_indices_to_fp16, y = valid_len_to_fp16)[name = tensor("op_246_cast_fp16")]; + tensor token_valid_to_fp16_dtype_0 = const()[name = tensor("token_valid_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_246_cast_fp16_to_fp16 = cast(dtype = token_valid_to_fp16_dtype_0, x = var_246_cast_fp16)[name = tensor("cast_468")]; + tensor var_252_cast_fp16 = mul(x = write_idx_1_cast_fp16, y = var_246_cast_fp16_to_fp16)[name = tensor("op_252_cast_fp16")]; + tensor var_253_to_fp16 = const()[name = tensor("op_253_to_fp16"), val = tensor(0x1p+0)]; + tensor var_255_cast_fp16 = sub(x = var_253_to_fp16, y = var_246_cast_fp16_to_fp16)[name = tensor("op_255_cast_fp16")]; + tensor var_256_to_fp16 = const()[name = tensor("op_256_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_257_cast_fp16 = mul(x = var_255_cast_fp16, y = var_256_to_fp16)[name = tensor("op_257_cast_fp16")]; + tensor write_idx_cast_fp16 = add(x = var_252_cast_fp16, y = var_257_cast_fp16)[name = tensor("write_idx_cast_fp16")]; + tensor var_262 = const()[name = tensor("op_262"), val = tensor([256, 1])]; + tensor var_263_cast_fp16 = reshape(shape = var_262, x = write_idx_cast_fp16)[name = tensor("op_263_cast_fp16")]; + tensor var_267_to_fp16 = const()[name = tensor("op_267_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(768)))]; + tensor var_268_cast_fp16 = equal(x = var_263_cast_fp16, y = var_267_to_fp16)[name = tensor("op_268_cast_fp16")]; + tensor var_278_axes_0 = const()[name = tensor("op_278_axes_0"), val = tensor([0])]; + tensor var_278_keep_dims_0 = const()[name = tensor("op_278_keep_dims_0"), val = tensor(false)]; + tensor assign_to_fp16_dtype_0 = const()[name = tensor("assign_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_268_cast_fp16_to_fp16 = cast(dtype = assign_to_fp16_dtype_0, x = var_268_cast_fp16)[name = tensor("cast_467")]; + tensor var_278_cast_fp16 = reduce_sum(axes = var_278_axes_0, keep_dims = var_278_keep_dims_0, x = var_268_cast_fp16_to_fp16)[name = tensor("op_278_cast_fp16")]; + tensor const_0_to_fp16 = const()[name = tensor("const_0_to_fp16"), val = tensor(-inf)]; + tensor var_280_to_fp16 = const()[name = tensor("op_280_to_fp16"), val = tensor(0x1p+0)]; + tensor clip_0_cast_fp16 = clip(alpha = const_0_to_fp16, beta = var_280_to_fp16, x = var_278_cast_fp16)[name = tensor("clip_0_cast_fp16")]; + tensor var_284 = const()[name = tensor("op_284"), val = tensor([1, 1])]; + tensor var_285_cast_fp16 = reshape(shape = var_284, x = position0_to_fp16)[name = tensor("op_285_cast_fp16")]; + tensor var_289_to_fp16 = const()[name = tensor("op_289_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1856)))]; + tensor q_positions_cast_fp16 = add(x = var_285_cast_fp16, y = var_289_to_fp16)[name = tensor("q_positions_cast_fp16")]; + tensor var_297_cast_fp16 = add(x = position0_to_fp16, y = valid_len_to_fp16)[name = tensor("op_297_cast_fp16")]; + tensor var_300 = const()[name = tensor("op_300"), val = tensor([1, 1])]; + tensor valid_end_cast_fp16 = reshape(shape = var_300, x = var_297_cast_fp16)[name = tensor("valid_end_cast_fp16")]; + tensor var_302_cast_fp16 = less(x = var_267_to_fp16, y = valid_end_cast_fp16)[name = tensor("op_302_cast_fp16")]; + tensor var_303_cast_fp16 = less_equal(x = var_267_to_fp16, y = q_positions_cast_fp16)[name = tensor("op_303_cast_fp16")]; + tensor var_304 = logical_and(x = var_302_cast_fp16, y = var_303_cast_fp16)[name = tensor("op_304")]; + tensor mask_to_fp16_dtype_0 = const()[name = tensor("mask_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_311_to_fp16 = const()[name = tensor("op_311_to_fp16"), val = tensor(0x1p+0)]; + tensor var_304_to_fp16 = cast(dtype = mask_to_fp16_dtype_0, x = var_304)[name = tensor("cast_466")]; + tensor var_312_cast_fp16 = sub(x = var_304_to_fp16, y = var_311_to_fp16)[name = tensor("op_312_cast_fp16")]; + tensor var_313_to_fp16 = const()[name = tensor("op_313_to_fp16"), val = tensor(0x1.388p+13)]; + tensor mask_bias_cast_fp16 = mul(x = var_312_cast_fp16, y = var_313_to_fp16)[name = tensor("mask_bias_cast_fp16")]; + tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; + tensor conditioning_to_fp16_dtype_0 = const()[name = tensor("conditioning_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor norm0_1_weight_to_fp16 = const()[name = tensor("norm0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2432)))]; + tensor norm0_1_bias_to_fp16 = const()[name = tensor("norm0_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4544)))]; + tensor var_316_to_fp16 = const()[name = tensor("op_316_to_fp16"), val = tensor(0x1.5p-17)]; + tensor conditioning_to_fp16 = cast(dtype = conditioning_to_fp16_dtype_0, x = conditioning)[name = tensor("cast_465")]; + tensor input_3_cast_fp16 = layer_norm(axes = input_3_axes_0, beta = norm0_1_bias_to_fp16, epsilon = var_316_to_fp16, gamma = norm0_1_weight_to_fp16, x = conditioning_to_fp16)[name = tensor("input_3_cast_fp16")]; + tensor attn0_in_proj_weight_to_fp16 = const()[name = tensor("attn0_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6656)))]; + tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6298176)))]; + tensor linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn0_in_proj_weight_to_fp16, x = input_3_cast_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor var_330 = const()[name = tensor("op_330"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_1_cast_fp16 = reshape(shape = var_330, x = linear_0_cast_fp16)[name = tensor("qkv_1_cast_fp16")]; + tensor x_1_begin_0 = const()[name = tensor("x_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_1_end_0 = const()[name = tensor("x_1_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_1_end_mask_0 = const()[name = tensor("x_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_1_squeeze_mask_0 = const()[name = tensor("x_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_1_cast_fp16 = slice_by_index(begin = x_1_begin_0, end = x_1_end_0, end_mask = x_1_end_mask_0, squeeze_mask = x_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("x_1_cast_fp16")]; + tensor x_3_begin_0 = const()[name = tensor("x_3_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_3_end_0 = const()[name = tensor("x_3_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_3_end_mask_0 = const()[name = tensor("x_3_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_3_squeeze_mask_0 = const()[name = tensor("x_3_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_3_cast_fp16 = slice_by_index(begin = x_3_begin_0, end = x_3_end_0, end_mask = x_3_end_mask_0, squeeze_mask = x_3_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("x_3_cast_fp16")]; + tensor v_1_begin_0 = const()[name = tensor("v_1_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_1_end_0 = const()[name = tensor("v_1_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_1_end_mask_0 = const()[name = tensor("v_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_1_squeeze_mask_0 = const()[name = tensor("v_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_1_cast_fp16 = slice_by_index(begin = v_1_begin_0, end = v_1_end_0, end_mask = v_1_end_mask_0, squeeze_mask = v_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("v_1_cast_fp16")]; + tensor var_375 = const()[name = tensor("op_375"), val = tensor([256, 16, 32, 2])]; + tensor pairs_1_cast_fp16 = reshape(shape = var_375, x = x_1_cast_fp16)[name = tensor("pairs_1_cast_fp16")]; + tensor xr_1_begin_0 = const()[name = tensor("xr_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_1_end_0 = const()[name = tensor("xr_1_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_1_end_mask_0 = const()[name = tensor("xr_1_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_1_squeeze_mask_0 = const()[name = tensor("xr_1_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_1_cast_fp16 = slice_by_index(begin = xr_1_begin_0, end = xr_1_end_0, end_mask = xr_1_end_mask_0, squeeze_mask = xr_1_squeeze_mask_0, x = pairs_1_cast_fp16)[name = tensor("xr_1_cast_fp16")]; + tensor xi_1_begin_0 = const()[name = tensor("xi_1_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_1_end_0 = const()[name = tensor("xi_1_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_1_end_mask_0 = const()[name = tensor("xi_1_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_1_squeeze_mask_0 = const()[name = tensor("xi_1_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_1_cast_fp16 = slice_by_index(begin = xi_1_begin_0, end = xi_1_end_0, end_mask = xi_1_end_mask_0, squeeze_mask = xi_1_squeeze_mask_0, x = pairs_1_cast_fp16)[name = tensor("xi_1_cast_fp16")]; + tensor var_383_cast_fp16 = mul(x = xr_1_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_383_cast_fp16")]; + tensor var_384_cast_fp16 = mul(x = xi_1_cast_fp16, y = roti_cast_fp16)[name = tensor("op_384_cast_fp16")]; + tensor xor__1_cast_fp16 = sub(x = var_383_cast_fp16, y = var_384_cast_fp16)[name = tensor("xor__1_cast_fp16")]; + tensor var_387_cast_fp16 = mul(x = xr_1_cast_fp16, y = roti_cast_fp16)[name = tensor("op_387_cast_fp16")]; + tensor var_388_cast_fp16 = mul(x = xi_1_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_388_cast_fp16")]; + tensor xoi_1_cast_fp16 = add(x = var_387_cast_fp16, y = var_388_cast_fp16)[name = tensor("xoi_1_cast_fp16")]; + tensor out_1_axis_0 = const()[name = tensor("out_1_axis_0"), val = tensor(-1)]; + tensor out_1_cast_fp16 = stack(axis = out_1_axis_0, values = (xor__1_cast_fp16, xoi_1_cast_fp16))[name = tensor("out_1_cast_fp16")]; + tensor var_398 = const()[name = tensor("op_398"), val = tensor([1, 256, 16, 64])]; + tensor q_1_cast_fp16 = reshape(shape = var_398, x = out_1_cast_fp16)[name = tensor("q_1_cast_fp16")]; + tensor var_404 = const()[name = tensor("op_404"), val = tensor([256, 16, 32, 2])]; + tensor pairs_3_cast_fp16 = reshape(shape = var_404, x = x_3_cast_fp16)[name = tensor("pairs_3_cast_fp16")]; + tensor xr_3_begin_0 = const()[name = tensor("xr_3_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_3_end_0 = const()[name = tensor("xr_3_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_3_end_mask_0 = const()[name = tensor("xr_3_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_3_squeeze_mask_0 = const()[name = tensor("xr_3_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_3_cast_fp16 = slice_by_index(begin = xr_3_begin_0, end = xr_3_end_0, end_mask = xr_3_end_mask_0, squeeze_mask = xr_3_squeeze_mask_0, x = pairs_3_cast_fp16)[name = tensor("xr_3_cast_fp16")]; + tensor xi_3_begin_0 = const()[name = tensor("xi_3_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_3_end_0 = const()[name = tensor("xi_3_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_3_end_mask_0 = const()[name = tensor("xi_3_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_3_squeeze_mask_0 = const()[name = tensor("xi_3_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_3_cast_fp16 = slice_by_index(begin = xi_3_begin_0, end = xi_3_end_0, end_mask = xi_3_end_mask_0, squeeze_mask = xi_3_squeeze_mask_0, x = pairs_3_cast_fp16)[name = tensor("xi_3_cast_fp16")]; + tensor var_412_cast_fp16 = mul(x = xr_3_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_412_cast_fp16")]; + tensor var_413_cast_fp16 = mul(x = xi_3_cast_fp16, y = roti_cast_fp16)[name = tensor("op_413_cast_fp16")]; + tensor xor__3_cast_fp16 = sub(x = var_412_cast_fp16, y = var_413_cast_fp16)[name = tensor("xor__3_cast_fp16")]; + tensor var_416_cast_fp16 = mul(x = xr_3_cast_fp16, y = roti_cast_fp16)[name = tensor("op_416_cast_fp16")]; + tensor var_417_cast_fp16 = mul(x = xi_3_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_417_cast_fp16")]; + tensor xoi_3_cast_fp16 = add(x = var_416_cast_fp16, y = var_417_cast_fp16)[name = tensor("xoi_3_cast_fp16")]; + tensor out_3_axis_0 = const()[name = tensor("out_3_axis_0"), val = tensor(-1)]; + tensor out_3_cast_fp16 = stack(axis = out_3_axis_0, values = (xor__3_cast_fp16, xoi_3_cast_fp16))[name = tensor("out_3_cast_fp16")]; + tensor var_434 = const()[name = tensor("op_434"), val = tensor([256, 1024])]; + tensor var_435_cast_fp16 = reshape(shape = var_434, x = out_3_cast_fp16)[name = tensor("op_435_cast_fp16")]; + tensor written_k_1_transpose_x_1 = const()[name = tensor("written_k_1_transpose_x_1"), val = tensor(true)]; + tensor written_k_1_transpose_y_1 = const()[name = tensor("written_k_1_transpose_y_1"), val = tensor(false)]; + tensor written_k_1_cast_fp16 = matmul(transpose_x = written_k_1_transpose_x_1, transpose_y = written_k_1_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_435_cast_fp16)[name = tensor("written_k_1_cast_fp16")]; + tensor var_442 = const()[name = tensor("op_442"), val = tensor([256, 1024])]; + tensor var_443_cast_fp16 = reshape(shape = var_442, x = v_1_cast_fp16)[name = tensor("op_443_cast_fp16")]; + tensor written_v_1_transpose_x_1 = const()[name = tensor("written_v_1_transpose_x_1"), val = tensor(true)]; + tensor written_v_1_transpose_y_1 = const()[name = tensor("written_v_1_transpose_y_1"), val = tensor(false)]; + tensor written_v_1_cast_fp16 = matmul(transpose_x = written_v_1_transpose_x_1, transpose_y = written_v_1_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_443_cast_fp16)[name = tensor("written_v_1_cast_fp16")]; + tensor var_445_to_fp16 = const()[name = tensor("op_445_to_fp16"), val = tensor(0x1p+0)]; + tensor var_447_cast_fp16 = sub(x = var_445_to_fp16, y = clip_0_cast_fp16)[name = tensor("op_447_cast_fp16")]; + tensor var_452 = const()[name = tensor("op_452"), val = tensor([1, 512, 1, 1])]; + tensor keep_1_cast_fp16 = reshape(shape = var_452, x = var_447_cast_fp16)[name = tensor("keep_1_cast_fp16")]; + tensor k_cache0_to_fp16_dtype_0 = const()[name = tensor("k_cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache0_to_fp16 = cast(dtype = k_cache0_to_fp16_dtype_0, x = k_cache0)[name = tensor("cast_464")]; + tensor var_454_cast_fp16 = mul(x = k_cache0_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_454_cast_fp16")]; + tensor var_459 = const()[name = tensor("op_459"), val = tensor([1, 512, 16, 64])]; + tensor var_460_cast_fp16 = reshape(shape = var_459, x = written_k_1_cast_fp16)[name = tensor("op_460_cast_fp16")]; + tensor new_k_cache_1_cast_fp16 = add(x = var_454_cast_fp16, y = var_460_cast_fp16)[name = tensor("new_k_cache_1_cast_fp16")]; + tensor new_k_cache_1_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_1_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache0_to_fp16_dtype_0 = const()[name = tensor("v_cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache0_to_fp16 = cast(dtype = v_cache0_to_fp16_dtype_0, x = v_cache0)[name = tensor("cast_462")]; + tensor var_463_cast_fp16 = mul(x = v_cache0_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_463_cast_fp16")]; + tensor var_468 = const()[name = tensor("op_468"), val = tensor([1, 512, 16, 64])]; + tensor var_469_cast_fp16 = reshape(shape = var_468, x = written_v_1_cast_fp16)[name = tensor("op_469_cast_fp16")]; + tensor new_v_cache_1_cast_fp16 = add(x = var_463_cast_fp16, y = var_469_cast_fp16)[name = tensor("new_v_cache_1_cast_fp16")]; + tensor new_v_cache_1_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_1_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_482 = const()[name = tensor("op_482"), val = tensor([0, 2, 1, 3])]; + tensor var_493_transpose_x_0 = const()[name = tensor("op_493_transpose_x_0"), val = tensor(false)]; + tensor var_493_transpose_y_0 = const()[name = tensor("op_493_transpose_y_0"), val = tensor(false)]; + tensor transpose_69_perm_0 = const()[name = tensor("transpose_69_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_70_perm_0 = const()[name = tensor("transpose_70_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_70 = transpose(perm = transpose_70_perm_0, x = new_k_cache_1_cast_fp16)[name = tensor("transpose_204")]; + tensor transpose_69 = transpose(perm = transpose_69_perm_0, x = q_1_cast_fp16)[name = tensor("transpose_205")]; + tensor var_493_cast_fp16 = matmul(transpose_x = var_493_transpose_x_0, transpose_y = var_493_transpose_y_0, x = transpose_69, y = transpose_70)[name = tensor("op_493_cast_fp16")]; + tensor var_494_to_fp16 = const()[name = tensor("op_494_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_1_cast_fp16 = mul(x = var_493_cast_fp16, y = var_494_to_fp16)[name = tensor("scores_1_cast_fp16")]; + tensor var_500 = const()[name = tensor("op_500"), val = tensor([1, 1, 256, 512])]; + tensor var_501_cast_fp16 = reshape(shape = var_500, x = mask_bias_cast_fp16)[name = tensor("op_501_cast_fp16")]; + tensor scores_3_cast_fp16 = add(x = scores_1_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_3_cast_fp16")]; + tensor var_504 = const()[name = tensor("op_504"), val = tensor(-1)]; + tensor weights_1_cast_fp16 = softmax(axis = var_504, x = scores_3_cast_fp16)[name = tensor("weights_1_cast_fp16")]; + tensor attn_1_transpose_x_0 = const()[name = tensor("attn_1_transpose_x_0"), val = tensor(false)]; + tensor attn_1_transpose_y_0 = const()[name = tensor("attn_1_transpose_y_0"), val = tensor(false)]; + tensor values_1_cast_fp16 = transpose(perm = var_482, x = new_v_cache_1_cast_fp16)[name = tensor("transpose_206")]; + tensor attn_1_cast_fp16 = matmul(transpose_x = attn_1_transpose_x_0, transpose_y = attn_1_transpose_y_0, x = weights_1_cast_fp16, y = values_1_cast_fp16)[name = tensor("attn_1_cast_fp16")]; + tensor var_512 = const()[name = tensor("op_512"), val = tensor([0, 2, 1, 3])]; + tensor var_517 = const()[name = tensor("op_517"), val = tensor([1, 256, 1024])]; + tensor var_513_cast_fp16 = transpose(perm = var_512, x = attn_1_cast_fp16)[name = tensor("transpose_203")]; + tensor input_5_cast_fp16 = reshape(shape = var_517, x = var_513_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor attn0_out_proj_weight_to_fp16 = const()[name = tensor("attn0_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6304384)))]; + tensor linear_1_bias_0_to_fp16 = const()[name = tensor("linear_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401600)))]; + tensor linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn0_out_proj_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("linear_1_cast_fp16")]; + tensor var_526_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_526_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_7_cast_fp16 = add(x = conditioning_to_fp16, y = linear_1_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor input_9_axes_0 = const()[name = tensor("input_9_axes_0"), val = tensor([-1])]; + tensor norm0_2_weight_to_fp16 = const()[name = tensor("norm0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8403712)))]; + tensor norm0_2_bias_to_fp16 = const()[name = tensor("norm0_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8405824)))]; + tensor var_530_to_fp16 = const()[name = tensor("op_530_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_9_cast_fp16 = layer_norm(axes = input_9_axes_0, beta = norm0_2_bias_to_fp16, epsilon = var_530_to_fp16, gamma = norm0_2_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor linear0_1_weight_to_fp16 = const()[name = tensor("linear0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8407936)))]; + tensor linear_2_bias_0_to_fp16 = const()[name = tensor("linear_2_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16796608)))]; + tensor linear_2_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear0_1_weight_to_fp16, x = input_9_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor input_11_mode_0 = const()[name = tensor("input_11_mode_0"), val = tensor("EXACT")]; + tensor input_11_cast_fp16 = gelu(mode = input_11_mode_0, x = linear_2_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor linear0_2_weight_to_fp16 = const()[name = tensor("linear0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16804864)))]; + tensor linear_3_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear0_2_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("linear_3_cast_fp16")]; + tensor input_13_cast_fp16 = add(x = input_7_cast_fp16, y = linear_3_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor input_15_axes_0 = const()[name = tensor("input_15_axes_0"), val = tensor([-1])]; + tensor norm1_1_weight_to_fp16 = const()[name = tensor("norm1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25193536)))]; + tensor norm1_1_bias_to_fp16 = const()[name = tensor("norm1_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25195648)))]; + tensor var_547_to_fp16 = const()[name = tensor("op_547_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_15_cast_fp16 = layer_norm(axes = input_15_axes_0, beta = norm1_1_bias_to_fp16, epsilon = var_547_to_fp16, gamma = norm1_1_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor attn1_in_proj_weight_to_fp16 = const()[name = tensor("attn1_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25197760)))]; + tensor linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn1_in_proj_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor var_561 = const()[name = tensor("op_561"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_3_cast_fp16 = reshape(shape = var_561, x = linear_4_cast_fp16)[name = tensor("qkv_3_cast_fp16")]; + tensor x_5_begin_0 = const()[name = tensor("x_5_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_5_end_0 = const()[name = tensor("x_5_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_5_end_mask_0 = const()[name = tensor("x_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_5_squeeze_mask_0 = const()[name = tensor("x_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_5_cast_fp16 = slice_by_index(begin = x_5_begin_0, end = x_5_end_0, end_mask = x_5_end_mask_0, squeeze_mask = x_5_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("x_5_cast_fp16")]; + tensor x_7_begin_0 = const()[name = tensor("x_7_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_7_end_0 = const()[name = tensor("x_7_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_7_end_mask_0 = const()[name = tensor("x_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_7_squeeze_mask_0 = const()[name = tensor("x_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_7_cast_fp16 = slice_by_index(begin = x_7_begin_0, end = x_7_end_0, end_mask = x_7_end_mask_0, squeeze_mask = x_7_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("x_7_cast_fp16")]; + tensor v_3_begin_0 = const()[name = tensor("v_3_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_3_end_0 = const()[name = tensor("v_3_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_3_end_mask_0 = const()[name = tensor("v_3_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_3_squeeze_mask_0 = const()[name = tensor("v_3_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_3_cast_fp16 = slice_by_index(begin = v_3_begin_0, end = v_3_end_0, end_mask = v_3_end_mask_0, squeeze_mask = v_3_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("v_3_cast_fp16")]; + tensor var_606 = const()[name = tensor("op_606"), val = tensor([256, 16, 32, 2])]; + tensor pairs_5_cast_fp16 = reshape(shape = var_606, x = x_5_cast_fp16)[name = tensor("pairs_5_cast_fp16")]; + tensor xr_5_begin_0 = const()[name = tensor("xr_5_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_5_end_0 = const()[name = tensor("xr_5_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_5_end_mask_0 = const()[name = tensor("xr_5_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_5_squeeze_mask_0 = const()[name = tensor("xr_5_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_5_cast_fp16 = slice_by_index(begin = xr_5_begin_0, end = xr_5_end_0, end_mask = xr_5_end_mask_0, squeeze_mask = xr_5_squeeze_mask_0, x = pairs_5_cast_fp16)[name = tensor("xr_5_cast_fp16")]; + tensor xi_5_begin_0 = const()[name = tensor("xi_5_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_5_end_0 = const()[name = tensor("xi_5_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_5_end_mask_0 = const()[name = tensor("xi_5_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_5_squeeze_mask_0 = const()[name = tensor("xi_5_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_5_cast_fp16 = slice_by_index(begin = xi_5_begin_0, end = xi_5_end_0, end_mask = xi_5_end_mask_0, squeeze_mask = xi_5_squeeze_mask_0, x = pairs_5_cast_fp16)[name = tensor("xi_5_cast_fp16")]; + tensor var_614_cast_fp16 = mul(x = xr_5_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_614_cast_fp16")]; + tensor var_615_cast_fp16 = mul(x = xi_5_cast_fp16, y = roti_cast_fp16)[name = tensor("op_615_cast_fp16")]; + tensor xor__5_cast_fp16 = sub(x = var_614_cast_fp16, y = var_615_cast_fp16)[name = tensor("xor__5_cast_fp16")]; + tensor var_618_cast_fp16 = mul(x = xr_5_cast_fp16, y = roti_cast_fp16)[name = tensor("op_618_cast_fp16")]; + tensor var_619_cast_fp16 = mul(x = xi_5_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_619_cast_fp16")]; + tensor xoi_5_cast_fp16 = add(x = var_618_cast_fp16, y = var_619_cast_fp16)[name = tensor("xoi_5_cast_fp16")]; + tensor out_5_axis_0 = const()[name = tensor("out_5_axis_0"), val = tensor(-1)]; + tensor out_5_cast_fp16 = stack(axis = out_5_axis_0, values = (xor__5_cast_fp16, xoi_5_cast_fp16))[name = tensor("out_5_cast_fp16")]; + tensor var_629 = const()[name = tensor("op_629"), val = tensor([1, 256, 16, 64])]; + tensor q_3_cast_fp16 = reshape(shape = var_629, x = out_5_cast_fp16)[name = tensor("q_3_cast_fp16")]; + tensor var_635 = const()[name = tensor("op_635"), val = tensor([256, 16, 32, 2])]; + tensor pairs_7_cast_fp16 = reshape(shape = var_635, x = x_7_cast_fp16)[name = tensor("pairs_7_cast_fp16")]; + tensor xr_7_begin_0 = const()[name = tensor("xr_7_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_7_end_0 = const()[name = tensor("xr_7_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_7_end_mask_0 = const()[name = tensor("xr_7_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_7_squeeze_mask_0 = const()[name = tensor("xr_7_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_7_cast_fp16 = slice_by_index(begin = xr_7_begin_0, end = xr_7_end_0, end_mask = xr_7_end_mask_0, squeeze_mask = xr_7_squeeze_mask_0, x = pairs_7_cast_fp16)[name = tensor("xr_7_cast_fp16")]; + tensor xi_7_begin_0 = const()[name = tensor("xi_7_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_7_end_0 = const()[name = tensor("xi_7_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_7_end_mask_0 = const()[name = tensor("xi_7_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_7_squeeze_mask_0 = const()[name = tensor("xi_7_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_7_cast_fp16 = slice_by_index(begin = xi_7_begin_0, end = xi_7_end_0, end_mask = xi_7_end_mask_0, squeeze_mask = xi_7_squeeze_mask_0, x = pairs_7_cast_fp16)[name = tensor("xi_7_cast_fp16")]; + tensor var_643_cast_fp16 = mul(x = xr_7_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_643_cast_fp16")]; + tensor var_644_cast_fp16 = mul(x = xi_7_cast_fp16, y = roti_cast_fp16)[name = tensor("op_644_cast_fp16")]; + tensor xor__7_cast_fp16 = sub(x = var_643_cast_fp16, y = var_644_cast_fp16)[name = tensor("xor__7_cast_fp16")]; + tensor var_647_cast_fp16 = mul(x = xr_7_cast_fp16, y = roti_cast_fp16)[name = tensor("op_647_cast_fp16")]; + tensor var_648_cast_fp16 = mul(x = xi_7_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_648_cast_fp16")]; + tensor xoi_7_cast_fp16 = add(x = var_647_cast_fp16, y = var_648_cast_fp16)[name = tensor("xoi_7_cast_fp16")]; + tensor out_7_axis_0 = const()[name = tensor("out_7_axis_0"), val = tensor(-1)]; + tensor out_7_cast_fp16 = stack(axis = out_7_axis_0, values = (xor__7_cast_fp16, xoi_7_cast_fp16))[name = tensor("out_7_cast_fp16")]; + tensor var_665 = const()[name = tensor("op_665"), val = tensor([256, 1024])]; + tensor var_666_cast_fp16 = reshape(shape = var_665, x = out_7_cast_fp16)[name = tensor("op_666_cast_fp16")]; + tensor written_k_3_transpose_x_1 = const()[name = tensor("written_k_3_transpose_x_1"), val = tensor(true)]; + tensor written_k_3_transpose_y_1 = const()[name = tensor("written_k_3_transpose_y_1"), val = tensor(false)]; + tensor written_k_3_cast_fp16 = matmul(transpose_x = written_k_3_transpose_x_1, transpose_y = written_k_3_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_666_cast_fp16)[name = tensor("written_k_3_cast_fp16")]; + tensor var_673 = const()[name = tensor("op_673"), val = tensor([256, 1024])]; + tensor var_674_cast_fp16 = reshape(shape = var_673, x = v_3_cast_fp16)[name = tensor("op_674_cast_fp16")]; + tensor written_v_3_transpose_x_1 = const()[name = tensor("written_v_3_transpose_x_1"), val = tensor(true)]; + tensor written_v_3_transpose_y_1 = const()[name = tensor("written_v_3_transpose_y_1"), val = tensor(false)]; + tensor written_v_3_cast_fp16 = matmul(transpose_x = written_v_3_transpose_x_1, transpose_y = written_v_3_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_674_cast_fp16)[name = tensor("written_v_3_cast_fp16")]; + tensor k_cache1_to_fp16_dtype_0 = const()[name = tensor("k_cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache1_to_fp16 = cast(dtype = k_cache1_to_fp16_dtype_0, x = k_cache1)[name = tensor("cast_459")]; + tensor var_685_cast_fp16 = mul(x = k_cache1_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_685_cast_fp16")]; + tensor var_690 = const()[name = tensor("op_690"), val = tensor([1, 512, 16, 64])]; + tensor var_691_cast_fp16 = reshape(shape = var_690, x = written_k_3_cast_fp16)[name = tensor("op_691_cast_fp16")]; + tensor new_k_cache_3_cast_fp16 = add(x = var_685_cast_fp16, y = var_691_cast_fp16)[name = tensor("new_k_cache_3_cast_fp16")]; + tensor new_k_cache_3_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_3_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache1_to_fp16_dtype_0 = const()[name = tensor("v_cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache1_to_fp16 = cast(dtype = v_cache1_to_fp16_dtype_0, x = v_cache1)[name = tensor("cast_457")]; + tensor var_694_cast_fp16 = mul(x = v_cache1_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_694_cast_fp16")]; + tensor var_699 = const()[name = tensor("op_699"), val = tensor([1, 512, 16, 64])]; + tensor var_700_cast_fp16 = reshape(shape = var_699, x = written_v_3_cast_fp16)[name = tensor("op_700_cast_fp16")]; + tensor new_v_cache_3_cast_fp16 = add(x = var_694_cast_fp16, y = var_700_cast_fp16)[name = tensor("new_v_cache_3_cast_fp16")]; + tensor new_v_cache_3_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_3_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_713 = const()[name = tensor("op_713"), val = tensor([0, 2, 1, 3])]; + tensor var_724_transpose_x_0 = const()[name = tensor("op_724_transpose_x_0"), val = tensor(false)]; + tensor var_724_transpose_y_0 = const()[name = tensor("op_724_transpose_y_0"), val = tensor(false)]; + tensor transpose_71_perm_0 = const()[name = tensor("transpose_71_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_72_perm_0 = const()[name = tensor("transpose_72_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_72 = transpose(perm = transpose_72_perm_0, x = new_k_cache_3_cast_fp16)[name = tensor("transpose_200")]; + tensor transpose_71 = transpose(perm = transpose_71_perm_0, x = q_3_cast_fp16)[name = tensor("transpose_201")]; + tensor var_724_cast_fp16 = matmul(transpose_x = var_724_transpose_x_0, transpose_y = var_724_transpose_y_0, x = transpose_71, y = transpose_72)[name = tensor("op_724_cast_fp16")]; + tensor var_725_to_fp16 = const()[name = tensor("op_725_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_5_cast_fp16 = mul(x = var_724_cast_fp16, y = var_725_to_fp16)[name = tensor("scores_5_cast_fp16")]; + tensor scores_7_cast_fp16 = add(x = scores_5_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_7_cast_fp16")]; + tensor var_735 = const()[name = tensor("op_735"), val = tensor(-1)]; + tensor weights_3_cast_fp16 = softmax(axis = var_735, x = scores_7_cast_fp16)[name = tensor("weights_3_cast_fp16")]; + tensor attn_3_transpose_x_0 = const()[name = tensor("attn_3_transpose_x_0"), val = tensor(false)]; + tensor attn_3_transpose_y_0 = const()[name = tensor("attn_3_transpose_y_0"), val = tensor(false)]; + tensor values_3_cast_fp16 = transpose(perm = var_713, x = new_v_cache_3_cast_fp16)[name = tensor("transpose_202")]; + tensor attn_3_cast_fp16 = matmul(transpose_x = attn_3_transpose_x_0, transpose_y = attn_3_transpose_y_0, x = weights_3_cast_fp16, y = values_3_cast_fp16)[name = tensor("attn_3_cast_fp16")]; + tensor var_743 = const()[name = tensor("op_743"), val = tensor([0, 2, 1, 3])]; + tensor var_748 = const()[name = tensor("op_748"), val = tensor([1, 256, 1024])]; + tensor var_744_cast_fp16 = transpose(perm = var_743, x = attn_3_cast_fp16)[name = tensor("transpose_199")]; + tensor input_17_cast_fp16 = reshape(shape = var_748, x = var_744_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor attn1_out_proj_weight_to_fp16 = const()[name = tensor("attn1_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31489280)))]; + tensor linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn1_out_proj_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor position1_to_fp16_dtype_0 = const()[name = tensor("position1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position1_to_fp16 = cast(dtype = position1_to_fp16_dtype_0, x = position1)[name = tensor("cast_455")]; + tensor var_757_cast_fp16 = add(x = position1_to_fp16, y = valid_len_to_fp16)[name = tensor("op_757_cast_fp16")]; + tensor var_757_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_757_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_19_cast_fp16 = add(x = input_13_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor input_21_axes_0 = const()[name = tensor("input_21_axes_0"), val = tensor([-1])]; + tensor norm1_2_weight_to_fp16 = const()[name = tensor("norm1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33586496)))]; + tensor norm1_2_bias_to_fp16 = const()[name = tensor("norm1_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33588608)))]; + tensor var_761_to_fp16 = const()[name = tensor("op_761_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_21_cast_fp16 = layer_norm(axes = input_21_axes_0, beta = norm1_2_bias_to_fp16, epsilon = var_761_to_fp16, gamma = norm1_2_weight_to_fp16, x = input_19_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor linear1_1_weight_to_fp16 = const()[name = tensor("linear1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33590720)))]; + tensor linear_6_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear1_1_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor input_23_mode_0 = const()[name = tensor("input_23_mode_0"), val = tensor("EXACT")]; + tensor input_23_cast_fp16 = gelu(mode = input_23_mode_0, x = linear_6_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor linear1_2_weight_to_fp16 = const()[name = tensor("linear1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41979392)))]; + tensor linear_7_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear1_2_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_25_cast_fp16 = add(x = input_19_cast_fp16, y = linear_7_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor input_27_axes_0 = const()[name = tensor("input_27_axes_0"), val = tensor([-1])]; + tensor norm2_1_weight_to_fp16 = const()[name = tensor("norm2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50368064)))]; + tensor norm2_1_bias_to_fp16 = const()[name = tensor("norm2_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50370176)))]; + tensor var_778_to_fp16 = const()[name = tensor("op_778_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = norm2_1_bias_to_fp16, epsilon = var_778_to_fp16, gamma = norm2_1_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor attn2_in_proj_weight_to_fp16 = const()[name = tensor("attn2_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50372288)))]; + tensor linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn2_in_proj_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor var_792 = const()[name = tensor("op_792"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_5_cast_fp16 = reshape(shape = var_792, x = linear_8_cast_fp16)[name = tensor("qkv_5_cast_fp16")]; + tensor x_9_begin_0 = const()[name = tensor("x_9_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_9_end_0 = const()[name = tensor("x_9_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_9_end_mask_0 = const()[name = tensor("x_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_9_squeeze_mask_0 = const()[name = tensor("x_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_9_cast_fp16 = slice_by_index(begin = x_9_begin_0, end = x_9_end_0, end_mask = x_9_end_mask_0, squeeze_mask = x_9_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("x_9_cast_fp16")]; + tensor x_11_begin_0 = const()[name = tensor("x_11_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_11_end_0 = const()[name = tensor("x_11_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_11_end_mask_0 = const()[name = tensor("x_11_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_11_squeeze_mask_0 = const()[name = tensor("x_11_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_11_cast_fp16 = slice_by_index(begin = x_11_begin_0, end = x_11_end_0, end_mask = x_11_end_mask_0, squeeze_mask = x_11_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("x_11_cast_fp16")]; + tensor v_5_begin_0 = const()[name = tensor("v_5_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_5_end_0 = const()[name = tensor("v_5_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_5_end_mask_0 = const()[name = tensor("v_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_5_squeeze_mask_0 = const()[name = tensor("v_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_5_cast_fp16 = slice_by_index(begin = v_5_begin_0, end = v_5_end_0, end_mask = v_5_end_mask_0, squeeze_mask = v_5_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("v_5_cast_fp16")]; + tensor var_837 = const()[name = tensor("op_837"), val = tensor([256, 16, 32, 2])]; + tensor pairs_9_cast_fp16 = reshape(shape = var_837, x = x_9_cast_fp16)[name = tensor("pairs_9_cast_fp16")]; + tensor xr_9_begin_0 = const()[name = tensor("xr_9_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_9_end_0 = const()[name = tensor("xr_9_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_9_end_mask_0 = const()[name = tensor("xr_9_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_9_squeeze_mask_0 = const()[name = tensor("xr_9_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_9_cast_fp16 = slice_by_index(begin = xr_9_begin_0, end = xr_9_end_0, end_mask = xr_9_end_mask_0, squeeze_mask = xr_9_squeeze_mask_0, x = pairs_9_cast_fp16)[name = tensor("xr_9_cast_fp16")]; + tensor xi_9_begin_0 = const()[name = tensor("xi_9_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_9_end_0 = const()[name = tensor("xi_9_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_9_end_mask_0 = const()[name = tensor("xi_9_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_9_squeeze_mask_0 = const()[name = tensor("xi_9_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_9_cast_fp16 = slice_by_index(begin = xi_9_begin_0, end = xi_9_end_0, end_mask = xi_9_end_mask_0, squeeze_mask = xi_9_squeeze_mask_0, x = pairs_9_cast_fp16)[name = tensor("xi_9_cast_fp16")]; + tensor var_845_cast_fp16 = mul(x = xr_9_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_845_cast_fp16")]; + tensor var_846_cast_fp16 = mul(x = xi_9_cast_fp16, y = roti_cast_fp16)[name = tensor("op_846_cast_fp16")]; + tensor xor__9_cast_fp16 = sub(x = var_845_cast_fp16, y = var_846_cast_fp16)[name = tensor("xor__9_cast_fp16")]; + tensor var_849_cast_fp16 = mul(x = xr_9_cast_fp16, y = roti_cast_fp16)[name = tensor("op_849_cast_fp16")]; + tensor var_850_cast_fp16 = mul(x = xi_9_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_850_cast_fp16")]; + tensor xoi_9_cast_fp16 = add(x = var_849_cast_fp16, y = var_850_cast_fp16)[name = tensor("xoi_9_cast_fp16")]; + tensor out_9_axis_0 = const()[name = tensor("out_9_axis_0"), val = tensor(-1)]; + tensor out_9_cast_fp16 = stack(axis = out_9_axis_0, values = (xor__9_cast_fp16, xoi_9_cast_fp16))[name = tensor("out_9_cast_fp16")]; + tensor var_860 = const()[name = tensor("op_860"), val = tensor([1, 256, 16, 64])]; + tensor q_5_cast_fp16 = reshape(shape = var_860, x = out_9_cast_fp16)[name = tensor("q_5_cast_fp16")]; + tensor var_866 = const()[name = tensor("op_866"), val = tensor([256, 16, 32, 2])]; + tensor pairs_11_cast_fp16 = reshape(shape = var_866, x = x_11_cast_fp16)[name = tensor("pairs_11_cast_fp16")]; + tensor xr_11_begin_0 = const()[name = tensor("xr_11_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_11_end_0 = const()[name = tensor("xr_11_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_11_end_mask_0 = const()[name = tensor("xr_11_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_11_squeeze_mask_0 = const()[name = tensor("xr_11_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_11_cast_fp16 = slice_by_index(begin = xr_11_begin_0, end = xr_11_end_0, end_mask = xr_11_end_mask_0, squeeze_mask = xr_11_squeeze_mask_0, x = pairs_11_cast_fp16)[name = tensor("xr_11_cast_fp16")]; + tensor xi_11_begin_0 = const()[name = tensor("xi_11_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_11_end_0 = const()[name = tensor("xi_11_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_11_end_mask_0 = const()[name = tensor("xi_11_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_11_squeeze_mask_0 = const()[name = tensor("xi_11_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_11_cast_fp16 = slice_by_index(begin = xi_11_begin_0, end = xi_11_end_0, end_mask = xi_11_end_mask_0, squeeze_mask = xi_11_squeeze_mask_0, x = pairs_11_cast_fp16)[name = tensor("xi_11_cast_fp16")]; + tensor var_874_cast_fp16 = mul(x = xr_11_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_874_cast_fp16")]; + tensor var_875_cast_fp16 = mul(x = xi_11_cast_fp16, y = roti_cast_fp16)[name = tensor("op_875_cast_fp16")]; + tensor xor__11_cast_fp16 = sub(x = var_874_cast_fp16, y = var_875_cast_fp16)[name = tensor("xor__11_cast_fp16")]; + tensor var_878_cast_fp16 = mul(x = xr_11_cast_fp16, y = roti_cast_fp16)[name = tensor("op_878_cast_fp16")]; + tensor var_879_cast_fp16 = mul(x = xi_11_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_879_cast_fp16")]; + tensor xoi_11_cast_fp16 = add(x = var_878_cast_fp16, y = var_879_cast_fp16)[name = tensor("xoi_11_cast_fp16")]; + tensor out_11_axis_0 = const()[name = tensor("out_11_axis_0"), val = tensor(-1)]; + tensor out_11_cast_fp16 = stack(axis = out_11_axis_0, values = (xor__11_cast_fp16, xoi_11_cast_fp16))[name = tensor("out_11_cast_fp16")]; + tensor var_896 = const()[name = tensor("op_896"), val = tensor([256, 1024])]; + tensor var_897_cast_fp16 = reshape(shape = var_896, x = out_11_cast_fp16)[name = tensor("op_897_cast_fp16")]; + tensor written_k_5_transpose_x_1 = const()[name = tensor("written_k_5_transpose_x_1"), val = tensor(true)]; + tensor written_k_5_transpose_y_1 = const()[name = tensor("written_k_5_transpose_y_1"), val = tensor(false)]; + tensor written_k_5_cast_fp16 = matmul(transpose_x = written_k_5_transpose_x_1, transpose_y = written_k_5_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_897_cast_fp16)[name = tensor("written_k_5_cast_fp16")]; + tensor var_904 = const()[name = tensor("op_904"), val = tensor([256, 1024])]; + tensor var_905_cast_fp16 = reshape(shape = var_904, x = v_5_cast_fp16)[name = tensor("op_905_cast_fp16")]; + tensor written_v_5_transpose_x_1 = const()[name = tensor("written_v_5_transpose_x_1"), val = tensor(true)]; + tensor written_v_5_transpose_y_1 = const()[name = tensor("written_v_5_transpose_y_1"), val = tensor(false)]; + tensor written_v_5_cast_fp16 = matmul(transpose_x = written_v_5_transpose_x_1, transpose_y = written_v_5_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_905_cast_fp16)[name = tensor("written_v_5_cast_fp16")]; + tensor k_cache2_to_fp16_dtype_0 = const()[name = tensor("k_cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache2_to_fp16 = cast(dtype = k_cache2_to_fp16_dtype_0, x = k_cache2)[name = tensor("cast_453")]; + tensor var_916_cast_fp16 = mul(x = k_cache2_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_916_cast_fp16")]; + tensor var_921 = const()[name = tensor("op_921"), val = tensor([1, 512, 16, 64])]; + tensor var_922_cast_fp16 = reshape(shape = var_921, x = written_k_5_cast_fp16)[name = tensor("op_922_cast_fp16")]; + tensor new_k_cache_5_cast_fp16 = add(x = var_916_cast_fp16, y = var_922_cast_fp16)[name = tensor("new_k_cache_5_cast_fp16")]; + tensor new_k_cache_5_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_5_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache2_to_fp16_dtype_0 = const()[name = tensor("v_cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache2_to_fp16 = cast(dtype = v_cache2_to_fp16_dtype_0, x = v_cache2)[name = tensor("cast_451")]; + tensor var_925_cast_fp16 = mul(x = v_cache2_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_925_cast_fp16")]; + tensor var_930 = const()[name = tensor("op_930"), val = tensor([1, 512, 16, 64])]; + tensor var_931_cast_fp16 = reshape(shape = var_930, x = written_v_5_cast_fp16)[name = tensor("op_931_cast_fp16")]; + tensor new_v_cache_5_cast_fp16 = add(x = var_925_cast_fp16, y = var_931_cast_fp16)[name = tensor("new_v_cache_5_cast_fp16")]; + tensor new_v_cache_5_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_5_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_944 = const()[name = tensor("op_944"), val = tensor([0, 2, 1, 3])]; + tensor var_955_transpose_x_0 = const()[name = tensor("op_955_transpose_x_0"), val = tensor(false)]; + tensor var_955_transpose_y_0 = const()[name = tensor("op_955_transpose_y_0"), val = tensor(false)]; + tensor transpose_73_perm_0 = const()[name = tensor("transpose_73_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_74_perm_0 = const()[name = tensor("transpose_74_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_74 = transpose(perm = transpose_74_perm_0, x = new_k_cache_5_cast_fp16)[name = tensor("transpose_196")]; + tensor transpose_73 = transpose(perm = transpose_73_perm_0, x = q_5_cast_fp16)[name = tensor("transpose_197")]; + tensor var_955_cast_fp16 = matmul(transpose_x = var_955_transpose_x_0, transpose_y = var_955_transpose_y_0, x = transpose_73, y = transpose_74)[name = tensor("op_955_cast_fp16")]; + tensor var_956_to_fp16 = const()[name = tensor("op_956_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_9_cast_fp16 = mul(x = var_955_cast_fp16, y = var_956_to_fp16)[name = tensor("scores_9_cast_fp16")]; + tensor scores_11_cast_fp16 = add(x = scores_9_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_11_cast_fp16")]; + tensor var_966 = const()[name = tensor("op_966"), val = tensor(-1)]; + tensor weights_5_cast_fp16 = softmax(axis = var_966, x = scores_11_cast_fp16)[name = tensor("weights_5_cast_fp16")]; + tensor attn_5_transpose_x_0 = const()[name = tensor("attn_5_transpose_x_0"), val = tensor(false)]; + tensor attn_5_transpose_y_0 = const()[name = tensor("attn_5_transpose_y_0"), val = tensor(false)]; + tensor values_5_cast_fp16 = transpose(perm = var_944, x = new_v_cache_5_cast_fp16)[name = tensor("transpose_198")]; + tensor attn_5_cast_fp16 = matmul(transpose_x = attn_5_transpose_x_0, transpose_y = attn_5_transpose_y_0, x = weights_5_cast_fp16, y = values_5_cast_fp16)[name = tensor("attn_5_cast_fp16")]; + tensor var_974 = const()[name = tensor("op_974"), val = tensor([0, 2, 1, 3])]; + tensor var_979 = const()[name = tensor("op_979"), val = tensor([1, 256, 1024])]; + tensor var_975_cast_fp16 = transpose(perm = var_974, x = attn_5_cast_fp16)[name = tensor("transpose_195")]; + tensor input_29_cast_fp16 = reshape(shape = var_979, x = var_975_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor attn2_out_proj_weight_to_fp16 = const()[name = tensor("attn2_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(56663808)))]; + tensor linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn2_out_proj_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor position2_to_fp16_dtype_0 = const()[name = tensor("position2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position2_to_fp16 = cast(dtype = position2_to_fp16_dtype_0, x = position2)[name = tensor("cast_449")]; + tensor var_988_cast_fp16 = add(x = position2_to_fp16, y = valid_len_to_fp16)[name = tensor("op_988_cast_fp16")]; + tensor var_988_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_988_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_31_cast_fp16 = add(x = input_25_cast_fp16, y = linear_9_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; + tensor norm2_2_weight_to_fp16 = const()[name = tensor("norm2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58761024)))]; + tensor norm2_2_bias_to_fp16 = const()[name = tensor("norm2_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58763136)))]; + tensor var_992_to_fp16 = const()[name = tensor("op_992_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_33_cast_fp16 = layer_norm(axes = input_33_axes_0, beta = norm2_2_bias_to_fp16, epsilon = var_992_to_fp16, gamma = norm2_2_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor linear2_1_weight_to_fp16 = const()[name = tensor("linear2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58765248)))]; + tensor linear_10_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear2_1_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor input_35_mode_0 = const()[name = tensor("input_35_mode_0"), val = tensor("EXACT")]; + tensor input_35_cast_fp16 = gelu(mode = input_35_mode_0, x = linear_10_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor linear2_2_weight_to_fp16 = const()[name = tensor("linear2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67153920)))]; + tensor linear_11_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear2_2_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor input_37_cast_fp16 = add(x = input_31_cast_fp16, y = linear_11_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor input_39_axes_0 = const()[name = tensor("input_39_axes_0"), val = tensor([-1])]; + tensor norm3_1_weight_to_fp16 = const()[name = tensor("norm3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75542592)))]; + tensor norm3_1_bias_to_fp16 = const()[name = tensor("norm3_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75544704)))]; + tensor var_1009_to_fp16 = const()[name = tensor("op_1009_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_39_cast_fp16 = layer_norm(axes = input_39_axes_0, beta = norm3_1_bias_to_fp16, epsilon = var_1009_to_fp16, gamma = norm3_1_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor attn3_in_proj_weight_to_fp16 = const()[name = tensor("attn3_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75546816)))]; + tensor linear_12_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn3_in_proj_weight_to_fp16, x = input_39_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor var_1023 = const()[name = tensor("op_1023"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_7_cast_fp16 = reshape(shape = var_1023, x = linear_12_cast_fp16)[name = tensor("qkv_7_cast_fp16")]; + tensor x_13_begin_0 = const()[name = tensor("x_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_13_end_0 = const()[name = tensor("x_13_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_13_end_mask_0 = const()[name = tensor("x_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_13_squeeze_mask_0 = const()[name = tensor("x_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_13_cast_fp16 = slice_by_index(begin = x_13_begin_0, end = x_13_end_0, end_mask = x_13_end_mask_0, squeeze_mask = x_13_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("x_13_cast_fp16")]; + tensor x_15_begin_0 = const()[name = tensor("x_15_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_15_end_0 = const()[name = tensor("x_15_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_15_end_mask_0 = const()[name = tensor("x_15_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_15_squeeze_mask_0 = const()[name = tensor("x_15_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_15_cast_fp16 = slice_by_index(begin = x_15_begin_0, end = x_15_end_0, end_mask = x_15_end_mask_0, squeeze_mask = x_15_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("x_15_cast_fp16")]; + tensor v_7_begin_0 = const()[name = tensor("v_7_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_7_end_0 = const()[name = tensor("v_7_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_7_end_mask_0 = const()[name = tensor("v_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_7_squeeze_mask_0 = const()[name = tensor("v_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_7_cast_fp16 = slice_by_index(begin = v_7_begin_0, end = v_7_end_0, end_mask = v_7_end_mask_0, squeeze_mask = v_7_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("v_7_cast_fp16")]; + tensor var_1068 = const()[name = tensor("op_1068"), val = tensor([256, 16, 32, 2])]; + tensor pairs_13_cast_fp16 = reshape(shape = var_1068, x = x_13_cast_fp16)[name = tensor("pairs_13_cast_fp16")]; + tensor xr_13_begin_0 = const()[name = tensor("xr_13_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_13_end_0 = const()[name = tensor("xr_13_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_13_end_mask_0 = const()[name = tensor("xr_13_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_13_squeeze_mask_0 = const()[name = tensor("xr_13_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_13_cast_fp16 = slice_by_index(begin = xr_13_begin_0, end = xr_13_end_0, end_mask = xr_13_end_mask_0, squeeze_mask = xr_13_squeeze_mask_0, x = pairs_13_cast_fp16)[name = tensor("xr_13_cast_fp16")]; + tensor xi_13_begin_0 = const()[name = tensor("xi_13_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_13_end_0 = const()[name = tensor("xi_13_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_13_end_mask_0 = const()[name = tensor("xi_13_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_13_squeeze_mask_0 = const()[name = tensor("xi_13_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_13_cast_fp16 = slice_by_index(begin = xi_13_begin_0, end = xi_13_end_0, end_mask = xi_13_end_mask_0, squeeze_mask = xi_13_squeeze_mask_0, x = pairs_13_cast_fp16)[name = tensor("xi_13_cast_fp16")]; + tensor var_1076_cast_fp16 = mul(x = xr_13_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1076_cast_fp16")]; + tensor var_1077_cast_fp16 = mul(x = xi_13_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1077_cast_fp16")]; + tensor xor__13_cast_fp16 = sub(x = var_1076_cast_fp16, y = var_1077_cast_fp16)[name = tensor("xor__13_cast_fp16")]; + tensor var_1080_cast_fp16 = mul(x = xr_13_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1080_cast_fp16")]; + tensor var_1081_cast_fp16 = mul(x = xi_13_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1081_cast_fp16")]; + tensor xoi_13_cast_fp16 = add(x = var_1080_cast_fp16, y = var_1081_cast_fp16)[name = tensor("xoi_13_cast_fp16")]; + tensor out_13_axis_0 = const()[name = tensor("out_13_axis_0"), val = tensor(-1)]; + tensor out_13_cast_fp16 = stack(axis = out_13_axis_0, values = (xor__13_cast_fp16, xoi_13_cast_fp16))[name = tensor("out_13_cast_fp16")]; + tensor var_1091 = const()[name = tensor("op_1091"), val = tensor([1, 256, 16, 64])]; + tensor q_7_cast_fp16 = reshape(shape = var_1091, x = out_13_cast_fp16)[name = tensor("q_7_cast_fp16")]; + tensor var_1097 = const()[name = tensor("op_1097"), val = tensor([256, 16, 32, 2])]; + tensor pairs_15_cast_fp16 = reshape(shape = var_1097, x = x_15_cast_fp16)[name = tensor("pairs_15_cast_fp16")]; + tensor xr_15_begin_0 = const()[name = tensor("xr_15_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_15_end_0 = const()[name = tensor("xr_15_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_15_end_mask_0 = const()[name = tensor("xr_15_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_15_squeeze_mask_0 = const()[name = tensor("xr_15_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_15_cast_fp16 = slice_by_index(begin = xr_15_begin_0, end = xr_15_end_0, end_mask = xr_15_end_mask_0, squeeze_mask = xr_15_squeeze_mask_0, x = pairs_15_cast_fp16)[name = tensor("xr_15_cast_fp16")]; + tensor xi_15_begin_0 = const()[name = tensor("xi_15_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_15_end_0 = const()[name = tensor("xi_15_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_15_end_mask_0 = const()[name = tensor("xi_15_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_15_squeeze_mask_0 = const()[name = tensor("xi_15_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_15_cast_fp16 = slice_by_index(begin = xi_15_begin_0, end = xi_15_end_0, end_mask = xi_15_end_mask_0, squeeze_mask = xi_15_squeeze_mask_0, x = pairs_15_cast_fp16)[name = tensor("xi_15_cast_fp16")]; + tensor var_1105_cast_fp16 = mul(x = xr_15_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1105_cast_fp16")]; + tensor var_1106_cast_fp16 = mul(x = xi_15_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1106_cast_fp16")]; + tensor xor__15_cast_fp16 = sub(x = var_1105_cast_fp16, y = var_1106_cast_fp16)[name = tensor("xor__15_cast_fp16")]; + tensor var_1109_cast_fp16 = mul(x = xr_15_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1109_cast_fp16")]; + tensor var_1110_cast_fp16 = mul(x = xi_15_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1110_cast_fp16")]; + tensor xoi_15_cast_fp16 = add(x = var_1109_cast_fp16, y = var_1110_cast_fp16)[name = tensor("xoi_15_cast_fp16")]; + tensor out_15_axis_0 = const()[name = tensor("out_15_axis_0"), val = tensor(-1)]; + tensor out_15_cast_fp16 = stack(axis = out_15_axis_0, values = (xor__15_cast_fp16, xoi_15_cast_fp16))[name = tensor("out_15_cast_fp16")]; + tensor var_1127 = const()[name = tensor("op_1127"), val = tensor([256, 1024])]; + tensor var_1128_cast_fp16 = reshape(shape = var_1127, x = out_15_cast_fp16)[name = tensor("op_1128_cast_fp16")]; + tensor written_k_7_transpose_x_1 = const()[name = tensor("written_k_7_transpose_x_1"), val = tensor(true)]; + tensor written_k_7_transpose_y_1 = const()[name = tensor("written_k_7_transpose_y_1"), val = tensor(false)]; + tensor written_k_7_cast_fp16 = matmul(transpose_x = written_k_7_transpose_x_1, transpose_y = written_k_7_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_1128_cast_fp16)[name = tensor("written_k_7_cast_fp16")]; + tensor var_1135 = const()[name = tensor("op_1135"), val = tensor([256, 1024])]; + tensor var_1136_cast_fp16 = reshape(shape = var_1135, x = v_7_cast_fp16)[name = tensor("op_1136_cast_fp16")]; + tensor written_v_7_transpose_x_1 = const()[name = tensor("written_v_7_transpose_x_1"), val = tensor(true)]; + tensor written_v_7_transpose_y_1 = const()[name = tensor("written_v_7_transpose_y_1"), val = tensor(false)]; + tensor written_v_7_cast_fp16 = matmul(transpose_x = written_v_7_transpose_x_1, transpose_y = written_v_7_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_1136_cast_fp16)[name = tensor("written_v_7_cast_fp16")]; + tensor k_cache3_to_fp16_dtype_0 = const()[name = tensor("k_cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache3_to_fp16 = cast(dtype = k_cache3_to_fp16_dtype_0, x = k_cache3)[name = tensor("cast_447")]; + tensor var_1147_cast_fp16 = mul(x = k_cache3_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_1147_cast_fp16")]; + tensor var_1152 = const()[name = tensor("op_1152"), val = tensor([1, 512, 16, 64])]; + tensor var_1153_cast_fp16 = reshape(shape = var_1152, x = written_k_7_cast_fp16)[name = tensor("op_1153_cast_fp16")]; + tensor new_k_cache_7_cast_fp16 = add(x = var_1147_cast_fp16, y = var_1153_cast_fp16)[name = tensor("new_k_cache_7_cast_fp16")]; + tensor new_k_cache_7_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_7_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache3_to_fp16_dtype_0 = const()[name = tensor("v_cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache3_to_fp16 = cast(dtype = v_cache3_to_fp16_dtype_0, x = v_cache3)[name = tensor("cast_445")]; + tensor var_1156_cast_fp16 = mul(x = v_cache3_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_1156_cast_fp16")]; + tensor var_1161 = const()[name = tensor("op_1161"), val = tensor([1, 512, 16, 64])]; + tensor var_1162_cast_fp16 = reshape(shape = var_1161, x = written_v_7_cast_fp16)[name = tensor("op_1162_cast_fp16")]; + tensor new_v_cache_7_cast_fp16 = add(x = var_1156_cast_fp16, y = var_1162_cast_fp16)[name = tensor("new_v_cache_7_cast_fp16")]; + tensor new_v_cache_7_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_7_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_1175 = const()[name = tensor("op_1175"), val = tensor([0, 2, 1, 3])]; + tensor var_1186_transpose_x_0 = const()[name = tensor("op_1186_transpose_x_0"), val = tensor(false)]; + tensor var_1186_transpose_y_0 = const()[name = tensor("op_1186_transpose_y_0"), val = tensor(false)]; + tensor transpose_75_perm_0 = const()[name = tensor("transpose_75_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_76_perm_0 = const()[name = tensor("transpose_76_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_76 = transpose(perm = transpose_76_perm_0, x = new_k_cache_7_cast_fp16)[name = tensor("transpose_192")]; + tensor transpose_75 = transpose(perm = transpose_75_perm_0, x = q_7_cast_fp16)[name = tensor("transpose_193")]; + tensor var_1186_cast_fp16 = matmul(transpose_x = var_1186_transpose_x_0, transpose_y = var_1186_transpose_y_0, x = transpose_75, y = transpose_76)[name = tensor("op_1186_cast_fp16")]; + tensor var_1187_to_fp16 = const()[name = tensor("op_1187_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_13_cast_fp16 = mul(x = var_1186_cast_fp16, y = var_1187_to_fp16)[name = tensor("scores_13_cast_fp16")]; + tensor scores_15_cast_fp16 = add(x = scores_13_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_15_cast_fp16")]; + tensor var_1197 = const()[name = tensor("op_1197"), val = tensor(-1)]; + tensor weights_7_cast_fp16 = softmax(axis = var_1197, x = scores_15_cast_fp16)[name = tensor("weights_7_cast_fp16")]; + tensor attn_7_transpose_x_0 = const()[name = tensor("attn_7_transpose_x_0"), val = tensor(false)]; + tensor attn_7_transpose_y_0 = const()[name = tensor("attn_7_transpose_y_0"), val = tensor(false)]; + tensor values_7_cast_fp16 = transpose(perm = var_1175, x = new_v_cache_7_cast_fp16)[name = tensor("transpose_194")]; + tensor attn_7_cast_fp16 = matmul(transpose_x = attn_7_transpose_x_0, transpose_y = attn_7_transpose_y_0, x = weights_7_cast_fp16, y = values_7_cast_fp16)[name = tensor("attn_7_cast_fp16")]; + tensor var_1205 = const()[name = tensor("op_1205"), val = tensor([0, 2, 1, 3])]; + tensor var_1210 = const()[name = tensor("op_1210"), val = tensor([1, 256, 1024])]; + tensor var_1206_cast_fp16 = transpose(perm = var_1205, x = attn_7_cast_fp16)[name = tensor("transpose_191")]; + tensor input_41_cast_fp16 = reshape(shape = var_1210, x = var_1206_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor attn3_out_proj_weight_to_fp16 = const()[name = tensor("attn3_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(81838336)))]; + tensor linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn3_out_proj_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor position3_to_fp16_dtype_0 = const()[name = tensor("position3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position3_to_fp16 = cast(dtype = position3_to_fp16_dtype_0, x = position3)[name = tensor("cast_443")]; + tensor var_1219_cast_fp16 = add(x = position3_to_fp16, y = valid_len_to_fp16)[name = tensor("op_1219_cast_fp16")]; + tensor var_1219_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1219_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_43_cast_fp16 = add(x = input_37_cast_fp16, y = linear_13_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; + tensor norm3_2_weight_to_fp16 = const()[name = tensor("norm3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(83935552)))]; + tensor norm3_2_bias_to_fp16 = const()[name = tensor("norm3_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(83937664)))]; + tensor var_1223_to_fp16 = const()[name = tensor("op_1223_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_45_cast_fp16 = layer_norm(axes = input_45_axes_0, beta = norm3_2_bias_to_fp16, epsilon = var_1223_to_fp16, gamma = norm3_2_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor linear3_1_weight_to_fp16 = const()[name = tensor("linear3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(83939776)))]; + tensor linear_14_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear3_1_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor input_47_mode_0 = const()[name = tensor("input_47_mode_0"), val = tensor("EXACT")]; + tensor input_47_cast_fp16 = gelu(mode = input_47_mode_0, x = linear_14_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor linear3_2_weight_to_fp16 = const()[name = tensor("linear3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(92328448)))]; + tensor linear_15_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear3_2_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor input_49_cast_fp16 = add(x = input_43_cast_fp16, y = linear_15_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor input_51_axes_0 = const()[name = tensor("input_51_axes_0"), val = tensor([-1])]; + tensor norm4_1_weight_to_fp16 = const()[name = tensor("norm4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100717120)))]; + tensor norm4_1_bias_to_fp16 = const()[name = tensor("norm4_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100719232)))]; + tensor var_1240_to_fp16 = const()[name = tensor("op_1240_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_51_cast_fp16 = layer_norm(axes = input_51_axes_0, beta = norm4_1_bias_to_fp16, epsilon = var_1240_to_fp16, gamma = norm4_1_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor attn4_in_proj_weight_to_fp16 = const()[name = tensor("attn4_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100721344)))]; + tensor linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn4_in_proj_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor var_1254 = const()[name = tensor("op_1254"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_9_cast_fp16 = reshape(shape = var_1254, x = linear_16_cast_fp16)[name = tensor("qkv_9_cast_fp16")]; + tensor x_17_begin_0 = const()[name = tensor("x_17_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_17_end_0 = const()[name = tensor("x_17_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_17_end_mask_0 = const()[name = tensor("x_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_17_squeeze_mask_0 = const()[name = tensor("x_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_17_cast_fp16 = slice_by_index(begin = x_17_begin_0, end = x_17_end_0, end_mask = x_17_end_mask_0, squeeze_mask = x_17_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("x_17_cast_fp16")]; + tensor x_19_begin_0 = const()[name = tensor("x_19_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_19_end_0 = const()[name = tensor("x_19_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_19_end_mask_0 = const()[name = tensor("x_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_19_squeeze_mask_0 = const()[name = tensor("x_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_19_cast_fp16 = slice_by_index(begin = x_19_begin_0, end = x_19_end_0, end_mask = x_19_end_mask_0, squeeze_mask = x_19_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("x_19_cast_fp16")]; + tensor v_9_begin_0 = const()[name = tensor("v_9_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_9_end_0 = const()[name = tensor("v_9_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_9_end_mask_0 = const()[name = tensor("v_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_9_squeeze_mask_0 = const()[name = tensor("v_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_9_cast_fp16 = slice_by_index(begin = v_9_begin_0, end = v_9_end_0, end_mask = v_9_end_mask_0, squeeze_mask = v_9_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("v_9_cast_fp16")]; + tensor var_1299 = const()[name = tensor("op_1299"), val = tensor([256, 16, 32, 2])]; + tensor pairs_17_cast_fp16 = reshape(shape = var_1299, x = x_17_cast_fp16)[name = tensor("pairs_17_cast_fp16")]; + tensor xr_17_begin_0 = const()[name = tensor("xr_17_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_17_end_0 = const()[name = tensor("xr_17_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_17_end_mask_0 = const()[name = tensor("xr_17_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_17_squeeze_mask_0 = const()[name = tensor("xr_17_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_17_cast_fp16 = slice_by_index(begin = xr_17_begin_0, end = xr_17_end_0, end_mask = xr_17_end_mask_0, squeeze_mask = xr_17_squeeze_mask_0, x = pairs_17_cast_fp16)[name = tensor("xr_17_cast_fp16")]; + tensor xi_17_begin_0 = const()[name = tensor("xi_17_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_17_end_0 = const()[name = tensor("xi_17_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_17_end_mask_0 = const()[name = tensor("xi_17_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_17_squeeze_mask_0 = const()[name = tensor("xi_17_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_17_cast_fp16 = slice_by_index(begin = xi_17_begin_0, end = xi_17_end_0, end_mask = xi_17_end_mask_0, squeeze_mask = xi_17_squeeze_mask_0, x = pairs_17_cast_fp16)[name = tensor("xi_17_cast_fp16")]; + tensor var_1307_cast_fp16 = mul(x = xr_17_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1307_cast_fp16")]; + tensor var_1308_cast_fp16 = mul(x = xi_17_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1308_cast_fp16")]; + tensor xor__17_cast_fp16 = sub(x = var_1307_cast_fp16, y = var_1308_cast_fp16)[name = tensor("xor__17_cast_fp16")]; + tensor var_1311_cast_fp16 = mul(x = xr_17_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1311_cast_fp16")]; + tensor var_1312_cast_fp16 = mul(x = xi_17_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1312_cast_fp16")]; + tensor xoi_17_cast_fp16 = add(x = var_1311_cast_fp16, y = var_1312_cast_fp16)[name = tensor("xoi_17_cast_fp16")]; + tensor out_17_axis_0 = const()[name = tensor("out_17_axis_0"), val = tensor(-1)]; + tensor out_17_cast_fp16 = stack(axis = out_17_axis_0, values = (xor__17_cast_fp16, xoi_17_cast_fp16))[name = tensor("out_17_cast_fp16")]; + tensor var_1322 = const()[name = tensor("op_1322"), val = tensor([1, 256, 16, 64])]; + tensor q_9_cast_fp16 = reshape(shape = var_1322, x = out_17_cast_fp16)[name = tensor("q_9_cast_fp16")]; + tensor var_1328 = const()[name = tensor("op_1328"), val = tensor([256, 16, 32, 2])]; + tensor pairs_19_cast_fp16 = reshape(shape = var_1328, x = x_19_cast_fp16)[name = tensor("pairs_19_cast_fp16")]; + tensor xr_19_begin_0 = const()[name = tensor("xr_19_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_19_end_0 = const()[name = tensor("xr_19_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_19_end_mask_0 = const()[name = tensor("xr_19_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_19_squeeze_mask_0 = const()[name = tensor("xr_19_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_19_cast_fp16 = slice_by_index(begin = xr_19_begin_0, end = xr_19_end_0, end_mask = xr_19_end_mask_0, squeeze_mask = xr_19_squeeze_mask_0, x = pairs_19_cast_fp16)[name = tensor("xr_19_cast_fp16")]; + tensor xi_19_begin_0 = const()[name = tensor("xi_19_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_19_end_0 = const()[name = tensor("xi_19_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_19_end_mask_0 = const()[name = tensor("xi_19_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_19_squeeze_mask_0 = const()[name = tensor("xi_19_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_19_cast_fp16 = slice_by_index(begin = xi_19_begin_0, end = xi_19_end_0, end_mask = xi_19_end_mask_0, squeeze_mask = xi_19_squeeze_mask_0, x = pairs_19_cast_fp16)[name = tensor("xi_19_cast_fp16")]; + tensor var_1336_cast_fp16 = mul(x = xr_19_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1336_cast_fp16")]; + tensor var_1337_cast_fp16 = mul(x = xi_19_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1337_cast_fp16")]; + tensor xor__19_cast_fp16 = sub(x = var_1336_cast_fp16, y = var_1337_cast_fp16)[name = tensor("xor__19_cast_fp16")]; + tensor var_1340_cast_fp16 = mul(x = xr_19_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1340_cast_fp16")]; + tensor var_1341_cast_fp16 = mul(x = xi_19_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1341_cast_fp16")]; + tensor xoi_19_cast_fp16 = add(x = var_1340_cast_fp16, y = var_1341_cast_fp16)[name = tensor("xoi_19_cast_fp16")]; + tensor out_19_axis_0 = const()[name = tensor("out_19_axis_0"), val = tensor(-1)]; + tensor out_19_cast_fp16 = stack(axis = out_19_axis_0, values = (xor__19_cast_fp16, xoi_19_cast_fp16))[name = tensor("out_19_cast_fp16")]; + tensor var_1358 = const()[name = tensor("op_1358"), val = tensor([256, 1024])]; + tensor var_1359_cast_fp16 = reshape(shape = var_1358, x = out_19_cast_fp16)[name = tensor("op_1359_cast_fp16")]; + tensor written_k_9_transpose_x_1 = const()[name = tensor("written_k_9_transpose_x_1"), val = tensor(true)]; + tensor written_k_9_transpose_y_1 = const()[name = tensor("written_k_9_transpose_y_1"), val = tensor(false)]; + tensor written_k_9_cast_fp16 = matmul(transpose_x = written_k_9_transpose_x_1, transpose_y = written_k_9_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_1359_cast_fp16)[name = tensor("written_k_9_cast_fp16")]; + tensor var_1366 = const()[name = tensor("op_1366"), val = tensor([256, 1024])]; + tensor var_1367_cast_fp16 = reshape(shape = var_1366, x = v_9_cast_fp16)[name = tensor("op_1367_cast_fp16")]; + tensor written_v_9_transpose_x_1 = const()[name = tensor("written_v_9_transpose_x_1"), val = tensor(true)]; + tensor written_v_9_transpose_y_1 = const()[name = tensor("written_v_9_transpose_y_1"), val = tensor(false)]; + tensor written_v_9_cast_fp16 = matmul(transpose_x = written_v_9_transpose_x_1, transpose_y = written_v_9_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_1367_cast_fp16)[name = tensor("written_v_9_cast_fp16")]; + tensor k_cache4_to_fp16_dtype_0 = const()[name = tensor("k_cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache4_to_fp16 = cast(dtype = k_cache4_to_fp16_dtype_0, x = k_cache4)[name = tensor("cast_441")]; + tensor var_1378_cast_fp16 = mul(x = k_cache4_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_1378_cast_fp16")]; + tensor var_1383 = const()[name = tensor("op_1383"), val = tensor([1, 512, 16, 64])]; + tensor var_1384_cast_fp16 = reshape(shape = var_1383, x = written_k_9_cast_fp16)[name = tensor("op_1384_cast_fp16")]; + tensor new_k_cache_9_cast_fp16 = add(x = var_1378_cast_fp16, y = var_1384_cast_fp16)[name = tensor("new_k_cache_9_cast_fp16")]; + tensor new_k_cache_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache4_to_fp16_dtype_0 = const()[name = tensor("v_cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache4_to_fp16 = cast(dtype = v_cache4_to_fp16_dtype_0, x = v_cache4)[name = tensor("cast_439")]; + tensor var_1387_cast_fp16 = mul(x = v_cache4_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_1387_cast_fp16")]; + tensor var_1392 = const()[name = tensor("op_1392"), val = tensor([1, 512, 16, 64])]; + tensor var_1393_cast_fp16 = reshape(shape = var_1392, x = written_v_9_cast_fp16)[name = tensor("op_1393_cast_fp16")]; + tensor new_v_cache_9_cast_fp16 = add(x = var_1387_cast_fp16, y = var_1393_cast_fp16)[name = tensor("new_v_cache_9_cast_fp16")]; + tensor new_v_cache_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_1406 = const()[name = tensor("op_1406"), val = tensor([0, 2, 1, 3])]; + tensor var_1417_transpose_x_0 = const()[name = tensor("op_1417_transpose_x_0"), val = tensor(false)]; + tensor var_1417_transpose_y_0 = const()[name = tensor("op_1417_transpose_y_0"), val = tensor(false)]; + tensor transpose_77_perm_0 = const()[name = tensor("transpose_77_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_78_perm_0 = const()[name = tensor("transpose_78_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_78 = transpose(perm = transpose_78_perm_0, x = new_k_cache_9_cast_fp16)[name = tensor("transpose_188")]; + tensor transpose_77 = transpose(perm = transpose_77_perm_0, x = q_9_cast_fp16)[name = tensor("transpose_189")]; + tensor var_1417_cast_fp16 = matmul(transpose_x = var_1417_transpose_x_0, transpose_y = var_1417_transpose_y_0, x = transpose_77, y = transpose_78)[name = tensor("op_1417_cast_fp16")]; + tensor var_1418_to_fp16 = const()[name = tensor("op_1418_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_17_cast_fp16 = mul(x = var_1417_cast_fp16, y = var_1418_to_fp16)[name = tensor("scores_17_cast_fp16")]; + tensor scores_19_cast_fp16 = add(x = scores_17_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_19_cast_fp16")]; + tensor var_1428 = const()[name = tensor("op_1428"), val = tensor(-1)]; + tensor weights_9_cast_fp16 = softmax(axis = var_1428, x = scores_19_cast_fp16)[name = tensor("weights_9_cast_fp16")]; + tensor attn_9_transpose_x_0 = const()[name = tensor("attn_9_transpose_x_0"), val = tensor(false)]; + tensor attn_9_transpose_y_0 = const()[name = tensor("attn_9_transpose_y_0"), val = tensor(false)]; + tensor values_9_cast_fp16 = transpose(perm = var_1406, x = new_v_cache_9_cast_fp16)[name = tensor("transpose_190")]; + tensor attn_9_cast_fp16 = matmul(transpose_x = attn_9_transpose_x_0, transpose_y = attn_9_transpose_y_0, x = weights_9_cast_fp16, y = values_9_cast_fp16)[name = tensor("attn_9_cast_fp16")]; + tensor var_1436 = const()[name = tensor("op_1436"), val = tensor([0, 2, 1, 3])]; + tensor var_1441 = const()[name = tensor("op_1441"), val = tensor([1, 256, 1024])]; + tensor var_1437_cast_fp16 = transpose(perm = var_1436, x = attn_9_cast_fp16)[name = tensor("transpose_187")]; + tensor input_53_cast_fp16 = reshape(shape = var_1441, x = var_1437_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor attn4_out_proj_weight_to_fp16 = const()[name = tensor("attn4_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(107012864)))]; + tensor linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn4_out_proj_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor position4_to_fp16_dtype_0 = const()[name = tensor("position4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position4_to_fp16 = cast(dtype = position4_to_fp16_dtype_0, x = position4)[name = tensor("cast_437")]; + tensor var_1450_cast_fp16 = add(x = position4_to_fp16, y = valid_len_to_fp16)[name = tensor("op_1450_cast_fp16")]; + tensor var_1450_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1450_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_55_cast_fp16 = add(x = input_49_cast_fp16, y = linear_17_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor input_57_axes_0 = const()[name = tensor("input_57_axes_0"), val = tensor([-1])]; + tensor norm4_2_weight_to_fp16 = const()[name = tensor("norm4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109110080)))]; + tensor norm4_2_bias_to_fp16 = const()[name = tensor("norm4_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109112192)))]; + tensor var_1454_to_fp16 = const()[name = tensor("op_1454_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_57_cast_fp16 = layer_norm(axes = input_57_axes_0, beta = norm4_2_bias_to_fp16, epsilon = var_1454_to_fp16, gamma = norm4_2_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor linear4_1_weight_to_fp16 = const()[name = tensor("linear4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109114304)))]; + tensor linear_18_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear4_1_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor input_59_mode_0 = const()[name = tensor("input_59_mode_0"), val = tensor("EXACT")]; + tensor input_59_cast_fp16 = gelu(mode = input_59_mode_0, x = linear_18_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor linear4_2_weight_to_fp16 = const()[name = tensor("linear4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(117502976)))]; + tensor linear_19_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear4_2_weight_to_fp16, x = input_59_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_61_cast_fp16 = add(x = input_55_cast_fp16, y = linear_19_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor input_63_axes_0 = const()[name = tensor("input_63_axes_0"), val = tensor([-1])]; + tensor norm5_1_weight_to_fp16 = const()[name = tensor("norm5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125891648)))]; + tensor norm5_1_bias_to_fp16 = const()[name = tensor("norm5_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125893760)))]; + tensor var_1471_to_fp16 = const()[name = tensor("op_1471_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_63_cast_fp16 = layer_norm(axes = input_63_axes_0, beta = norm5_1_bias_to_fp16, epsilon = var_1471_to_fp16, gamma = norm5_1_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor attn5_in_proj_weight_to_fp16 = const()[name = tensor("attn5_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125895872)))]; + tensor linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn5_in_proj_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor var_1485 = const()[name = tensor("op_1485"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_11_cast_fp16 = reshape(shape = var_1485, x = linear_20_cast_fp16)[name = tensor("qkv_11_cast_fp16")]; + tensor x_21_begin_0 = const()[name = tensor("x_21_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_21_end_0 = const()[name = tensor("x_21_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_21_end_mask_0 = const()[name = tensor("x_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_21_squeeze_mask_0 = const()[name = tensor("x_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_21_cast_fp16 = slice_by_index(begin = x_21_begin_0, end = x_21_end_0, end_mask = x_21_end_mask_0, squeeze_mask = x_21_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("x_21_cast_fp16")]; + tensor x_23_begin_0 = const()[name = tensor("x_23_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_23_end_0 = const()[name = tensor("x_23_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_23_end_mask_0 = const()[name = tensor("x_23_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_23_squeeze_mask_0 = const()[name = tensor("x_23_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_23_cast_fp16 = slice_by_index(begin = x_23_begin_0, end = x_23_end_0, end_mask = x_23_end_mask_0, squeeze_mask = x_23_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("x_23_cast_fp16")]; + tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_11_cast_fp16 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("v_11_cast_fp16")]; + tensor var_1530 = const()[name = tensor("op_1530"), val = tensor([256, 16, 32, 2])]; + tensor pairs_21_cast_fp16 = reshape(shape = var_1530, x = x_21_cast_fp16)[name = tensor("pairs_21_cast_fp16")]; + tensor xr_21_begin_0 = const()[name = tensor("xr_21_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_21_end_0 = const()[name = tensor("xr_21_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_21_end_mask_0 = const()[name = tensor("xr_21_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_21_squeeze_mask_0 = const()[name = tensor("xr_21_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_21_cast_fp16 = slice_by_index(begin = xr_21_begin_0, end = xr_21_end_0, end_mask = xr_21_end_mask_0, squeeze_mask = xr_21_squeeze_mask_0, x = pairs_21_cast_fp16)[name = tensor("xr_21_cast_fp16")]; + tensor xi_21_begin_0 = const()[name = tensor("xi_21_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_21_end_0 = const()[name = tensor("xi_21_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_21_end_mask_0 = const()[name = tensor("xi_21_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_21_squeeze_mask_0 = const()[name = tensor("xi_21_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_21_cast_fp16 = slice_by_index(begin = xi_21_begin_0, end = xi_21_end_0, end_mask = xi_21_end_mask_0, squeeze_mask = xi_21_squeeze_mask_0, x = pairs_21_cast_fp16)[name = tensor("xi_21_cast_fp16")]; + tensor var_1538_cast_fp16 = mul(x = xr_21_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1538_cast_fp16")]; + tensor var_1539_cast_fp16 = mul(x = xi_21_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1539_cast_fp16")]; + tensor xor__21_cast_fp16 = sub(x = var_1538_cast_fp16, y = var_1539_cast_fp16)[name = tensor("xor__21_cast_fp16")]; + tensor var_1542_cast_fp16 = mul(x = xr_21_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1542_cast_fp16")]; + tensor var_1543_cast_fp16 = mul(x = xi_21_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1543_cast_fp16")]; + tensor xoi_21_cast_fp16 = add(x = var_1542_cast_fp16, y = var_1543_cast_fp16)[name = tensor("xoi_21_cast_fp16")]; + tensor out_21_axis_0 = const()[name = tensor("out_21_axis_0"), val = tensor(-1)]; + tensor out_21_cast_fp16 = stack(axis = out_21_axis_0, values = (xor__21_cast_fp16, xoi_21_cast_fp16))[name = tensor("out_21_cast_fp16")]; + tensor var_1553 = const()[name = tensor("op_1553"), val = tensor([1, 256, 16, 64])]; + tensor q_11_cast_fp16 = reshape(shape = var_1553, x = out_21_cast_fp16)[name = tensor("q_11_cast_fp16")]; + tensor var_1559 = const()[name = tensor("op_1559"), val = tensor([256, 16, 32, 2])]; + tensor pairs_23_cast_fp16 = reshape(shape = var_1559, x = x_23_cast_fp16)[name = tensor("pairs_23_cast_fp16")]; + tensor xr_23_begin_0 = const()[name = tensor("xr_23_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_23_end_0 = const()[name = tensor("xr_23_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_23_end_mask_0 = const()[name = tensor("xr_23_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_23_squeeze_mask_0 = const()[name = tensor("xr_23_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_23_cast_fp16 = slice_by_index(begin = xr_23_begin_0, end = xr_23_end_0, end_mask = xr_23_end_mask_0, squeeze_mask = xr_23_squeeze_mask_0, x = pairs_23_cast_fp16)[name = tensor("xr_23_cast_fp16")]; + tensor xi_23_begin_0 = const()[name = tensor("xi_23_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_23_end_0 = const()[name = tensor("xi_23_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_23_end_mask_0 = const()[name = tensor("xi_23_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_23_squeeze_mask_0 = const()[name = tensor("xi_23_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_23_cast_fp16 = slice_by_index(begin = xi_23_begin_0, end = xi_23_end_0, end_mask = xi_23_end_mask_0, squeeze_mask = xi_23_squeeze_mask_0, x = pairs_23_cast_fp16)[name = tensor("xi_23_cast_fp16")]; + tensor var_1567_cast_fp16 = mul(x = xr_23_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1567_cast_fp16")]; + tensor var_1568_cast_fp16 = mul(x = xi_23_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1568_cast_fp16")]; + tensor xor__23_cast_fp16 = sub(x = var_1567_cast_fp16, y = var_1568_cast_fp16)[name = tensor("xor__23_cast_fp16")]; + tensor var_1571_cast_fp16 = mul(x = xr_23_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1571_cast_fp16")]; + tensor var_1572_cast_fp16 = mul(x = xi_23_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1572_cast_fp16")]; + tensor xoi_23_cast_fp16 = add(x = var_1571_cast_fp16, y = var_1572_cast_fp16)[name = tensor("xoi_23_cast_fp16")]; + tensor out_23_axis_0 = const()[name = tensor("out_23_axis_0"), val = tensor(-1)]; + tensor out_23_cast_fp16 = stack(axis = out_23_axis_0, values = (xor__23_cast_fp16, xoi_23_cast_fp16))[name = tensor("out_23_cast_fp16")]; + tensor var_1589 = const()[name = tensor("op_1589"), val = tensor([256, 1024])]; + tensor var_1590_cast_fp16 = reshape(shape = var_1589, x = out_23_cast_fp16)[name = tensor("op_1590_cast_fp16")]; + tensor written_k_11_transpose_x_1 = const()[name = tensor("written_k_11_transpose_x_1"), val = tensor(true)]; + tensor written_k_11_transpose_y_1 = const()[name = tensor("written_k_11_transpose_y_1"), val = tensor(false)]; + tensor written_k_11_cast_fp16 = matmul(transpose_x = written_k_11_transpose_x_1, transpose_y = written_k_11_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_1590_cast_fp16)[name = tensor("written_k_11_cast_fp16")]; + tensor var_1597 = const()[name = tensor("op_1597"), val = tensor([256, 1024])]; + tensor var_1598_cast_fp16 = reshape(shape = var_1597, x = v_11_cast_fp16)[name = tensor("op_1598_cast_fp16")]; + tensor written_v_11_transpose_x_1 = const()[name = tensor("written_v_11_transpose_x_1"), val = tensor(true)]; + tensor written_v_11_transpose_y_1 = const()[name = tensor("written_v_11_transpose_y_1"), val = tensor(false)]; + tensor written_v_11_cast_fp16 = matmul(transpose_x = written_v_11_transpose_x_1, transpose_y = written_v_11_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_1598_cast_fp16)[name = tensor("written_v_11_cast_fp16")]; + tensor k_cache5_to_fp16_dtype_0 = const()[name = tensor("k_cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache5_to_fp16 = cast(dtype = k_cache5_to_fp16_dtype_0, x = k_cache5)[name = tensor("cast_435")]; + tensor var_1609_cast_fp16 = mul(x = k_cache5_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_1609_cast_fp16")]; + tensor var_1614 = const()[name = tensor("op_1614"), val = tensor([1, 512, 16, 64])]; + tensor var_1615_cast_fp16 = reshape(shape = var_1614, x = written_k_11_cast_fp16)[name = tensor("op_1615_cast_fp16")]; + tensor new_k_cache_11_cast_fp16 = add(x = var_1609_cast_fp16, y = var_1615_cast_fp16)[name = tensor("new_k_cache_11_cast_fp16")]; + tensor new_k_cache_11_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_11_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache5_to_fp16_dtype_0 = const()[name = tensor("v_cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache5_to_fp16 = cast(dtype = v_cache5_to_fp16_dtype_0, x = v_cache5)[name = tensor("cast_433")]; + tensor var_1618_cast_fp16 = mul(x = v_cache5_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_1618_cast_fp16")]; + tensor var_1623 = const()[name = tensor("op_1623"), val = tensor([1, 512, 16, 64])]; + tensor var_1624_cast_fp16 = reshape(shape = var_1623, x = written_v_11_cast_fp16)[name = tensor("op_1624_cast_fp16")]; + tensor new_v_cache_11_cast_fp16 = add(x = var_1618_cast_fp16, y = var_1624_cast_fp16)[name = tensor("new_v_cache_11_cast_fp16")]; + tensor new_v_cache_11_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_11_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_1637 = const()[name = tensor("op_1637"), val = tensor([0, 2, 1, 3])]; + tensor var_1648_transpose_x_0 = const()[name = tensor("op_1648_transpose_x_0"), val = tensor(false)]; + tensor var_1648_transpose_y_0 = const()[name = tensor("op_1648_transpose_y_0"), val = tensor(false)]; + tensor transpose_79_perm_0 = const()[name = tensor("transpose_79_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_80_perm_0 = const()[name = tensor("transpose_80_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_80 = transpose(perm = transpose_80_perm_0, x = new_k_cache_11_cast_fp16)[name = tensor("transpose_184")]; + tensor transpose_79 = transpose(perm = transpose_79_perm_0, x = q_11_cast_fp16)[name = tensor("transpose_185")]; + tensor var_1648_cast_fp16 = matmul(transpose_x = var_1648_transpose_x_0, transpose_y = var_1648_transpose_y_0, x = transpose_79, y = transpose_80)[name = tensor("op_1648_cast_fp16")]; + tensor var_1649_to_fp16 = const()[name = tensor("op_1649_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_21_cast_fp16 = mul(x = var_1648_cast_fp16, y = var_1649_to_fp16)[name = tensor("scores_21_cast_fp16")]; + tensor scores_23_cast_fp16 = add(x = scores_21_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_23_cast_fp16")]; + tensor var_1659 = const()[name = tensor("op_1659"), val = tensor(-1)]; + tensor weights_11_cast_fp16 = softmax(axis = var_1659, x = scores_23_cast_fp16)[name = tensor("weights_11_cast_fp16")]; + tensor attn_11_transpose_x_0 = const()[name = tensor("attn_11_transpose_x_0"), val = tensor(false)]; + tensor attn_11_transpose_y_0 = const()[name = tensor("attn_11_transpose_y_0"), val = tensor(false)]; + tensor values_11_cast_fp16 = transpose(perm = var_1637, x = new_v_cache_11_cast_fp16)[name = tensor("transpose_186")]; + tensor attn_11_cast_fp16 = matmul(transpose_x = attn_11_transpose_x_0, transpose_y = attn_11_transpose_y_0, x = weights_11_cast_fp16, y = values_11_cast_fp16)[name = tensor("attn_11_cast_fp16")]; + tensor var_1667 = const()[name = tensor("op_1667"), val = tensor([0, 2, 1, 3])]; + tensor var_1672 = const()[name = tensor("op_1672"), val = tensor([1, 256, 1024])]; + tensor var_1668_cast_fp16 = transpose(perm = var_1667, x = attn_11_cast_fp16)[name = tensor("transpose_183")]; + tensor input_65_cast_fp16 = reshape(shape = var_1672, x = var_1668_cast_fp16)[name = tensor("input_65_cast_fp16")]; + tensor attn5_out_proj_weight_to_fp16 = const()[name = tensor("attn5_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132187392)))]; + tensor linear_21_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn5_out_proj_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("linear_21_cast_fp16")]; + tensor position5_to_fp16_dtype_0 = const()[name = tensor("position5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position5_to_fp16 = cast(dtype = position5_to_fp16_dtype_0, x = position5)[name = tensor("cast_431")]; + tensor var_1681_cast_fp16 = add(x = position5_to_fp16, y = valid_len_to_fp16)[name = tensor("op_1681_cast_fp16")]; + tensor var_1681_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1681_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_67_cast_fp16 = add(x = input_61_cast_fp16, y = linear_21_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor input_69_axes_0 = const()[name = tensor("input_69_axes_0"), val = tensor([-1])]; + tensor norm5_2_weight_to_fp16 = const()[name = tensor("norm5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134284608)))]; + tensor norm5_2_bias_to_fp16 = const()[name = tensor("norm5_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134286720)))]; + tensor var_1685_to_fp16 = const()[name = tensor("op_1685_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_69_cast_fp16 = layer_norm(axes = input_69_axes_0, beta = norm5_2_bias_to_fp16, epsilon = var_1685_to_fp16, gamma = norm5_2_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("input_69_cast_fp16")]; + tensor linear5_1_weight_to_fp16 = const()[name = tensor("linear5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134288832)))]; + tensor linear_22_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear5_1_weight_to_fp16, x = input_69_cast_fp16)[name = tensor("linear_22_cast_fp16")]; + tensor input_71_mode_0 = const()[name = tensor("input_71_mode_0"), val = tensor("EXACT")]; + tensor input_71_cast_fp16 = gelu(mode = input_71_mode_0, x = linear_22_cast_fp16)[name = tensor("input_71_cast_fp16")]; + tensor linear5_2_weight_to_fp16 = const()[name = tensor("linear5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142677504)))]; + tensor linear_23_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear5_2_weight_to_fp16, x = input_71_cast_fp16)[name = tensor("linear_23_cast_fp16")]; + tensor input_73_cast_fp16 = add(x = input_67_cast_fp16, y = linear_23_cast_fp16)[name = tensor("input_73_cast_fp16")]; + tensor input_75_axes_0 = const()[name = tensor("input_75_axes_0"), val = tensor([-1])]; + tensor norm6_1_weight_to_fp16 = const()[name = tensor("norm6_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151066176)))]; + tensor norm6_1_bias_to_fp16 = const()[name = tensor("norm6_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151068288)))]; + tensor var_1702_to_fp16 = const()[name = tensor("op_1702_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_75_cast_fp16 = layer_norm(axes = input_75_axes_0, beta = norm6_1_bias_to_fp16, epsilon = var_1702_to_fp16, gamma = norm6_1_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("input_75_cast_fp16")]; + tensor attn6_in_proj_weight_to_fp16 = const()[name = tensor("attn6_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151070400)))]; + tensor linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn6_in_proj_weight_to_fp16, x = input_75_cast_fp16)[name = tensor("linear_24_cast_fp16")]; + tensor var_1716 = const()[name = tensor("op_1716"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_13_cast_fp16 = reshape(shape = var_1716, x = linear_24_cast_fp16)[name = tensor("qkv_13_cast_fp16")]; + tensor x_25_begin_0 = const()[name = tensor("x_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_25_end_0 = const()[name = tensor("x_25_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_25_end_mask_0 = const()[name = tensor("x_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_25_squeeze_mask_0 = const()[name = tensor("x_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_25_cast_fp16 = slice_by_index(begin = x_25_begin_0, end = x_25_end_0, end_mask = x_25_end_mask_0, squeeze_mask = x_25_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("x_25_cast_fp16")]; + tensor x_27_begin_0 = const()[name = tensor("x_27_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_27_end_0 = const()[name = tensor("x_27_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_27_end_mask_0 = const()[name = tensor("x_27_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_27_squeeze_mask_0 = const()[name = tensor("x_27_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_27_cast_fp16 = slice_by_index(begin = x_27_begin_0, end = x_27_end_0, end_mask = x_27_end_mask_0, squeeze_mask = x_27_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("x_27_cast_fp16")]; + tensor v_13_begin_0 = const()[name = tensor("v_13_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_13_end_0 = const()[name = tensor("v_13_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_13_end_mask_0 = const()[name = tensor("v_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_13_squeeze_mask_0 = const()[name = tensor("v_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_13_cast_fp16 = slice_by_index(begin = v_13_begin_0, end = v_13_end_0, end_mask = v_13_end_mask_0, squeeze_mask = v_13_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("v_13_cast_fp16")]; + tensor var_1761 = const()[name = tensor("op_1761"), val = tensor([256, 16, 32, 2])]; + tensor pairs_25_cast_fp16 = reshape(shape = var_1761, x = x_25_cast_fp16)[name = tensor("pairs_25_cast_fp16")]; + tensor xr_25_begin_0 = const()[name = tensor("xr_25_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_25_end_0 = const()[name = tensor("xr_25_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_25_end_mask_0 = const()[name = tensor("xr_25_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_25_squeeze_mask_0 = const()[name = tensor("xr_25_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_25_cast_fp16 = slice_by_index(begin = xr_25_begin_0, end = xr_25_end_0, end_mask = xr_25_end_mask_0, squeeze_mask = xr_25_squeeze_mask_0, x = pairs_25_cast_fp16)[name = tensor("xr_25_cast_fp16")]; + tensor xi_25_begin_0 = const()[name = tensor("xi_25_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_25_end_0 = const()[name = tensor("xi_25_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_25_end_mask_0 = const()[name = tensor("xi_25_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_25_squeeze_mask_0 = const()[name = tensor("xi_25_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_25_cast_fp16 = slice_by_index(begin = xi_25_begin_0, end = xi_25_end_0, end_mask = xi_25_end_mask_0, squeeze_mask = xi_25_squeeze_mask_0, x = pairs_25_cast_fp16)[name = tensor("xi_25_cast_fp16")]; + tensor var_1769_cast_fp16 = mul(x = xr_25_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1769_cast_fp16")]; + tensor var_1770_cast_fp16 = mul(x = xi_25_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1770_cast_fp16")]; + tensor xor__25_cast_fp16 = sub(x = var_1769_cast_fp16, y = var_1770_cast_fp16)[name = tensor("xor__25_cast_fp16")]; + tensor var_1773_cast_fp16 = mul(x = xr_25_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1773_cast_fp16")]; + tensor var_1774_cast_fp16 = mul(x = xi_25_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1774_cast_fp16")]; + tensor xoi_25_cast_fp16 = add(x = var_1773_cast_fp16, y = var_1774_cast_fp16)[name = tensor("xoi_25_cast_fp16")]; + tensor out_25_axis_0 = const()[name = tensor("out_25_axis_0"), val = tensor(-1)]; + tensor out_25_cast_fp16 = stack(axis = out_25_axis_0, values = (xor__25_cast_fp16, xoi_25_cast_fp16))[name = tensor("out_25_cast_fp16")]; + tensor var_1784 = const()[name = tensor("op_1784"), val = tensor([1, 256, 16, 64])]; + tensor q_13_cast_fp16 = reshape(shape = var_1784, x = out_25_cast_fp16)[name = tensor("q_13_cast_fp16")]; + tensor var_1790 = const()[name = tensor("op_1790"), val = tensor([256, 16, 32, 2])]; + tensor pairs_27_cast_fp16 = reshape(shape = var_1790, x = x_27_cast_fp16)[name = tensor("pairs_27_cast_fp16")]; + tensor xr_27_begin_0 = const()[name = tensor("xr_27_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_27_end_0 = const()[name = tensor("xr_27_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_27_end_mask_0 = const()[name = tensor("xr_27_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_27_squeeze_mask_0 = const()[name = tensor("xr_27_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_27_cast_fp16 = slice_by_index(begin = xr_27_begin_0, end = xr_27_end_0, end_mask = xr_27_end_mask_0, squeeze_mask = xr_27_squeeze_mask_0, x = pairs_27_cast_fp16)[name = tensor("xr_27_cast_fp16")]; + tensor xi_27_begin_0 = const()[name = tensor("xi_27_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_27_end_0 = const()[name = tensor("xi_27_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_27_end_mask_0 = const()[name = tensor("xi_27_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_27_squeeze_mask_0 = const()[name = tensor("xi_27_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_27_cast_fp16 = slice_by_index(begin = xi_27_begin_0, end = xi_27_end_0, end_mask = xi_27_end_mask_0, squeeze_mask = xi_27_squeeze_mask_0, x = pairs_27_cast_fp16)[name = tensor("xi_27_cast_fp16")]; + tensor var_1798_cast_fp16 = mul(x = xr_27_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1798_cast_fp16")]; + tensor var_1799_cast_fp16 = mul(x = xi_27_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1799_cast_fp16")]; + tensor xor__27_cast_fp16 = sub(x = var_1798_cast_fp16, y = var_1799_cast_fp16)[name = tensor("xor__27_cast_fp16")]; + tensor var_1802_cast_fp16 = mul(x = xr_27_cast_fp16, y = roti_cast_fp16)[name = tensor("op_1802_cast_fp16")]; + tensor var_1803_cast_fp16 = mul(x = xi_27_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_1803_cast_fp16")]; + tensor xoi_27_cast_fp16 = add(x = var_1802_cast_fp16, y = var_1803_cast_fp16)[name = tensor("xoi_27_cast_fp16")]; + tensor out_27_axis_0 = const()[name = tensor("out_27_axis_0"), val = tensor(-1)]; + tensor out_27_cast_fp16 = stack(axis = out_27_axis_0, values = (xor__27_cast_fp16, xoi_27_cast_fp16))[name = tensor("out_27_cast_fp16")]; + tensor var_1820 = const()[name = tensor("op_1820"), val = tensor([256, 1024])]; + tensor var_1821_cast_fp16 = reshape(shape = var_1820, x = out_27_cast_fp16)[name = tensor("op_1821_cast_fp16")]; + tensor written_k_13_transpose_x_1 = const()[name = tensor("written_k_13_transpose_x_1"), val = tensor(true)]; + tensor written_k_13_transpose_y_1 = const()[name = tensor("written_k_13_transpose_y_1"), val = tensor(false)]; + tensor written_k_13_cast_fp16 = matmul(transpose_x = written_k_13_transpose_x_1, transpose_y = written_k_13_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_1821_cast_fp16)[name = tensor("written_k_13_cast_fp16")]; + tensor var_1828 = const()[name = tensor("op_1828"), val = tensor([256, 1024])]; + tensor var_1829_cast_fp16 = reshape(shape = var_1828, x = v_13_cast_fp16)[name = tensor("op_1829_cast_fp16")]; + tensor written_v_13_transpose_x_1 = const()[name = tensor("written_v_13_transpose_x_1"), val = tensor(true)]; + tensor written_v_13_transpose_y_1 = const()[name = tensor("written_v_13_transpose_y_1"), val = tensor(false)]; + tensor written_v_13_cast_fp16 = matmul(transpose_x = written_v_13_transpose_x_1, transpose_y = written_v_13_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_1829_cast_fp16)[name = tensor("written_v_13_cast_fp16")]; + tensor k_cache6_to_fp16_dtype_0 = const()[name = tensor("k_cache6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache6_to_fp16 = cast(dtype = k_cache6_to_fp16_dtype_0, x = k_cache6)[name = tensor("cast_429")]; + tensor var_1840_cast_fp16 = mul(x = k_cache6_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_1840_cast_fp16")]; + tensor var_1845 = const()[name = tensor("op_1845"), val = tensor([1, 512, 16, 64])]; + tensor var_1846_cast_fp16 = reshape(shape = var_1845, x = written_k_13_cast_fp16)[name = tensor("op_1846_cast_fp16")]; + tensor new_k_cache_13_cast_fp16 = add(x = var_1840_cast_fp16, y = var_1846_cast_fp16)[name = tensor("new_k_cache_13_cast_fp16")]; + tensor new_k_cache_13_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_13_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache6_to_fp16_dtype_0 = const()[name = tensor("v_cache6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache6_to_fp16 = cast(dtype = v_cache6_to_fp16_dtype_0, x = v_cache6)[name = tensor("cast_427")]; + tensor var_1849_cast_fp16 = mul(x = v_cache6_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_1849_cast_fp16")]; + tensor var_1854 = const()[name = tensor("op_1854"), val = tensor([1, 512, 16, 64])]; + tensor var_1855_cast_fp16 = reshape(shape = var_1854, x = written_v_13_cast_fp16)[name = tensor("op_1855_cast_fp16")]; + tensor new_v_cache_13_cast_fp16 = add(x = var_1849_cast_fp16, y = var_1855_cast_fp16)[name = tensor("new_v_cache_13_cast_fp16")]; + tensor new_v_cache_13_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_13_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_1868 = const()[name = tensor("op_1868"), val = tensor([0, 2, 1, 3])]; + tensor var_1879_transpose_x_0 = const()[name = tensor("op_1879_transpose_x_0"), val = tensor(false)]; + tensor var_1879_transpose_y_0 = const()[name = tensor("op_1879_transpose_y_0"), val = tensor(false)]; + tensor transpose_81_perm_0 = const()[name = tensor("transpose_81_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_82_perm_0 = const()[name = tensor("transpose_82_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_82 = transpose(perm = transpose_82_perm_0, x = new_k_cache_13_cast_fp16)[name = tensor("transpose_180")]; + tensor transpose_81 = transpose(perm = transpose_81_perm_0, x = q_13_cast_fp16)[name = tensor("transpose_181")]; + tensor var_1879_cast_fp16 = matmul(transpose_x = var_1879_transpose_x_0, transpose_y = var_1879_transpose_y_0, x = transpose_81, y = transpose_82)[name = tensor("op_1879_cast_fp16")]; + tensor var_1880_to_fp16 = const()[name = tensor("op_1880_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_25_cast_fp16 = mul(x = var_1879_cast_fp16, y = var_1880_to_fp16)[name = tensor("scores_25_cast_fp16")]; + tensor scores_27_cast_fp16 = add(x = scores_25_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_27_cast_fp16")]; + tensor var_1890 = const()[name = tensor("op_1890"), val = tensor(-1)]; + tensor weights_13_cast_fp16 = softmax(axis = var_1890, x = scores_27_cast_fp16)[name = tensor("weights_13_cast_fp16")]; + tensor attn_13_transpose_x_0 = const()[name = tensor("attn_13_transpose_x_0"), val = tensor(false)]; + tensor attn_13_transpose_y_0 = const()[name = tensor("attn_13_transpose_y_0"), val = tensor(false)]; + tensor values_13_cast_fp16 = transpose(perm = var_1868, x = new_v_cache_13_cast_fp16)[name = tensor("transpose_182")]; + tensor attn_13_cast_fp16 = matmul(transpose_x = attn_13_transpose_x_0, transpose_y = attn_13_transpose_y_0, x = weights_13_cast_fp16, y = values_13_cast_fp16)[name = tensor("attn_13_cast_fp16")]; + tensor var_1898 = const()[name = tensor("op_1898"), val = tensor([0, 2, 1, 3])]; + tensor var_1903 = const()[name = tensor("op_1903"), val = tensor([1, 256, 1024])]; + tensor var_1899_cast_fp16 = transpose(perm = var_1898, x = attn_13_cast_fp16)[name = tensor("transpose_179")]; + tensor input_77_cast_fp16 = reshape(shape = var_1903, x = var_1899_cast_fp16)[name = tensor("input_77_cast_fp16")]; + tensor attn6_out_proj_weight_to_fp16 = const()[name = tensor("attn6_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(157361920)))]; + tensor linear_25_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn6_out_proj_weight_to_fp16, x = input_77_cast_fp16)[name = tensor("linear_25_cast_fp16")]; + tensor position6_to_fp16_dtype_0 = const()[name = tensor("position6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position6_to_fp16 = cast(dtype = position6_to_fp16_dtype_0, x = position6)[name = tensor("cast_425")]; + tensor var_1912_cast_fp16 = add(x = position6_to_fp16, y = valid_len_to_fp16)[name = tensor("op_1912_cast_fp16")]; + tensor var_1912_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1912_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_79_cast_fp16 = add(x = input_73_cast_fp16, y = linear_25_cast_fp16)[name = tensor("input_79_cast_fp16")]; + tensor input_81_axes_0 = const()[name = tensor("input_81_axes_0"), val = tensor([-1])]; + tensor norm6_2_weight_to_fp16 = const()[name = tensor("norm6_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159459136)))]; + tensor norm6_2_bias_to_fp16 = const()[name = tensor("norm6_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159461248)))]; + tensor var_1916_to_fp16 = const()[name = tensor("op_1916_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_81_cast_fp16 = layer_norm(axes = input_81_axes_0, beta = norm6_2_bias_to_fp16, epsilon = var_1916_to_fp16, gamma = norm6_2_weight_to_fp16, x = input_79_cast_fp16)[name = tensor("input_81_cast_fp16")]; + tensor linear6_1_weight_to_fp16 = const()[name = tensor("linear6_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159463360)))]; + tensor linear_26_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear6_1_weight_to_fp16, x = input_81_cast_fp16)[name = tensor("linear_26_cast_fp16")]; + tensor input_83_mode_0 = const()[name = tensor("input_83_mode_0"), val = tensor("EXACT")]; + tensor input_83_cast_fp16 = gelu(mode = input_83_mode_0, x = linear_26_cast_fp16)[name = tensor("input_83_cast_fp16")]; + tensor linear6_2_weight_to_fp16 = const()[name = tensor("linear6_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(167852032)))]; + tensor linear_27_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear6_2_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_27_cast_fp16")]; + tensor input_85_cast_fp16 = add(x = input_79_cast_fp16, y = linear_27_cast_fp16)[name = tensor("input_85_cast_fp16")]; + tensor input_87_axes_0 = const()[name = tensor("input_87_axes_0"), val = tensor([-1])]; + tensor norm7_1_weight_to_fp16 = const()[name = tensor("norm7_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176240704)))]; + tensor norm7_1_bias_to_fp16 = const()[name = tensor("norm7_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176242816)))]; + tensor var_1933_to_fp16 = const()[name = tensor("op_1933_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_87_cast_fp16 = layer_norm(axes = input_87_axes_0, beta = norm7_1_bias_to_fp16, epsilon = var_1933_to_fp16, gamma = norm7_1_weight_to_fp16, x = input_85_cast_fp16)[name = tensor("input_87_cast_fp16")]; + tensor attn7_in_proj_weight_to_fp16 = const()[name = tensor("attn7_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176244928)))]; + tensor linear_28_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn7_in_proj_weight_to_fp16, x = input_87_cast_fp16)[name = tensor("linear_28_cast_fp16")]; + tensor var_1947 = const()[name = tensor("op_1947"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_15_cast_fp16 = reshape(shape = var_1947, x = linear_28_cast_fp16)[name = tensor("qkv_15_cast_fp16")]; + tensor x_29_begin_0 = const()[name = tensor("x_29_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_29_end_0 = const()[name = tensor("x_29_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_29_end_mask_0 = const()[name = tensor("x_29_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_29_squeeze_mask_0 = const()[name = tensor("x_29_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_29_cast_fp16 = slice_by_index(begin = x_29_begin_0, end = x_29_end_0, end_mask = x_29_end_mask_0, squeeze_mask = x_29_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("x_29_cast_fp16")]; + tensor x_31_begin_0 = const()[name = tensor("x_31_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_31_end_0 = const()[name = tensor("x_31_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_31_end_mask_0 = const()[name = tensor("x_31_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_31_squeeze_mask_0 = const()[name = tensor("x_31_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_31_cast_fp16 = slice_by_index(begin = x_31_begin_0, end = x_31_end_0, end_mask = x_31_end_mask_0, squeeze_mask = x_31_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("x_31_cast_fp16")]; + tensor v_15_begin_0 = const()[name = tensor("v_15_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_15_end_0 = const()[name = tensor("v_15_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_15_end_mask_0 = const()[name = tensor("v_15_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_15_squeeze_mask_0 = const()[name = tensor("v_15_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_15_cast_fp16 = slice_by_index(begin = v_15_begin_0, end = v_15_end_0, end_mask = v_15_end_mask_0, squeeze_mask = v_15_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("v_15_cast_fp16")]; + tensor var_1992 = const()[name = tensor("op_1992"), val = tensor([256, 16, 32, 2])]; + tensor pairs_29_cast_fp16 = reshape(shape = var_1992, x = x_29_cast_fp16)[name = tensor("pairs_29_cast_fp16")]; + tensor xr_29_begin_0 = const()[name = tensor("xr_29_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_29_end_0 = const()[name = tensor("xr_29_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_29_end_mask_0 = const()[name = tensor("xr_29_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_29_squeeze_mask_0 = const()[name = tensor("xr_29_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_29_cast_fp16 = slice_by_index(begin = xr_29_begin_0, end = xr_29_end_0, end_mask = xr_29_end_mask_0, squeeze_mask = xr_29_squeeze_mask_0, x = pairs_29_cast_fp16)[name = tensor("xr_29_cast_fp16")]; + tensor xi_29_begin_0 = const()[name = tensor("xi_29_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_29_end_0 = const()[name = tensor("xi_29_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_29_end_mask_0 = const()[name = tensor("xi_29_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_29_squeeze_mask_0 = const()[name = tensor("xi_29_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_29_cast_fp16 = slice_by_index(begin = xi_29_begin_0, end = xi_29_end_0, end_mask = xi_29_end_mask_0, squeeze_mask = xi_29_squeeze_mask_0, x = pairs_29_cast_fp16)[name = tensor("xi_29_cast_fp16")]; + tensor var_2000_cast_fp16 = mul(x = xr_29_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2000_cast_fp16")]; + tensor var_2001_cast_fp16 = mul(x = xi_29_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2001_cast_fp16")]; + tensor xor__29_cast_fp16 = sub(x = var_2000_cast_fp16, y = var_2001_cast_fp16)[name = tensor("xor__29_cast_fp16")]; + tensor var_2004_cast_fp16 = mul(x = xr_29_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2004_cast_fp16")]; + tensor var_2005_cast_fp16 = mul(x = xi_29_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2005_cast_fp16")]; + tensor xoi_29_cast_fp16 = add(x = var_2004_cast_fp16, y = var_2005_cast_fp16)[name = tensor("xoi_29_cast_fp16")]; + tensor out_29_axis_0 = const()[name = tensor("out_29_axis_0"), val = tensor(-1)]; + tensor out_29_cast_fp16 = stack(axis = out_29_axis_0, values = (xor__29_cast_fp16, xoi_29_cast_fp16))[name = tensor("out_29_cast_fp16")]; + tensor var_2015 = const()[name = tensor("op_2015"), val = tensor([1, 256, 16, 64])]; + tensor q_15_cast_fp16 = reshape(shape = var_2015, x = out_29_cast_fp16)[name = tensor("q_15_cast_fp16")]; + tensor var_2021 = const()[name = tensor("op_2021"), val = tensor([256, 16, 32, 2])]; + tensor pairs_31_cast_fp16 = reshape(shape = var_2021, x = x_31_cast_fp16)[name = tensor("pairs_31_cast_fp16")]; + tensor xr_31_begin_0 = const()[name = tensor("xr_31_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_31_end_0 = const()[name = tensor("xr_31_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_31_end_mask_0 = const()[name = tensor("xr_31_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_31_squeeze_mask_0 = const()[name = tensor("xr_31_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_31_cast_fp16 = slice_by_index(begin = xr_31_begin_0, end = xr_31_end_0, end_mask = xr_31_end_mask_0, squeeze_mask = xr_31_squeeze_mask_0, x = pairs_31_cast_fp16)[name = tensor("xr_31_cast_fp16")]; + tensor xi_31_begin_0 = const()[name = tensor("xi_31_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_31_end_0 = const()[name = tensor("xi_31_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_31_end_mask_0 = const()[name = tensor("xi_31_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_31_squeeze_mask_0 = const()[name = tensor("xi_31_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_31_cast_fp16 = slice_by_index(begin = xi_31_begin_0, end = xi_31_end_0, end_mask = xi_31_end_mask_0, squeeze_mask = xi_31_squeeze_mask_0, x = pairs_31_cast_fp16)[name = tensor("xi_31_cast_fp16")]; + tensor var_2029_cast_fp16 = mul(x = xr_31_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2029_cast_fp16")]; + tensor var_2030_cast_fp16 = mul(x = xi_31_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2030_cast_fp16")]; + tensor xor__31_cast_fp16 = sub(x = var_2029_cast_fp16, y = var_2030_cast_fp16)[name = tensor("xor__31_cast_fp16")]; + tensor var_2033_cast_fp16 = mul(x = xr_31_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2033_cast_fp16")]; + tensor var_2034_cast_fp16 = mul(x = xi_31_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2034_cast_fp16")]; + tensor xoi_31_cast_fp16 = add(x = var_2033_cast_fp16, y = var_2034_cast_fp16)[name = tensor("xoi_31_cast_fp16")]; + tensor out_31_axis_0 = const()[name = tensor("out_31_axis_0"), val = tensor(-1)]; + tensor out_31_cast_fp16 = stack(axis = out_31_axis_0, values = (xor__31_cast_fp16, xoi_31_cast_fp16))[name = tensor("out_31_cast_fp16")]; + tensor var_2051 = const()[name = tensor("op_2051"), val = tensor([256, 1024])]; + tensor var_2052_cast_fp16 = reshape(shape = var_2051, x = out_31_cast_fp16)[name = tensor("op_2052_cast_fp16")]; + tensor written_k_15_transpose_x_1 = const()[name = tensor("written_k_15_transpose_x_1"), val = tensor(true)]; + tensor written_k_15_transpose_y_1 = const()[name = tensor("written_k_15_transpose_y_1"), val = tensor(false)]; + tensor written_k_15_cast_fp16 = matmul(transpose_x = written_k_15_transpose_x_1, transpose_y = written_k_15_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_2052_cast_fp16)[name = tensor("written_k_15_cast_fp16")]; + tensor var_2059 = const()[name = tensor("op_2059"), val = tensor([256, 1024])]; + tensor var_2060_cast_fp16 = reshape(shape = var_2059, x = v_15_cast_fp16)[name = tensor("op_2060_cast_fp16")]; + tensor written_v_15_transpose_x_1 = const()[name = tensor("written_v_15_transpose_x_1"), val = tensor(true)]; + tensor written_v_15_transpose_y_1 = const()[name = tensor("written_v_15_transpose_y_1"), val = tensor(false)]; + tensor written_v_15_cast_fp16 = matmul(transpose_x = written_v_15_transpose_x_1, transpose_y = written_v_15_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_2060_cast_fp16)[name = tensor("written_v_15_cast_fp16")]; + tensor k_cache7_to_fp16_dtype_0 = const()[name = tensor("k_cache7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache7_to_fp16 = cast(dtype = k_cache7_to_fp16_dtype_0, x = k_cache7)[name = tensor("cast_423")]; + tensor var_2071_cast_fp16 = mul(x = k_cache7_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_2071_cast_fp16")]; + tensor var_2076 = const()[name = tensor("op_2076"), val = tensor([1, 512, 16, 64])]; + tensor var_2077_cast_fp16 = reshape(shape = var_2076, x = written_k_15_cast_fp16)[name = tensor("op_2077_cast_fp16")]; + tensor new_k_cache_15_cast_fp16 = add(x = var_2071_cast_fp16, y = var_2077_cast_fp16)[name = tensor("new_k_cache_15_cast_fp16")]; + tensor new_k_cache_15_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_15_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache7_to_fp16_dtype_0 = const()[name = tensor("v_cache7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache7_to_fp16 = cast(dtype = v_cache7_to_fp16_dtype_0, x = v_cache7)[name = tensor("cast_421")]; + tensor var_2080_cast_fp16 = mul(x = v_cache7_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_2080_cast_fp16")]; + tensor var_2085 = const()[name = tensor("op_2085"), val = tensor([1, 512, 16, 64])]; + tensor var_2086_cast_fp16 = reshape(shape = var_2085, x = written_v_15_cast_fp16)[name = tensor("op_2086_cast_fp16")]; + tensor new_v_cache_15_cast_fp16 = add(x = var_2080_cast_fp16, y = var_2086_cast_fp16)[name = tensor("new_v_cache_15_cast_fp16")]; + tensor new_v_cache_15_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_15_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_2099 = const()[name = tensor("op_2099"), val = tensor([0, 2, 1, 3])]; + tensor var_2110_transpose_x_0 = const()[name = tensor("op_2110_transpose_x_0"), val = tensor(false)]; + tensor var_2110_transpose_y_0 = const()[name = tensor("op_2110_transpose_y_0"), val = tensor(false)]; + tensor transpose_83_perm_0 = const()[name = tensor("transpose_83_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_84_perm_0 = const()[name = tensor("transpose_84_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_84 = transpose(perm = transpose_84_perm_0, x = new_k_cache_15_cast_fp16)[name = tensor("transpose_176")]; + tensor transpose_83 = transpose(perm = transpose_83_perm_0, x = q_15_cast_fp16)[name = tensor("transpose_177")]; + tensor var_2110_cast_fp16 = matmul(transpose_x = var_2110_transpose_x_0, transpose_y = var_2110_transpose_y_0, x = transpose_83, y = transpose_84)[name = tensor("op_2110_cast_fp16")]; + tensor var_2111_to_fp16 = const()[name = tensor("op_2111_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_29_cast_fp16 = mul(x = var_2110_cast_fp16, y = var_2111_to_fp16)[name = tensor("scores_29_cast_fp16")]; + tensor scores_31_cast_fp16 = add(x = scores_29_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_31_cast_fp16")]; + tensor var_2121 = const()[name = tensor("op_2121"), val = tensor(-1)]; + tensor weights_15_cast_fp16 = softmax(axis = var_2121, x = scores_31_cast_fp16)[name = tensor("weights_15_cast_fp16")]; + tensor attn_15_transpose_x_0 = const()[name = tensor("attn_15_transpose_x_0"), val = tensor(false)]; + tensor attn_15_transpose_y_0 = const()[name = tensor("attn_15_transpose_y_0"), val = tensor(false)]; + tensor values_15_cast_fp16 = transpose(perm = var_2099, x = new_v_cache_15_cast_fp16)[name = tensor("transpose_178")]; + tensor attn_15_cast_fp16 = matmul(transpose_x = attn_15_transpose_x_0, transpose_y = attn_15_transpose_y_0, x = weights_15_cast_fp16, y = values_15_cast_fp16)[name = tensor("attn_15_cast_fp16")]; + tensor var_2129 = const()[name = tensor("op_2129"), val = tensor([0, 2, 1, 3])]; + tensor var_2134 = const()[name = tensor("op_2134"), val = tensor([1, 256, 1024])]; + tensor var_2130_cast_fp16 = transpose(perm = var_2129, x = attn_15_cast_fp16)[name = tensor("transpose_175")]; + tensor input_89_cast_fp16 = reshape(shape = var_2134, x = var_2130_cast_fp16)[name = tensor("input_89_cast_fp16")]; + tensor attn7_out_proj_weight_to_fp16 = const()[name = tensor("attn7_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(182536448)))]; + tensor linear_29_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn7_out_proj_weight_to_fp16, x = input_89_cast_fp16)[name = tensor("linear_29_cast_fp16")]; + tensor position7_to_fp16_dtype_0 = const()[name = tensor("position7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position7_to_fp16 = cast(dtype = position7_to_fp16_dtype_0, x = position7)[name = tensor("cast_419")]; + tensor var_2143_cast_fp16 = add(x = position7_to_fp16, y = valid_len_to_fp16)[name = tensor("op_2143_cast_fp16")]; + tensor var_2143_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2143_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_91_cast_fp16 = add(x = input_85_cast_fp16, y = linear_29_cast_fp16)[name = tensor("input_91_cast_fp16")]; + tensor input_93_axes_0 = const()[name = tensor("input_93_axes_0"), val = tensor([-1])]; + tensor norm7_2_weight_to_fp16 = const()[name = tensor("norm7_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184633664)))]; + tensor norm7_2_bias_to_fp16 = const()[name = tensor("norm7_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184635776)))]; + tensor var_2147_to_fp16 = const()[name = tensor("op_2147_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_93_cast_fp16 = layer_norm(axes = input_93_axes_0, beta = norm7_2_bias_to_fp16, epsilon = var_2147_to_fp16, gamma = norm7_2_weight_to_fp16, x = input_91_cast_fp16)[name = tensor("input_93_cast_fp16")]; + tensor linear7_1_weight_to_fp16 = const()[name = tensor("linear7_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184637888)))]; + tensor linear_30_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear7_1_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("linear_30_cast_fp16")]; + tensor input_95_mode_0 = const()[name = tensor("input_95_mode_0"), val = tensor("EXACT")]; + tensor input_95_cast_fp16 = gelu(mode = input_95_mode_0, x = linear_30_cast_fp16)[name = tensor("input_95_cast_fp16")]; + tensor linear7_2_weight_to_fp16 = const()[name = tensor("linear7_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(193026560)))]; + tensor linear_31_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear7_2_weight_to_fp16, x = input_95_cast_fp16)[name = tensor("linear_31_cast_fp16")]; + tensor input_97_cast_fp16 = add(x = input_91_cast_fp16, y = linear_31_cast_fp16)[name = tensor("input_97_cast_fp16")]; + tensor input_99_axes_0 = const()[name = tensor("input_99_axes_0"), val = tensor([-1])]; + tensor norm8_1_weight_to_fp16 = const()[name = tensor("norm8_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201415232)))]; + tensor norm8_1_bias_to_fp16 = const()[name = tensor("norm8_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201417344)))]; + tensor var_2164_to_fp16 = const()[name = tensor("op_2164_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_99_cast_fp16 = layer_norm(axes = input_99_axes_0, beta = norm8_1_bias_to_fp16, epsilon = var_2164_to_fp16, gamma = norm8_1_weight_to_fp16, x = input_97_cast_fp16)[name = tensor("input_99_cast_fp16")]; + tensor attn8_in_proj_weight_to_fp16 = const()[name = tensor("attn8_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201419456)))]; + tensor linear_32_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn8_in_proj_weight_to_fp16, x = input_99_cast_fp16)[name = tensor("linear_32_cast_fp16")]; + tensor var_2178 = const()[name = tensor("op_2178"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_17_cast_fp16 = reshape(shape = var_2178, x = linear_32_cast_fp16)[name = tensor("qkv_17_cast_fp16")]; + tensor x_33_begin_0 = const()[name = tensor("x_33_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_33_end_0 = const()[name = tensor("x_33_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_33_end_mask_0 = const()[name = tensor("x_33_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_33_squeeze_mask_0 = const()[name = tensor("x_33_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_33_cast_fp16 = slice_by_index(begin = x_33_begin_0, end = x_33_end_0, end_mask = x_33_end_mask_0, squeeze_mask = x_33_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("x_33_cast_fp16")]; + tensor x_35_begin_0 = const()[name = tensor("x_35_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_35_end_0 = const()[name = tensor("x_35_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_35_end_mask_0 = const()[name = tensor("x_35_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_35_squeeze_mask_0 = const()[name = tensor("x_35_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_35_cast_fp16 = slice_by_index(begin = x_35_begin_0, end = x_35_end_0, end_mask = x_35_end_mask_0, squeeze_mask = x_35_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("x_35_cast_fp16")]; + tensor v_17_begin_0 = const()[name = tensor("v_17_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_17_end_0 = const()[name = tensor("v_17_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_17_end_mask_0 = const()[name = tensor("v_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_17_squeeze_mask_0 = const()[name = tensor("v_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_17_cast_fp16 = slice_by_index(begin = v_17_begin_0, end = v_17_end_0, end_mask = v_17_end_mask_0, squeeze_mask = v_17_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("v_17_cast_fp16")]; + tensor var_2223 = const()[name = tensor("op_2223"), val = tensor([256, 16, 32, 2])]; + tensor pairs_33_cast_fp16 = reshape(shape = var_2223, x = x_33_cast_fp16)[name = tensor("pairs_33_cast_fp16")]; + tensor xr_33_begin_0 = const()[name = tensor("xr_33_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_33_end_0 = const()[name = tensor("xr_33_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_33_end_mask_0 = const()[name = tensor("xr_33_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_33_squeeze_mask_0 = const()[name = tensor("xr_33_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_33_cast_fp16 = slice_by_index(begin = xr_33_begin_0, end = xr_33_end_0, end_mask = xr_33_end_mask_0, squeeze_mask = xr_33_squeeze_mask_0, x = pairs_33_cast_fp16)[name = tensor("xr_33_cast_fp16")]; + tensor xi_33_begin_0 = const()[name = tensor("xi_33_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_33_end_0 = const()[name = tensor("xi_33_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_33_end_mask_0 = const()[name = tensor("xi_33_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_33_squeeze_mask_0 = const()[name = tensor("xi_33_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_33_cast_fp16 = slice_by_index(begin = xi_33_begin_0, end = xi_33_end_0, end_mask = xi_33_end_mask_0, squeeze_mask = xi_33_squeeze_mask_0, x = pairs_33_cast_fp16)[name = tensor("xi_33_cast_fp16")]; + tensor var_2231_cast_fp16 = mul(x = xr_33_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2231_cast_fp16")]; + tensor var_2232_cast_fp16 = mul(x = xi_33_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2232_cast_fp16")]; + tensor xor__33_cast_fp16 = sub(x = var_2231_cast_fp16, y = var_2232_cast_fp16)[name = tensor("xor__33_cast_fp16")]; + tensor var_2235_cast_fp16 = mul(x = xr_33_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2235_cast_fp16")]; + tensor var_2236_cast_fp16 = mul(x = xi_33_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2236_cast_fp16")]; + tensor xoi_33_cast_fp16 = add(x = var_2235_cast_fp16, y = var_2236_cast_fp16)[name = tensor("xoi_33_cast_fp16")]; + tensor out_33_axis_0 = const()[name = tensor("out_33_axis_0"), val = tensor(-1)]; + tensor out_33_cast_fp16 = stack(axis = out_33_axis_0, values = (xor__33_cast_fp16, xoi_33_cast_fp16))[name = tensor("out_33_cast_fp16")]; + tensor var_2246 = const()[name = tensor("op_2246"), val = tensor([1, 256, 16, 64])]; + tensor q_17_cast_fp16 = reshape(shape = var_2246, x = out_33_cast_fp16)[name = tensor("q_17_cast_fp16")]; + tensor var_2252 = const()[name = tensor("op_2252"), val = tensor([256, 16, 32, 2])]; + tensor pairs_35_cast_fp16 = reshape(shape = var_2252, x = x_35_cast_fp16)[name = tensor("pairs_35_cast_fp16")]; + tensor xr_35_begin_0 = const()[name = tensor("xr_35_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_35_end_0 = const()[name = tensor("xr_35_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_35_end_mask_0 = const()[name = tensor("xr_35_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_35_squeeze_mask_0 = const()[name = tensor("xr_35_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_35_cast_fp16 = slice_by_index(begin = xr_35_begin_0, end = xr_35_end_0, end_mask = xr_35_end_mask_0, squeeze_mask = xr_35_squeeze_mask_0, x = pairs_35_cast_fp16)[name = tensor("xr_35_cast_fp16")]; + tensor xi_35_begin_0 = const()[name = tensor("xi_35_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_35_end_0 = const()[name = tensor("xi_35_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_35_end_mask_0 = const()[name = tensor("xi_35_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_35_squeeze_mask_0 = const()[name = tensor("xi_35_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_35_cast_fp16 = slice_by_index(begin = xi_35_begin_0, end = xi_35_end_0, end_mask = xi_35_end_mask_0, squeeze_mask = xi_35_squeeze_mask_0, x = pairs_35_cast_fp16)[name = tensor("xi_35_cast_fp16")]; + tensor var_2260_cast_fp16 = mul(x = xr_35_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2260_cast_fp16")]; + tensor var_2261_cast_fp16 = mul(x = xi_35_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2261_cast_fp16")]; + tensor xor__35_cast_fp16 = sub(x = var_2260_cast_fp16, y = var_2261_cast_fp16)[name = tensor("xor__35_cast_fp16")]; + tensor var_2264_cast_fp16 = mul(x = xr_35_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2264_cast_fp16")]; + tensor var_2265_cast_fp16 = mul(x = xi_35_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2265_cast_fp16")]; + tensor xoi_35_cast_fp16 = add(x = var_2264_cast_fp16, y = var_2265_cast_fp16)[name = tensor("xoi_35_cast_fp16")]; + tensor out_35_axis_0 = const()[name = tensor("out_35_axis_0"), val = tensor(-1)]; + tensor out_35_cast_fp16 = stack(axis = out_35_axis_0, values = (xor__35_cast_fp16, xoi_35_cast_fp16))[name = tensor("out_35_cast_fp16")]; + tensor var_2282 = const()[name = tensor("op_2282"), val = tensor([256, 1024])]; + tensor var_2283_cast_fp16 = reshape(shape = var_2282, x = out_35_cast_fp16)[name = tensor("op_2283_cast_fp16")]; + tensor written_k_17_transpose_x_1 = const()[name = tensor("written_k_17_transpose_x_1"), val = tensor(true)]; + tensor written_k_17_transpose_y_1 = const()[name = tensor("written_k_17_transpose_y_1"), val = tensor(false)]; + tensor written_k_17_cast_fp16 = matmul(transpose_x = written_k_17_transpose_x_1, transpose_y = written_k_17_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_2283_cast_fp16)[name = tensor("written_k_17_cast_fp16")]; + tensor var_2290 = const()[name = tensor("op_2290"), val = tensor([256, 1024])]; + tensor var_2291_cast_fp16 = reshape(shape = var_2290, x = v_17_cast_fp16)[name = tensor("op_2291_cast_fp16")]; + tensor written_v_17_transpose_x_1 = const()[name = tensor("written_v_17_transpose_x_1"), val = tensor(true)]; + tensor written_v_17_transpose_y_1 = const()[name = tensor("written_v_17_transpose_y_1"), val = tensor(false)]; + tensor written_v_17_cast_fp16 = matmul(transpose_x = written_v_17_transpose_x_1, transpose_y = written_v_17_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_2291_cast_fp16)[name = tensor("written_v_17_cast_fp16")]; + tensor k_cache8_to_fp16_dtype_0 = const()[name = tensor("k_cache8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache8_to_fp16 = cast(dtype = k_cache8_to_fp16_dtype_0, x = k_cache8)[name = tensor("cast_417")]; + tensor var_2302_cast_fp16 = mul(x = k_cache8_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_2302_cast_fp16")]; + tensor var_2307 = const()[name = tensor("op_2307"), val = tensor([1, 512, 16, 64])]; + tensor var_2308_cast_fp16 = reshape(shape = var_2307, x = written_k_17_cast_fp16)[name = tensor("op_2308_cast_fp16")]; + tensor new_k_cache_17_cast_fp16 = add(x = var_2302_cast_fp16, y = var_2308_cast_fp16)[name = tensor("new_k_cache_17_cast_fp16")]; + tensor new_k_cache_17_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_17_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache8_to_fp16_dtype_0 = const()[name = tensor("v_cache8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache8_to_fp16 = cast(dtype = v_cache8_to_fp16_dtype_0, x = v_cache8)[name = tensor("cast_415")]; + tensor var_2311_cast_fp16 = mul(x = v_cache8_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_2311_cast_fp16")]; + tensor var_2316 = const()[name = tensor("op_2316"), val = tensor([1, 512, 16, 64])]; + tensor var_2317_cast_fp16 = reshape(shape = var_2316, x = written_v_17_cast_fp16)[name = tensor("op_2317_cast_fp16")]; + tensor new_v_cache_17_cast_fp16 = add(x = var_2311_cast_fp16, y = var_2317_cast_fp16)[name = tensor("new_v_cache_17_cast_fp16")]; + tensor new_v_cache_17_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_17_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_2330 = const()[name = tensor("op_2330"), val = tensor([0, 2, 1, 3])]; + tensor var_2341_transpose_x_0 = const()[name = tensor("op_2341_transpose_x_0"), val = tensor(false)]; + tensor var_2341_transpose_y_0 = const()[name = tensor("op_2341_transpose_y_0"), val = tensor(false)]; + tensor transpose_85_perm_0 = const()[name = tensor("transpose_85_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_86_perm_0 = const()[name = tensor("transpose_86_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_86 = transpose(perm = transpose_86_perm_0, x = new_k_cache_17_cast_fp16)[name = tensor("transpose_172")]; + tensor transpose_85 = transpose(perm = transpose_85_perm_0, x = q_17_cast_fp16)[name = tensor("transpose_173")]; + tensor var_2341_cast_fp16 = matmul(transpose_x = var_2341_transpose_x_0, transpose_y = var_2341_transpose_y_0, x = transpose_85, y = transpose_86)[name = tensor("op_2341_cast_fp16")]; + tensor var_2342_to_fp16 = const()[name = tensor("op_2342_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_33_cast_fp16 = mul(x = var_2341_cast_fp16, y = var_2342_to_fp16)[name = tensor("scores_33_cast_fp16")]; + tensor scores_35_cast_fp16 = add(x = scores_33_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_35_cast_fp16")]; + tensor var_2352 = const()[name = tensor("op_2352"), val = tensor(-1)]; + tensor weights_17_cast_fp16 = softmax(axis = var_2352, x = scores_35_cast_fp16)[name = tensor("weights_17_cast_fp16")]; + tensor attn_17_transpose_x_0 = const()[name = tensor("attn_17_transpose_x_0"), val = tensor(false)]; + tensor attn_17_transpose_y_0 = const()[name = tensor("attn_17_transpose_y_0"), val = tensor(false)]; + tensor values_17_cast_fp16 = transpose(perm = var_2330, x = new_v_cache_17_cast_fp16)[name = tensor("transpose_174")]; + tensor attn_17_cast_fp16 = matmul(transpose_x = attn_17_transpose_x_0, transpose_y = attn_17_transpose_y_0, x = weights_17_cast_fp16, y = values_17_cast_fp16)[name = tensor("attn_17_cast_fp16")]; + tensor var_2360 = const()[name = tensor("op_2360"), val = tensor([0, 2, 1, 3])]; + tensor var_2365 = const()[name = tensor("op_2365"), val = tensor([1, 256, 1024])]; + tensor var_2361_cast_fp16 = transpose(perm = var_2360, x = attn_17_cast_fp16)[name = tensor("transpose_171")]; + tensor input_101_cast_fp16 = reshape(shape = var_2365, x = var_2361_cast_fp16)[name = tensor("input_101_cast_fp16")]; + tensor attn8_out_proj_weight_to_fp16 = const()[name = tensor("attn8_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(207710976)))]; + tensor linear_33_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn8_out_proj_weight_to_fp16, x = input_101_cast_fp16)[name = tensor("linear_33_cast_fp16")]; + tensor position8_to_fp16_dtype_0 = const()[name = tensor("position8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position8_to_fp16 = cast(dtype = position8_to_fp16_dtype_0, x = position8)[name = tensor("cast_413")]; + tensor var_2374_cast_fp16 = add(x = position8_to_fp16, y = valid_len_to_fp16)[name = tensor("op_2374_cast_fp16")]; + tensor var_2374_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2374_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_103_cast_fp16 = add(x = input_97_cast_fp16, y = linear_33_cast_fp16)[name = tensor("input_103_cast_fp16")]; + tensor input_105_axes_0 = const()[name = tensor("input_105_axes_0"), val = tensor([-1])]; + tensor norm8_2_weight_to_fp16 = const()[name = tensor("norm8_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209808192)))]; + tensor norm8_2_bias_to_fp16 = const()[name = tensor("norm8_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209810304)))]; + tensor var_2378_to_fp16 = const()[name = tensor("op_2378_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_105_cast_fp16 = layer_norm(axes = input_105_axes_0, beta = norm8_2_bias_to_fp16, epsilon = var_2378_to_fp16, gamma = norm8_2_weight_to_fp16, x = input_103_cast_fp16)[name = tensor("input_105_cast_fp16")]; + tensor linear8_1_weight_to_fp16 = const()[name = tensor("linear8_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209812416)))]; + tensor linear_34_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear8_1_weight_to_fp16, x = input_105_cast_fp16)[name = tensor("linear_34_cast_fp16")]; + tensor input_107_mode_0 = const()[name = tensor("input_107_mode_0"), val = tensor("EXACT")]; + tensor input_107_cast_fp16 = gelu(mode = input_107_mode_0, x = linear_34_cast_fp16)[name = tensor("input_107_cast_fp16")]; + tensor linear8_2_weight_to_fp16 = const()[name = tensor("linear8_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(218201088)))]; + tensor linear_35_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear8_2_weight_to_fp16, x = input_107_cast_fp16)[name = tensor("linear_35_cast_fp16")]; + tensor input_109_cast_fp16 = add(x = input_103_cast_fp16, y = linear_35_cast_fp16)[name = tensor("input_109_cast_fp16")]; + tensor input_111_axes_0 = const()[name = tensor("input_111_axes_0"), val = tensor([-1])]; + tensor norm9_1_weight_to_fp16 = const()[name = tensor("norm9_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226589760)))]; + tensor norm9_1_bias_to_fp16 = const()[name = tensor("norm9_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226591872)))]; + tensor var_2395_to_fp16 = const()[name = tensor("op_2395_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_111_cast_fp16 = layer_norm(axes = input_111_axes_0, beta = norm9_1_bias_to_fp16, epsilon = var_2395_to_fp16, gamma = norm9_1_weight_to_fp16, x = input_109_cast_fp16)[name = tensor("input_111_cast_fp16")]; + tensor attn9_in_proj_weight_to_fp16 = const()[name = tensor("attn9_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226593984)))]; + tensor linear_36_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn9_in_proj_weight_to_fp16, x = input_111_cast_fp16)[name = tensor("linear_36_cast_fp16")]; + tensor var_2409 = const()[name = tensor("op_2409"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_19_cast_fp16 = reshape(shape = var_2409, x = linear_36_cast_fp16)[name = tensor("qkv_19_cast_fp16")]; + tensor x_37_begin_0 = const()[name = tensor("x_37_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_37_end_0 = const()[name = tensor("x_37_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_37_end_mask_0 = const()[name = tensor("x_37_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_37_squeeze_mask_0 = const()[name = tensor("x_37_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_37_cast_fp16 = slice_by_index(begin = x_37_begin_0, end = x_37_end_0, end_mask = x_37_end_mask_0, squeeze_mask = x_37_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("x_37_cast_fp16")]; + tensor x_39_begin_0 = const()[name = tensor("x_39_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_39_end_0 = const()[name = tensor("x_39_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_39_end_mask_0 = const()[name = tensor("x_39_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_39_squeeze_mask_0 = const()[name = tensor("x_39_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_39_cast_fp16 = slice_by_index(begin = x_39_begin_0, end = x_39_end_0, end_mask = x_39_end_mask_0, squeeze_mask = x_39_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("x_39_cast_fp16")]; + tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_19_cast_fp16 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("v_19_cast_fp16")]; + tensor var_2454 = const()[name = tensor("op_2454"), val = tensor([256, 16, 32, 2])]; + tensor pairs_37_cast_fp16 = reshape(shape = var_2454, x = x_37_cast_fp16)[name = tensor("pairs_37_cast_fp16")]; + tensor xr_37_begin_0 = const()[name = tensor("xr_37_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_37_end_0 = const()[name = tensor("xr_37_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_37_end_mask_0 = const()[name = tensor("xr_37_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_37_squeeze_mask_0 = const()[name = tensor("xr_37_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_37_cast_fp16 = slice_by_index(begin = xr_37_begin_0, end = xr_37_end_0, end_mask = xr_37_end_mask_0, squeeze_mask = xr_37_squeeze_mask_0, x = pairs_37_cast_fp16)[name = tensor("xr_37_cast_fp16")]; + tensor xi_37_begin_0 = const()[name = tensor("xi_37_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_37_end_0 = const()[name = tensor("xi_37_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_37_end_mask_0 = const()[name = tensor("xi_37_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_37_squeeze_mask_0 = const()[name = tensor("xi_37_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_37_cast_fp16 = slice_by_index(begin = xi_37_begin_0, end = xi_37_end_0, end_mask = xi_37_end_mask_0, squeeze_mask = xi_37_squeeze_mask_0, x = pairs_37_cast_fp16)[name = tensor("xi_37_cast_fp16")]; + tensor var_2462_cast_fp16 = mul(x = xr_37_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2462_cast_fp16")]; + tensor var_2463_cast_fp16 = mul(x = xi_37_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2463_cast_fp16")]; + tensor xor__37_cast_fp16 = sub(x = var_2462_cast_fp16, y = var_2463_cast_fp16)[name = tensor("xor__37_cast_fp16")]; + tensor var_2466_cast_fp16 = mul(x = xr_37_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2466_cast_fp16")]; + tensor var_2467_cast_fp16 = mul(x = xi_37_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2467_cast_fp16")]; + tensor xoi_37_cast_fp16 = add(x = var_2466_cast_fp16, y = var_2467_cast_fp16)[name = tensor("xoi_37_cast_fp16")]; + tensor out_37_axis_0 = const()[name = tensor("out_37_axis_0"), val = tensor(-1)]; + tensor out_37_cast_fp16 = stack(axis = out_37_axis_0, values = (xor__37_cast_fp16, xoi_37_cast_fp16))[name = tensor("out_37_cast_fp16")]; + tensor var_2477 = const()[name = tensor("op_2477"), val = tensor([1, 256, 16, 64])]; + tensor q_19_cast_fp16 = reshape(shape = var_2477, x = out_37_cast_fp16)[name = tensor("q_19_cast_fp16")]; + tensor var_2483 = const()[name = tensor("op_2483"), val = tensor([256, 16, 32, 2])]; + tensor pairs_39_cast_fp16 = reshape(shape = var_2483, x = x_39_cast_fp16)[name = tensor("pairs_39_cast_fp16")]; + tensor xr_39_begin_0 = const()[name = tensor("xr_39_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_39_end_0 = const()[name = tensor("xr_39_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_39_end_mask_0 = const()[name = tensor("xr_39_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_39_squeeze_mask_0 = const()[name = tensor("xr_39_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_39_cast_fp16 = slice_by_index(begin = xr_39_begin_0, end = xr_39_end_0, end_mask = xr_39_end_mask_0, squeeze_mask = xr_39_squeeze_mask_0, x = pairs_39_cast_fp16)[name = tensor("xr_39_cast_fp16")]; + tensor xi_39_begin_0 = const()[name = tensor("xi_39_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_39_end_0 = const()[name = tensor("xi_39_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_39_end_mask_0 = const()[name = tensor("xi_39_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_39_squeeze_mask_0 = const()[name = tensor("xi_39_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_39_cast_fp16 = slice_by_index(begin = xi_39_begin_0, end = xi_39_end_0, end_mask = xi_39_end_mask_0, squeeze_mask = xi_39_squeeze_mask_0, x = pairs_39_cast_fp16)[name = tensor("xi_39_cast_fp16")]; + tensor var_2491_cast_fp16 = mul(x = xr_39_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2491_cast_fp16")]; + tensor var_2492_cast_fp16 = mul(x = xi_39_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2492_cast_fp16")]; + tensor xor__39_cast_fp16 = sub(x = var_2491_cast_fp16, y = var_2492_cast_fp16)[name = tensor("xor__39_cast_fp16")]; + tensor var_2495_cast_fp16 = mul(x = xr_39_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2495_cast_fp16")]; + tensor var_2496_cast_fp16 = mul(x = xi_39_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2496_cast_fp16")]; + tensor xoi_39_cast_fp16 = add(x = var_2495_cast_fp16, y = var_2496_cast_fp16)[name = tensor("xoi_39_cast_fp16")]; + tensor out_39_axis_0 = const()[name = tensor("out_39_axis_0"), val = tensor(-1)]; + tensor out_39_cast_fp16 = stack(axis = out_39_axis_0, values = (xor__39_cast_fp16, xoi_39_cast_fp16))[name = tensor("out_39_cast_fp16")]; + tensor var_2513 = const()[name = tensor("op_2513"), val = tensor([256, 1024])]; + tensor var_2514_cast_fp16 = reshape(shape = var_2513, x = out_39_cast_fp16)[name = tensor("op_2514_cast_fp16")]; + tensor written_k_19_transpose_x_1 = const()[name = tensor("written_k_19_transpose_x_1"), val = tensor(true)]; + tensor written_k_19_transpose_y_1 = const()[name = tensor("written_k_19_transpose_y_1"), val = tensor(false)]; + tensor written_k_19_cast_fp16 = matmul(transpose_x = written_k_19_transpose_x_1, transpose_y = written_k_19_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_2514_cast_fp16)[name = tensor("written_k_19_cast_fp16")]; + tensor var_2521 = const()[name = tensor("op_2521"), val = tensor([256, 1024])]; + tensor var_2522_cast_fp16 = reshape(shape = var_2521, x = v_19_cast_fp16)[name = tensor("op_2522_cast_fp16")]; + tensor written_v_19_transpose_x_1 = const()[name = tensor("written_v_19_transpose_x_1"), val = tensor(true)]; + tensor written_v_19_transpose_y_1 = const()[name = tensor("written_v_19_transpose_y_1"), val = tensor(false)]; + tensor written_v_19_cast_fp16 = matmul(transpose_x = written_v_19_transpose_x_1, transpose_y = written_v_19_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_2522_cast_fp16)[name = tensor("written_v_19_cast_fp16")]; + tensor k_cache9_to_fp16_dtype_0 = const()[name = tensor("k_cache9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache9_to_fp16 = cast(dtype = k_cache9_to_fp16_dtype_0, x = k_cache9)[name = tensor("cast_411")]; + tensor var_2533_cast_fp16 = mul(x = k_cache9_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_2533_cast_fp16")]; + tensor var_2538 = const()[name = tensor("op_2538"), val = tensor([1, 512, 16, 64])]; + tensor var_2539_cast_fp16 = reshape(shape = var_2538, x = written_k_19_cast_fp16)[name = tensor("op_2539_cast_fp16")]; + tensor new_k_cache_19_cast_fp16 = add(x = var_2533_cast_fp16, y = var_2539_cast_fp16)[name = tensor("new_k_cache_19_cast_fp16")]; + tensor new_k_cache_19_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_19_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache9_to_fp16_dtype_0 = const()[name = tensor("v_cache9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache9_to_fp16 = cast(dtype = v_cache9_to_fp16_dtype_0, x = v_cache9)[name = tensor("cast_409")]; + tensor var_2542_cast_fp16 = mul(x = v_cache9_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_2542_cast_fp16")]; + tensor var_2547 = const()[name = tensor("op_2547"), val = tensor([1, 512, 16, 64])]; + tensor var_2548_cast_fp16 = reshape(shape = var_2547, x = written_v_19_cast_fp16)[name = tensor("op_2548_cast_fp16")]; + tensor new_v_cache_19_cast_fp16 = add(x = var_2542_cast_fp16, y = var_2548_cast_fp16)[name = tensor("new_v_cache_19_cast_fp16")]; + tensor new_v_cache_19_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_19_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_2561 = const()[name = tensor("op_2561"), val = tensor([0, 2, 1, 3])]; + tensor var_2572_transpose_x_0 = const()[name = tensor("op_2572_transpose_x_0"), val = tensor(false)]; + tensor var_2572_transpose_y_0 = const()[name = tensor("op_2572_transpose_y_0"), val = tensor(false)]; + tensor transpose_87_perm_0 = const()[name = tensor("transpose_87_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_88_perm_0 = const()[name = tensor("transpose_88_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_88 = transpose(perm = transpose_88_perm_0, x = new_k_cache_19_cast_fp16)[name = tensor("transpose_168")]; + tensor transpose_87 = transpose(perm = transpose_87_perm_0, x = q_19_cast_fp16)[name = tensor("transpose_169")]; + tensor var_2572_cast_fp16 = matmul(transpose_x = var_2572_transpose_x_0, transpose_y = var_2572_transpose_y_0, x = transpose_87, y = transpose_88)[name = tensor("op_2572_cast_fp16")]; + tensor var_2573_to_fp16 = const()[name = tensor("op_2573_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_37_cast_fp16 = mul(x = var_2572_cast_fp16, y = var_2573_to_fp16)[name = tensor("scores_37_cast_fp16")]; + tensor scores_39_cast_fp16 = add(x = scores_37_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_39_cast_fp16")]; + tensor var_2583 = const()[name = tensor("op_2583"), val = tensor(-1)]; + tensor weights_19_cast_fp16 = softmax(axis = var_2583, x = scores_39_cast_fp16)[name = tensor("weights_19_cast_fp16")]; + tensor attn_19_transpose_x_0 = const()[name = tensor("attn_19_transpose_x_0"), val = tensor(false)]; + tensor attn_19_transpose_y_0 = const()[name = tensor("attn_19_transpose_y_0"), val = tensor(false)]; + tensor values_19_cast_fp16 = transpose(perm = var_2561, x = new_v_cache_19_cast_fp16)[name = tensor("transpose_170")]; + tensor attn_19_cast_fp16 = matmul(transpose_x = attn_19_transpose_x_0, transpose_y = attn_19_transpose_y_0, x = weights_19_cast_fp16, y = values_19_cast_fp16)[name = tensor("attn_19_cast_fp16")]; + tensor var_2591 = const()[name = tensor("op_2591"), val = tensor([0, 2, 1, 3])]; + tensor var_2596 = const()[name = tensor("op_2596"), val = tensor([1, 256, 1024])]; + tensor var_2592_cast_fp16 = transpose(perm = var_2591, x = attn_19_cast_fp16)[name = tensor("transpose_167")]; + tensor input_113_cast_fp16 = reshape(shape = var_2596, x = var_2592_cast_fp16)[name = tensor("input_113_cast_fp16")]; + tensor attn9_out_proj_weight_to_fp16 = const()[name = tensor("attn9_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(232885504)))]; + tensor linear_37_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn9_out_proj_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("linear_37_cast_fp16")]; + tensor position9_to_fp16_dtype_0 = const()[name = tensor("position9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position9_to_fp16 = cast(dtype = position9_to_fp16_dtype_0, x = position9)[name = tensor("cast_407")]; + tensor var_2605_cast_fp16 = add(x = position9_to_fp16, y = valid_len_to_fp16)[name = tensor("op_2605_cast_fp16")]; + tensor var_2605_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2605_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_115_cast_fp16 = add(x = input_109_cast_fp16, y = linear_37_cast_fp16)[name = tensor("input_115_cast_fp16")]; + tensor input_117_axes_0 = const()[name = tensor("input_117_axes_0"), val = tensor([-1])]; + tensor norm9_2_weight_to_fp16 = const()[name = tensor("norm9_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(234982720)))]; + tensor norm9_2_bias_to_fp16 = const()[name = tensor("norm9_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(234984832)))]; + tensor var_2609_to_fp16 = const()[name = tensor("op_2609_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_117_cast_fp16 = layer_norm(axes = input_117_axes_0, beta = norm9_2_bias_to_fp16, epsilon = var_2609_to_fp16, gamma = norm9_2_weight_to_fp16, x = input_115_cast_fp16)[name = tensor("input_117_cast_fp16")]; + tensor linear9_1_weight_to_fp16 = const()[name = tensor("linear9_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(234986944)))]; + tensor linear_38_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear9_1_weight_to_fp16, x = input_117_cast_fp16)[name = tensor("linear_38_cast_fp16")]; + tensor input_119_mode_0 = const()[name = tensor("input_119_mode_0"), val = tensor("EXACT")]; + tensor input_119_cast_fp16 = gelu(mode = input_119_mode_0, x = linear_38_cast_fp16)[name = tensor("input_119_cast_fp16")]; + tensor linear9_2_weight_to_fp16 = const()[name = tensor("linear9_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(243375616)))]; + tensor linear_39_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear9_2_weight_to_fp16, x = input_119_cast_fp16)[name = tensor("linear_39_cast_fp16")]; + tensor input_121_cast_fp16 = add(x = input_115_cast_fp16, y = linear_39_cast_fp16)[name = tensor("input_121_cast_fp16")]; + tensor input_123_axes_0 = const()[name = tensor("input_123_axes_0"), val = tensor([-1])]; + tensor norm10_1_weight_to_fp16 = const()[name = tensor("norm10_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(251764288)))]; + tensor norm10_1_bias_to_fp16 = const()[name = tensor("norm10_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(251766400)))]; + tensor var_2626_to_fp16 = const()[name = tensor("op_2626_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_123_cast_fp16 = layer_norm(axes = input_123_axes_0, beta = norm10_1_bias_to_fp16, epsilon = var_2626_to_fp16, gamma = norm10_1_weight_to_fp16, x = input_121_cast_fp16)[name = tensor("input_123_cast_fp16")]; + tensor attn10_in_proj_weight_to_fp16 = const()[name = tensor("attn10_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(251768512)))]; + tensor linear_40_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn10_in_proj_weight_to_fp16, x = input_123_cast_fp16)[name = tensor("linear_40_cast_fp16")]; + tensor var_2640 = const()[name = tensor("op_2640"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_21_cast_fp16 = reshape(shape = var_2640, x = linear_40_cast_fp16)[name = tensor("qkv_21_cast_fp16")]; + tensor x_41_begin_0 = const()[name = tensor("x_41_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_41_end_0 = const()[name = tensor("x_41_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_41_end_mask_0 = const()[name = tensor("x_41_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_41_squeeze_mask_0 = const()[name = tensor("x_41_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_41_cast_fp16 = slice_by_index(begin = x_41_begin_0, end = x_41_end_0, end_mask = x_41_end_mask_0, squeeze_mask = x_41_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("x_41_cast_fp16")]; + tensor x_43_begin_0 = const()[name = tensor("x_43_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_43_end_0 = const()[name = tensor("x_43_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_43_end_mask_0 = const()[name = tensor("x_43_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_43_squeeze_mask_0 = const()[name = tensor("x_43_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_43_cast_fp16 = slice_by_index(begin = x_43_begin_0, end = x_43_end_0, end_mask = x_43_end_mask_0, squeeze_mask = x_43_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("x_43_cast_fp16")]; + tensor v_21_begin_0 = const()[name = tensor("v_21_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_21_end_0 = const()[name = tensor("v_21_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_21_end_mask_0 = const()[name = tensor("v_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_21_squeeze_mask_0 = const()[name = tensor("v_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_21_cast_fp16 = slice_by_index(begin = v_21_begin_0, end = v_21_end_0, end_mask = v_21_end_mask_0, squeeze_mask = v_21_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("v_21_cast_fp16")]; + tensor var_2685 = const()[name = tensor("op_2685"), val = tensor([256, 16, 32, 2])]; + tensor pairs_41_cast_fp16 = reshape(shape = var_2685, x = x_41_cast_fp16)[name = tensor("pairs_41_cast_fp16")]; + tensor xr_41_begin_0 = const()[name = tensor("xr_41_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_41_end_0 = const()[name = tensor("xr_41_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_41_end_mask_0 = const()[name = tensor("xr_41_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_41_squeeze_mask_0 = const()[name = tensor("xr_41_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_41_cast_fp16 = slice_by_index(begin = xr_41_begin_0, end = xr_41_end_0, end_mask = xr_41_end_mask_0, squeeze_mask = xr_41_squeeze_mask_0, x = pairs_41_cast_fp16)[name = tensor("xr_41_cast_fp16")]; + tensor xi_41_begin_0 = const()[name = tensor("xi_41_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_41_end_0 = const()[name = tensor("xi_41_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_41_end_mask_0 = const()[name = tensor("xi_41_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_41_squeeze_mask_0 = const()[name = tensor("xi_41_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_41_cast_fp16 = slice_by_index(begin = xi_41_begin_0, end = xi_41_end_0, end_mask = xi_41_end_mask_0, squeeze_mask = xi_41_squeeze_mask_0, x = pairs_41_cast_fp16)[name = tensor("xi_41_cast_fp16")]; + tensor var_2693_cast_fp16 = mul(x = xr_41_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2693_cast_fp16")]; + tensor var_2694_cast_fp16 = mul(x = xi_41_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2694_cast_fp16")]; + tensor xor__41_cast_fp16 = sub(x = var_2693_cast_fp16, y = var_2694_cast_fp16)[name = tensor("xor__41_cast_fp16")]; + tensor var_2697_cast_fp16 = mul(x = xr_41_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2697_cast_fp16")]; + tensor var_2698_cast_fp16 = mul(x = xi_41_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2698_cast_fp16")]; + tensor xoi_41_cast_fp16 = add(x = var_2697_cast_fp16, y = var_2698_cast_fp16)[name = tensor("xoi_41_cast_fp16")]; + tensor out_41_axis_0 = const()[name = tensor("out_41_axis_0"), val = tensor(-1)]; + tensor out_41_cast_fp16 = stack(axis = out_41_axis_0, values = (xor__41_cast_fp16, xoi_41_cast_fp16))[name = tensor("out_41_cast_fp16")]; + tensor var_2708 = const()[name = tensor("op_2708"), val = tensor([1, 256, 16, 64])]; + tensor q_21_cast_fp16 = reshape(shape = var_2708, x = out_41_cast_fp16)[name = tensor("q_21_cast_fp16")]; + tensor var_2714 = const()[name = tensor("op_2714"), val = tensor([256, 16, 32, 2])]; + tensor pairs_43_cast_fp16 = reshape(shape = var_2714, x = x_43_cast_fp16)[name = tensor("pairs_43_cast_fp16")]; + tensor xr_43_begin_0 = const()[name = tensor("xr_43_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_43_end_0 = const()[name = tensor("xr_43_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_43_end_mask_0 = const()[name = tensor("xr_43_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_43_squeeze_mask_0 = const()[name = tensor("xr_43_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_43_cast_fp16 = slice_by_index(begin = xr_43_begin_0, end = xr_43_end_0, end_mask = xr_43_end_mask_0, squeeze_mask = xr_43_squeeze_mask_0, x = pairs_43_cast_fp16)[name = tensor("xr_43_cast_fp16")]; + tensor xi_43_begin_0 = const()[name = tensor("xi_43_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_43_end_0 = const()[name = tensor("xi_43_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_43_end_mask_0 = const()[name = tensor("xi_43_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_43_squeeze_mask_0 = const()[name = tensor("xi_43_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_43_cast_fp16 = slice_by_index(begin = xi_43_begin_0, end = xi_43_end_0, end_mask = xi_43_end_mask_0, squeeze_mask = xi_43_squeeze_mask_0, x = pairs_43_cast_fp16)[name = tensor("xi_43_cast_fp16")]; + tensor var_2722_cast_fp16 = mul(x = xr_43_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2722_cast_fp16")]; + tensor var_2723_cast_fp16 = mul(x = xi_43_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2723_cast_fp16")]; + tensor xor__43_cast_fp16 = sub(x = var_2722_cast_fp16, y = var_2723_cast_fp16)[name = tensor("xor__43_cast_fp16")]; + tensor var_2726_cast_fp16 = mul(x = xr_43_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2726_cast_fp16")]; + tensor var_2727_cast_fp16 = mul(x = xi_43_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2727_cast_fp16")]; + tensor xoi_43_cast_fp16 = add(x = var_2726_cast_fp16, y = var_2727_cast_fp16)[name = tensor("xoi_43_cast_fp16")]; + tensor out_43_axis_0 = const()[name = tensor("out_43_axis_0"), val = tensor(-1)]; + tensor out_43_cast_fp16 = stack(axis = out_43_axis_0, values = (xor__43_cast_fp16, xoi_43_cast_fp16))[name = tensor("out_43_cast_fp16")]; + tensor var_2744 = const()[name = tensor("op_2744"), val = tensor([256, 1024])]; + tensor var_2745_cast_fp16 = reshape(shape = var_2744, x = out_43_cast_fp16)[name = tensor("op_2745_cast_fp16")]; + tensor written_k_21_transpose_x_1 = const()[name = tensor("written_k_21_transpose_x_1"), val = tensor(true)]; + tensor written_k_21_transpose_y_1 = const()[name = tensor("written_k_21_transpose_y_1"), val = tensor(false)]; + tensor written_k_21_cast_fp16 = matmul(transpose_x = written_k_21_transpose_x_1, transpose_y = written_k_21_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_2745_cast_fp16)[name = tensor("written_k_21_cast_fp16")]; + tensor var_2752 = const()[name = tensor("op_2752"), val = tensor([256, 1024])]; + tensor var_2753_cast_fp16 = reshape(shape = var_2752, x = v_21_cast_fp16)[name = tensor("op_2753_cast_fp16")]; + tensor written_v_21_transpose_x_1 = const()[name = tensor("written_v_21_transpose_x_1"), val = tensor(true)]; + tensor written_v_21_transpose_y_1 = const()[name = tensor("written_v_21_transpose_y_1"), val = tensor(false)]; + tensor written_v_21_cast_fp16 = matmul(transpose_x = written_v_21_transpose_x_1, transpose_y = written_v_21_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_2753_cast_fp16)[name = tensor("written_v_21_cast_fp16")]; + tensor k_cache10_to_fp16_dtype_0 = const()[name = tensor("k_cache10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache10_to_fp16 = cast(dtype = k_cache10_to_fp16_dtype_0, x = k_cache10)[name = tensor("cast_405")]; + tensor var_2764_cast_fp16 = mul(x = k_cache10_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_2764_cast_fp16")]; + tensor var_2769 = const()[name = tensor("op_2769"), val = tensor([1, 512, 16, 64])]; + tensor var_2770_cast_fp16 = reshape(shape = var_2769, x = written_k_21_cast_fp16)[name = tensor("op_2770_cast_fp16")]; + tensor new_k_cache_21_cast_fp16 = add(x = var_2764_cast_fp16, y = var_2770_cast_fp16)[name = tensor("new_k_cache_21_cast_fp16")]; + tensor new_k_cache_21_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_21_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache10_to_fp16_dtype_0 = const()[name = tensor("v_cache10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache10_to_fp16 = cast(dtype = v_cache10_to_fp16_dtype_0, x = v_cache10)[name = tensor("cast_403")]; + tensor var_2773_cast_fp16 = mul(x = v_cache10_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_2773_cast_fp16")]; + tensor var_2778 = const()[name = tensor("op_2778"), val = tensor([1, 512, 16, 64])]; + tensor var_2779_cast_fp16 = reshape(shape = var_2778, x = written_v_21_cast_fp16)[name = tensor("op_2779_cast_fp16")]; + tensor new_v_cache_21_cast_fp16 = add(x = var_2773_cast_fp16, y = var_2779_cast_fp16)[name = tensor("new_v_cache_21_cast_fp16")]; + tensor new_v_cache_21_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_21_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_2792 = const()[name = tensor("op_2792"), val = tensor([0, 2, 1, 3])]; + tensor var_2803_transpose_x_0 = const()[name = tensor("op_2803_transpose_x_0"), val = tensor(false)]; + tensor var_2803_transpose_y_0 = const()[name = tensor("op_2803_transpose_y_0"), val = tensor(false)]; + tensor transpose_89_perm_0 = const()[name = tensor("transpose_89_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_90_perm_0 = const()[name = tensor("transpose_90_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_90 = transpose(perm = transpose_90_perm_0, x = new_k_cache_21_cast_fp16)[name = tensor("transpose_164")]; + tensor transpose_89 = transpose(perm = transpose_89_perm_0, x = q_21_cast_fp16)[name = tensor("transpose_165")]; + tensor var_2803_cast_fp16 = matmul(transpose_x = var_2803_transpose_x_0, transpose_y = var_2803_transpose_y_0, x = transpose_89, y = transpose_90)[name = tensor("op_2803_cast_fp16")]; + tensor var_2804_to_fp16 = const()[name = tensor("op_2804_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_41_cast_fp16 = mul(x = var_2803_cast_fp16, y = var_2804_to_fp16)[name = tensor("scores_41_cast_fp16")]; + tensor scores_43_cast_fp16 = add(x = scores_41_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_43_cast_fp16")]; + tensor var_2814 = const()[name = tensor("op_2814"), val = tensor(-1)]; + tensor weights_21_cast_fp16 = softmax(axis = var_2814, x = scores_43_cast_fp16)[name = tensor("weights_21_cast_fp16")]; + tensor attn_21_transpose_x_0 = const()[name = tensor("attn_21_transpose_x_0"), val = tensor(false)]; + tensor attn_21_transpose_y_0 = const()[name = tensor("attn_21_transpose_y_0"), val = tensor(false)]; + tensor values_21_cast_fp16 = transpose(perm = var_2792, x = new_v_cache_21_cast_fp16)[name = tensor("transpose_166")]; + tensor attn_21_cast_fp16 = matmul(transpose_x = attn_21_transpose_x_0, transpose_y = attn_21_transpose_y_0, x = weights_21_cast_fp16, y = values_21_cast_fp16)[name = tensor("attn_21_cast_fp16")]; + tensor var_2822 = const()[name = tensor("op_2822"), val = tensor([0, 2, 1, 3])]; + tensor var_2827 = const()[name = tensor("op_2827"), val = tensor([1, 256, 1024])]; + tensor var_2823_cast_fp16 = transpose(perm = var_2822, x = attn_21_cast_fp16)[name = tensor("transpose_163")]; + tensor input_125_cast_fp16 = reshape(shape = var_2827, x = var_2823_cast_fp16)[name = tensor("input_125_cast_fp16")]; + tensor attn10_out_proj_weight_to_fp16 = const()[name = tensor("attn10_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(258060032)))]; + tensor linear_41_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn10_out_proj_weight_to_fp16, x = input_125_cast_fp16)[name = tensor("linear_41_cast_fp16")]; + tensor position10_to_fp16_dtype_0 = const()[name = tensor("position10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position10_to_fp16 = cast(dtype = position10_to_fp16_dtype_0, x = position10)[name = tensor("cast_401")]; + tensor var_2836_cast_fp16 = add(x = position10_to_fp16, y = valid_len_to_fp16)[name = tensor("op_2836_cast_fp16")]; + tensor var_2836_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2836_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_127_cast_fp16 = add(x = input_121_cast_fp16, y = linear_41_cast_fp16)[name = tensor("input_127_cast_fp16")]; + tensor input_129_axes_0 = const()[name = tensor("input_129_axes_0"), val = tensor([-1])]; + tensor norm10_2_weight_to_fp16 = const()[name = tensor("norm10_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260157248)))]; + tensor norm10_2_bias_to_fp16 = const()[name = tensor("norm10_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260159360)))]; + tensor var_2840_to_fp16 = const()[name = tensor("op_2840_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_129_cast_fp16 = layer_norm(axes = input_129_axes_0, beta = norm10_2_bias_to_fp16, epsilon = var_2840_to_fp16, gamma = norm10_2_weight_to_fp16, x = input_127_cast_fp16)[name = tensor("input_129_cast_fp16")]; + tensor linear10_1_weight_to_fp16 = const()[name = tensor("linear10_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260161472)))]; + tensor linear_42_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear10_1_weight_to_fp16, x = input_129_cast_fp16)[name = tensor("linear_42_cast_fp16")]; + tensor input_131_mode_0 = const()[name = tensor("input_131_mode_0"), val = tensor("EXACT")]; + tensor input_131_cast_fp16 = gelu(mode = input_131_mode_0, x = linear_42_cast_fp16)[name = tensor("input_131_cast_fp16")]; + tensor linear10_2_weight_to_fp16 = const()[name = tensor("linear10_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(268550144)))]; + tensor linear_43_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear10_2_weight_to_fp16, x = input_131_cast_fp16)[name = tensor("linear_43_cast_fp16")]; + tensor input_133_cast_fp16 = add(x = input_127_cast_fp16, y = linear_43_cast_fp16)[name = tensor("input_133_cast_fp16")]; + tensor input_135_axes_0 = const()[name = tensor("input_135_axes_0"), val = tensor([-1])]; + tensor norm11_1_weight_to_fp16 = const()[name = tensor("norm11_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(276938816)))]; + tensor norm11_1_bias_to_fp16 = const()[name = tensor("norm11_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(276940928)))]; + tensor var_2857_to_fp16 = const()[name = tensor("op_2857_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_135_cast_fp16 = layer_norm(axes = input_135_axes_0, beta = norm11_1_bias_to_fp16, epsilon = var_2857_to_fp16, gamma = norm11_1_weight_to_fp16, x = input_133_cast_fp16)[name = tensor("input_135_cast_fp16")]; + tensor attn11_in_proj_weight_to_fp16 = const()[name = tensor("attn11_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(276943040)))]; + tensor linear_44_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn11_in_proj_weight_to_fp16, x = input_135_cast_fp16)[name = tensor("linear_44_cast_fp16")]; + tensor var_2871 = const()[name = tensor("op_2871"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_23_cast_fp16 = reshape(shape = var_2871, x = linear_44_cast_fp16)[name = tensor("qkv_23_cast_fp16")]; + tensor x_45_begin_0 = const()[name = tensor("x_45_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_45_end_0 = const()[name = tensor("x_45_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_45_end_mask_0 = const()[name = tensor("x_45_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_45_squeeze_mask_0 = const()[name = tensor("x_45_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_45_cast_fp16 = slice_by_index(begin = x_45_begin_0, end = x_45_end_0, end_mask = x_45_end_mask_0, squeeze_mask = x_45_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("x_45_cast_fp16")]; + tensor x_47_begin_0 = const()[name = tensor("x_47_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_47_end_0 = const()[name = tensor("x_47_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_47_end_mask_0 = const()[name = tensor("x_47_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_47_squeeze_mask_0 = const()[name = tensor("x_47_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_47_cast_fp16 = slice_by_index(begin = x_47_begin_0, end = x_47_end_0, end_mask = x_47_end_mask_0, squeeze_mask = x_47_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("x_47_cast_fp16")]; + tensor v_23_begin_0 = const()[name = tensor("v_23_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_23_end_0 = const()[name = tensor("v_23_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_23_end_mask_0 = const()[name = tensor("v_23_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_23_squeeze_mask_0 = const()[name = tensor("v_23_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_23_cast_fp16 = slice_by_index(begin = v_23_begin_0, end = v_23_end_0, end_mask = v_23_end_mask_0, squeeze_mask = v_23_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("v_23_cast_fp16")]; + tensor var_2916 = const()[name = tensor("op_2916"), val = tensor([256, 16, 32, 2])]; + tensor pairs_45_cast_fp16 = reshape(shape = var_2916, x = x_45_cast_fp16)[name = tensor("pairs_45_cast_fp16")]; + tensor xr_45_begin_0 = const()[name = tensor("xr_45_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_45_end_0 = const()[name = tensor("xr_45_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_45_end_mask_0 = const()[name = tensor("xr_45_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_45_squeeze_mask_0 = const()[name = tensor("xr_45_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_45_cast_fp16 = slice_by_index(begin = xr_45_begin_0, end = xr_45_end_0, end_mask = xr_45_end_mask_0, squeeze_mask = xr_45_squeeze_mask_0, x = pairs_45_cast_fp16)[name = tensor("xr_45_cast_fp16")]; + tensor xi_45_begin_0 = const()[name = tensor("xi_45_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_45_end_0 = const()[name = tensor("xi_45_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_45_end_mask_0 = const()[name = tensor("xi_45_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_45_squeeze_mask_0 = const()[name = tensor("xi_45_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_45_cast_fp16 = slice_by_index(begin = xi_45_begin_0, end = xi_45_end_0, end_mask = xi_45_end_mask_0, squeeze_mask = xi_45_squeeze_mask_0, x = pairs_45_cast_fp16)[name = tensor("xi_45_cast_fp16")]; + tensor var_2924_cast_fp16 = mul(x = xr_45_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2924_cast_fp16")]; + tensor var_2925_cast_fp16 = mul(x = xi_45_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2925_cast_fp16")]; + tensor xor__45_cast_fp16 = sub(x = var_2924_cast_fp16, y = var_2925_cast_fp16)[name = tensor("xor__45_cast_fp16")]; + tensor var_2928_cast_fp16 = mul(x = xr_45_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2928_cast_fp16")]; + tensor var_2929_cast_fp16 = mul(x = xi_45_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2929_cast_fp16")]; + tensor xoi_45_cast_fp16 = add(x = var_2928_cast_fp16, y = var_2929_cast_fp16)[name = tensor("xoi_45_cast_fp16")]; + tensor out_45_axis_0 = const()[name = tensor("out_45_axis_0"), val = tensor(-1)]; + tensor out_45_cast_fp16 = stack(axis = out_45_axis_0, values = (xor__45_cast_fp16, xoi_45_cast_fp16))[name = tensor("out_45_cast_fp16")]; + tensor var_2939 = const()[name = tensor("op_2939"), val = tensor([1, 256, 16, 64])]; + tensor q_23_cast_fp16 = reshape(shape = var_2939, x = out_45_cast_fp16)[name = tensor("q_23_cast_fp16")]; + tensor var_2945 = const()[name = tensor("op_2945"), val = tensor([256, 16, 32, 2])]; + tensor pairs_47_cast_fp16 = reshape(shape = var_2945, x = x_47_cast_fp16)[name = tensor("pairs_47_cast_fp16")]; + tensor xr_47_begin_0 = const()[name = tensor("xr_47_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_47_end_0 = const()[name = tensor("xr_47_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_47_end_mask_0 = const()[name = tensor("xr_47_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_47_squeeze_mask_0 = const()[name = tensor("xr_47_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_47_cast_fp16 = slice_by_index(begin = xr_47_begin_0, end = xr_47_end_0, end_mask = xr_47_end_mask_0, squeeze_mask = xr_47_squeeze_mask_0, x = pairs_47_cast_fp16)[name = tensor("xr_47_cast_fp16")]; + tensor xi_47_begin_0 = const()[name = tensor("xi_47_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_47_end_0 = const()[name = tensor("xi_47_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_47_end_mask_0 = const()[name = tensor("xi_47_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_47_squeeze_mask_0 = const()[name = tensor("xi_47_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_47_cast_fp16 = slice_by_index(begin = xi_47_begin_0, end = xi_47_end_0, end_mask = xi_47_end_mask_0, squeeze_mask = xi_47_squeeze_mask_0, x = pairs_47_cast_fp16)[name = tensor("xi_47_cast_fp16")]; + tensor var_2953_cast_fp16 = mul(x = xr_47_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2953_cast_fp16")]; + tensor var_2954_cast_fp16 = mul(x = xi_47_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2954_cast_fp16")]; + tensor xor__47_cast_fp16 = sub(x = var_2953_cast_fp16, y = var_2954_cast_fp16)[name = tensor("xor__47_cast_fp16")]; + tensor var_2957_cast_fp16 = mul(x = xr_47_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2957_cast_fp16")]; + tensor var_2958_cast_fp16 = mul(x = xi_47_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2958_cast_fp16")]; + tensor xoi_47_cast_fp16 = add(x = var_2957_cast_fp16, y = var_2958_cast_fp16)[name = tensor("xoi_47_cast_fp16")]; + tensor out_47_axis_0 = const()[name = tensor("out_47_axis_0"), val = tensor(-1)]; + tensor out_47_cast_fp16 = stack(axis = out_47_axis_0, values = (xor__47_cast_fp16, xoi_47_cast_fp16))[name = tensor("out_47_cast_fp16")]; + tensor var_2975 = const()[name = tensor("op_2975"), val = tensor([256, 1024])]; + tensor var_2976_cast_fp16 = reshape(shape = var_2975, x = out_47_cast_fp16)[name = tensor("op_2976_cast_fp16")]; + tensor written_k_23_transpose_x_1 = const()[name = tensor("written_k_23_transpose_x_1"), val = tensor(true)]; + tensor written_k_23_transpose_y_1 = const()[name = tensor("written_k_23_transpose_y_1"), val = tensor(false)]; + tensor written_k_23_cast_fp16 = matmul(transpose_x = written_k_23_transpose_x_1, transpose_y = written_k_23_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_2976_cast_fp16)[name = tensor("written_k_23_cast_fp16")]; + tensor var_2983 = const()[name = tensor("op_2983"), val = tensor([256, 1024])]; + tensor var_2984_cast_fp16 = reshape(shape = var_2983, x = v_23_cast_fp16)[name = tensor("op_2984_cast_fp16")]; + tensor written_v_23_transpose_x_1 = const()[name = tensor("written_v_23_transpose_x_1"), val = tensor(true)]; + tensor written_v_23_transpose_y_1 = const()[name = tensor("written_v_23_transpose_y_1"), val = tensor(false)]; + tensor written_v_23_cast_fp16 = matmul(transpose_x = written_v_23_transpose_x_1, transpose_y = written_v_23_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_2984_cast_fp16)[name = tensor("written_v_23_cast_fp16")]; + tensor k_cache11_to_fp16_dtype_0 = const()[name = tensor("k_cache11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache11_to_fp16 = cast(dtype = k_cache11_to_fp16_dtype_0, x = k_cache11)[name = tensor("cast_399")]; + tensor var_2995_cast_fp16 = mul(x = k_cache11_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_2995_cast_fp16")]; + tensor var_3000 = const()[name = tensor("op_3000"), val = tensor([1, 512, 16, 64])]; + tensor var_3001_cast_fp16 = reshape(shape = var_3000, x = written_k_23_cast_fp16)[name = tensor("op_3001_cast_fp16")]; + tensor new_k_cache_23_cast_fp16 = add(x = var_2995_cast_fp16, y = var_3001_cast_fp16)[name = tensor("new_k_cache_23_cast_fp16")]; + tensor new_k_cache_23_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_23_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache11_to_fp16_dtype_0 = const()[name = tensor("v_cache11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache11_to_fp16 = cast(dtype = v_cache11_to_fp16_dtype_0, x = v_cache11)[name = tensor("cast_397")]; + tensor var_3004_cast_fp16 = mul(x = v_cache11_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_3004_cast_fp16")]; + tensor var_3009 = const()[name = tensor("op_3009"), val = tensor([1, 512, 16, 64])]; + tensor var_3010_cast_fp16 = reshape(shape = var_3009, x = written_v_23_cast_fp16)[name = tensor("op_3010_cast_fp16")]; + tensor new_v_cache_23_cast_fp16 = add(x = var_3004_cast_fp16, y = var_3010_cast_fp16)[name = tensor("new_v_cache_23_cast_fp16")]; + tensor new_v_cache_23_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_23_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_3023 = const()[name = tensor("op_3023"), val = tensor([0, 2, 1, 3])]; + tensor var_3034_transpose_x_0 = const()[name = tensor("op_3034_transpose_x_0"), val = tensor(false)]; + tensor var_3034_transpose_y_0 = const()[name = tensor("op_3034_transpose_y_0"), val = tensor(false)]; + tensor transpose_91_perm_0 = const()[name = tensor("transpose_91_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_92_perm_0 = const()[name = tensor("transpose_92_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_92 = transpose(perm = transpose_92_perm_0, x = new_k_cache_23_cast_fp16)[name = tensor("transpose_160")]; + tensor transpose_91 = transpose(perm = transpose_91_perm_0, x = q_23_cast_fp16)[name = tensor("transpose_161")]; + tensor var_3034_cast_fp16 = matmul(transpose_x = var_3034_transpose_x_0, transpose_y = var_3034_transpose_y_0, x = transpose_91, y = transpose_92)[name = tensor("op_3034_cast_fp16")]; + tensor var_3035_to_fp16 = const()[name = tensor("op_3035_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_45_cast_fp16 = mul(x = var_3034_cast_fp16, y = var_3035_to_fp16)[name = tensor("scores_45_cast_fp16")]; + tensor scores_47_cast_fp16 = add(x = scores_45_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_47_cast_fp16")]; + tensor var_3045 = const()[name = tensor("op_3045"), val = tensor(-1)]; + tensor weights_23_cast_fp16 = softmax(axis = var_3045, x = scores_47_cast_fp16)[name = tensor("weights_23_cast_fp16")]; + tensor attn_23_transpose_x_0 = const()[name = tensor("attn_23_transpose_x_0"), val = tensor(false)]; + tensor attn_23_transpose_y_0 = const()[name = tensor("attn_23_transpose_y_0"), val = tensor(false)]; + tensor values_23_cast_fp16 = transpose(perm = var_3023, x = new_v_cache_23_cast_fp16)[name = tensor("transpose_162")]; + tensor attn_23_cast_fp16 = matmul(transpose_x = attn_23_transpose_x_0, transpose_y = attn_23_transpose_y_0, x = weights_23_cast_fp16, y = values_23_cast_fp16)[name = tensor("attn_23_cast_fp16")]; + tensor var_3053 = const()[name = tensor("op_3053"), val = tensor([0, 2, 1, 3])]; + tensor var_3058 = const()[name = tensor("op_3058"), val = tensor([1, 256, 1024])]; + tensor var_3054_cast_fp16 = transpose(perm = var_3053, x = attn_23_cast_fp16)[name = tensor("transpose_159")]; + tensor input_137_cast_fp16 = reshape(shape = var_3058, x = var_3054_cast_fp16)[name = tensor("input_137_cast_fp16")]; + tensor attn11_out_proj_weight_to_fp16 = const()[name = tensor("attn11_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(283234560)))]; + tensor linear_45_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn11_out_proj_weight_to_fp16, x = input_137_cast_fp16)[name = tensor("linear_45_cast_fp16")]; + tensor position11_to_fp16_dtype_0 = const()[name = tensor("position11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position11_to_fp16 = cast(dtype = position11_to_fp16_dtype_0, x = position11)[name = tensor("cast_395")]; + tensor var_3067_cast_fp16 = add(x = position11_to_fp16, y = valid_len_to_fp16)[name = tensor("op_3067_cast_fp16")]; + tensor var_3067_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3067_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_139_cast_fp16 = add(x = input_133_cast_fp16, y = linear_45_cast_fp16)[name = tensor("input_139_cast_fp16")]; + tensor input_141_axes_0 = const()[name = tensor("input_141_axes_0"), val = tensor([-1])]; + tensor norm11_2_weight_to_fp16 = const()[name = tensor("norm11_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285331776)))]; + tensor norm11_2_bias_to_fp16 = const()[name = tensor("norm11_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285333888)))]; + tensor var_3071_to_fp16 = const()[name = tensor("op_3071_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_141_cast_fp16 = layer_norm(axes = input_141_axes_0, beta = norm11_2_bias_to_fp16, epsilon = var_3071_to_fp16, gamma = norm11_2_weight_to_fp16, x = input_139_cast_fp16)[name = tensor("input_141_cast_fp16")]; + tensor linear11_1_weight_to_fp16 = const()[name = tensor("linear11_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285336000)))]; + tensor linear_46_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear11_1_weight_to_fp16, x = input_141_cast_fp16)[name = tensor("linear_46_cast_fp16")]; + tensor input_143_mode_0 = const()[name = tensor("input_143_mode_0"), val = tensor("EXACT")]; + tensor input_143_cast_fp16 = gelu(mode = input_143_mode_0, x = linear_46_cast_fp16)[name = tensor("input_143_cast_fp16")]; + tensor linear11_2_weight_to_fp16 = const()[name = tensor("linear11_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293724672)))]; + tensor linear_47_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear11_2_weight_to_fp16, x = input_143_cast_fp16)[name = tensor("linear_47_cast_fp16")]; + tensor input_145_cast_fp16 = add(x = input_139_cast_fp16, y = linear_47_cast_fp16)[name = tensor("input_145_cast_fp16")]; + tensor input_147_axes_0 = const()[name = tensor("input_147_axes_0"), val = tensor([-1])]; + tensor norm12_1_weight_to_fp16 = const()[name = tensor("norm12_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302113344)))]; + tensor norm12_1_bias_to_fp16 = const()[name = tensor("norm12_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302115456)))]; + tensor var_3088_to_fp16 = const()[name = tensor("op_3088_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_147_cast_fp16 = layer_norm(axes = input_147_axes_0, beta = norm12_1_bias_to_fp16, epsilon = var_3088_to_fp16, gamma = norm12_1_weight_to_fp16, x = input_145_cast_fp16)[name = tensor("input_147_cast_fp16")]; + tensor attn12_in_proj_weight_to_fp16 = const()[name = tensor("attn12_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302117568)))]; + tensor linear_48_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn12_in_proj_weight_to_fp16, x = input_147_cast_fp16)[name = tensor("linear_48_cast_fp16")]; + tensor var_3102 = const()[name = tensor("op_3102"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_25_cast_fp16 = reshape(shape = var_3102, x = linear_48_cast_fp16)[name = tensor("qkv_25_cast_fp16")]; + tensor x_49_begin_0 = const()[name = tensor("x_49_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_49_end_0 = const()[name = tensor("x_49_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_49_end_mask_0 = const()[name = tensor("x_49_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_49_squeeze_mask_0 = const()[name = tensor("x_49_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_49_cast_fp16 = slice_by_index(begin = x_49_begin_0, end = x_49_end_0, end_mask = x_49_end_mask_0, squeeze_mask = x_49_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("x_49_cast_fp16")]; + tensor x_51_begin_0 = const()[name = tensor("x_51_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_51_end_0 = const()[name = tensor("x_51_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_51_end_mask_0 = const()[name = tensor("x_51_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_51_squeeze_mask_0 = const()[name = tensor("x_51_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_51_cast_fp16 = slice_by_index(begin = x_51_begin_0, end = x_51_end_0, end_mask = x_51_end_mask_0, squeeze_mask = x_51_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("x_51_cast_fp16")]; + tensor v_25_begin_0 = const()[name = tensor("v_25_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_25_end_0 = const()[name = tensor("v_25_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_25_end_mask_0 = const()[name = tensor("v_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_25_squeeze_mask_0 = const()[name = tensor("v_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_25_cast_fp16 = slice_by_index(begin = v_25_begin_0, end = v_25_end_0, end_mask = v_25_end_mask_0, squeeze_mask = v_25_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("v_25_cast_fp16")]; + tensor var_3147 = const()[name = tensor("op_3147"), val = tensor([256, 16, 32, 2])]; + tensor pairs_49_cast_fp16 = reshape(shape = var_3147, x = x_49_cast_fp16)[name = tensor("pairs_49_cast_fp16")]; + tensor xr_49_begin_0 = const()[name = tensor("xr_49_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_49_end_0 = const()[name = tensor("xr_49_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_49_end_mask_0 = const()[name = tensor("xr_49_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_49_squeeze_mask_0 = const()[name = tensor("xr_49_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_49_cast_fp16 = slice_by_index(begin = xr_49_begin_0, end = xr_49_end_0, end_mask = xr_49_end_mask_0, squeeze_mask = xr_49_squeeze_mask_0, x = pairs_49_cast_fp16)[name = tensor("xr_49_cast_fp16")]; + tensor xi_49_begin_0 = const()[name = tensor("xi_49_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_49_end_0 = const()[name = tensor("xi_49_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_49_end_mask_0 = const()[name = tensor("xi_49_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_49_squeeze_mask_0 = const()[name = tensor("xi_49_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_49_cast_fp16 = slice_by_index(begin = xi_49_begin_0, end = xi_49_end_0, end_mask = xi_49_end_mask_0, squeeze_mask = xi_49_squeeze_mask_0, x = pairs_49_cast_fp16)[name = tensor("xi_49_cast_fp16")]; + tensor var_3155_cast_fp16 = mul(x = xr_49_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3155_cast_fp16")]; + tensor var_3156_cast_fp16 = mul(x = xi_49_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3156_cast_fp16")]; + tensor xor__49_cast_fp16 = sub(x = var_3155_cast_fp16, y = var_3156_cast_fp16)[name = tensor("xor__49_cast_fp16")]; + tensor var_3159_cast_fp16 = mul(x = xr_49_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3159_cast_fp16")]; + tensor var_3160_cast_fp16 = mul(x = xi_49_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3160_cast_fp16")]; + tensor xoi_49_cast_fp16 = add(x = var_3159_cast_fp16, y = var_3160_cast_fp16)[name = tensor("xoi_49_cast_fp16")]; + tensor out_49_axis_0 = const()[name = tensor("out_49_axis_0"), val = tensor(-1)]; + tensor out_49_cast_fp16 = stack(axis = out_49_axis_0, values = (xor__49_cast_fp16, xoi_49_cast_fp16))[name = tensor("out_49_cast_fp16")]; + tensor var_3170 = const()[name = tensor("op_3170"), val = tensor([1, 256, 16, 64])]; + tensor q_25_cast_fp16 = reshape(shape = var_3170, x = out_49_cast_fp16)[name = tensor("q_25_cast_fp16")]; + tensor var_3176 = const()[name = tensor("op_3176"), val = tensor([256, 16, 32, 2])]; + tensor pairs_51_cast_fp16 = reshape(shape = var_3176, x = x_51_cast_fp16)[name = tensor("pairs_51_cast_fp16")]; + tensor xr_51_begin_0 = const()[name = tensor("xr_51_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_51_end_0 = const()[name = tensor("xr_51_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_51_end_mask_0 = const()[name = tensor("xr_51_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_51_squeeze_mask_0 = const()[name = tensor("xr_51_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_51_cast_fp16 = slice_by_index(begin = xr_51_begin_0, end = xr_51_end_0, end_mask = xr_51_end_mask_0, squeeze_mask = xr_51_squeeze_mask_0, x = pairs_51_cast_fp16)[name = tensor("xr_51_cast_fp16")]; + tensor xi_51_begin_0 = const()[name = tensor("xi_51_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_51_end_0 = const()[name = tensor("xi_51_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_51_end_mask_0 = const()[name = tensor("xi_51_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_51_squeeze_mask_0 = const()[name = tensor("xi_51_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_51_cast_fp16 = slice_by_index(begin = xi_51_begin_0, end = xi_51_end_0, end_mask = xi_51_end_mask_0, squeeze_mask = xi_51_squeeze_mask_0, x = pairs_51_cast_fp16)[name = tensor("xi_51_cast_fp16")]; + tensor var_3184_cast_fp16 = mul(x = xr_51_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3184_cast_fp16")]; + tensor var_3185_cast_fp16 = mul(x = xi_51_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3185_cast_fp16")]; + tensor xor__51_cast_fp16 = sub(x = var_3184_cast_fp16, y = var_3185_cast_fp16)[name = tensor("xor__51_cast_fp16")]; + tensor var_3188_cast_fp16 = mul(x = xr_51_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3188_cast_fp16")]; + tensor var_3189_cast_fp16 = mul(x = xi_51_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3189_cast_fp16")]; + tensor xoi_51_cast_fp16 = add(x = var_3188_cast_fp16, y = var_3189_cast_fp16)[name = tensor("xoi_51_cast_fp16")]; + tensor out_51_axis_0 = const()[name = tensor("out_51_axis_0"), val = tensor(-1)]; + tensor out_51_cast_fp16 = stack(axis = out_51_axis_0, values = (xor__51_cast_fp16, xoi_51_cast_fp16))[name = tensor("out_51_cast_fp16")]; + tensor var_3206 = const()[name = tensor("op_3206"), val = tensor([256, 1024])]; + tensor var_3207_cast_fp16 = reshape(shape = var_3206, x = out_51_cast_fp16)[name = tensor("op_3207_cast_fp16")]; + tensor written_k_25_transpose_x_1 = const()[name = tensor("written_k_25_transpose_x_1"), val = tensor(true)]; + tensor written_k_25_transpose_y_1 = const()[name = tensor("written_k_25_transpose_y_1"), val = tensor(false)]; + tensor written_k_25_cast_fp16 = matmul(transpose_x = written_k_25_transpose_x_1, transpose_y = written_k_25_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_3207_cast_fp16)[name = tensor("written_k_25_cast_fp16")]; + tensor var_3214 = const()[name = tensor("op_3214"), val = tensor([256, 1024])]; + tensor var_3215_cast_fp16 = reshape(shape = var_3214, x = v_25_cast_fp16)[name = tensor("op_3215_cast_fp16")]; + tensor written_v_25_transpose_x_1 = const()[name = tensor("written_v_25_transpose_x_1"), val = tensor(true)]; + tensor written_v_25_transpose_y_1 = const()[name = tensor("written_v_25_transpose_y_1"), val = tensor(false)]; + tensor written_v_25_cast_fp16 = matmul(transpose_x = written_v_25_transpose_x_1, transpose_y = written_v_25_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_3215_cast_fp16)[name = tensor("written_v_25_cast_fp16")]; + tensor k_cache12_to_fp16_dtype_0 = const()[name = tensor("k_cache12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache12_to_fp16 = cast(dtype = k_cache12_to_fp16_dtype_0, x = k_cache12)[name = tensor("cast_393")]; + tensor var_3226_cast_fp16 = mul(x = k_cache12_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_3226_cast_fp16")]; + tensor var_3231 = const()[name = tensor("op_3231"), val = tensor([1, 512, 16, 64])]; + tensor var_3232_cast_fp16 = reshape(shape = var_3231, x = written_k_25_cast_fp16)[name = tensor("op_3232_cast_fp16")]; + tensor new_k_cache_25_cast_fp16 = add(x = var_3226_cast_fp16, y = var_3232_cast_fp16)[name = tensor("new_k_cache_25_cast_fp16")]; + tensor new_k_cache_25_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_25_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache12_to_fp16_dtype_0 = const()[name = tensor("v_cache12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache12_to_fp16 = cast(dtype = v_cache12_to_fp16_dtype_0, x = v_cache12)[name = tensor("cast_391")]; + tensor var_3235_cast_fp16 = mul(x = v_cache12_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_3235_cast_fp16")]; + tensor var_3240 = const()[name = tensor("op_3240"), val = tensor([1, 512, 16, 64])]; + tensor var_3241_cast_fp16 = reshape(shape = var_3240, x = written_v_25_cast_fp16)[name = tensor("op_3241_cast_fp16")]; + tensor new_v_cache_25_cast_fp16 = add(x = var_3235_cast_fp16, y = var_3241_cast_fp16)[name = tensor("new_v_cache_25_cast_fp16")]; + tensor new_v_cache_25_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_25_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_3254 = const()[name = tensor("op_3254"), val = tensor([0, 2, 1, 3])]; + tensor var_3265_transpose_x_0 = const()[name = tensor("op_3265_transpose_x_0"), val = tensor(false)]; + tensor var_3265_transpose_y_0 = const()[name = tensor("op_3265_transpose_y_0"), val = tensor(false)]; + tensor transpose_93_perm_0 = const()[name = tensor("transpose_93_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_94_perm_0 = const()[name = tensor("transpose_94_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_94 = transpose(perm = transpose_94_perm_0, x = new_k_cache_25_cast_fp16)[name = tensor("transpose_156")]; + tensor transpose_93 = transpose(perm = transpose_93_perm_0, x = q_25_cast_fp16)[name = tensor("transpose_157")]; + tensor var_3265_cast_fp16 = matmul(transpose_x = var_3265_transpose_x_0, transpose_y = var_3265_transpose_y_0, x = transpose_93, y = transpose_94)[name = tensor("op_3265_cast_fp16")]; + tensor var_3266_to_fp16 = const()[name = tensor("op_3266_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_49_cast_fp16 = mul(x = var_3265_cast_fp16, y = var_3266_to_fp16)[name = tensor("scores_49_cast_fp16")]; + tensor scores_51_cast_fp16 = add(x = scores_49_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_51_cast_fp16")]; + tensor var_3276 = const()[name = tensor("op_3276"), val = tensor(-1)]; + tensor weights_25_cast_fp16 = softmax(axis = var_3276, x = scores_51_cast_fp16)[name = tensor("weights_25_cast_fp16")]; + tensor attn_25_transpose_x_0 = const()[name = tensor("attn_25_transpose_x_0"), val = tensor(false)]; + tensor attn_25_transpose_y_0 = const()[name = tensor("attn_25_transpose_y_0"), val = tensor(false)]; + tensor values_25_cast_fp16 = transpose(perm = var_3254, x = new_v_cache_25_cast_fp16)[name = tensor("transpose_158")]; + tensor attn_25_cast_fp16 = matmul(transpose_x = attn_25_transpose_x_0, transpose_y = attn_25_transpose_y_0, x = weights_25_cast_fp16, y = values_25_cast_fp16)[name = tensor("attn_25_cast_fp16")]; + tensor var_3284 = const()[name = tensor("op_3284"), val = tensor([0, 2, 1, 3])]; + tensor var_3289 = const()[name = tensor("op_3289"), val = tensor([1, 256, 1024])]; + tensor var_3285_cast_fp16 = transpose(perm = var_3284, x = attn_25_cast_fp16)[name = tensor("transpose_155")]; + tensor input_149_cast_fp16 = reshape(shape = var_3289, x = var_3285_cast_fp16)[name = tensor("input_149_cast_fp16")]; + tensor attn12_out_proj_weight_to_fp16 = const()[name = tensor("attn12_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(308409088)))]; + tensor linear_49_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn12_out_proj_weight_to_fp16, x = input_149_cast_fp16)[name = tensor("linear_49_cast_fp16")]; + tensor position12_to_fp16_dtype_0 = const()[name = tensor("position12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position12_to_fp16 = cast(dtype = position12_to_fp16_dtype_0, x = position12)[name = tensor("cast_389")]; + tensor var_3298_cast_fp16 = add(x = position12_to_fp16, y = valid_len_to_fp16)[name = tensor("op_3298_cast_fp16")]; + tensor var_3298_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3298_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_151_cast_fp16 = add(x = input_145_cast_fp16, y = linear_49_cast_fp16)[name = tensor("input_151_cast_fp16")]; + tensor input_153_axes_0 = const()[name = tensor("input_153_axes_0"), val = tensor([-1])]; + tensor norm12_2_weight_to_fp16 = const()[name = tensor("norm12_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(310506304)))]; + tensor norm12_2_bias_to_fp16 = const()[name = tensor("norm12_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(310508416)))]; + tensor var_3302_to_fp16 = const()[name = tensor("op_3302_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_153_cast_fp16 = layer_norm(axes = input_153_axes_0, beta = norm12_2_bias_to_fp16, epsilon = var_3302_to_fp16, gamma = norm12_2_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("input_153_cast_fp16")]; + tensor linear12_1_weight_to_fp16 = const()[name = tensor("linear12_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(310510528)))]; + tensor linear_50_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear12_1_weight_to_fp16, x = input_153_cast_fp16)[name = tensor("linear_50_cast_fp16")]; + tensor input_155_mode_0 = const()[name = tensor("input_155_mode_0"), val = tensor("EXACT")]; + tensor input_155_cast_fp16 = gelu(mode = input_155_mode_0, x = linear_50_cast_fp16)[name = tensor("input_155_cast_fp16")]; + tensor linear12_2_weight_to_fp16 = const()[name = tensor("linear12_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(318899200)))]; + tensor linear_51_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear12_2_weight_to_fp16, x = input_155_cast_fp16)[name = tensor("linear_51_cast_fp16")]; + tensor input_157_cast_fp16 = add(x = input_151_cast_fp16, y = linear_51_cast_fp16)[name = tensor("input_157_cast_fp16")]; + tensor input_159_axes_0 = const()[name = tensor("input_159_axes_0"), val = tensor([-1])]; + tensor norm13_1_weight_to_fp16 = const()[name = tensor("norm13_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(327287872)))]; + tensor norm13_1_bias_to_fp16 = const()[name = tensor("norm13_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(327289984)))]; + tensor var_3319_to_fp16 = const()[name = tensor("op_3319_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_159_cast_fp16 = layer_norm(axes = input_159_axes_0, beta = norm13_1_bias_to_fp16, epsilon = var_3319_to_fp16, gamma = norm13_1_weight_to_fp16, x = input_157_cast_fp16)[name = tensor("input_159_cast_fp16")]; + tensor attn13_in_proj_weight_to_fp16 = const()[name = tensor("attn13_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(327292096)))]; + tensor linear_52_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn13_in_proj_weight_to_fp16, x = input_159_cast_fp16)[name = tensor("linear_52_cast_fp16")]; + tensor var_3333 = const()[name = tensor("op_3333"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_27_cast_fp16 = reshape(shape = var_3333, x = linear_52_cast_fp16)[name = tensor("qkv_27_cast_fp16")]; + tensor x_53_begin_0 = const()[name = tensor("x_53_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_53_end_0 = const()[name = tensor("x_53_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_53_end_mask_0 = const()[name = tensor("x_53_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_53_squeeze_mask_0 = const()[name = tensor("x_53_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_53_cast_fp16 = slice_by_index(begin = x_53_begin_0, end = x_53_end_0, end_mask = x_53_end_mask_0, squeeze_mask = x_53_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("x_53_cast_fp16")]; + tensor x_55_begin_0 = const()[name = tensor("x_55_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_55_end_0 = const()[name = tensor("x_55_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_55_end_mask_0 = const()[name = tensor("x_55_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_55_squeeze_mask_0 = const()[name = tensor("x_55_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_55_cast_fp16 = slice_by_index(begin = x_55_begin_0, end = x_55_end_0, end_mask = x_55_end_mask_0, squeeze_mask = x_55_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("x_55_cast_fp16")]; + tensor v_27_begin_0 = const()[name = tensor("v_27_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_27_end_0 = const()[name = tensor("v_27_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_27_end_mask_0 = const()[name = tensor("v_27_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_27_squeeze_mask_0 = const()[name = tensor("v_27_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_27_cast_fp16 = slice_by_index(begin = v_27_begin_0, end = v_27_end_0, end_mask = v_27_end_mask_0, squeeze_mask = v_27_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("v_27_cast_fp16")]; + tensor var_3378 = const()[name = tensor("op_3378"), val = tensor([256, 16, 32, 2])]; + tensor pairs_53_cast_fp16 = reshape(shape = var_3378, x = x_53_cast_fp16)[name = tensor("pairs_53_cast_fp16")]; + tensor xr_53_begin_0 = const()[name = tensor("xr_53_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_53_end_0 = const()[name = tensor("xr_53_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_53_end_mask_0 = const()[name = tensor("xr_53_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_53_squeeze_mask_0 = const()[name = tensor("xr_53_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_53_cast_fp16 = slice_by_index(begin = xr_53_begin_0, end = xr_53_end_0, end_mask = xr_53_end_mask_0, squeeze_mask = xr_53_squeeze_mask_0, x = pairs_53_cast_fp16)[name = tensor("xr_53_cast_fp16")]; + tensor xi_53_begin_0 = const()[name = tensor("xi_53_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_53_end_0 = const()[name = tensor("xi_53_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_53_end_mask_0 = const()[name = tensor("xi_53_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_53_squeeze_mask_0 = const()[name = tensor("xi_53_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_53_cast_fp16 = slice_by_index(begin = xi_53_begin_0, end = xi_53_end_0, end_mask = xi_53_end_mask_0, squeeze_mask = xi_53_squeeze_mask_0, x = pairs_53_cast_fp16)[name = tensor("xi_53_cast_fp16")]; + tensor var_3386_cast_fp16 = mul(x = xr_53_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3386_cast_fp16")]; + tensor var_3387_cast_fp16 = mul(x = xi_53_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3387_cast_fp16")]; + tensor xor__53_cast_fp16 = sub(x = var_3386_cast_fp16, y = var_3387_cast_fp16)[name = tensor("xor__53_cast_fp16")]; + tensor var_3390_cast_fp16 = mul(x = xr_53_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3390_cast_fp16")]; + tensor var_3391_cast_fp16 = mul(x = xi_53_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3391_cast_fp16")]; + tensor xoi_53_cast_fp16 = add(x = var_3390_cast_fp16, y = var_3391_cast_fp16)[name = tensor("xoi_53_cast_fp16")]; + tensor out_53_axis_0 = const()[name = tensor("out_53_axis_0"), val = tensor(-1)]; + tensor out_53_cast_fp16 = stack(axis = out_53_axis_0, values = (xor__53_cast_fp16, xoi_53_cast_fp16))[name = tensor("out_53_cast_fp16")]; + tensor var_3401 = const()[name = tensor("op_3401"), val = tensor([1, 256, 16, 64])]; + tensor q_27_cast_fp16 = reshape(shape = var_3401, x = out_53_cast_fp16)[name = tensor("q_27_cast_fp16")]; + tensor var_3407 = const()[name = tensor("op_3407"), val = tensor([256, 16, 32, 2])]; + tensor pairs_55_cast_fp16 = reshape(shape = var_3407, x = x_55_cast_fp16)[name = tensor("pairs_55_cast_fp16")]; + tensor xr_55_begin_0 = const()[name = tensor("xr_55_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_55_end_0 = const()[name = tensor("xr_55_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_55_end_mask_0 = const()[name = tensor("xr_55_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_55_squeeze_mask_0 = const()[name = tensor("xr_55_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_55_cast_fp16 = slice_by_index(begin = xr_55_begin_0, end = xr_55_end_0, end_mask = xr_55_end_mask_0, squeeze_mask = xr_55_squeeze_mask_0, x = pairs_55_cast_fp16)[name = tensor("xr_55_cast_fp16")]; + tensor xi_55_begin_0 = const()[name = tensor("xi_55_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_55_end_0 = const()[name = tensor("xi_55_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_55_end_mask_0 = const()[name = tensor("xi_55_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_55_squeeze_mask_0 = const()[name = tensor("xi_55_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_55_cast_fp16 = slice_by_index(begin = xi_55_begin_0, end = xi_55_end_0, end_mask = xi_55_end_mask_0, squeeze_mask = xi_55_squeeze_mask_0, x = pairs_55_cast_fp16)[name = tensor("xi_55_cast_fp16")]; + tensor var_3415_cast_fp16 = mul(x = xr_55_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3415_cast_fp16")]; + tensor var_3416_cast_fp16 = mul(x = xi_55_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3416_cast_fp16")]; + tensor xor__55_cast_fp16 = sub(x = var_3415_cast_fp16, y = var_3416_cast_fp16)[name = tensor("xor__55_cast_fp16")]; + tensor var_3419_cast_fp16 = mul(x = xr_55_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3419_cast_fp16")]; + tensor var_3420_cast_fp16 = mul(x = xi_55_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3420_cast_fp16")]; + tensor xoi_55_cast_fp16 = add(x = var_3419_cast_fp16, y = var_3420_cast_fp16)[name = tensor("xoi_55_cast_fp16")]; + tensor out_55_axis_0 = const()[name = tensor("out_55_axis_0"), val = tensor(-1)]; + tensor out_55_cast_fp16 = stack(axis = out_55_axis_0, values = (xor__55_cast_fp16, xoi_55_cast_fp16))[name = tensor("out_55_cast_fp16")]; + tensor var_3437 = const()[name = tensor("op_3437"), val = tensor([256, 1024])]; + tensor var_3438_cast_fp16 = reshape(shape = var_3437, x = out_55_cast_fp16)[name = tensor("op_3438_cast_fp16")]; + tensor written_k_27_transpose_x_1 = const()[name = tensor("written_k_27_transpose_x_1"), val = tensor(true)]; + tensor written_k_27_transpose_y_1 = const()[name = tensor("written_k_27_transpose_y_1"), val = tensor(false)]; + tensor written_k_27_cast_fp16 = matmul(transpose_x = written_k_27_transpose_x_1, transpose_y = written_k_27_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_3438_cast_fp16)[name = tensor("written_k_27_cast_fp16")]; + tensor var_3445 = const()[name = tensor("op_3445"), val = tensor([256, 1024])]; + tensor var_3446_cast_fp16 = reshape(shape = var_3445, x = v_27_cast_fp16)[name = tensor("op_3446_cast_fp16")]; + tensor written_v_27_transpose_x_1 = const()[name = tensor("written_v_27_transpose_x_1"), val = tensor(true)]; + tensor written_v_27_transpose_y_1 = const()[name = tensor("written_v_27_transpose_y_1"), val = tensor(false)]; + tensor written_v_27_cast_fp16 = matmul(transpose_x = written_v_27_transpose_x_1, transpose_y = written_v_27_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_3446_cast_fp16)[name = tensor("written_v_27_cast_fp16")]; + tensor k_cache13_to_fp16_dtype_0 = const()[name = tensor("k_cache13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache13_to_fp16 = cast(dtype = k_cache13_to_fp16_dtype_0, x = k_cache13)[name = tensor("cast_387")]; + tensor var_3457_cast_fp16 = mul(x = k_cache13_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_3457_cast_fp16")]; + tensor var_3462 = const()[name = tensor("op_3462"), val = tensor([1, 512, 16, 64])]; + tensor var_3463_cast_fp16 = reshape(shape = var_3462, x = written_k_27_cast_fp16)[name = tensor("op_3463_cast_fp16")]; + tensor new_k_cache_27_cast_fp16 = add(x = var_3457_cast_fp16, y = var_3463_cast_fp16)[name = tensor("new_k_cache_27_cast_fp16")]; + tensor new_k_cache_27_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_27_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache13_to_fp16_dtype_0 = const()[name = tensor("v_cache13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache13_to_fp16 = cast(dtype = v_cache13_to_fp16_dtype_0, x = v_cache13)[name = tensor("cast_385")]; + tensor var_3466_cast_fp16 = mul(x = v_cache13_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_3466_cast_fp16")]; + tensor var_3471 = const()[name = tensor("op_3471"), val = tensor([1, 512, 16, 64])]; + tensor var_3472_cast_fp16 = reshape(shape = var_3471, x = written_v_27_cast_fp16)[name = tensor("op_3472_cast_fp16")]; + tensor new_v_cache_27_cast_fp16 = add(x = var_3466_cast_fp16, y = var_3472_cast_fp16)[name = tensor("new_v_cache_27_cast_fp16")]; + tensor new_v_cache_27_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_27_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_3485 = const()[name = tensor("op_3485"), val = tensor([0, 2, 1, 3])]; + tensor var_3496_transpose_x_0 = const()[name = tensor("op_3496_transpose_x_0"), val = tensor(false)]; + tensor var_3496_transpose_y_0 = const()[name = tensor("op_3496_transpose_y_0"), val = tensor(false)]; + tensor transpose_95_perm_0 = const()[name = tensor("transpose_95_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_96_perm_0 = const()[name = tensor("transpose_96_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_96 = transpose(perm = transpose_96_perm_0, x = new_k_cache_27_cast_fp16)[name = tensor("transpose_152")]; + tensor transpose_95 = transpose(perm = transpose_95_perm_0, x = q_27_cast_fp16)[name = tensor("transpose_153")]; + tensor var_3496_cast_fp16 = matmul(transpose_x = var_3496_transpose_x_0, transpose_y = var_3496_transpose_y_0, x = transpose_95, y = transpose_96)[name = tensor("op_3496_cast_fp16")]; + tensor var_3497_to_fp16 = const()[name = tensor("op_3497_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_53_cast_fp16 = mul(x = var_3496_cast_fp16, y = var_3497_to_fp16)[name = tensor("scores_53_cast_fp16")]; + tensor scores_55_cast_fp16 = add(x = scores_53_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_55_cast_fp16")]; + tensor var_3507 = const()[name = tensor("op_3507"), val = tensor(-1)]; + tensor weights_27_cast_fp16 = softmax(axis = var_3507, x = scores_55_cast_fp16)[name = tensor("weights_27_cast_fp16")]; + tensor attn_27_transpose_x_0 = const()[name = tensor("attn_27_transpose_x_0"), val = tensor(false)]; + tensor attn_27_transpose_y_0 = const()[name = tensor("attn_27_transpose_y_0"), val = tensor(false)]; + tensor values_27_cast_fp16 = transpose(perm = var_3485, x = new_v_cache_27_cast_fp16)[name = tensor("transpose_154")]; + tensor attn_27_cast_fp16 = matmul(transpose_x = attn_27_transpose_x_0, transpose_y = attn_27_transpose_y_0, x = weights_27_cast_fp16, y = values_27_cast_fp16)[name = tensor("attn_27_cast_fp16")]; + tensor var_3515 = const()[name = tensor("op_3515"), val = tensor([0, 2, 1, 3])]; + tensor var_3520 = const()[name = tensor("op_3520"), val = tensor([1, 256, 1024])]; + tensor var_3516_cast_fp16 = transpose(perm = var_3515, x = attn_27_cast_fp16)[name = tensor("transpose_151")]; + tensor input_161_cast_fp16 = reshape(shape = var_3520, x = var_3516_cast_fp16)[name = tensor("input_161_cast_fp16")]; + tensor attn13_out_proj_weight_to_fp16 = const()[name = tensor("attn13_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(333583616)))]; + tensor linear_53_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn13_out_proj_weight_to_fp16, x = input_161_cast_fp16)[name = tensor("linear_53_cast_fp16")]; + tensor position13_to_fp16_dtype_0 = const()[name = tensor("position13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position13_to_fp16 = cast(dtype = position13_to_fp16_dtype_0, x = position13)[name = tensor("cast_383")]; + tensor var_3529_cast_fp16 = add(x = position13_to_fp16, y = valid_len_to_fp16)[name = tensor("op_3529_cast_fp16")]; + tensor var_3529_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3529_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_163_cast_fp16 = add(x = input_157_cast_fp16, y = linear_53_cast_fp16)[name = tensor("input_163_cast_fp16")]; + tensor input_165_axes_0 = const()[name = tensor("input_165_axes_0"), val = tensor([-1])]; + tensor norm13_2_weight_to_fp16 = const()[name = tensor("norm13_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335680832)))]; + tensor norm13_2_bias_to_fp16 = const()[name = tensor("norm13_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335682944)))]; + tensor var_3533_to_fp16 = const()[name = tensor("op_3533_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_165_cast_fp16 = layer_norm(axes = input_165_axes_0, beta = norm13_2_bias_to_fp16, epsilon = var_3533_to_fp16, gamma = norm13_2_weight_to_fp16, x = input_163_cast_fp16)[name = tensor("input_165_cast_fp16")]; + tensor linear13_1_weight_to_fp16 = const()[name = tensor("linear13_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335685056)))]; + tensor linear_54_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear13_1_weight_to_fp16, x = input_165_cast_fp16)[name = tensor("linear_54_cast_fp16")]; + tensor input_167_mode_0 = const()[name = tensor("input_167_mode_0"), val = tensor("EXACT")]; + tensor input_167_cast_fp16 = gelu(mode = input_167_mode_0, x = linear_54_cast_fp16)[name = tensor("input_167_cast_fp16")]; + tensor linear13_2_weight_to_fp16 = const()[name = tensor("linear13_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(344073728)))]; + tensor linear_55_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear13_2_weight_to_fp16, x = input_167_cast_fp16)[name = tensor("linear_55_cast_fp16")]; + tensor input_169_cast_fp16 = add(x = input_163_cast_fp16, y = linear_55_cast_fp16)[name = tensor("input_169_cast_fp16")]; + tensor input_171_axes_0 = const()[name = tensor("input_171_axes_0"), val = tensor([-1])]; + tensor norm14_1_weight_to_fp16 = const()[name = tensor("norm14_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(352462400)))]; + tensor norm14_1_bias_to_fp16 = const()[name = tensor("norm14_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(352464512)))]; + tensor var_3550_to_fp16 = const()[name = tensor("op_3550_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_171_cast_fp16 = layer_norm(axes = input_171_axes_0, beta = norm14_1_bias_to_fp16, epsilon = var_3550_to_fp16, gamma = norm14_1_weight_to_fp16, x = input_169_cast_fp16)[name = tensor("input_171_cast_fp16")]; + tensor attn14_in_proj_weight_to_fp16 = const()[name = tensor("attn14_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(352466624)))]; + tensor linear_56_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn14_in_proj_weight_to_fp16, x = input_171_cast_fp16)[name = tensor("linear_56_cast_fp16")]; + tensor var_3564 = const()[name = tensor("op_3564"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_29_cast_fp16 = reshape(shape = var_3564, x = linear_56_cast_fp16)[name = tensor("qkv_29_cast_fp16")]; + tensor x_57_begin_0 = const()[name = tensor("x_57_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_57_end_0 = const()[name = tensor("x_57_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_57_end_mask_0 = const()[name = tensor("x_57_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_57_squeeze_mask_0 = const()[name = tensor("x_57_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_57_cast_fp16 = slice_by_index(begin = x_57_begin_0, end = x_57_end_0, end_mask = x_57_end_mask_0, squeeze_mask = x_57_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("x_57_cast_fp16")]; + tensor x_59_begin_0 = const()[name = tensor("x_59_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_59_end_0 = const()[name = tensor("x_59_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_59_end_mask_0 = const()[name = tensor("x_59_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_59_squeeze_mask_0 = const()[name = tensor("x_59_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_59_cast_fp16 = slice_by_index(begin = x_59_begin_0, end = x_59_end_0, end_mask = x_59_end_mask_0, squeeze_mask = x_59_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("x_59_cast_fp16")]; + tensor v_29_begin_0 = const()[name = tensor("v_29_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_29_end_0 = const()[name = tensor("v_29_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_29_end_mask_0 = const()[name = tensor("v_29_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_29_squeeze_mask_0 = const()[name = tensor("v_29_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_29_cast_fp16 = slice_by_index(begin = v_29_begin_0, end = v_29_end_0, end_mask = v_29_end_mask_0, squeeze_mask = v_29_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("v_29_cast_fp16")]; + tensor var_3609 = const()[name = tensor("op_3609"), val = tensor([256, 16, 32, 2])]; + tensor pairs_57_cast_fp16 = reshape(shape = var_3609, x = x_57_cast_fp16)[name = tensor("pairs_57_cast_fp16")]; + tensor xr_57_begin_0 = const()[name = tensor("xr_57_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_57_end_0 = const()[name = tensor("xr_57_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_57_end_mask_0 = const()[name = tensor("xr_57_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_57_squeeze_mask_0 = const()[name = tensor("xr_57_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_57_cast_fp16 = slice_by_index(begin = xr_57_begin_0, end = xr_57_end_0, end_mask = xr_57_end_mask_0, squeeze_mask = xr_57_squeeze_mask_0, x = pairs_57_cast_fp16)[name = tensor("xr_57_cast_fp16")]; + tensor xi_57_begin_0 = const()[name = tensor("xi_57_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_57_end_0 = const()[name = tensor("xi_57_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_57_end_mask_0 = const()[name = tensor("xi_57_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_57_squeeze_mask_0 = const()[name = tensor("xi_57_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_57_cast_fp16 = slice_by_index(begin = xi_57_begin_0, end = xi_57_end_0, end_mask = xi_57_end_mask_0, squeeze_mask = xi_57_squeeze_mask_0, x = pairs_57_cast_fp16)[name = tensor("xi_57_cast_fp16")]; + tensor var_3617_cast_fp16 = mul(x = xr_57_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3617_cast_fp16")]; + tensor var_3618_cast_fp16 = mul(x = xi_57_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3618_cast_fp16")]; + tensor xor__57_cast_fp16 = sub(x = var_3617_cast_fp16, y = var_3618_cast_fp16)[name = tensor("xor__57_cast_fp16")]; + tensor var_3621_cast_fp16 = mul(x = xr_57_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3621_cast_fp16")]; + tensor var_3622_cast_fp16 = mul(x = xi_57_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3622_cast_fp16")]; + tensor xoi_57_cast_fp16 = add(x = var_3621_cast_fp16, y = var_3622_cast_fp16)[name = tensor("xoi_57_cast_fp16")]; + tensor out_57_axis_0 = const()[name = tensor("out_57_axis_0"), val = tensor(-1)]; + tensor out_57_cast_fp16 = stack(axis = out_57_axis_0, values = (xor__57_cast_fp16, xoi_57_cast_fp16))[name = tensor("out_57_cast_fp16")]; + tensor var_3632 = const()[name = tensor("op_3632"), val = tensor([1, 256, 16, 64])]; + tensor q_29_cast_fp16 = reshape(shape = var_3632, x = out_57_cast_fp16)[name = tensor("q_29_cast_fp16")]; + tensor var_3638 = const()[name = tensor("op_3638"), val = tensor([256, 16, 32, 2])]; + tensor pairs_59_cast_fp16 = reshape(shape = var_3638, x = x_59_cast_fp16)[name = tensor("pairs_59_cast_fp16")]; + tensor xr_59_begin_0 = const()[name = tensor("xr_59_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_59_end_0 = const()[name = tensor("xr_59_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_59_end_mask_0 = const()[name = tensor("xr_59_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_59_squeeze_mask_0 = const()[name = tensor("xr_59_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_59_cast_fp16 = slice_by_index(begin = xr_59_begin_0, end = xr_59_end_0, end_mask = xr_59_end_mask_0, squeeze_mask = xr_59_squeeze_mask_0, x = pairs_59_cast_fp16)[name = tensor("xr_59_cast_fp16")]; + tensor xi_59_begin_0 = const()[name = tensor("xi_59_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_59_end_0 = const()[name = tensor("xi_59_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_59_end_mask_0 = const()[name = tensor("xi_59_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_59_squeeze_mask_0 = const()[name = tensor("xi_59_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_59_cast_fp16 = slice_by_index(begin = xi_59_begin_0, end = xi_59_end_0, end_mask = xi_59_end_mask_0, squeeze_mask = xi_59_squeeze_mask_0, x = pairs_59_cast_fp16)[name = tensor("xi_59_cast_fp16")]; + tensor var_3646_cast_fp16 = mul(x = xr_59_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3646_cast_fp16")]; + tensor var_3647_cast_fp16 = mul(x = xi_59_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3647_cast_fp16")]; + tensor xor__59_cast_fp16 = sub(x = var_3646_cast_fp16, y = var_3647_cast_fp16)[name = tensor("xor__59_cast_fp16")]; + tensor var_3650_cast_fp16 = mul(x = xr_59_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3650_cast_fp16")]; + tensor var_3651_cast_fp16 = mul(x = xi_59_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3651_cast_fp16")]; + tensor xoi_59_cast_fp16 = add(x = var_3650_cast_fp16, y = var_3651_cast_fp16)[name = tensor("xoi_59_cast_fp16")]; + tensor out_59_axis_0 = const()[name = tensor("out_59_axis_0"), val = tensor(-1)]; + tensor out_59_cast_fp16 = stack(axis = out_59_axis_0, values = (xor__59_cast_fp16, xoi_59_cast_fp16))[name = tensor("out_59_cast_fp16")]; + tensor var_3668 = const()[name = tensor("op_3668"), val = tensor([256, 1024])]; + tensor var_3669_cast_fp16 = reshape(shape = var_3668, x = out_59_cast_fp16)[name = tensor("op_3669_cast_fp16")]; + tensor written_k_29_transpose_x_1 = const()[name = tensor("written_k_29_transpose_x_1"), val = tensor(true)]; + tensor written_k_29_transpose_y_1 = const()[name = tensor("written_k_29_transpose_y_1"), val = tensor(false)]; + tensor written_k_29_cast_fp16 = matmul(transpose_x = written_k_29_transpose_x_1, transpose_y = written_k_29_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_3669_cast_fp16)[name = tensor("written_k_29_cast_fp16")]; + tensor var_3676 = const()[name = tensor("op_3676"), val = tensor([256, 1024])]; + tensor var_3677_cast_fp16 = reshape(shape = var_3676, x = v_29_cast_fp16)[name = tensor("op_3677_cast_fp16")]; + tensor written_v_29_transpose_x_1 = const()[name = tensor("written_v_29_transpose_x_1"), val = tensor(true)]; + tensor written_v_29_transpose_y_1 = const()[name = tensor("written_v_29_transpose_y_1"), val = tensor(false)]; + tensor written_v_29_cast_fp16 = matmul(transpose_x = written_v_29_transpose_x_1, transpose_y = written_v_29_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_3677_cast_fp16)[name = tensor("written_v_29_cast_fp16")]; + tensor k_cache14_to_fp16_dtype_0 = const()[name = tensor("k_cache14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache14_to_fp16 = cast(dtype = k_cache14_to_fp16_dtype_0, x = k_cache14)[name = tensor("cast_381")]; + tensor var_3688_cast_fp16 = mul(x = k_cache14_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_3688_cast_fp16")]; + tensor var_3693 = const()[name = tensor("op_3693"), val = tensor([1, 512, 16, 64])]; + tensor var_3694_cast_fp16 = reshape(shape = var_3693, x = written_k_29_cast_fp16)[name = tensor("op_3694_cast_fp16")]; + tensor new_k_cache_29_cast_fp16 = add(x = var_3688_cast_fp16, y = var_3694_cast_fp16)[name = tensor("new_k_cache_29_cast_fp16")]; + tensor new_k_cache_29_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_29_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache14_to_fp16_dtype_0 = const()[name = tensor("v_cache14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache14_to_fp16 = cast(dtype = v_cache14_to_fp16_dtype_0, x = v_cache14)[name = tensor("cast_379")]; + tensor var_3697_cast_fp16 = mul(x = v_cache14_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_3697_cast_fp16")]; + tensor var_3702 = const()[name = tensor("op_3702"), val = tensor([1, 512, 16, 64])]; + tensor var_3703_cast_fp16 = reshape(shape = var_3702, x = written_v_29_cast_fp16)[name = tensor("op_3703_cast_fp16")]; + tensor new_v_cache_29_cast_fp16 = add(x = var_3697_cast_fp16, y = var_3703_cast_fp16)[name = tensor("new_v_cache_29_cast_fp16")]; + tensor new_v_cache_29_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_29_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_3716 = const()[name = tensor("op_3716"), val = tensor([0, 2, 1, 3])]; + tensor var_3727_transpose_x_0 = const()[name = tensor("op_3727_transpose_x_0"), val = tensor(false)]; + tensor var_3727_transpose_y_0 = const()[name = tensor("op_3727_transpose_y_0"), val = tensor(false)]; + tensor transpose_97_perm_0 = const()[name = tensor("transpose_97_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_98_perm_0 = const()[name = tensor("transpose_98_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_98 = transpose(perm = transpose_98_perm_0, x = new_k_cache_29_cast_fp16)[name = tensor("transpose_148")]; + tensor transpose_97 = transpose(perm = transpose_97_perm_0, x = q_29_cast_fp16)[name = tensor("transpose_149")]; + tensor var_3727_cast_fp16 = matmul(transpose_x = var_3727_transpose_x_0, transpose_y = var_3727_transpose_y_0, x = transpose_97, y = transpose_98)[name = tensor("op_3727_cast_fp16")]; + tensor var_3728_to_fp16 = const()[name = tensor("op_3728_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_57_cast_fp16 = mul(x = var_3727_cast_fp16, y = var_3728_to_fp16)[name = tensor("scores_57_cast_fp16")]; + tensor scores_59_cast_fp16 = add(x = scores_57_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_59_cast_fp16")]; + tensor var_3738 = const()[name = tensor("op_3738"), val = tensor(-1)]; + tensor weights_29_cast_fp16 = softmax(axis = var_3738, x = scores_59_cast_fp16)[name = tensor("weights_29_cast_fp16")]; + tensor attn_29_transpose_x_0 = const()[name = tensor("attn_29_transpose_x_0"), val = tensor(false)]; + tensor attn_29_transpose_y_0 = const()[name = tensor("attn_29_transpose_y_0"), val = tensor(false)]; + tensor values_29_cast_fp16 = transpose(perm = var_3716, x = new_v_cache_29_cast_fp16)[name = tensor("transpose_150")]; + tensor attn_29_cast_fp16 = matmul(transpose_x = attn_29_transpose_x_0, transpose_y = attn_29_transpose_y_0, x = weights_29_cast_fp16, y = values_29_cast_fp16)[name = tensor("attn_29_cast_fp16")]; + tensor var_3746 = const()[name = tensor("op_3746"), val = tensor([0, 2, 1, 3])]; + tensor var_3751 = const()[name = tensor("op_3751"), val = tensor([1, 256, 1024])]; + tensor var_3747_cast_fp16 = transpose(perm = var_3746, x = attn_29_cast_fp16)[name = tensor("transpose_147")]; + tensor input_173_cast_fp16 = reshape(shape = var_3751, x = var_3747_cast_fp16)[name = tensor("input_173_cast_fp16")]; + tensor attn14_out_proj_weight_to_fp16 = const()[name = tensor("attn14_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(358758144)))]; + tensor linear_57_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn14_out_proj_weight_to_fp16, x = input_173_cast_fp16)[name = tensor("linear_57_cast_fp16")]; + tensor position14_to_fp16_dtype_0 = const()[name = tensor("position14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position14_to_fp16 = cast(dtype = position14_to_fp16_dtype_0, x = position14)[name = tensor("cast_377")]; + tensor var_3760_cast_fp16 = add(x = position14_to_fp16, y = valid_len_to_fp16)[name = tensor("op_3760_cast_fp16")]; + tensor var_3760_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3760_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_175_cast_fp16 = add(x = input_169_cast_fp16, y = linear_57_cast_fp16)[name = tensor("input_175_cast_fp16")]; + tensor input_177_axes_0 = const()[name = tensor("input_177_axes_0"), val = tensor([-1])]; + tensor norm14_2_weight_to_fp16 = const()[name = tensor("norm14_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(360855360)))]; + tensor norm14_2_bias_to_fp16 = const()[name = tensor("norm14_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(360857472)))]; + tensor var_3764_to_fp16 = const()[name = tensor("op_3764_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_177_cast_fp16 = layer_norm(axes = input_177_axes_0, beta = norm14_2_bias_to_fp16, epsilon = var_3764_to_fp16, gamma = norm14_2_weight_to_fp16, x = input_175_cast_fp16)[name = tensor("input_177_cast_fp16")]; + tensor linear14_1_weight_to_fp16 = const()[name = tensor("linear14_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(360859584)))]; + tensor linear_58_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear14_1_weight_to_fp16, x = input_177_cast_fp16)[name = tensor("linear_58_cast_fp16")]; + tensor input_179_mode_0 = const()[name = tensor("input_179_mode_0"), val = tensor("EXACT")]; + tensor input_179_cast_fp16 = gelu(mode = input_179_mode_0, x = linear_58_cast_fp16)[name = tensor("input_179_cast_fp16")]; + tensor linear14_2_weight_to_fp16 = const()[name = tensor("linear14_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(369248256)))]; + tensor linear_59_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear14_2_weight_to_fp16, x = input_179_cast_fp16)[name = tensor("linear_59_cast_fp16")]; + tensor input_181_cast_fp16 = add(x = input_175_cast_fp16, y = linear_59_cast_fp16)[name = tensor("input_181_cast_fp16")]; + tensor input_183_axes_0 = const()[name = tensor("input_183_axes_0"), val = tensor([-1])]; + tensor norm15_1_weight_to_fp16 = const()[name = tensor("norm15_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(377636928)))]; + tensor norm15_1_bias_to_fp16 = const()[name = tensor("norm15_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(377639040)))]; + tensor var_3781_to_fp16 = const()[name = tensor("op_3781_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_183_cast_fp16 = layer_norm(axes = input_183_axes_0, beta = norm15_1_bias_to_fp16, epsilon = var_3781_to_fp16, gamma = norm15_1_weight_to_fp16, x = input_181_cast_fp16)[name = tensor("input_183_cast_fp16")]; + tensor attn15_in_proj_weight_to_fp16 = const()[name = tensor("attn15_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(377641152)))]; + tensor linear_60_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn15_in_proj_weight_to_fp16, x = input_183_cast_fp16)[name = tensor("linear_60_cast_fp16")]; + tensor var_3795 = const()[name = tensor("op_3795"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_31_cast_fp16 = reshape(shape = var_3795, x = linear_60_cast_fp16)[name = tensor("qkv_31_cast_fp16")]; + tensor x_61_begin_0 = const()[name = tensor("x_61_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_61_end_0 = const()[name = tensor("x_61_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_61_end_mask_0 = const()[name = tensor("x_61_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_61_squeeze_mask_0 = const()[name = tensor("x_61_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_61_cast_fp16 = slice_by_index(begin = x_61_begin_0, end = x_61_end_0, end_mask = x_61_end_mask_0, squeeze_mask = x_61_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("x_61_cast_fp16")]; + tensor x_63_begin_0 = const()[name = tensor("x_63_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_63_end_0 = const()[name = tensor("x_63_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_63_end_mask_0 = const()[name = tensor("x_63_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_63_squeeze_mask_0 = const()[name = tensor("x_63_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_63_cast_fp16 = slice_by_index(begin = x_63_begin_0, end = x_63_end_0, end_mask = x_63_end_mask_0, squeeze_mask = x_63_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("x_63_cast_fp16")]; + tensor v_31_begin_0 = const()[name = tensor("v_31_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_31_end_0 = const()[name = tensor("v_31_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_31_end_mask_0 = const()[name = tensor("v_31_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_31_squeeze_mask_0 = const()[name = tensor("v_31_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_31_cast_fp16 = slice_by_index(begin = v_31_begin_0, end = v_31_end_0, end_mask = v_31_end_mask_0, squeeze_mask = v_31_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("v_31_cast_fp16")]; + tensor var_3840 = const()[name = tensor("op_3840"), val = tensor([256, 16, 32, 2])]; + tensor pairs_61_cast_fp16 = reshape(shape = var_3840, x = x_61_cast_fp16)[name = tensor("pairs_61_cast_fp16")]; + tensor xr_61_begin_0 = const()[name = tensor("xr_61_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_61_end_0 = const()[name = tensor("xr_61_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_61_end_mask_0 = const()[name = tensor("xr_61_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_61_squeeze_mask_0 = const()[name = tensor("xr_61_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_61_cast_fp16 = slice_by_index(begin = xr_61_begin_0, end = xr_61_end_0, end_mask = xr_61_end_mask_0, squeeze_mask = xr_61_squeeze_mask_0, x = pairs_61_cast_fp16)[name = tensor("xr_61_cast_fp16")]; + tensor xi_61_begin_0 = const()[name = tensor("xi_61_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_61_end_0 = const()[name = tensor("xi_61_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_61_end_mask_0 = const()[name = tensor("xi_61_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_61_squeeze_mask_0 = const()[name = tensor("xi_61_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_61_cast_fp16 = slice_by_index(begin = xi_61_begin_0, end = xi_61_end_0, end_mask = xi_61_end_mask_0, squeeze_mask = xi_61_squeeze_mask_0, x = pairs_61_cast_fp16)[name = tensor("xi_61_cast_fp16")]; + tensor var_3848_cast_fp16 = mul(x = xr_61_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3848_cast_fp16")]; + tensor var_3849_cast_fp16 = mul(x = xi_61_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3849_cast_fp16")]; + tensor xor__61_cast_fp16 = sub(x = var_3848_cast_fp16, y = var_3849_cast_fp16)[name = tensor("xor__61_cast_fp16")]; + tensor var_3852_cast_fp16 = mul(x = xr_61_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3852_cast_fp16")]; + tensor var_3853_cast_fp16 = mul(x = xi_61_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3853_cast_fp16")]; + tensor xoi_61_cast_fp16 = add(x = var_3852_cast_fp16, y = var_3853_cast_fp16)[name = tensor("xoi_61_cast_fp16")]; + tensor out_61_axis_0 = const()[name = tensor("out_61_axis_0"), val = tensor(-1)]; + tensor out_61_cast_fp16 = stack(axis = out_61_axis_0, values = (xor__61_cast_fp16, xoi_61_cast_fp16))[name = tensor("out_61_cast_fp16")]; + tensor var_3863 = const()[name = tensor("op_3863"), val = tensor([1, 256, 16, 64])]; + tensor q_31_cast_fp16 = reshape(shape = var_3863, x = out_61_cast_fp16)[name = tensor("q_31_cast_fp16")]; + tensor var_3869 = const()[name = tensor("op_3869"), val = tensor([256, 16, 32, 2])]; + tensor pairs_63_cast_fp16 = reshape(shape = var_3869, x = x_63_cast_fp16)[name = tensor("pairs_63_cast_fp16")]; + tensor xr_63_begin_0 = const()[name = tensor("xr_63_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_63_end_0 = const()[name = tensor("xr_63_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_63_end_mask_0 = const()[name = tensor("xr_63_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_63_squeeze_mask_0 = const()[name = tensor("xr_63_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_63_cast_fp16 = slice_by_index(begin = xr_63_begin_0, end = xr_63_end_0, end_mask = xr_63_end_mask_0, squeeze_mask = xr_63_squeeze_mask_0, x = pairs_63_cast_fp16)[name = tensor("xr_63_cast_fp16")]; + tensor xi_63_begin_0 = const()[name = tensor("xi_63_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_63_end_0 = const()[name = tensor("xi_63_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_63_end_mask_0 = const()[name = tensor("xi_63_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_63_squeeze_mask_0 = const()[name = tensor("xi_63_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_63_cast_fp16 = slice_by_index(begin = xi_63_begin_0, end = xi_63_end_0, end_mask = xi_63_end_mask_0, squeeze_mask = xi_63_squeeze_mask_0, x = pairs_63_cast_fp16)[name = tensor("xi_63_cast_fp16")]; + tensor var_3877_cast_fp16 = mul(x = xr_63_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3877_cast_fp16")]; + tensor var_3878_cast_fp16 = mul(x = xi_63_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3878_cast_fp16")]; + tensor xor__63_cast_fp16 = sub(x = var_3877_cast_fp16, y = var_3878_cast_fp16)[name = tensor("xor__63_cast_fp16")]; + tensor var_3881_cast_fp16 = mul(x = xr_63_cast_fp16, y = roti_cast_fp16)[name = tensor("op_3881_cast_fp16")]; + tensor var_3882_cast_fp16 = mul(x = xi_63_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_3882_cast_fp16")]; + tensor xoi_63_cast_fp16 = add(x = var_3881_cast_fp16, y = var_3882_cast_fp16)[name = tensor("xoi_63_cast_fp16")]; + tensor out_63_axis_0 = const()[name = tensor("out_63_axis_0"), val = tensor(-1)]; + tensor out_63_cast_fp16 = stack(axis = out_63_axis_0, values = (xor__63_cast_fp16, xoi_63_cast_fp16))[name = tensor("out_63_cast_fp16")]; + tensor var_3899 = const()[name = tensor("op_3899"), val = tensor([256, 1024])]; + tensor var_3900_cast_fp16 = reshape(shape = var_3899, x = out_63_cast_fp16)[name = tensor("op_3900_cast_fp16")]; + tensor written_k_31_transpose_x_1 = const()[name = tensor("written_k_31_transpose_x_1"), val = tensor(true)]; + tensor written_k_31_transpose_y_1 = const()[name = tensor("written_k_31_transpose_y_1"), val = tensor(false)]; + tensor written_k_31_cast_fp16 = matmul(transpose_x = written_k_31_transpose_x_1, transpose_y = written_k_31_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_3900_cast_fp16)[name = tensor("written_k_31_cast_fp16")]; + tensor var_3907 = const()[name = tensor("op_3907"), val = tensor([256, 1024])]; + tensor var_3908_cast_fp16 = reshape(shape = var_3907, x = v_31_cast_fp16)[name = tensor("op_3908_cast_fp16")]; + tensor written_v_31_transpose_x_1 = const()[name = tensor("written_v_31_transpose_x_1"), val = tensor(true)]; + tensor written_v_31_transpose_y_1 = const()[name = tensor("written_v_31_transpose_y_1"), val = tensor(false)]; + tensor written_v_31_cast_fp16 = matmul(transpose_x = written_v_31_transpose_x_1, transpose_y = written_v_31_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_3908_cast_fp16)[name = tensor("written_v_31_cast_fp16")]; + tensor k_cache15_to_fp16_dtype_0 = const()[name = tensor("k_cache15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache15_to_fp16 = cast(dtype = k_cache15_to_fp16_dtype_0, x = k_cache15)[name = tensor("cast_375")]; + tensor var_3919_cast_fp16 = mul(x = k_cache15_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_3919_cast_fp16")]; + tensor var_3924 = const()[name = tensor("op_3924"), val = tensor([1, 512, 16, 64])]; + tensor var_3925_cast_fp16 = reshape(shape = var_3924, x = written_k_31_cast_fp16)[name = tensor("op_3925_cast_fp16")]; + tensor new_k_cache_31_cast_fp16 = add(x = var_3919_cast_fp16, y = var_3925_cast_fp16)[name = tensor("new_k_cache_31_cast_fp16")]; + tensor new_k_cache_31_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_31_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache15_to_fp16_dtype_0 = const()[name = tensor("v_cache15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache15_to_fp16 = cast(dtype = v_cache15_to_fp16_dtype_0, x = v_cache15)[name = tensor("cast_373")]; + tensor var_3928_cast_fp16 = mul(x = v_cache15_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_3928_cast_fp16")]; + tensor var_3933 = const()[name = tensor("op_3933"), val = tensor([1, 512, 16, 64])]; + tensor var_3934_cast_fp16 = reshape(shape = var_3933, x = written_v_31_cast_fp16)[name = tensor("op_3934_cast_fp16")]; + tensor new_v_cache_31_cast_fp16 = add(x = var_3928_cast_fp16, y = var_3934_cast_fp16)[name = tensor("new_v_cache_31_cast_fp16")]; + tensor new_v_cache_31_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_31_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_3947 = const()[name = tensor("op_3947"), val = tensor([0, 2, 1, 3])]; + tensor var_3958_transpose_x_0 = const()[name = tensor("op_3958_transpose_x_0"), val = tensor(false)]; + tensor var_3958_transpose_y_0 = const()[name = tensor("op_3958_transpose_y_0"), val = tensor(false)]; + tensor transpose_99_perm_0 = const()[name = tensor("transpose_99_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_100_perm_0 = const()[name = tensor("transpose_100_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_100 = transpose(perm = transpose_100_perm_0, x = new_k_cache_31_cast_fp16)[name = tensor("transpose_144")]; + tensor transpose_99 = transpose(perm = transpose_99_perm_0, x = q_31_cast_fp16)[name = tensor("transpose_145")]; + tensor var_3958_cast_fp16 = matmul(transpose_x = var_3958_transpose_x_0, transpose_y = var_3958_transpose_y_0, x = transpose_99, y = transpose_100)[name = tensor("op_3958_cast_fp16")]; + tensor var_3959_to_fp16 = const()[name = tensor("op_3959_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_61_cast_fp16 = mul(x = var_3958_cast_fp16, y = var_3959_to_fp16)[name = tensor("scores_61_cast_fp16")]; + tensor scores_63_cast_fp16 = add(x = scores_61_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_63_cast_fp16")]; + tensor var_3969 = const()[name = tensor("op_3969"), val = tensor(-1)]; + tensor weights_31_cast_fp16 = softmax(axis = var_3969, x = scores_63_cast_fp16)[name = tensor("weights_31_cast_fp16")]; + tensor attn_31_transpose_x_0 = const()[name = tensor("attn_31_transpose_x_0"), val = tensor(false)]; + tensor attn_31_transpose_y_0 = const()[name = tensor("attn_31_transpose_y_0"), val = tensor(false)]; + tensor values_31_cast_fp16 = transpose(perm = var_3947, x = new_v_cache_31_cast_fp16)[name = tensor("transpose_146")]; + tensor attn_31_cast_fp16 = matmul(transpose_x = attn_31_transpose_x_0, transpose_y = attn_31_transpose_y_0, x = weights_31_cast_fp16, y = values_31_cast_fp16)[name = tensor("attn_31_cast_fp16")]; + tensor var_3977 = const()[name = tensor("op_3977"), val = tensor([0, 2, 1, 3])]; + tensor var_3982 = const()[name = tensor("op_3982"), val = tensor([1, 256, 1024])]; + tensor var_3978_cast_fp16 = transpose(perm = var_3977, x = attn_31_cast_fp16)[name = tensor("transpose_143")]; + tensor input_185_cast_fp16 = reshape(shape = var_3982, x = var_3978_cast_fp16)[name = tensor("input_185_cast_fp16")]; + tensor attn15_out_proj_weight_to_fp16 = const()[name = tensor("attn15_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(383932672)))]; + tensor linear_61_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn15_out_proj_weight_to_fp16, x = input_185_cast_fp16)[name = tensor("linear_61_cast_fp16")]; + tensor position15_to_fp16_dtype_0 = const()[name = tensor("position15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position15_to_fp16 = cast(dtype = position15_to_fp16_dtype_0, x = position15)[name = tensor("cast_371")]; + tensor var_3991_cast_fp16 = add(x = position15_to_fp16, y = valid_len_to_fp16)[name = tensor("op_3991_cast_fp16")]; + tensor var_3991_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3991_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_187_cast_fp16 = add(x = input_181_cast_fp16, y = linear_61_cast_fp16)[name = tensor("input_187_cast_fp16")]; + tensor input_189_axes_0 = const()[name = tensor("input_189_axes_0"), val = tensor([-1])]; + tensor norm15_2_weight_to_fp16 = const()[name = tensor("norm15_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(386029888)))]; + tensor norm15_2_bias_to_fp16 = const()[name = tensor("norm15_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(386032000)))]; + tensor var_3995_to_fp16 = const()[name = tensor("op_3995_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_189_cast_fp16 = layer_norm(axes = input_189_axes_0, beta = norm15_2_bias_to_fp16, epsilon = var_3995_to_fp16, gamma = norm15_2_weight_to_fp16, x = input_187_cast_fp16)[name = tensor("input_189_cast_fp16")]; + tensor linear15_1_weight_to_fp16 = const()[name = tensor("linear15_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(386034112)))]; + tensor linear_62_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear15_1_weight_to_fp16, x = input_189_cast_fp16)[name = tensor("linear_62_cast_fp16")]; + tensor input_191_mode_0 = const()[name = tensor("input_191_mode_0"), val = tensor("EXACT")]; + tensor input_191_cast_fp16 = gelu(mode = input_191_mode_0, x = linear_62_cast_fp16)[name = tensor("input_191_cast_fp16")]; + tensor linear15_2_weight_to_fp16 = const()[name = tensor("linear15_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(394422784)))]; + tensor linear_63_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear15_2_weight_to_fp16, x = input_191_cast_fp16)[name = tensor("linear_63_cast_fp16")]; + tensor input_193_cast_fp16 = add(x = input_187_cast_fp16, y = linear_63_cast_fp16)[name = tensor("input_193_cast_fp16")]; + tensor input_195_axes_0 = const()[name = tensor("input_195_axes_0"), val = tensor([-1])]; + tensor norm16_1_weight_to_fp16 = const()[name = tensor("norm16_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(402811456)))]; + tensor norm16_1_bias_to_fp16 = const()[name = tensor("norm16_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(402813568)))]; + tensor var_4012_to_fp16 = const()[name = tensor("op_4012_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_195_cast_fp16 = layer_norm(axes = input_195_axes_0, beta = norm16_1_bias_to_fp16, epsilon = var_4012_to_fp16, gamma = norm16_1_weight_to_fp16, x = input_193_cast_fp16)[name = tensor("input_195_cast_fp16")]; + tensor attn16_in_proj_weight_to_fp16 = const()[name = tensor("attn16_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(402815680)))]; + tensor linear_64_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn16_in_proj_weight_to_fp16, x = input_195_cast_fp16)[name = tensor("linear_64_cast_fp16")]; + tensor var_4026 = const()[name = tensor("op_4026"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_33_cast_fp16 = reshape(shape = var_4026, x = linear_64_cast_fp16)[name = tensor("qkv_33_cast_fp16")]; + tensor x_65_begin_0 = const()[name = tensor("x_65_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_65_end_0 = const()[name = tensor("x_65_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_65_end_mask_0 = const()[name = tensor("x_65_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_65_squeeze_mask_0 = const()[name = tensor("x_65_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_65_cast_fp16 = slice_by_index(begin = x_65_begin_0, end = x_65_end_0, end_mask = x_65_end_mask_0, squeeze_mask = x_65_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("x_65_cast_fp16")]; + tensor x_67_begin_0 = const()[name = tensor("x_67_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_67_end_0 = const()[name = tensor("x_67_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_67_end_mask_0 = const()[name = tensor("x_67_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_67_squeeze_mask_0 = const()[name = tensor("x_67_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_67_cast_fp16 = slice_by_index(begin = x_67_begin_0, end = x_67_end_0, end_mask = x_67_end_mask_0, squeeze_mask = x_67_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("x_67_cast_fp16")]; + tensor v_33_begin_0 = const()[name = tensor("v_33_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_33_end_0 = const()[name = tensor("v_33_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_33_end_mask_0 = const()[name = tensor("v_33_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_33_squeeze_mask_0 = const()[name = tensor("v_33_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_33_cast_fp16 = slice_by_index(begin = v_33_begin_0, end = v_33_end_0, end_mask = v_33_end_mask_0, squeeze_mask = v_33_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("v_33_cast_fp16")]; + tensor var_4071 = const()[name = tensor("op_4071"), val = tensor([256, 16, 32, 2])]; + tensor pairs_65_cast_fp16 = reshape(shape = var_4071, x = x_65_cast_fp16)[name = tensor("pairs_65_cast_fp16")]; + tensor xr_65_begin_0 = const()[name = tensor("xr_65_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_65_end_0 = const()[name = tensor("xr_65_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_65_end_mask_0 = const()[name = tensor("xr_65_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_65_squeeze_mask_0 = const()[name = tensor("xr_65_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_65_cast_fp16 = slice_by_index(begin = xr_65_begin_0, end = xr_65_end_0, end_mask = xr_65_end_mask_0, squeeze_mask = xr_65_squeeze_mask_0, x = pairs_65_cast_fp16)[name = tensor("xr_65_cast_fp16")]; + tensor xi_65_begin_0 = const()[name = tensor("xi_65_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_65_end_0 = const()[name = tensor("xi_65_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_65_end_mask_0 = const()[name = tensor("xi_65_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_65_squeeze_mask_0 = const()[name = tensor("xi_65_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_65_cast_fp16 = slice_by_index(begin = xi_65_begin_0, end = xi_65_end_0, end_mask = xi_65_end_mask_0, squeeze_mask = xi_65_squeeze_mask_0, x = pairs_65_cast_fp16)[name = tensor("xi_65_cast_fp16")]; + tensor var_4079_cast_fp16 = mul(x = xr_65_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4079_cast_fp16")]; + tensor var_4080_cast_fp16 = mul(x = xi_65_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4080_cast_fp16")]; + tensor xor__65_cast_fp16 = sub(x = var_4079_cast_fp16, y = var_4080_cast_fp16)[name = tensor("xor__65_cast_fp16")]; + tensor var_4083_cast_fp16 = mul(x = xr_65_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4083_cast_fp16")]; + tensor var_4084_cast_fp16 = mul(x = xi_65_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4084_cast_fp16")]; + tensor xoi_65_cast_fp16 = add(x = var_4083_cast_fp16, y = var_4084_cast_fp16)[name = tensor("xoi_65_cast_fp16")]; + tensor out_65_axis_0 = const()[name = tensor("out_65_axis_0"), val = tensor(-1)]; + tensor out_65_cast_fp16 = stack(axis = out_65_axis_0, values = (xor__65_cast_fp16, xoi_65_cast_fp16))[name = tensor("out_65_cast_fp16")]; + tensor var_4094 = const()[name = tensor("op_4094"), val = tensor([1, 256, 16, 64])]; + tensor q_33_cast_fp16 = reshape(shape = var_4094, x = out_65_cast_fp16)[name = tensor("q_33_cast_fp16")]; + tensor var_4100 = const()[name = tensor("op_4100"), val = tensor([256, 16, 32, 2])]; + tensor pairs_67_cast_fp16 = reshape(shape = var_4100, x = x_67_cast_fp16)[name = tensor("pairs_67_cast_fp16")]; + tensor xr_67_begin_0 = const()[name = tensor("xr_67_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_67_end_0 = const()[name = tensor("xr_67_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_67_end_mask_0 = const()[name = tensor("xr_67_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_67_squeeze_mask_0 = const()[name = tensor("xr_67_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_67_cast_fp16 = slice_by_index(begin = xr_67_begin_0, end = xr_67_end_0, end_mask = xr_67_end_mask_0, squeeze_mask = xr_67_squeeze_mask_0, x = pairs_67_cast_fp16)[name = tensor("xr_67_cast_fp16")]; + tensor xi_67_begin_0 = const()[name = tensor("xi_67_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_67_end_0 = const()[name = tensor("xi_67_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_67_end_mask_0 = const()[name = tensor("xi_67_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_67_squeeze_mask_0 = const()[name = tensor("xi_67_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_67_cast_fp16 = slice_by_index(begin = xi_67_begin_0, end = xi_67_end_0, end_mask = xi_67_end_mask_0, squeeze_mask = xi_67_squeeze_mask_0, x = pairs_67_cast_fp16)[name = tensor("xi_67_cast_fp16")]; + tensor var_4108_cast_fp16 = mul(x = xr_67_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4108_cast_fp16")]; + tensor var_4109_cast_fp16 = mul(x = xi_67_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4109_cast_fp16")]; + tensor xor__67_cast_fp16 = sub(x = var_4108_cast_fp16, y = var_4109_cast_fp16)[name = tensor("xor__67_cast_fp16")]; + tensor var_4112_cast_fp16 = mul(x = xr_67_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4112_cast_fp16")]; + tensor var_4113_cast_fp16 = mul(x = xi_67_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4113_cast_fp16")]; + tensor xoi_67_cast_fp16 = add(x = var_4112_cast_fp16, y = var_4113_cast_fp16)[name = tensor("xoi_67_cast_fp16")]; + tensor out_67_axis_0 = const()[name = tensor("out_67_axis_0"), val = tensor(-1)]; + tensor out_67_cast_fp16 = stack(axis = out_67_axis_0, values = (xor__67_cast_fp16, xoi_67_cast_fp16))[name = tensor("out_67_cast_fp16")]; + tensor var_4130 = const()[name = tensor("op_4130"), val = tensor([256, 1024])]; + tensor var_4131_cast_fp16 = reshape(shape = var_4130, x = out_67_cast_fp16)[name = tensor("op_4131_cast_fp16")]; + tensor written_k_33_transpose_x_1 = const()[name = tensor("written_k_33_transpose_x_1"), val = tensor(true)]; + tensor written_k_33_transpose_y_1 = const()[name = tensor("written_k_33_transpose_y_1"), val = tensor(false)]; + tensor written_k_33_cast_fp16 = matmul(transpose_x = written_k_33_transpose_x_1, transpose_y = written_k_33_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_4131_cast_fp16)[name = tensor("written_k_33_cast_fp16")]; + tensor var_4138 = const()[name = tensor("op_4138"), val = tensor([256, 1024])]; + tensor var_4139_cast_fp16 = reshape(shape = var_4138, x = v_33_cast_fp16)[name = tensor("op_4139_cast_fp16")]; + tensor written_v_33_transpose_x_1 = const()[name = tensor("written_v_33_transpose_x_1"), val = tensor(true)]; + tensor written_v_33_transpose_y_1 = const()[name = tensor("written_v_33_transpose_y_1"), val = tensor(false)]; + tensor written_v_33_cast_fp16 = matmul(transpose_x = written_v_33_transpose_x_1, transpose_y = written_v_33_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_4139_cast_fp16)[name = tensor("written_v_33_cast_fp16")]; + tensor k_cache16_to_fp16_dtype_0 = const()[name = tensor("k_cache16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache16_to_fp16 = cast(dtype = k_cache16_to_fp16_dtype_0, x = k_cache16)[name = tensor("cast_369")]; + tensor var_4150_cast_fp16 = mul(x = k_cache16_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_4150_cast_fp16")]; + tensor var_4155 = const()[name = tensor("op_4155"), val = tensor([1, 512, 16, 64])]; + tensor var_4156_cast_fp16 = reshape(shape = var_4155, x = written_k_33_cast_fp16)[name = tensor("op_4156_cast_fp16")]; + tensor new_k_cache_33_cast_fp16 = add(x = var_4150_cast_fp16, y = var_4156_cast_fp16)[name = tensor("new_k_cache_33_cast_fp16")]; + tensor new_k_cache_33_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_33_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache16_to_fp16_dtype_0 = const()[name = tensor("v_cache16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache16_to_fp16 = cast(dtype = v_cache16_to_fp16_dtype_0, x = v_cache16)[name = tensor("cast_367")]; + tensor var_4159_cast_fp16 = mul(x = v_cache16_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_4159_cast_fp16")]; + tensor var_4164 = const()[name = tensor("op_4164"), val = tensor([1, 512, 16, 64])]; + tensor var_4165_cast_fp16 = reshape(shape = var_4164, x = written_v_33_cast_fp16)[name = tensor("op_4165_cast_fp16")]; + tensor new_v_cache_33_cast_fp16 = add(x = var_4159_cast_fp16, y = var_4165_cast_fp16)[name = tensor("new_v_cache_33_cast_fp16")]; + tensor new_v_cache_33_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_33_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_4178 = const()[name = tensor("op_4178"), val = tensor([0, 2, 1, 3])]; + tensor var_4189_transpose_x_0 = const()[name = tensor("op_4189_transpose_x_0"), val = tensor(false)]; + tensor var_4189_transpose_y_0 = const()[name = tensor("op_4189_transpose_y_0"), val = tensor(false)]; + tensor transpose_101_perm_0 = const()[name = tensor("transpose_101_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_102_perm_0 = const()[name = tensor("transpose_102_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_102 = transpose(perm = transpose_102_perm_0, x = new_k_cache_33_cast_fp16)[name = tensor("transpose_140")]; + tensor transpose_101 = transpose(perm = transpose_101_perm_0, x = q_33_cast_fp16)[name = tensor("transpose_141")]; + tensor var_4189_cast_fp16 = matmul(transpose_x = var_4189_transpose_x_0, transpose_y = var_4189_transpose_y_0, x = transpose_101, y = transpose_102)[name = tensor("op_4189_cast_fp16")]; + tensor var_4190_to_fp16 = const()[name = tensor("op_4190_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_65_cast_fp16 = mul(x = var_4189_cast_fp16, y = var_4190_to_fp16)[name = tensor("scores_65_cast_fp16")]; + tensor scores_67_cast_fp16 = add(x = scores_65_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_67_cast_fp16")]; + tensor var_4200 = const()[name = tensor("op_4200"), val = tensor(-1)]; + tensor weights_33_cast_fp16 = softmax(axis = var_4200, x = scores_67_cast_fp16)[name = tensor("weights_33_cast_fp16")]; + tensor attn_33_transpose_x_0 = const()[name = tensor("attn_33_transpose_x_0"), val = tensor(false)]; + tensor attn_33_transpose_y_0 = const()[name = tensor("attn_33_transpose_y_0"), val = tensor(false)]; + tensor values_33_cast_fp16 = transpose(perm = var_4178, x = new_v_cache_33_cast_fp16)[name = tensor("transpose_142")]; + tensor attn_33_cast_fp16 = matmul(transpose_x = attn_33_transpose_x_0, transpose_y = attn_33_transpose_y_0, x = weights_33_cast_fp16, y = values_33_cast_fp16)[name = tensor("attn_33_cast_fp16")]; + tensor var_4208 = const()[name = tensor("op_4208"), val = tensor([0, 2, 1, 3])]; + tensor var_4213 = const()[name = tensor("op_4213"), val = tensor([1, 256, 1024])]; + tensor var_4209_cast_fp16 = transpose(perm = var_4208, x = attn_33_cast_fp16)[name = tensor("transpose_139")]; + tensor input_197_cast_fp16 = reshape(shape = var_4213, x = var_4209_cast_fp16)[name = tensor("input_197_cast_fp16")]; + tensor attn16_out_proj_weight_to_fp16 = const()[name = tensor("attn16_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(409107200)))]; + tensor linear_65_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn16_out_proj_weight_to_fp16, x = input_197_cast_fp16)[name = tensor("linear_65_cast_fp16")]; + tensor position16_to_fp16_dtype_0 = const()[name = tensor("position16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position16_to_fp16 = cast(dtype = position16_to_fp16_dtype_0, x = position16)[name = tensor("cast_365")]; + tensor var_4222_cast_fp16 = add(x = position16_to_fp16, y = valid_len_to_fp16)[name = tensor("op_4222_cast_fp16")]; + tensor var_4222_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4222_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_199_cast_fp16 = add(x = input_193_cast_fp16, y = linear_65_cast_fp16)[name = tensor("input_199_cast_fp16")]; + tensor input_201_axes_0 = const()[name = tensor("input_201_axes_0"), val = tensor([-1])]; + tensor norm16_2_weight_to_fp16 = const()[name = tensor("norm16_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411204416)))]; + tensor norm16_2_bias_to_fp16 = const()[name = tensor("norm16_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411206528)))]; + tensor var_4226_to_fp16 = const()[name = tensor("op_4226_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_201_cast_fp16 = layer_norm(axes = input_201_axes_0, beta = norm16_2_bias_to_fp16, epsilon = var_4226_to_fp16, gamma = norm16_2_weight_to_fp16, x = input_199_cast_fp16)[name = tensor("input_201_cast_fp16")]; + tensor linear16_1_weight_to_fp16 = const()[name = tensor("linear16_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411208640)))]; + tensor linear_66_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear16_1_weight_to_fp16, x = input_201_cast_fp16)[name = tensor("linear_66_cast_fp16")]; + tensor input_203_mode_0 = const()[name = tensor("input_203_mode_0"), val = tensor("EXACT")]; + tensor input_203_cast_fp16 = gelu(mode = input_203_mode_0, x = linear_66_cast_fp16)[name = tensor("input_203_cast_fp16")]; + tensor linear16_2_weight_to_fp16 = const()[name = tensor("linear16_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(419597312)))]; + tensor linear_67_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear16_2_weight_to_fp16, x = input_203_cast_fp16)[name = tensor("linear_67_cast_fp16")]; + tensor input_205_cast_fp16 = add(x = input_199_cast_fp16, y = linear_67_cast_fp16)[name = tensor("input_205_cast_fp16")]; + tensor input_207_axes_0 = const()[name = tensor("input_207_axes_0"), val = tensor([-1])]; + tensor norm17_1_weight_to_fp16 = const()[name = tensor("norm17_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(427985984)))]; + tensor norm17_1_bias_to_fp16 = const()[name = tensor("norm17_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(427988096)))]; + tensor var_4243_to_fp16 = const()[name = tensor("op_4243_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_207_cast_fp16 = layer_norm(axes = input_207_axes_0, beta = norm17_1_bias_to_fp16, epsilon = var_4243_to_fp16, gamma = norm17_1_weight_to_fp16, x = input_205_cast_fp16)[name = tensor("input_207_cast_fp16")]; + tensor attn17_in_proj_weight_to_fp16 = const()[name = tensor("attn17_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(427990208)))]; + tensor linear_68_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn17_in_proj_weight_to_fp16, x = input_207_cast_fp16)[name = tensor("linear_68_cast_fp16")]; + tensor var_4257 = const()[name = tensor("op_4257"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_35_cast_fp16 = reshape(shape = var_4257, x = linear_68_cast_fp16)[name = tensor("qkv_35_cast_fp16")]; + tensor x_69_begin_0 = const()[name = tensor("x_69_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_69_end_0 = const()[name = tensor("x_69_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_69_end_mask_0 = const()[name = tensor("x_69_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_69_squeeze_mask_0 = const()[name = tensor("x_69_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_69_cast_fp16 = slice_by_index(begin = x_69_begin_0, end = x_69_end_0, end_mask = x_69_end_mask_0, squeeze_mask = x_69_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("x_69_cast_fp16")]; + tensor x_71_begin_0 = const()[name = tensor("x_71_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_71_end_0 = const()[name = tensor("x_71_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_71_end_mask_0 = const()[name = tensor("x_71_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_71_squeeze_mask_0 = const()[name = tensor("x_71_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_71_cast_fp16 = slice_by_index(begin = x_71_begin_0, end = x_71_end_0, end_mask = x_71_end_mask_0, squeeze_mask = x_71_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("x_71_cast_fp16")]; + tensor v_35_begin_0 = const()[name = tensor("v_35_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_35_end_0 = const()[name = tensor("v_35_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_35_end_mask_0 = const()[name = tensor("v_35_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_35_squeeze_mask_0 = const()[name = tensor("v_35_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_35_cast_fp16 = slice_by_index(begin = v_35_begin_0, end = v_35_end_0, end_mask = v_35_end_mask_0, squeeze_mask = v_35_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("v_35_cast_fp16")]; + tensor var_4302 = const()[name = tensor("op_4302"), val = tensor([256, 16, 32, 2])]; + tensor pairs_69_cast_fp16 = reshape(shape = var_4302, x = x_69_cast_fp16)[name = tensor("pairs_69_cast_fp16")]; + tensor xr_69_begin_0 = const()[name = tensor("xr_69_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_69_end_0 = const()[name = tensor("xr_69_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_69_end_mask_0 = const()[name = tensor("xr_69_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_69_squeeze_mask_0 = const()[name = tensor("xr_69_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_69_cast_fp16 = slice_by_index(begin = xr_69_begin_0, end = xr_69_end_0, end_mask = xr_69_end_mask_0, squeeze_mask = xr_69_squeeze_mask_0, x = pairs_69_cast_fp16)[name = tensor("xr_69_cast_fp16")]; + tensor xi_69_begin_0 = const()[name = tensor("xi_69_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_69_end_0 = const()[name = tensor("xi_69_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_69_end_mask_0 = const()[name = tensor("xi_69_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_69_squeeze_mask_0 = const()[name = tensor("xi_69_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_69_cast_fp16 = slice_by_index(begin = xi_69_begin_0, end = xi_69_end_0, end_mask = xi_69_end_mask_0, squeeze_mask = xi_69_squeeze_mask_0, x = pairs_69_cast_fp16)[name = tensor("xi_69_cast_fp16")]; + tensor var_4310_cast_fp16 = mul(x = xr_69_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4310_cast_fp16")]; + tensor var_4311_cast_fp16 = mul(x = xi_69_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4311_cast_fp16")]; + tensor xor__69_cast_fp16 = sub(x = var_4310_cast_fp16, y = var_4311_cast_fp16)[name = tensor("xor__69_cast_fp16")]; + tensor var_4314_cast_fp16 = mul(x = xr_69_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4314_cast_fp16")]; + tensor var_4315_cast_fp16 = mul(x = xi_69_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4315_cast_fp16")]; + tensor xoi_69_cast_fp16 = add(x = var_4314_cast_fp16, y = var_4315_cast_fp16)[name = tensor("xoi_69_cast_fp16")]; + tensor out_69_axis_0 = const()[name = tensor("out_69_axis_0"), val = tensor(-1)]; + tensor out_69_cast_fp16 = stack(axis = out_69_axis_0, values = (xor__69_cast_fp16, xoi_69_cast_fp16))[name = tensor("out_69_cast_fp16")]; + tensor var_4325 = const()[name = tensor("op_4325"), val = tensor([1, 256, 16, 64])]; + tensor q_35_cast_fp16 = reshape(shape = var_4325, x = out_69_cast_fp16)[name = tensor("q_35_cast_fp16")]; + tensor var_4331 = const()[name = tensor("op_4331"), val = tensor([256, 16, 32, 2])]; + tensor pairs_71_cast_fp16 = reshape(shape = var_4331, x = x_71_cast_fp16)[name = tensor("pairs_71_cast_fp16")]; + tensor xr_71_begin_0 = const()[name = tensor("xr_71_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_71_end_0 = const()[name = tensor("xr_71_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_71_end_mask_0 = const()[name = tensor("xr_71_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_71_squeeze_mask_0 = const()[name = tensor("xr_71_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_71_cast_fp16 = slice_by_index(begin = xr_71_begin_0, end = xr_71_end_0, end_mask = xr_71_end_mask_0, squeeze_mask = xr_71_squeeze_mask_0, x = pairs_71_cast_fp16)[name = tensor("xr_71_cast_fp16")]; + tensor xi_71_begin_0 = const()[name = tensor("xi_71_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_71_end_0 = const()[name = tensor("xi_71_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_71_end_mask_0 = const()[name = tensor("xi_71_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_71_squeeze_mask_0 = const()[name = tensor("xi_71_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_71_cast_fp16 = slice_by_index(begin = xi_71_begin_0, end = xi_71_end_0, end_mask = xi_71_end_mask_0, squeeze_mask = xi_71_squeeze_mask_0, x = pairs_71_cast_fp16)[name = tensor("xi_71_cast_fp16")]; + tensor var_4339_cast_fp16 = mul(x = xr_71_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4339_cast_fp16")]; + tensor var_4340_cast_fp16 = mul(x = xi_71_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4340_cast_fp16")]; + tensor xor__71_cast_fp16 = sub(x = var_4339_cast_fp16, y = var_4340_cast_fp16)[name = tensor("xor__71_cast_fp16")]; + tensor var_4343_cast_fp16 = mul(x = xr_71_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4343_cast_fp16")]; + tensor var_4344_cast_fp16 = mul(x = xi_71_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4344_cast_fp16")]; + tensor xoi_71_cast_fp16 = add(x = var_4343_cast_fp16, y = var_4344_cast_fp16)[name = tensor("xoi_71_cast_fp16")]; + tensor out_71_axis_0 = const()[name = tensor("out_71_axis_0"), val = tensor(-1)]; + tensor out_71_cast_fp16 = stack(axis = out_71_axis_0, values = (xor__71_cast_fp16, xoi_71_cast_fp16))[name = tensor("out_71_cast_fp16")]; + tensor var_4361 = const()[name = tensor("op_4361"), val = tensor([256, 1024])]; + tensor var_4362_cast_fp16 = reshape(shape = var_4361, x = out_71_cast_fp16)[name = tensor("op_4362_cast_fp16")]; + tensor written_k_35_transpose_x_1 = const()[name = tensor("written_k_35_transpose_x_1"), val = tensor(true)]; + tensor written_k_35_transpose_y_1 = const()[name = tensor("written_k_35_transpose_y_1"), val = tensor(false)]; + tensor written_k_35_cast_fp16 = matmul(transpose_x = written_k_35_transpose_x_1, transpose_y = written_k_35_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_4362_cast_fp16)[name = tensor("written_k_35_cast_fp16")]; + tensor var_4369 = const()[name = tensor("op_4369"), val = tensor([256, 1024])]; + tensor var_4370_cast_fp16 = reshape(shape = var_4369, x = v_35_cast_fp16)[name = tensor("op_4370_cast_fp16")]; + tensor written_v_35_transpose_x_1 = const()[name = tensor("written_v_35_transpose_x_1"), val = tensor(true)]; + tensor written_v_35_transpose_y_1 = const()[name = tensor("written_v_35_transpose_y_1"), val = tensor(false)]; + tensor written_v_35_cast_fp16 = matmul(transpose_x = written_v_35_transpose_x_1, transpose_y = written_v_35_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_4370_cast_fp16)[name = tensor("written_v_35_cast_fp16")]; + tensor k_cache17_to_fp16_dtype_0 = const()[name = tensor("k_cache17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache17_to_fp16 = cast(dtype = k_cache17_to_fp16_dtype_0, x = k_cache17)[name = tensor("cast_363")]; + tensor var_4381_cast_fp16 = mul(x = k_cache17_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_4381_cast_fp16")]; + tensor var_4386 = const()[name = tensor("op_4386"), val = tensor([1, 512, 16, 64])]; + tensor var_4387_cast_fp16 = reshape(shape = var_4386, x = written_k_35_cast_fp16)[name = tensor("op_4387_cast_fp16")]; + tensor new_k_cache_35_cast_fp16 = add(x = var_4381_cast_fp16, y = var_4387_cast_fp16)[name = tensor("new_k_cache_35_cast_fp16")]; + tensor new_k_cache_35_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_35_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache17_to_fp16_dtype_0 = const()[name = tensor("v_cache17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache17_to_fp16 = cast(dtype = v_cache17_to_fp16_dtype_0, x = v_cache17)[name = tensor("cast_361")]; + tensor var_4390_cast_fp16 = mul(x = v_cache17_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_4390_cast_fp16")]; + tensor var_4395 = const()[name = tensor("op_4395"), val = tensor([1, 512, 16, 64])]; + tensor var_4396_cast_fp16 = reshape(shape = var_4395, x = written_v_35_cast_fp16)[name = tensor("op_4396_cast_fp16")]; + tensor new_v_cache_35_cast_fp16 = add(x = var_4390_cast_fp16, y = var_4396_cast_fp16)[name = tensor("new_v_cache_35_cast_fp16")]; + tensor new_v_cache_35_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_35_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_4409 = const()[name = tensor("op_4409"), val = tensor([0, 2, 1, 3])]; + tensor var_4420_transpose_x_0 = const()[name = tensor("op_4420_transpose_x_0"), val = tensor(false)]; + tensor var_4420_transpose_y_0 = const()[name = tensor("op_4420_transpose_y_0"), val = tensor(false)]; + tensor transpose_103_perm_0 = const()[name = tensor("transpose_103_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_104_perm_0 = const()[name = tensor("transpose_104_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_104 = transpose(perm = transpose_104_perm_0, x = new_k_cache_35_cast_fp16)[name = tensor("transpose_136")]; + tensor transpose_103 = transpose(perm = transpose_103_perm_0, x = q_35_cast_fp16)[name = tensor("transpose_137")]; + tensor var_4420_cast_fp16 = matmul(transpose_x = var_4420_transpose_x_0, transpose_y = var_4420_transpose_y_0, x = transpose_103, y = transpose_104)[name = tensor("op_4420_cast_fp16")]; + tensor var_4421_to_fp16 = const()[name = tensor("op_4421_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_69_cast_fp16 = mul(x = var_4420_cast_fp16, y = var_4421_to_fp16)[name = tensor("scores_69_cast_fp16")]; + tensor scores_71_cast_fp16 = add(x = scores_69_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_71_cast_fp16")]; + tensor var_4431 = const()[name = tensor("op_4431"), val = tensor(-1)]; + tensor weights_35_cast_fp16 = softmax(axis = var_4431, x = scores_71_cast_fp16)[name = tensor("weights_35_cast_fp16")]; + tensor attn_35_transpose_x_0 = const()[name = tensor("attn_35_transpose_x_0"), val = tensor(false)]; + tensor attn_35_transpose_y_0 = const()[name = tensor("attn_35_transpose_y_0"), val = tensor(false)]; + tensor values_35_cast_fp16 = transpose(perm = var_4409, x = new_v_cache_35_cast_fp16)[name = tensor("transpose_138")]; + tensor attn_35_cast_fp16 = matmul(transpose_x = attn_35_transpose_x_0, transpose_y = attn_35_transpose_y_0, x = weights_35_cast_fp16, y = values_35_cast_fp16)[name = tensor("attn_35_cast_fp16")]; + tensor var_4439 = const()[name = tensor("op_4439"), val = tensor([0, 2, 1, 3])]; + tensor var_4444 = const()[name = tensor("op_4444"), val = tensor([1, 256, 1024])]; + tensor var_4440_cast_fp16 = transpose(perm = var_4439, x = attn_35_cast_fp16)[name = tensor("transpose_135")]; + tensor input_209_cast_fp16 = reshape(shape = var_4444, x = var_4440_cast_fp16)[name = tensor("input_209_cast_fp16")]; + tensor attn17_out_proj_weight_to_fp16 = const()[name = tensor("attn17_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(434281728)))]; + tensor linear_69_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn17_out_proj_weight_to_fp16, x = input_209_cast_fp16)[name = tensor("linear_69_cast_fp16")]; + tensor position17_to_fp16_dtype_0 = const()[name = tensor("position17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position17_to_fp16 = cast(dtype = position17_to_fp16_dtype_0, x = position17)[name = tensor("cast_359")]; + tensor var_4453_cast_fp16 = add(x = position17_to_fp16, y = valid_len_to_fp16)[name = tensor("op_4453_cast_fp16")]; + tensor var_4453_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4453_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_211_cast_fp16 = add(x = input_205_cast_fp16, y = linear_69_cast_fp16)[name = tensor("input_211_cast_fp16")]; + tensor input_213_axes_0 = const()[name = tensor("input_213_axes_0"), val = tensor([-1])]; + tensor norm17_2_weight_to_fp16 = const()[name = tensor("norm17_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(436378944)))]; + tensor norm17_2_bias_to_fp16 = const()[name = tensor("norm17_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(436381056)))]; + tensor var_4457_to_fp16 = const()[name = tensor("op_4457_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_213_cast_fp16 = layer_norm(axes = input_213_axes_0, beta = norm17_2_bias_to_fp16, epsilon = var_4457_to_fp16, gamma = norm17_2_weight_to_fp16, x = input_211_cast_fp16)[name = tensor("input_213_cast_fp16")]; + tensor linear17_1_weight_to_fp16 = const()[name = tensor("linear17_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(436383168)))]; + tensor linear_70_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear17_1_weight_to_fp16, x = input_213_cast_fp16)[name = tensor("linear_70_cast_fp16")]; + tensor input_215_mode_0 = const()[name = tensor("input_215_mode_0"), val = tensor("EXACT")]; + tensor input_215_cast_fp16 = gelu(mode = input_215_mode_0, x = linear_70_cast_fp16)[name = tensor("input_215_cast_fp16")]; + tensor linear17_2_weight_to_fp16 = const()[name = tensor("linear17_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(444771840)))]; + tensor linear_71_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear17_2_weight_to_fp16, x = input_215_cast_fp16)[name = tensor("linear_71_cast_fp16")]; + tensor input_217_cast_fp16 = add(x = input_211_cast_fp16, y = linear_71_cast_fp16)[name = tensor("input_217_cast_fp16")]; + tensor input_219_axes_0 = const()[name = tensor("input_219_axes_0"), val = tensor([-1])]; + tensor norm18_1_weight_to_fp16 = const()[name = tensor("norm18_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(453160512)))]; + tensor norm18_1_bias_to_fp16 = const()[name = tensor("norm18_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(453162624)))]; + tensor var_4474_to_fp16 = const()[name = tensor("op_4474_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_219_cast_fp16 = layer_norm(axes = input_219_axes_0, beta = norm18_1_bias_to_fp16, epsilon = var_4474_to_fp16, gamma = norm18_1_weight_to_fp16, x = input_217_cast_fp16)[name = tensor("input_219_cast_fp16")]; + tensor attn18_in_proj_weight_to_fp16 = const()[name = tensor("attn18_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(453164736)))]; + tensor linear_72_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn18_in_proj_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_72_cast_fp16")]; + tensor var_4488 = const()[name = tensor("op_4488"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_37_cast_fp16 = reshape(shape = var_4488, x = linear_72_cast_fp16)[name = tensor("qkv_37_cast_fp16")]; + tensor x_73_begin_0 = const()[name = tensor("x_73_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_73_end_0 = const()[name = tensor("x_73_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_73_end_mask_0 = const()[name = tensor("x_73_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_73_squeeze_mask_0 = const()[name = tensor("x_73_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_73_cast_fp16 = slice_by_index(begin = x_73_begin_0, end = x_73_end_0, end_mask = x_73_end_mask_0, squeeze_mask = x_73_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("x_73_cast_fp16")]; + tensor x_75_begin_0 = const()[name = tensor("x_75_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_75_end_0 = const()[name = tensor("x_75_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_75_end_mask_0 = const()[name = tensor("x_75_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_75_squeeze_mask_0 = const()[name = tensor("x_75_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_75_cast_fp16 = slice_by_index(begin = x_75_begin_0, end = x_75_end_0, end_mask = x_75_end_mask_0, squeeze_mask = x_75_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("x_75_cast_fp16")]; + tensor v_37_begin_0 = const()[name = tensor("v_37_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_37_end_0 = const()[name = tensor("v_37_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_37_end_mask_0 = const()[name = tensor("v_37_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_37_squeeze_mask_0 = const()[name = tensor("v_37_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_37_cast_fp16 = slice_by_index(begin = v_37_begin_0, end = v_37_end_0, end_mask = v_37_end_mask_0, squeeze_mask = v_37_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("v_37_cast_fp16")]; + tensor var_4533 = const()[name = tensor("op_4533"), val = tensor([256, 16, 32, 2])]; + tensor pairs_73_cast_fp16 = reshape(shape = var_4533, x = x_73_cast_fp16)[name = tensor("pairs_73_cast_fp16")]; + tensor xr_73_begin_0 = const()[name = tensor("xr_73_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_73_end_0 = const()[name = tensor("xr_73_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_73_end_mask_0 = const()[name = tensor("xr_73_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_73_squeeze_mask_0 = const()[name = tensor("xr_73_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_73_cast_fp16 = slice_by_index(begin = xr_73_begin_0, end = xr_73_end_0, end_mask = xr_73_end_mask_0, squeeze_mask = xr_73_squeeze_mask_0, x = pairs_73_cast_fp16)[name = tensor("xr_73_cast_fp16")]; + tensor xi_73_begin_0 = const()[name = tensor("xi_73_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_73_end_0 = const()[name = tensor("xi_73_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_73_end_mask_0 = const()[name = tensor("xi_73_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_73_squeeze_mask_0 = const()[name = tensor("xi_73_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_73_cast_fp16 = slice_by_index(begin = xi_73_begin_0, end = xi_73_end_0, end_mask = xi_73_end_mask_0, squeeze_mask = xi_73_squeeze_mask_0, x = pairs_73_cast_fp16)[name = tensor("xi_73_cast_fp16")]; + tensor var_4541_cast_fp16 = mul(x = xr_73_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4541_cast_fp16")]; + tensor var_4542_cast_fp16 = mul(x = xi_73_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4542_cast_fp16")]; + tensor xor__73_cast_fp16 = sub(x = var_4541_cast_fp16, y = var_4542_cast_fp16)[name = tensor("xor__73_cast_fp16")]; + tensor var_4545_cast_fp16 = mul(x = xr_73_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4545_cast_fp16")]; + tensor var_4546_cast_fp16 = mul(x = xi_73_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4546_cast_fp16")]; + tensor xoi_73_cast_fp16 = add(x = var_4545_cast_fp16, y = var_4546_cast_fp16)[name = tensor("xoi_73_cast_fp16")]; + tensor out_73_axis_0 = const()[name = tensor("out_73_axis_0"), val = tensor(-1)]; + tensor out_73_cast_fp16 = stack(axis = out_73_axis_0, values = (xor__73_cast_fp16, xoi_73_cast_fp16))[name = tensor("out_73_cast_fp16")]; + tensor var_4556 = const()[name = tensor("op_4556"), val = tensor([1, 256, 16, 64])]; + tensor q_37_cast_fp16 = reshape(shape = var_4556, x = out_73_cast_fp16)[name = tensor("q_37_cast_fp16")]; + tensor var_4562 = const()[name = tensor("op_4562"), val = tensor([256, 16, 32, 2])]; + tensor pairs_75_cast_fp16 = reshape(shape = var_4562, x = x_75_cast_fp16)[name = tensor("pairs_75_cast_fp16")]; + tensor xr_75_begin_0 = const()[name = tensor("xr_75_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_75_end_0 = const()[name = tensor("xr_75_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_75_end_mask_0 = const()[name = tensor("xr_75_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_75_squeeze_mask_0 = const()[name = tensor("xr_75_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_75_cast_fp16 = slice_by_index(begin = xr_75_begin_0, end = xr_75_end_0, end_mask = xr_75_end_mask_0, squeeze_mask = xr_75_squeeze_mask_0, x = pairs_75_cast_fp16)[name = tensor("xr_75_cast_fp16")]; + tensor xi_75_begin_0 = const()[name = tensor("xi_75_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_75_end_0 = const()[name = tensor("xi_75_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_75_end_mask_0 = const()[name = tensor("xi_75_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_75_squeeze_mask_0 = const()[name = tensor("xi_75_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_75_cast_fp16 = slice_by_index(begin = xi_75_begin_0, end = xi_75_end_0, end_mask = xi_75_end_mask_0, squeeze_mask = xi_75_squeeze_mask_0, x = pairs_75_cast_fp16)[name = tensor("xi_75_cast_fp16")]; + tensor var_4570_cast_fp16 = mul(x = xr_75_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4570_cast_fp16")]; + tensor var_4571_cast_fp16 = mul(x = xi_75_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4571_cast_fp16")]; + tensor xor__75_cast_fp16 = sub(x = var_4570_cast_fp16, y = var_4571_cast_fp16)[name = tensor("xor__75_cast_fp16")]; + tensor var_4574_cast_fp16 = mul(x = xr_75_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4574_cast_fp16")]; + tensor var_4575_cast_fp16 = mul(x = xi_75_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4575_cast_fp16")]; + tensor xoi_75_cast_fp16 = add(x = var_4574_cast_fp16, y = var_4575_cast_fp16)[name = tensor("xoi_75_cast_fp16")]; + tensor out_75_axis_0 = const()[name = tensor("out_75_axis_0"), val = tensor(-1)]; + tensor out_75_cast_fp16 = stack(axis = out_75_axis_0, values = (xor__75_cast_fp16, xoi_75_cast_fp16))[name = tensor("out_75_cast_fp16")]; + tensor var_4592 = const()[name = tensor("op_4592"), val = tensor([256, 1024])]; + tensor var_4593_cast_fp16 = reshape(shape = var_4592, x = out_75_cast_fp16)[name = tensor("op_4593_cast_fp16")]; + tensor written_k_37_transpose_x_1 = const()[name = tensor("written_k_37_transpose_x_1"), val = tensor(true)]; + tensor written_k_37_transpose_y_1 = const()[name = tensor("written_k_37_transpose_y_1"), val = tensor(false)]; + tensor written_k_37_cast_fp16 = matmul(transpose_x = written_k_37_transpose_x_1, transpose_y = written_k_37_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_4593_cast_fp16)[name = tensor("written_k_37_cast_fp16")]; + tensor var_4600 = const()[name = tensor("op_4600"), val = tensor([256, 1024])]; + tensor var_4601_cast_fp16 = reshape(shape = var_4600, x = v_37_cast_fp16)[name = tensor("op_4601_cast_fp16")]; + tensor written_v_37_transpose_x_1 = const()[name = tensor("written_v_37_transpose_x_1"), val = tensor(true)]; + tensor written_v_37_transpose_y_1 = const()[name = tensor("written_v_37_transpose_y_1"), val = tensor(false)]; + tensor written_v_37_cast_fp16 = matmul(transpose_x = written_v_37_transpose_x_1, transpose_y = written_v_37_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_4601_cast_fp16)[name = tensor("written_v_37_cast_fp16")]; + tensor k_cache18_to_fp16_dtype_0 = const()[name = tensor("k_cache18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache18_to_fp16 = cast(dtype = k_cache18_to_fp16_dtype_0, x = k_cache18)[name = tensor("cast_357")]; + tensor var_4612_cast_fp16 = mul(x = k_cache18_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_4612_cast_fp16")]; + tensor var_4617 = const()[name = tensor("op_4617"), val = tensor([1, 512, 16, 64])]; + tensor var_4618_cast_fp16 = reshape(shape = var_4617, x = written_k_37_cast_fp16)[name = tensor("op_4618_cast_fp16")]; + tensor new_k_cache_37_cast_fp16 = add(x = var_4612_cast_fp16, y = var_4618_cast_fp16)[name = tensor("new_k_cache_37_cast_fp16")]; + tensor new_k_cache_37_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_37_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache18_to_fp16_dtype_0 = const()[name = tensor("v_cache18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache18_to_fp16 = cast(dtype = v_cache18_to_fp16_dtype_0, x = v_cache18)[name = tensor("cast_355")]; + tensor var_4621_cast_fp16 = mul(x = v_cache18_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_4621_cast_fp16")]; + tensor var_4626 = const()[name = tensor("op_4626"), val = tensor([1, 512, 16, 64])]; + tensor var_4627_cast_fp16 = reshape(shape = var_4626, x = written_v_37_cast_fp16)[name = tensor("op_4627_cast_fp16")]; + tensor new_v_cache_37_cast_fp16 = add(x = var_4621_cast_fp16, y = var_4627_cast_fp16)[name = tensor("new_v_cache_37_cast_fp16")]; + tensor new_v_cache_37_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_37_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_4640 = const()[name = tensor("op_4640"), val = tensor([0, 2, 1, 3])]; + tensor var_4651_transpose_x_0 = const()[name = tensor("op_4651_transpose_x_0"), val = tensor(false)]; + tensor var_4651_transpose_y_0 = const()[name = tensor("op_4651_transpose_y_0"), val = tensor(false)]; + tensor transpose_105_perm_0 = const()[name = tensor("transpose_105_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_106_perm_0 = const()[name = tensor("transpose_106_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_106 = transpose(perm = transpose_106_perm_0, x = new_k_cache_37_cast_fp16)[name = tensor("transpose_132")]; + tensor transpose_105 = transpose(perm = transpose_105_perm_0, x = q_37_cast_fp16)[name = tensor("transpose_133")]; + tensor var_4651_cast_fp16 = matmul(transpose_x = var_4651_transpose_x_0, transpose_y = var_4651_transpose_y_0, x = transpose_105, y = transpose_106)[name = tensor("op_4651_cast_fp16")]; + tensor var_4652_to_fp16 = const()[name = tensor("op_4652_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_73_cast_fp16 = mul(x = var_4651_cast_fp16, y = var_4652_to_fp16)[name = tensor("scores_73_cast_fp16")]; + tensor scores_75_cast_fp16 = add(x = scores_73_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_75_cast_fp16")]; + tensor var_4662 = const()[name = tensor("op_4662"), val = tensor(-1)]; + tensor weights_37_cast_fp16 = softmax(axis = var_4662, x = scores_75_cast_fp16)[name = tensor("weights_37_cast_fp16")]; + tensor attn_37_transpose_x_0 = const()[name = tensor("attn_37_transpose_x_0"), val = tensor(false)]; + tensor attn_37_transpose_y_0 = const()[name = tensor("attn_37_transpose_y_0"), val = tensor(false)]; + tensor values_37_cast_fp16 = transpose(perm = var_4640, x = new_v_cache_37_cast_fp16)[name = tensor("transpose_134")]; + tensor attn_37_cast_fp16 = matmul(transpose_x = attn_37_transpose_x_0, transpose_y = attn_37_transpose_y_0, x = weights_37_cast_fp16, y = values_37_cast_fp16)[name = tensor("attn_37_cast_fp16")]; + tensor var_4670 = const()[name = tensor("op_4670"), val = tensor([0, 2, 1, 3])]; + tensor var_4675 = const()[name = tensor("op_4675"), val = tensor([1, 256, 1024])]; + tensor var_4671_cast_fp16 = transpose(perm = var_4670, x = attn_37_cast_fp16)[name = tensor("transpose_131")]; + tensor input_221_cast_fp16 = reshape(shape = var_4675, x = var_4671_cast_fp16)[name = tensor("input_221_cast_fp16")]; + tensor attn18_out_proj_weight_to_fp16 = const()[name = tensor("attn18_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(459456256)))]; + tensor linear_73_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn18_out_proj_weight_to_fp16, x = input_221_cast_fp16)[name = tensor("linear_73_cast_fp16")]; + tensor position18_to_fp16_dtype_0 = const()[name = tensor("position18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position18_to_fp16 = cast(dtype = position18_to_fp16_dtype_0, x = position18)[name = tensor("cast_353")]; + tensor var_4684_cast_fp16 = add(x = position18_to_fp16, y = valid_len_to_fp16)[name = tensor("op_4684_cast_fp16")]; + tensor var_4684_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4684_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_223_cast_fp16 = add(x = input_217_cast_fp16, y = linear_73_cast_fp16)[name = tensor("input_223_cast_fp16")]; + tensor input_225_axes_0 = const()[name = tensor("input_225_axes_0"), val = tensor([-1])]; + tensor norm18_2_weight_to_fp16 = const()[name = tensor("norm18_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(461553472)))]; + tensor norm18_2_bias_to_fp16 = const()[name = tensor("norm18_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(461555584)))]; + tensor var_4688_to_fp16 = const()[name = tensor("op_4688_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_225_cast_fp16 = layer_norm(axes = input_225_axes_0, beta = norm18_2_bias_to_fp16, epsilon = var_4688_to_fp16, gamma = norm18_2_weight_to_fp16, x = input_223_cast_fp16)[name = tensor("input_225_cast_fp16")]; + tensor linear18_1_weight_to_fp16 = const()[name = tensor("linear18_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(461557696)))]; + tensor linear_74_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear18_1_weight_to_fp16, x = input_225_cast_fp16)[name = tensor("linear_74_cast_fp16")]; + tensor input_227_mode_0 = const()[name = tensor("input_227_mode_0"), val = tensor("EXACT")]; + tensor input_227_cast_fp16 = gelu(mode = input_227_mode_0, x = linear_74_cast_fp16)[name = tensor("input_227_cast_fp16")]; + tensor linear18_2_weight_to_fp16 = const()[name = tensor("linear18_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(469946368)))]; + tensor linear_75_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear18_2_weight_to_fp16, x = input_227_cast_fp16)[name = tensor("linear_75_cast_fp16")]; + tensor input_229_cast_fp16 = add(x = input_223_cast_fp16, y = linear_75_cast_fp16)[name = tensor("input_229_cast_fp16")]; + tensor input_231_axes_0 = const()[name = tensor("input_231_axes_0"), val = tensor([-1])]; + tensor norm19_1_weight_to_fp16 = const()[name = tensor("norm19_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(478335040)))]; + tensor norm19_1_bias_to_fp16 = const()[name = tensor("norm19_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(478337152)))]; + tensor var_4705_to_fp16 = const()[name = tensor("op_4705_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_231_cast_fp16 = layer_norm(axes = input_231_axes_0, beta = norm19_1_bias_to_fp16, epsilon = var_4705_to_fp16, gamma = norm19_1_weight_to_fp16, x = input_229_cast_fp16)[name = tensor("input_231_cast_fp16")]; + tensor attn19_in_proj_weight_to_fp16 = const()[name = tensor("attn19_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(478339264)))]; + tensor linear_76_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn19_in_proj_weight_to_fp16, x = input_231_cast_fp16)[name = tensor("linear_76_cast_fp16")]; + tensor var_4719 = const()[name = tensor("op_4719"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_39_cast_fp16 = reshape(shape = var_4719, x = linear_76_cast_fp16)[name = tensor("qkv_39_cast_fp16")]; + tensor x_77_begin_0 = const()[name = tensor("x_77_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_77_end_0 = const()[name = tensor("x_77_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_77_end_mask_0 = const()[name = tensor("x_77_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_77_squeeze_mask_0 = const()[name = tensor("x_77_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_77_cast_fp16 = slice_by_index(begin = x_77_begin_0, end = x_77_end_0, end_mask = x_77_end_mask_0, squeeze_mask = x_77_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("x_77_cast_fp16")]; + tensor x_79_begin_0 = const()[name = tensor("x_79_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_79_end_0 = const()[name = tensor("x_79_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_79_end_mask_0 = const()[name = tensor("x_79_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_79_squeeze_mask_0 = const()[name = tensor("x_79_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_79_cast_fp16 = slice_by_index(begin = x_79_begin_0, end = x_79_end_0, end_mask = x_79_end_mask_0, squeeze_mask = x_79_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("x_79_cast_fp16")]; + tensor v_39_begin_0 = const()[name = tensor("v_39_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_39_end_0 = const()[name = tensor("v_39_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_39_end_mask_0 = const()[name = tensor("v_39_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_39_squeeze_mask_0 = const()[name = tensor("v_39_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_39_cast_fp16 = slice_by_index(begin = v_39_begin_0, end = v_39_end_0, end_mask = v_39_end_mask_0, squeeze_mask = v_39_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("v_39_cast_fp16")]; + tensor var_4764 = const()[name = tensor("op_4764"), val = tensor([256, 16, 32, 2])]; + tensor pairs_77_cast_fp16 = reshape(shape = var_4764, x = x_77_cast_fp16)[name = tensor("pairs_77_cast_fp16")]; + tensor xr_77_begin_0 = const()[name = tensor("xr_77_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_77_end_0 = const()[name = tensor("xr_77_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_77_end_mask_0 = const()[name = tensor("xr_77_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_77_squeeze_mask_0 = const()[name = tensor("xr_77_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_77_cast_fp16 = slice_by_index(begin = xr_77_begin_0, end = xr_77_end_0, end_mask = xr_77_end_mask_0, squeeze_mask = xr_77_squeeze_mask_0, x = pairs_77_cast_fp16)[name = tensor("xr_77_cast_fp16")]; + tensor xi_77_begin_0 = const()[name = tensor("xi_77_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_77_end_0 = const()[name = tensor("xi_77_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_77_end_mask_0 = const()[name = tensor("xi_77_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_77_squeeze_mask_0 = const()[name = tensor("xi_77_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_77_cast_fp16 = slice_by_index(begin = xi_77_begin_0, end = xi_77_end_0, end_mask = xi_77_end_mask_0, squeeze_mask = xi_77_squeeze_mask_0, x = pairs_77_cast_fp16)[name = tensor("xi_77_cast_fp16")]; + tensor var_4772_cast_fp16 = mul(x = xr_77_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4772_cast_fp16")]; + tensor var_4773_cast_fp16 = mul(x = xi_77_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4773_cast_fp16")]; + tensor xor__77_cast_fp16 = sub(x = var_4772_cast_fp16, y = var_4773_cast_fp16)[name = tensor("xor__77_cast_fp16")]; + tensor var_4776_cast_fp16 = mul(x = xr_77_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4776_cast_fp16")]; + tensor var_4777_cast_fp16 = mul(x = xi_77_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4777_cast_fp16")]; + tensor xoi_77_cast_fp16 = add(x = var_4776_cast_fp16, y = var_4777_cast_fp16)[name = tensor("xoi_77_cast_fp16")]; + tensor out_77_axis_0 = const()[name = tensor("out_77_axis_0"), val = tensor(-1)]; + tensor out_77_cast_fp16 = stack(axis = out_77_axis_0, values = (xor__77_cast_fp16, xoi_77_cast_fp16))[name = tensor("out_77_cast_fp16")]; + tensor var_4787 = const()[name = tensor("op_4787"), val = tensor([1, 256, 16, 64])]; + tensor q_39_cast_fp16 = reshape(shape = var_4787, x = out_77_cast_fp16)[name = tensor("q_39_cast_fp16")]; + tensor var_4793 = const()[name = tensor("op_4793"), val = tensor([256, 16, 32, 2])]; + tensor pairs_79_cast_fp16 = reshape(shape = var_4793, x = x_79_cast_fp16)[name = tensor("pairs_79_cast_fp16")]; + tensor xr_79_begin_0 = const()[name = tensor("xr_79_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_79_end_0 = const()[name = tensor("xr_79_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_79_end_mask_0 = const()[name = tensor("xr_79_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_79_squeeze_mask_0 = const()[name = tensor("xr_79_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_79_cast_fp16 = slice_by_index(begin = xr_79_begin_0, end = xr_79_end_0, end_mask = xr_79_end_mask_0, squeeze_mask = xr_79_squeeze_mask_0, x = pairs_79_cast_fp16)[name = tensor("xr_79_cast_fp16")]; + tensor xi_79_begin_0 = const()[name = tensor("xi_79_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_79_end_0 = const()[name = tensor("xi_79_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_79_end_mask_0 = const()[name = tensor("xi_79_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_79_squeeze_mask_0 = const()[name = tensor("xi_79_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_79_cast_fp16 = slice_by_index(begin = xi_79_begin_0, end = xi_79_end_0, end_mask = xi_79_end_mask_0, squeeze_mask = xi_79_squeeze_mask_0, x = pairs_79_cast_fp16)[name = tensor("xi_79_cast_fp16")]; + tensor var_4801_cast_fp16 = mul(x = xr_79_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4801_cast_fp16")]; + tensor var_4802_cast_fp16 = mul(x = xi_79_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4802_cast_fp16")]; + tensor xor__79_cast_fp16 = sub(x = var_4801_cast_fp16, y = var_4802_cast_fp16)[name = tensor("xor__79_cast_fp16")]; + tensor var_4805_cast_fp16 = mul(x = xr_79_cast_fp16, y = roti_cast_fp16)[name = tensor("op_4805_cast_fp16")]; + tensor var_4806_cast_fp16 = mul(x = xi_79_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_4806_cast_fp16")]; + tensor xoi_79_cast_fp16 = add(x = var_4805_cast_fp16, y = var_4806_cast_fp16)[name = tensor("xoi_79_cast_fp16")]; + tensor out_79_axis_0 = const()[name = tensor("out_79_axis_0"), val = tensor(-1)]; + tensor out_79_cast_fp16 = stack(axis = out_79_axis_0, values = (xor__79_cast_fp16, xoi_79_cast_fp16))[name = tensor("out_79_cast_fp16")]; + tensor var_4823 = const()[name = tensor("op_4823"), val = tensor([256, 1024])]; + tensor var_4824_cast_fp16 = reshape(shape = var_4823, x = out_79_cast_fp16)[name = tensor("op_4824_cast_fp16")]; + tensor written_k_39_transpose_x_1 = const()[name = tensor("written_k_39_transpose_x_1"), val = tensor(true)]; + tensor written_k_39_transpose_y_1 = const()[name = tensor("written_k_39_transpose_y_1"), val = tensor(false)]; + tensor written_k_39_cast_fp16 = matmul(transpose_x = written_k_39_transpose_x_1, transpose_y = written_k_39_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_4824_cast_fp16)[name = tensor("written_k_39_cast_fp16")]; + tensor var_4831 = const()[name = tensor("op_4831"), val = tensor([256, 1024])]; + tensor var_4832_cast_fp16 = reshape(shape = var_4831, x = v_39_cast_fp16)[name = tensor("op_4832_cast_fp16")]; + tensor written_v_39_transpose_x_1 = const()[name = tensor("written_v_39_transpose_x_1"), val = tensor(true)]; + tensor written_v_39_transpose_y_1 = const()[name = tensor("written_v_39_transpose_y_1"), val = tensor(false)]; + tensor written_v_39_cast_fp16 = matmul(transpose_x = written_v_39_transpose_x_1, transpose_y = written_v_39_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_4832_cast_fp16)[name = tensor("written_v_39_cast_fp16")]; + tensor k_cache19_to_fp16_dtype_0 = const()[name = tensor("k_cache19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache19_to_fp16 = cast(dtype = k_cache19_to_fp16_dtype_0, x = k_cache19)[name = tensor("cast_351")]; + tensor var_4843_cast_fp16 = mul(x = k_cache19_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_4843_cast_fp16")]; + tensor var_4848 = const()[name = tensor("op_4848"), val = tensor([1, 512, 16, 64])]; + tensor var_4849_cast_fp16 = reshape(shape = var_4848, x = written_k_39_cast_fp16)[name = tensor("op_4849_cast_fp16")]; + tensor new_k_cache_39_cast_fp16 = add(x = var_4843_cast_fp16, y = var_4849_cast_fp16)[name = tensor("new_k_cache_39_cast_fp16")]; + tensor new_k_cache_39_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_39_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache19_to_fp16_dtype_0 = const()[name = tensor("v_cache19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache19_to_fp16 = cast(dtype = v_cache19_to_fp16_dtype_0, x = v_cache19)[name = tensor("cast_349")]; + tensor var_4852_cast_fp16 = mul(x = v_cache19_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_4852_cast_fp16")]; + tensor var_4857 = const()[name = tensor("op_4857"), val = tensor([1, 512, 16, 64])]; + tensor var_4858_cast_fp16 = reshape(shape = var_4857, x = written_v_39_cast_fp16)[name = tensor("op_4858_cast_fp16")]; + tensor new_v_cache_39_cast_fp16 = add(x = var_4852_cast_fp16, y = var_4858_cast_fp16)[name = tensor("new_v_cache_39_cast_fp16")]; + tensor new_v_cache_39_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_39_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_4871 = const()[name = tensor("op_4871"), val = tensor([0, 2, 1, 3])]; + tensor var_4882_transpose_x_0 = const()[name = tensor("op_4882_transpose_x_0"), val = tensor(false)]; + tensor var_4882_transpose_y_0 = const()[name = tensor("op_4882_transpose_y_0"), val = tensor(false)]; + tensor transpose_107_perm_0 = const()[name = tensor("transpose_107_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_108_perm_0 = const()[name = tensor("transpose_108_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_108 = transpose(perm = transpose_108_perm_0, x = new_k_cache_39_cast_fp16)[name = tensor("transpose_128")]; + tensor transpose_107 = transpose(perm = transpose_107_perm_0, x = q_39_cast_fp16)[name = tensor("transpose_129")]; + tensor var_4882_cast_fp16 = matmul(transpose_x = var_4882_transpose_x_0, transpose_y = var_4882_transpose_y_0, x = transpose_107, y = transpose_108)[name = tensor("op_4882_cast_fp16")]; + tensor var_4883_to_fp16 = const()[name = tensor("op_4883_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_77_cast_fp16 = mul(x = var_4882_cast_fp16, y = var_4883_to_fp16)[name = tensor("scores_77_cast_fp16")]; + tensor scores_79_cast_fp16 = add(x = scores_77_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_79_cast_fp16")]; + tensor var_4893 = const()[name = tensor("op_4893"), val = tensor(-1)]; + tensor weights_39_cast_fp16 = softmax(axis = var_4893, x = scores_79_cast_fp16)[name = tensor("weights_39_cast_fp16")]; + tensor attn_39_transpose_x_0 = const()[name = tensor("attn_39_transpose_x_0"), val = tensor(false)]; + tensor attn_39_transpose_y_0 = const()[name = tensor("attn_39_transpose_y_0"), val = tensor(false)]; + tensor values_39_cast_fp16 = transpose(perm = var_4871, x = new_v_cache_39_cast_fp16)[name = tensor("transpose_130")]; + tensor attn_39_cast_fp16 = matmul(transpose_x = attn_39_transpose_x_0, transpose_y = attn_39_transpose_y_0, x = weights_39_cast_fp16, y = values_39_cast_fp16)[name = tensor("attn_39_cast_fp16")]; + tensor var_4901 = const()[name = tensor("op_4901"), val = tensor([0, 2, 1, 3])]; + tensor var_4906 = const()[name = tensor("op_4906"), val = tensor([1, 256, 1024])]; + tensor var_4902_cast_fp16 = transpose(perm = var_4901, x = attn_39_cast_fp16)[name = tensor("transpose_127")]; + tensor input_233_cast_fp16 = reshape(shape = var_4906, x = var_4902_cast_fp16)[name = tensor("input_233_cast_fp16")]; + tensor attn19_out_proj_weight_to_fp16 = const()[name = tensor("attn19_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(484630784)))]; + tensor linear_77_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn19_out_proj_weight_to_fp16, x = input_233_cast_fp16)[name = tensor("linear_77_cast_fp16")]; + tensor position19_to_fp16_dtype_0 = const()[name = tensor("position19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position19_to_fp16 = cast(dtype = position19_to_fp16_dtype_0, x = position19)[name = tensor("cast_347")]; + tensor var_4915_cast_fp16 = add(x = position19_to_fp16, y = valid_len_to_fp16)[name = tensor("op_4915_cast_fp16")]; + tensor var_4915_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4915_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_235_cast_fp16 = add(x = input_229_cast_fp16, y = linear_77_cast_fp16)[name = tensor("input_235_cast_fp16")]; + tensor input_237_axes_0 = const()[name = tensor("input_237_axes_0"), val = tensor([-1])]; + tensor norm19_2_weight_to_fp16 = const()[name = tensor("norm19_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486728000)))]; + tensor norm19_2_bias_to_fp16 = const()[name = tensor("norm19_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486730112)))]; + tensor var_4919_to_fp16 = const()[name = tensor("op_4919_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_237_cast_fp16 = layer_norm(axes = input_237_axes_0, beta = norm19_2_bias_to_fp16, epsilon = var_4919_to_fp16, gamma = norm19_2_weight_to_fp16, x = input_235_cast_fp16)[name = tensor("input_237_cast_fp16")]; + tensor linear19_1_weight_to_fp16 = const()[name = tensor("linear19_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486732224)))]; + tensor linear_78_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear19_1_weight_to_fp16, x = input_237_cast_fp16)[name = tensor("linear_78_cast_fp16")]; + tensor input_239_mode_0 = const()[name = tensor("input_239_mode_0"), val = tensor("EXACT")]; + tensor input_239_cast_fp16 = gelu(mode = input_239_mode_0, x = linear_78_cast_fp16)[name = tensor("input_239_cast_fp16")]; + tensor linear19_2_weight_to_fp16 = const()[name = tensor("linear19_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(495120896)))]; + tensor linear_79_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear19_2_weight_to_fp16, x = input_239_cast_fp16)[name = tensor("linear_79_cast_fp16")]; + tensor input_241_cast_fp16 = add(x = input_235_cast_fp16, y = linear_79_cast_fp16)[name = tensor("input_241_cast_fp16")]; + tensor input_243_axes_0 = const()[name = tensor("input_243_axes_0"), val = tensor([-1])]; + tensor norm20_1_weight_to_fp16 = const()[name = tensor("norm20_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(503509568)))]; + tensor norm20_1_bias_to_fp16 = const()[name = tensor("norm20_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(503511680)))]; + tensor var_4936_to_fp16 = const()[name = tensor("op_4936_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_243_cast_fp16 = layer_norm(axes = input_243_axes_0, beta = norm20_1_bias_to_fp16, epsilon = var_4936_to_fp16, gamma = norm20_1_weight_to_fp16, x = input_241_cast_fp16)[name = tensor("input_243_cast_fp16")]; + tensor attn20_in_proj_weight_to_fp16 = const()[name = tensor("attn20_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(503513792)))]; + tensor linear_80_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn20_in_proj_weight_to_fp16, x = input_243_cast_fp16)[name = tensor("linear_80_cast_fp16")]; + tensor var_4950 = const()[name = tensor("op_4950"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_41_cast_fp16 = reshape(shape = var_4950, x = linear_80_cast_fp16)[name = tensor("qkv_41_cast_fp16")]; + tensor x_81_begin_0 = const()[name = tensor("x_81_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_81_end_0 = const()[name = tensor("x_81_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_81_end_mask_0 = const()[name = tensor("x_81_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_81_squeeze_mask_0 = const()[name = tensor("x_81_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_81_cast_fp16 = slice_by_index(begin = x_81_begin_0, end = x_81_end_0, end_mask = x_81_end_mask_0, squeeze_mask = x_81_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("x_81_cast_fp16")]; + tensor x_83_begin_0 = const()[name = tensor("x_83_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_83_end_0 = const()[name = tensor("x_83_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_83_end_mask_0 = const()[name = tensor("x_83_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_83_squeeze_mask_0 = const()[name = tensor("x_83_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_83_cast_fp16 = slice_by_index(begin = x_83_begin_0, end = x_83_end_0, end_mask = x_83_end_mask_0, squeeze_mask = x_83_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("x_83_cast_fp16")]; + tensor v_41_begin_0 = const()[name = tensor("v_41_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_41_end_0 = const()[name = tensor("v_41_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_41_end_mask_0 = const()[name = tensor("v_41_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_41_squeeze_mask_0 = const()[name = tensor("v_41_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_41_cast_fp16 = slice_by_index(begin = v_41_begin_0, end = v_41_end_0, end_mask = v_41_end_mask_0, squeeze_mask = v_41_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("v_41_cast_fp16")]; + tensor var_4995 = const()[name = tensor("op_4995"), val = tensor([256, 16, 32, 2])]; + tensor pairs_81_cast_fp16 = reshape(shape = var_4995, x = x_81_cast_fp16)[name = tensor("pairs_81_cast_fp16")]; + tensor xr_81_begin_0 = const()[name = tensor("xr_81_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_81_end_0 = const()[name = tensor("xr_81_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_81_end_mask_0 = const()[name = tensor("xr_81_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_81_squeeze_mask_0 = const()[name = tensor("xr_81_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_81_cast_fp16 = slice_by_index(begin = xr_81_begin_0, end = xr_81_end_0, end_mask = xr_81_end_mask_0, squeeze_mask = xr_81_squeeze_mask_0, x = pairs_81_cast_fp16)[name = tensor("xr_81_cast_fp16")]; + tensor xi_81_begin_0 = const()[name = tensor("xi_81_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_81_end_0 = const()[name = tensor("xi_81_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_81_end_mask_0 = const()[name = tensor("xi_81_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_81_squeeze_mask_0 = const()[name = tensor("xi_81_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_81_cast_fp16 = slice_by_index(begin = xi_81_begin_0, end = xi_81_end_0, end_mask = xi_81_end_mask_0, squeeze_mask = xi_81_squeeze_mask_0, x = pairs_81_cast_fp16)[name = tensor("xi_81_cast_fp16")]; + tensor var_5003_cast_fp16 = mul(x = xr_81_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5003_cast_fp16")]; + tensor var_5004_cast_fp16 = mul(x = xi_81_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5004_cast_fp16")]; + tensor xor__81_cast_fp16 = sub(x = var_5003_cast_fp16, y = var_5004_cast_fp16)[name = tensor("xor__81_cast_fp16")]; + tensor var_5007_cast_fp16 = mul(x = xr_81_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5007_cast_fp16")]; + tensor var_5008_cast_fp16 = mul(x = xi_81_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5008_cast_fp16")]; + tensor xoi_81_cast_fp16 = add(x = var_5007_cast_fp16, y = var_5008_cast_fp16)[name = tensor("xoi_81_cast_fp16")]; + tensor out_81_axis_0 = const()[name = tensor("out_81_axis_0"), val = tensor(-1)]; + tensor out_81_cast_fp16 = stack(axis = out_81_axis_0, values = (xor__81_cast_fp16, xoi_81_cast_fp16))[name = tensor("out_81_cast_fp16")]; + tensor var_5018 = const()[name = tensor("op_5018"), val = tensor([1, 256, 16, 64])]; + tensor q_41_cast_fp16 = reshape(shape = var_5018, x = out_81_cast_fp16)[name = tensor("q_41_cast_fp16")]; + tensor var_5024 = const()[name = tensor("op_5024"), val = tensor([256, 16, 32, 2])]; + tensor pairs_83_cast_fp16 = reshape(shape = var_5024, x = x_83_cast_fp16)[name = tensor("pairs_83_cast_fp16")]; + tensor xr_83_begin_0 = const()[name = tensor("xr_83_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_83_end_0 = const()[name = tensor("xr_83_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_83_end_mask_0 = const()[name = tensor("xr_83_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_83_squeeze_mask_0 = const()[name = tensor("xr_83_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_83_cast_fp16 = slice_by_index(begin = xr_83_begin_0, end = xr_83_end_0, end_mask = xr_83_end_mask_0, squeeze_mask = xr_83_squeeze_mask_0, x = pairs_83_cast_fp16)[name = tensor("xr_83_cast_fp16")]; + tensor xi_83_begin_0 = const()[name = tensor("xi_83_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_83_end_0 = const()[name = tensor("xi_83_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_83_end_mask_0 = const()[name = tensor("xi_83_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_83_squeeze_mask_0 = const()[name = tensor("xi_83_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_83_cast_fp16 = slice_by_index(begin = xi_83_begin_0, end = xi_83_end_0, end_mask = xi_83_end_mask_0, squeeze_mask = xi_83_squeeze_mask_0, x = pairs_83_cast_fp16)[name = tensor("xi_83_cast_fp16")]; + tensor var_5032_cast_fp16 = mul(x = xr_83_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5032_cast_fp16")]; + tensor var_5033_cast_fp16 = mul(x = xi_83_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5033_cast_fp16")]; + tensor xor__83_cast_fp16 = sub(x = var_5032_cast_fp16, y = var_5033_cast_fp16)[name = tensor("xor__83_cast_fp16")]; + tensor var_5036_cast_fp16 = mul(x = xr_83_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5036_cast_fp16")]; + tensor var_5037_cast_fp16 = mul(x = xi_83_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5037_cast_fp16")]; + tensor xoi_83_cast_fp16 = add(x = var_5036_cast_fp16, y = var_5037_cast_fp16)[name = tensor("xoi_83_cast_fp16")]; + tensor out_83_axis_0 = const()[name = tensor("out_83_axis_0"), val = tensor(-1)]; + tensor out_83_cast_fp16 = stack(axis = out_83_axis_0, values = (xor__83_cast_fp16, xoi_83_cast_fp16))[name = tensor("out_83_cast_fp16")]; + tensor var_5054 = const()[name = tensor("op_5054"), val = tensor([256, 1024])]; + tensor var_5055_cast_fp16 = reshape(shape = var_5054, x = out_83_cast_fp16)[name = tensor("op_5055_cast_fp16")]; + tensor written_k_41_transpose_x_1 = const()[name = tensor("written_k_41_transpose_x_1"), val = tensor(true)]; + tensor written_k_41_transpose_y_1 = const()[name = tensor("written_k_41_transpose_y_1"), val = tensor(false)]; + tensor written_k_41_cast_fp16 = matmul(transpose_x = written_k_41_transpose_x_1, transpose_y = written_k_41_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_5055_cast_fp16)[name = tensor("written_k_41_cast_fp16")]; + tensor var_5062 = const()[name = tensor("op_5062"), val = tensor([256, 1024])]; + tensor var_5063_cast_fp16 = reshape(shape = var_5062, x = v_41_cast_fp16)[name = tensor("op_5063_cast_fp16")]; + tensor written_v_41_transpose_x_1 = const()[name = tensor("written_v_41_transpose_x_1"), val = tensor(true)]; + tensor written_v_41_transpose_y_1 = const()[name = tensor("written_v_41_transpose_y_1"), val = tensor(false)]; + tensor written_v_41_cast_fp16 = matmul(transpose_x = written_v_41_transpose_x_1, transpose_y = written_v_41_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_5063_cast_fp16)[name = tensor("written_v_41_cast_fp16")]; + tensor k_cache20_to_fp16_dtype_0 = const()[name = tensor("k_cache20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache20_to_fp16 = cast(dtype = k_cache20_to_fp16_dtype_0, x = k_cache20)[name = tensor("cast_345")]; + tensor var_5074_cast_fp16 = mul(x = k_cache20_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_5074_cast_fp16")]; + tensor var_5079 = const()[name = tensor("op_5079"), val = tensor([1, 512, 16, 64])]; + tensor var_5080_cast_fp16 = reshape(shape = var_5079, x = written_k_41_cast_fp16)[name = tensor("op_5080_cast_fp16")]; + tensor new_k_cache_41_cast_fp16 = add(x = var_5074_cast_fp16, y = var_5080_cast_fp16)[name = tensor("new_k_cache_41_cast_fp16")]; + tensor new_k_cache_41_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_41_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache20_to_fp16_dtype_0 = const()[name = tensor("v_cache20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache20_to_fp16 = cast(dtype = v_cache20_to_fp16_dtype_0, x = v_cache20)[name = tensor("cast_343")]; + tensor var_5083_cast_fp16 = mul(x = v_cache20_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_5083_cast_fp16")]; + tensor var_5088 = const()[name = tensor("op_5088"), val = tensor([1, 512, 16, 64])]; + tensor var_5089_cast_fp16 = reshape(shape = var_5088, x = written_v_41_cast_fp16)[name = tensor("op_5089_cast_fp16")]; + tensor new_v_cache_41_cast_fp16 = add(x = var_5083_cast_fp16, y = var_5089_cast_fp16)[name = tensor("new_v_cache_41_cast_fp16")]; + tensor new_v_cache_41_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_41_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_5102 = const()[name = tensor("op_5102"), val = tensor([0, 2, 1, 3])]; + tensor var_5113_transpose_x_0 = const()[name = tensor("op_5113_transpose_x_0"), val = tensor(false)]; + tensor var_5113_transpose_y_0 = const()[name = tensor("op_5113_transpose_y_0"), val = tensor(false)]; + tensor transpose_109_perm_0 = const()[name = tensor("transpose_109_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_110_perm_0 = const()[name = tensor("transpose_110_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_110 = transpose(perm = transpose_110_perm_0, x = new_k_cache_41_cast_fp16)[name = tensor("transpose_124")]; + tensor transpose_109 = transpose(perm = transpose_109_perm_0, x = q_41_cast_fp16)[name = tensor("transpose_125")]; + tensor var_5113_cast_fp16 = matmul(transpose_x = var_5113_transpose_x_0, transpose_y = var_5113_transpose_y_0, x = transpose_109, y = transpose_110)[name = tensor("op_5113_cast_fp16")]; + tensor var_5114_to_fp16 = const()[name = tensor("op_5114_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_81_cast_fp16 = mul(x = var_5113_cast_fp16, y = var_5114_to_fp16)[name = tensor("scores_81_cast_fp16")]; + tensor scores_83_cast_fp16 = add(x = scores_81_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_83_cast_fp16")]; + tensor var_5124 = const()[name = tensor("op_5124"), val = tensor(-1)]; + tensor weights_41_cast_fp16 = softmax(axis = var_5124, x = scores_83_cast_fp16)[name = tensor("weights_41_cast_fp16")]; + tensor attn_41_transpose_x_0 = const()[name = tensor("attn_41_transpose_x_0"), val = tensor(false)]; + tensor attn_41_transpose_y_0 = const()[name = tensor("attn_41_transpose_y_0"), val = tensor(false)]; + tensor values_41_cast_fp16 = transpose(perm = var_5102, x = new_v_cache_41_cast_fp16)[name = tensor("transpose_126")]; + tensor attn_41_cast_fp16 = matmul(transpose_x = attn_41_transpose_x_0, transpose_y = attn_41_transpose_y_0, x = weights_41_cast_fp16, y = values_41_cast_fp16)[name = tensor("attn_41_cast_fp16")]; + tensor var_5132 = const()[name = tensor("op_5132"), val = tensor([0, 2, 1, 3])]; + tensor var_5137 = const()[name = tensor("op_5137"), val = tensor([1, 256, 1024])]; + tensor var_5133_cast_fp16 = transpose(perm = var_5132, x = attn_41_cast_fp16)[name = tensor("transpose_123")]; + tensor input_245_cast_fp16 = reshape(shape = var_5137, x = var_5133_cast_fp16)[name = tensor("input_245_cast_fp16")]; + tensor attn20_out_proj_weight_to_fp16 = const()[name = tensor("attn20_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(509805312)))]; + tensor linear_81_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn20_out_proj_weight_to_fp16, x = input_245_cast_fp16)[name = tensor("linear_81_cast_fp16")]; + tensor position20_to_fp16_dtype_0 = const()[name = tensor("position20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position20_to_fp16 = cast(dtype = position20_to_fp16_dtype_0, x = position20)[name = tensor("cast_341")]; + tensor var_5146_cast_fp16 = add(x = position20_to_fp16, y = valid_len_to_fp16)[name = tensor("op_5146_cast_fp16")]; + tensor var_5146_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5146_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_247_cast_fp16 = add(x = input_241_cast_fp16, y = linear_81_cast_fp16)[name = tensor("input_247_cast_fp16")]; + tensor input_249_axes_0 = const()[name = tensor("input_249_axes_0"), val = tensor([-1])]; + tensor norm20_2_weight_to_fp16 = const()[name = tensor("norm20_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(511902528)))]; + tensor norm20_2_bias_to_fp16 = const()[name = tensor("norm20_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(511904640)))]; + tensor var_5150_to_fp16 = const()[name = tensor("op_5150_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_249_cast_fp16 = layer_norm(axes = input_249_axes_0, beta = norm20_2_bias_to_fp16, epsilon = var_5150_to_fp16, gamma = norm20_2_weight_to_fp16, x = input_247_cast_fp16)[name = tensor("input_249_cast_fp16")]; + tensor linear20_1_weight_to_fp16 = const()[name = tensor("linear20_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(511906752)))]; + tensor linear_82_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear20_1_weight_to_fp16, x = input_249_cast_fp16)[name = tensor("linear_82_cast_fp16")]; + tensor input_251_mode_0 = const()[name = tensor("input_251_mode_0"), val = tensor("EXACT")]; + tensor input_251_cast_fp16 = gelu(mode = input_251_mode_0, x = linear_82_cast_fp16)[name = tensor("input_251_cast_fp16")]; + tensor linear20_2_weight_to_fp16 = const()[name = tensor("linear20_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(520295424)))]; + tensor linear_83_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear20_2_weight_to_fp16, x = input_251_cast_fp16)[name = tensor("linear_83_cast_fp16")]; + tensor input_253_cast_fp16 = add(x = input_247_cast_fp16, y = linear_83_cast_fp16)[name = tensor("input_253_cast_fp16")]; + tensor input_255_axes_0 = const()[name = tensor("input_255_axes_0"), val = tensor([-1])]; + tensor norm21_1_weight_to_fp16 = const()[name = tensor("norm21_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(528684096)))]; + tensor norm21_1_bias_to_fp16 = const()[name = tensor("norm21_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(528686208)))]; + tensor var_5167_to_fp16 = const()[name = tensor("op_5167_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_255_cast_fp16 = layer_norm(axes = input_255_axes_0, beta = norm21_1_bias_to_fp16, epsilon = var_5167_to_fp16, gamma = norm21_1_weight_to_fp16, x = input_253_cast_fp16)[name = tensor("input_255_cast_fp16")]; + tensor attn21_in_proj_weight_to_fp16 = const()[name = tensor("attn21_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(528688320)))]; + tensor linear_84_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn21_in_proj_weight_to_fp16, x = input_255_cast_fp16)[name = tensor("linear_84_cast_fp16")]; + tensor var_5181 = const()[name = tensor("op_5181"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_43_cast_fp16 = reshape(shape = var_5181, x = linear_84_cast_fp16)[name = tensor("qkv_43_cast_fp16")]; + tensor x_85_begin_0 = const()[name = tensor("x_85_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_85_end_0 = const()[name = tensor("x_85_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_85_end_mask_0 = const()[name = tensor("x_85_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_85_squeeze_mask_0 = const()[name = tensor("x_85_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_85_cast_fp16 = slice_by_index(begin = x_85_begin_0, end = x_85_end_0, end_mask = x_85_end_mask_0, squeeze_mask = x_85_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("x_85_cast_fp16")]; + tensor x_87_begin_0 = const()[name = tensor("x_87_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_87_end_0 = const()[name = tensor("x_87_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_87_end_mask_0 = const()[name = tensor("x_87_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_87_squeeze_mask_0 = const()[name = tensor("x_87_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_87_cast_fp16 = slice_by_index(begin = x_87_begin_0, end = x_87_end_0, end_mask = x_87_end_mask_0, squeeze_mask = x_87_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("x_87_cast_fp16")]; + tensor v_43_begin_0 = const()[name = tensor("v_43_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_43_end_0 = const()[name = tensor("v_43_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_43_end_mask_0 = const()[name = tensor("v_43_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_43_squeeze_mask_0 = const()[name = tensor("v_43_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_43_cast_fp16 = slice_by_index(begin = v_43_begin_0, end = v_43_end_0, end_mask = v_43_end_mask_0, squeeze_mask = v_43_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("v_43_cast_fp16")]; + tensor var_5226 = const()[name = tensor("op_5226"), val = tensor([256, 16, 32, 2])]; + tensor pairs_85_cast_fp16 = reshape(shape = var_5226, x = x_85_cast_fp16)[name = tensor("pairs_85_cast_fp16")]; + tensor xr_85_begin_0 = const()[name = tensor("xr_85_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_85_end_0 = const()[name = tensor("xr_85_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_85_end_mask_0 = const()[name = tensor("xr_85_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_85_squeeze_mask_0 = const()[name = tensor("xr_85_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_85_cast_fp16 = slice_by_index(begin = xr_85_begin_0, end = xr_85_end_0, end_mask = xr_85_end_mask_0, squeeze_mask = xr_85_squeeze_mask_0, x = pairs_85_cast_fp16)[name = tensor("xr_85_cast_fp16")]; + tensor xi_85_begin_0 = const()[name = tensor("xi_85_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_85_end_0 = const()[name = tensor("xi_85_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_85_end_mask_0 = const()[name = tensor("xi_85_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_85_squeeze_mask_0 = const()[name = tensor("xi_85_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_85_cast_fp16 = slice_by_index(begin = xi_85_begin_0, end = xi_85_end_0, end_mask = xi_85_end_mask_0, squeeze_mask = xi_85_squeeze_mask_0, x = pairs_85_cast_fp16)[name = tensor("xi_85_cast_fp16")]; + tensor var_5234_cast_fp16 = mul(x = xr_85_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5234_cast_fp16")]; + tensor var_5235_cast_fp16 = mul(x = xi_85_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5235_cast_fp16")]; + tensor xor__85_cast_fp16 = sub(x = var_5234_cast_fp16, y = var_5235_cast_fp16)[name = tensor("xor__85_cast_fp16")]; + tensor var_5238_cast_fp16 = mul(x = xr_85_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5238_cast_fp16")]; + tensor var_5239_cast_fp16 = mul(x = xi_85_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5239_cast_fp16")]; + tensor xoi_85_cast_fp16 = add(x = var_5238_cast_fp16, y = var_5239_cast_fp16)[name = tensor("xoi_85_cast_fp16")]; + tensor out_85_axis_0 = const()[name = tensor("out_85_axis_0"), val = tensor(-1)]; + tensor out_85_cast_fp16 = stack(axis = out_85_axis_0, values = (xor__85_cast_fp16, xoi_85_cast_fp16))[name = tensor("out_85_cast_fp16")]; + tensor var_5249 = const()[name = tensor("op_5249"), val = tensor([1, 256, 16, 64])]; + tensor q_43_cast_fp16 = reshape(shape = var_5249, x = out_85_cast_fp16)[name = tensor("q_43_cast_fp16")]; + tensor var_5255 = const()[name = tensor("op_5255"), val = tensor([256, 16, 32, 2])]; + tensor pairs_87_cast_fp16 = reshape(shape = var_5255, x = x_87_cast_fp16)[name = tensor("pairs_87_cast_fp16")]; + tensor xr_87_begin_0 = const()[name = tensor("xr_87_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_87_end_0 = const()[name = tensor("xr_87_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_87_end_mask_0 = const()[name = tensor("xr_87_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_87_squeeze_mask_0 = const()[name = tensor("xr_87_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_87_cast_fp16 = slice_by_index(begin = xr_87_begin_0, end = xr_87_end_0, end_mask = xr_87_end_mask_0, squeeze_mask = xr_87_squeeze_mask_0, x = pairs_87_cast_fp16)[name = tensor("xr_87_cast_fp16")]; + tensor xi_87_begin_0 = const()[name = tensor("xi_87_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_87_end_0 = const()[name = tensor("xi_87_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_87_end_mask_0 = const()[name = tensor("xi_87_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_87_squeeze_mask_0 = const()[name = tensor("xi_87_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_87_cast_fp16 = slice_by_index(begin = xi_87_begin_0, end = xi_87_end_0, end_mask = xi_87_end_mask_0, squeeze_mask = xi_87_squeeze_mask_0, x = pairs_87_cast_fp16)[name = tensor("xi_87_cast_fp16")]; + tensor var_5263_cast_fp16 = mul(x = xr_87_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5263_cast_fp16")]; + tensor var_5264_cast_fp16 = mul(x = xi_87_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5264_cast_fp16")]; + tensor xor__87_cast_fp16 = sub(x = var_5263_cast_fp16, y = var_5264_cast_fp16)[name = tensor("xor__87_cast_fp16")]; + tensor var_5267_cast_fp16 = mul(x = xr_87_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5267_cast_fp16")]; + tensor var_5268_cast_fp16 = mul(x = xi_87_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5268_cast_fp16")]; + tensor xoi_87_cast_fp16 = add(x = var_5267_cast_fp16, y = var_5268_cast_fp16)[name = tensor("xoi_87_cast_fp16")]; + tensor out_87_axis_0 = const()[name = tensor("out_87_axis_0"), val = tensor(-1)]; + tensor out_87_cast_fp16 = stack(axis = out_87_axis_0, values = (xor__87_cast_fp16, xoi_87_cast_fp16))[name = tensor("out_87_cast_fp16")]; + tensor var_5285 = const()[name = tensor("op_5285"), val = tensor([256, 1024])]; + tensor var_5286_cast_fp16 = reshape(shape = var_5285, x = out_87_cast_fp16)[name = tensor("op_5286_cast_fp16")]; + tensor written_k_43_transpose_x_1 = const()[name = tensor("written_k_43_transpose_x_1"), val = tensor(true)]; + tensor written_k_43_transpose_y_1 = const()[name = tensor("written_k_43_transpose_y_1"), val = tensor(false)]; + tensor written_k_43_cast_fp16 = matmul(transpose_x = written_k_43_transpose_x_1, transpose_y = written_k_43_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_5286_cast_fp16)[name = tensor("written_k_43_cast_fp16")]; + tensor var_5293 = const()[name = tensor("op_5293"), val = tensor([256, 1024])]; + tensor var_5294_cast_fp16 = reshape(shape = var_5293, x = v_43_cast_fp16)[name = tensor("op_5294_cast_fp16")]; + tensor written_v_43_transpose_x_1 = const()[name = tensor("written_v_43_transpose_x_1"), val = tensor(true)]; + tensor written_v_43_transpose_y_1 = const()[name = tensor("written_v_43_transpose_y_1"), val = tensor(false)]; + tensor written_v_43_cast_fp16 = matmul(transpose_x = written_v_43_transpose_x_1, transpose_y = written_v_43_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_5294_cast_fp16)[name = tensor("written_v_43_cast_fp16")]; + tensor k_cache21_to_fp16_dtype_0 = const()[name = tensor("k_cache21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache21_to_fp16 = cast(dtype = k_cache21_to_fp16_dtype_0, x = k_cache21)[name = tensor("cast_339")]; + tensor var_5305_cast_fp16 = mul(x = k_cache21_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_5305_cast_fp16")]; + tensor var_5310 = const()[name = tensor("op_5310"), val = tensor([1, 512, 16, 64])]; + tensor var_5311_cast_fp16 = reshape(shape = var_5310, x = written_k_43_cast_fp16)[name = tensor("op_5311_cast_fp16")]; + tensor new_k_cache_43_cast_fp16 = add(x = var_5305_cast_fp16, y = var_5311_cast_fp16)[name = tensor("new_k_cache_43_cast_fp16")]; + tensor new_k_cache_43_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_43_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache21_to_fp16_dtype_0 = const()[name = tensor("v_cache21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache21_to_fp16 = cast(dtype = v_cache21_to_fp16_dtype_0, x = v_cache21)[name = tensor("cast_337")]; + tensor var_5314_cast_fp16 = mul(x = v_cache21_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_5314_cast_fp16")]; + tensor var_5319 = const()[name = tensor("op_5319"), val = tensor([1, 512, 16, 64])]; + tensor var_5320_cast_fp16 = reshape(shape = var_5319, x = written_v_43_cast_fp16)[name = tensor("op_5320_cast_fp16")]; + tensor new_v_cache_43_cast_fp16 = add(x = var_5314_cast_fp16, y = var_5320_cast_fp16)[name = tensor("new_v_cache_43_cast_fp16")]; + tensor new_v_cache_43_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_43_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_5333 = const()[name = tensor("op_5333"), val = tensor([0, 2, 1, 3])]; + tensor var_5344_transpose_x_0 = const()[name = tensor("op_5344_transpose_x_0"), val = tensor(false)]; + tensor var_5344_transpose_y_0 = const()[name = tensor("op_5344_transpose_y_0"), val = tensor(false)]; + tensor transpose_111_perm_0 = const()[name = tensor("transpose_111_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_112_perm_0 = const()[name = tensor("transpose_112_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_112 = transpose(perm = transpose_112_perm_0, x = new_k_cache_43_cast_fp16)[name = tensor("transpose_120")]; + tensor transpose_111 = transpose(perm = transpose_111_perm_0, x = q_43_cast_fp16)[name = tensor("transpose_121")]; + tensor var_5344_cast_fp16 = matmul(transpose_x = var_5344_transpose_x_0, transpose_y = var_5344_transpose_y_0, x = transpose_111, y = transpose_112)[name = tensor("op_5344_cast_fp16")]; + tensor var_5345_to_fp16 = const()[name = tensor("op_5345_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_85_cast_fp16 = mul(x = var_5344_cast_fp16, y = var_5345_to_fp16)[name = tensor("scores_85_cast_fp16")]; + tensor scores_87_cast_fp16 = add(x = scores_85_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_87_cast_fp16")]; + tensor var_5355 = const()[name = tensor("op_5355"), val = tensor(-1)]; + tensor weights_43_cast_fp16 = softmax(axis = var_5355, x = scores_87_cast_fp16)[name = tensor("weights_43_cast_fp16")]; + tensor attn_43_transpose_x_0 = const()[name = tensor("attn_43_transpose_x_0"), val = tensor(false)]; + tensor attn_43_transpose_y_0 = const()[name = tensor("attn_43_transpose_y_0"), val = tensor(false)]; + tensor values_43_cast_fp16 = transpose(perm = var_5333, x = new_v_cache_43_cast_fp16)[name = tensor("transpose_122")]; + tensor attn_43_cast_fp16 = matmul(transpose_x = attn_43_transpose_x_0, transpose_y = attn_43_transpose_y_0, x = weights_43_cast_fp16, y = values_43_cast_fp16)[name = tensor("attn_43_cast_fp16")]; + tensor var_5363 = const()[name = tensor("op_5363"), val = tensor([0, 2, 1, 3])]; + tensor var_5368 = const()[name = tensor("op_5368"), val = tensor([1, 256, 1024])]; + tensor var_5364_cast_fp16 = transpose(perm = var_5363, x = attn_43_cast_fp16)[name = tensor("transpose_119")]; + tensor input_257_cast_fp16 = reshape(shape = var_5368, x = var_5364_cast_fp16)[name = tensor("input_257_cast_fp16")]; + tensor attn21_out_proj_weight_to_fp16 = const()[name = tensor("attn21_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(534979840)))]; + tensor linear_85_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn21_out_proj_weight_to_fp16, x = input_257_cast_fp16)[name = tensor("linear_85_cast_fp16")]; + tensor position21_to_fp16_dtype_0 = const()[name = tensor("position21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position21_to_fp16 = cast(dtype = position21_to_fp16_dtype_0, x = position21)[name = tensor("cast_335")]; + tensor var_5377_cast_fp16 = add(x = position21_to_fp16, y = valid_len_to_fp16)[name = tensor("op_5377_cast_fp16")]; + tensor var_5377_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5377_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_259_cast_fp16 = add(x = input_253_cast_fp16, y = linear_85_cast_fp16)[name = tensor("input_259_cast_fp16")]; + tensor input_261_axes_0 = const()[name = tensor("input_261_axes_0"), val = tensor([-1])]; + tensor norm21_2_weight_to_fp16 = const()[name = tensor("norm21_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(537077056)))]; + tensor norm21_2_bias_to_fp16 = const()[name = tensor("norm21_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(537079168)))]; + tensor var_5381_to_fp16 = const()[name = tensor("op_5381_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_261_cast_fp16 = layer_norm(axes = input_261_axes_0, beta = norm21_2_bias_to_fp16, epsilon = var_5381_to_fp16, gamma = norm21_2_weight_to_fp16, x = input_259_cast_fp16)[name = tensor("input_261_cast_fp16")]; + tensor linear21_1_weight_to_fp16 = const()[name = tensor("linear21_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(537081280)))]; + tensor linear_86_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear21_1_weight_to_fp16, x = input_261_cast_fp16)[name = tensor("linear_86_cast_fp16")]; + tensor input_263_mode_0 = const()[name = tensor("input_263_mode_0"), val = tensor("EXACT")]; + tensor input_263_cast_fp16 = gelu(mode = input_263_mode_0, x = linear_86_cast_fp16)[name = tensor("input_263_cast_fp16")]; + tensor linear21_2_weight_to_fp16 = const()[name = tensor("linear21_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(545469952)))]; + tensor linear_87_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear21_2_weight_to_fp16, x = input_263_cast_fp16)[name = tensor("linear_87_cast_fp16")]; + tensor input_265_cast_fp16 = add(x = input_259_cast_fp16, y = linear_87_cast_fp16)[name = tensor("input_265_cast_fp16")]; + tensor input_267_axes_0 = const()[name = tensor("input_267_axes_0"), val = tensor([-1])]; + tensor norm22_1_weight_to_fp16 = const()[name = tensor("norm22_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(553858624)))]; + tensor norm22_1_bias_to_fp16 = const()[name = tensor("norm22_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(553860736)))]; + tensor var_5398_to_fp16 = const()[name = tensor("op_5398_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_267_cast_fp16 = layer_norm(axes = input_267_axes_0, beta = norm22_1_bias_to_fp16, epsilon = var_5398_to_fp16, gamma = norm22_1_weight_to_fp16, x = input_265_cast_fp16)[name = tensor("input_267_cast_fp16")]; + tensor attn22_in_proj_weight_to_fp16 = const()[name = tensor("attn22_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(553862848)))]; + tensor linear_88_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn22_in_proj_weight_to_fp16, x = input_267_cast_fp16)[name = tensor("linear_88_cast_fp16")]; + tensor var_5412 = const()[name = tensor("op_5412"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_45_cast_fp16 = reshape(shape = var_5412, x = linear_88_cast_fp16)[name = tensor("qkv_45_cast_fp16")]; + tensor x_89_begin_0 = const()[name = tensor("x_89_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor x_89_end_0 = const()[name = tensor("x_89_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor x_89_end_mask_0 = const()[name = tensor("x_89_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_89_squeeze_mask_0 = const()[name = tensor("x_89_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_89_cast_fp16 = slice_by_index(begin = x_89_begin_0, end = x_89_end_0, end_mask = x_89_end_mask_0, squeeze_mask = x_89_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("x_89_cast_fp16")]; + tensor x_91_begin_0 = const()[name = tensor("x_91_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_91_end_0 = const()[name = tensor("x_91_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_91_end_mask_0 = const()[name = tensor("x_91_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_91_squeeze_mask_0 = const()[name = tensor("x_91_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_91_cast_fp16 = slice_by_index(begin = x_91_begin_0, end = x_91_end_0, end_mask = x_91_end_mask_0, squeeze_mask = x_91_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("x_91_cast_fp16")]; + tensor v_45_begin_0 = const()[name = tensor("v_45_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_45_end_0 = const()[name = tensor("v_45_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_45_end_mask_0 = const()[name = tensor("v_45_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_45_squeeze_mask_0 = const()[name = tensor("v_45_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_45_cast_fp16 = slice_by_index(begin = v_45_begin_0, end = v_45_end_0, end_mask = v_45_end_mask_0, squeeze_mask = v_45_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("v_45_cast_fp16")]; + tensor var_5457 = const()[name = tensor("op_5457"), val = tensor([256, 16, 32, 2])]; + tensor pairs_89_cast_fp16 = reshape(shape = var_5457, x = x_89_cast_fp16)[name = tensor("pairs_89_cast_fp16")]; + tensor xr_89_begin_0 = const()[name = tensor("xr_89_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_89_end_0 = const()[name = tensor("xr_89_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_89_end_mask_0 = const()[name = tensor("xr_89_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_89_squeeze_mask_0 = const()[name = tensor("xr_89_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_89_cast_fp16 = slice_by_index(begin = xr_89_begin_0, end = xr_89_end_0, end_mask = xr_89_end_mask_0, squeeze_mask = xr_89_squeeze_mask_0, x = pairs_89_cast_fp16)[name = tensor("xr_89_cast_fp16")]; + tensor xi_89_begin_0 = const()[name = tensor("xi_89_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_89_end_0 = const()[name = tensor("xi_89_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_89_end_mask_0 = const()[name = tensor("xi_89_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_89_squeeze_mask_0 = const()[name = tensor("xi_89_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_89_cast_fp16 = slice_by_index(begin = xi_89_begin_0, end = xi_89_end_0, end_mask = xi_89_end_mask_0, squeeze_mask = xi_89_squeeze_mask_0, x = pairs_89_cast_fp16)[name = tensor("xi_89_cast_fp16")]; + tensor var_5465_cast_fp16 = mul(x = xr_89_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5465_cast_fp16")]; + tensor var_5466_cast_fp16 = mul(x = xi_89_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5466_cast_fp16")]; + tensor xor__89_cast_fp16 = sub(x = var_5465_cast_fp16, y = var_5466_cast_fp16)[name = tensor("xor__89_cast_fp16")]; + tensor var_5469_cast_fp16 = mul(x = xr_89_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5469_cast_fp16")]; + tensor var_5470_cast_fp16 = mul(x = xi_89_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5470_cast_fp16")]; + tensor xoi_89_cast_fp16 = add(x = var_5469_cast_fp16, y = var_5470_cast_fp16)[name = tensor("xoi_89_cast_fp16")]; + tensor out_89_axis_0 = const()[name = tensor("out_89_axis_0"), val = tensor(-1)]; + tensor out_89_cast_fp16 = stack(axis = out_89_axis_0, values = (xor__89_cast_fp16, xoi_89_cast_fp16))[name = tensor("out_89_cast_fp16")]; + tensor var_5480 = const()[name = tensor("op_5480"), val = tensor([1, 256, 16, 64])]; + tensor q_45_cast_fp16 = reshape(shape = var_5480, x = out_89_cast_fp16)[name = tensor("q_45_cast_fp16")]; + tensor var_5486 = const()[name = tensor("op_5486"), val = tensor([256, 16, 32, 2])]; + tensor pairs_91_cast_fp16 = reshape(shape = var_5486, x = x_91_cast_fp16)[name = tensor("pairs_91_cast_fp16")]; + tensor xr_91_begin_0 = const()[name = tensor("xr_91_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_91_end_0 = const()[name = tensor("xr_91_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_91_end_mask_0 = const()[name = tensor("xr_91_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_91_squeeze_mask_0 = const()[name = tensor("xr_91_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_91_cast_fp16 = slice_by_index(begin = xr_91_begin_0, end = xr_91_end_0, end_mask = xr_91_end_mask_0, squeeze_mask = xr_91_squeeze_mask_0, x = pairs_91_cast_fp16)[name = tensor("xr_91_cast_fp16")]; + tensor xi_91_begin_0 = const()[name = tensor("xi_91_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_91_end_0 = const()[name = tensor("xi_91_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_91_end_mask_0 = const()[name = tensor("xi_91_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_91_squeeze_mask_0 = const()[name = tensor("xi_91_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_91_cast_fp16 = slice_by_index(begin = xi_91_begin_0, end = xi_91_end_0, end_mask = xi_91_end_mask_0, squeeze_mask = xi_91_squeeze_mask_0, x = pairs_91_cast_fp16)[name = tensor("xi_91_cast_fp16")]; + tensor var_5494_cast_fp16 = mul(x = xr_91_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5494_cast_fp16")]; + tensor var_5495_cast_fp16 = mul(x = xi_91_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5495_cast_fp16")]; + tensor xor__91_cast_fp16 = sub(x = var_5494_cast_fp16, y = var_5495_cast_fp16)[name = tensor("xor__91_cast_fp16")]; + tensor var_5498_cast_fp16 = mul(x = xr_91_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5498_cast_fp16")]; + tensor var_5499_cast_fp16 = mul(x = xi_91_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5499_cast_fp16")]; + tensor xoi_91_cast_fp16 = add(x = var_5498_cast_fp16, y = var_5499_cast_fp16)[name = tensor("xoi_91_cast_fp16")]; + tensor out_91_axis_0 = const()[name = tensor("out_91_axis_0"), val = tensor(-1)]; + tensor out_91_cast_fp16 = stack(axis = out_91_axis_0, values = (xor__91_cast_fp16, xoi_91_cast_fp16))[name = tensor("out_91_cast_fp16")]; + tensor var_5516 = const()[name = tensor("op_5516"), val = tensor([256, 1024])]; + tensor var_5517_cast_fp16 = reshape(shape = var_5516, x = out_91_cast_fp16)[name = tensor("op_5517_cast_fp16")]; + tensor written_k_45_transpose_x_1 = const()[name = tensor("written_k_45_transpose_x_1"), val = tensor(true)]; + tensor written_k_45_transpose_y_1 = const()[name = tensor("written_k_45_transpose_y_1"), val = tensor(false)]; + tensor written_k_45_cast_fp16 = matmul(transpose_x = written_k_45_transpose_x_1, transpose_y = written_k_45_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_5517_cast_fp16)[name = tensor("written_k_45_cast_fp16")]; + tensor var_5524 = const()[name = tensor("op_5524"), val = tensor([256, 1024])]; + tensor var_5525_cast_fp16 = reshape(shape = var_5524, x = v_45_cast_fp16)[name = tensor("op_5525_cast_fp16")]; + tensor written_v_45_transpose_x_1 = const()[name = tensor("written_v_45_transpose_x_1"), val = tensor(true)]; + tensor written_v_45_transpose_y_1 = const()[name = tensor("written_v_45_transpose_y_1"), val = tensor(false)]; + tensor written_v_45_cast_fp16 = matmul(transpose_x = written_v_45_transpose_x_1, transpose_y = written_v_45_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_5525_cast_fp16)[name = tensor("written_v_45_cast_fp16")]; + tensor k_cache22_to_fp16_dtype_0 = const()[name = tensor("k_cache22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache22_to_fp16 = cast(dtype = k_cache22_to_fp16_dtype_0, x = k_cache22)[name = tensor("cast_333")]; + tensor var_5536_cast_fp16 = mul(x = k_cache22_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_5536_cast_fp16")]; + tensor var_5541 = const()[name = tensor("op_5541"), val = tensor([1, 512, 16, 64])]; + tensor var_5542_cast_fp16 = reshape(shape = var_5541, x = written_k_45_cast_fp16)[name = tensor("op_5542_cast_fp16")]; + tensor new_k_cache_45_cast_fp16 = add(x = var_5536_cast_fp16, y = var_5542_cast_fp16)[name = tensor("new_k_cache_45_cast_fp16")]; + tensor new_k_cache_45_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_45_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache22_to_fp16_dtype_0 = const()[name = tensor("v_cache22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache22_to_fp16 = cast(dtype = v_cache22_to_fp16_dtype_0, x = v_cache22)[name = tensor("cast_331")]; + tensor var_5545_cast_fp16 = mul(x = v_cache22_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_5545_cast_fp16")]; + tensor var_5550 = const()[name = tensor("op_5550"), val = tensor([1, 512, 16, 64])]; + tensor var_5551_cast_fp16 = reshape(shape = var_5550, x = written_v_45_cast_fp16)[name = tensor("op_5551_cast_fp16")]; + tensor new_v_cache_45_cast_fp16 = add(x = var_5545_cast_fp16, y = var_5551_cast_fp16)[name = tensor("new_v_cache_45_cast_fp16")]; + tensor new_v_cache_45_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_45_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_5564 = const()[name = tensor("op_5564"), val = tensor([0, 2, 1, 3])]; + tensor var_5575_transpose_x_0 = const()[name = tensor("op_5575_transpose_x_0"), val = tensor(false)]; + tensor var_5575_transpose_y_0 = const()[name = tensor("op_5575_transpose_y_0"), val = tensor(false)]; + tensor transpose_113_perm_0 = const()[name = tensor("transpose_113_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_114_perm_0 = const()[name = tensor("transpose_114_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_114 = transpose(perm = transpose_114_perm_0, x = new_k_cache_45_cast_fp16)[name = tensor("transpose_116")]; + tensor transpose_113 = transpose(perm = transpose_113_perm_0, x = q_45_cast_fp16)[name = tensor("transpose_117")]; + tensor var_5575_cast_fp16 = matmul(transpose_x = var_5575_transpose_x_0, transpose_y = var_5575_transpose_y_0, x = transpose_113, y = transpose_114)[name = tensor("op_5575_cast_fp16")]; + tensor var_5576_to_fp16 = const()[name = tensor("op_5576_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_89_cast_fp16 = mul(x = var_5575_cast_fp16, y = var_5576_to_fp16)[name = tensor("scores_89_cast_fp16")]; + tensor scores_91_cast_fp16 = add(x = scores_89_cast_fp16, y = var_501_cast_fp16)[name = tensor("scores_91_cast_fp16")]; + tensor var_5586 = const()[name = tensor("op_5586"), val = tensor(-1)]; + tensor weights_45_cast_fp16 = softmax(axis = var_5586, x = scores_91_cast_fp16)[name = tensor("weights_45_cast_fp16")]; + tensor attn_45_transpose_x_0 = const()[name = tensor("attn_45_transpose_x_0"), val = tensor(false)]; + tensor attn_45_transpose_y_0 = const()[name = tensor("attn_45_transpose_y_0"), val = tensor(false)]; + tensor values_45_cast_fp16 = transpose(perm = var_5564, x = new_v_cache_45_cast_fp16)[name = tensor("transpose_118")]; + tensor attn_45_cast_fp16 = matmul(transpose_x = attn_45_transpose_x_0, transpose_y = attn_45_transpose_y_0, x = weights_45_cast_fp16, y = values_45_cast_fp16)[name = tensor("attn_45_cast_fp16")]; + tensor var_5594 = const()[name = tensor("op_5594"), val = tensor([0, 2, 1, 3])]; + tensor var_5599 = const()[name = tensor("op_5599"), val = tensor([1, 256, 1024])]; + tensor var_5595_cast_fp16 = transpose(perm = var_5594, x = attn_45_cast_fp16)[name = tensor("transpose_115")]; + tensor input_269_cast_fp16 = reshape(shape = var_5599, x = var_5595_cast_fp16)[name = tensor("input_269_cast_fp16")]; + tensor attn22_out_proj_weight_to_fp16 = const()[name = tensor("attn22_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(560154368)))]; + tensor linear_89_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn22_out_proj_weight_to_fp16, x = input_269_cast_fp16)[name = tensor("linear_89_cast_fp16")]; + tensor position22_to_fp16_dtype_0 = const()[name = tensor("position22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position22_to_fp16 = cast(dtype = position22_to_fp16_dtype_0, x = position22)[name = tensor("cast_329")]; + tensor var_5608_cast_fp16 = add(x = position22_to_fp16, y = valid_len_to_fp16)[name = tensor("op_5608_cast_fp16")]; + tensor var_5608_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5608_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_271_cast_fp16 = add(x = input_265_cast_fp16, y = linear_89_cast_fp16)[name = tensor("input_271_cast_fp16")]; + tensor input_273_axes_0 = const()[name = tensor("input_273_axes_0"), val = tensor([-1])]; + tensor norm22_2_weight_to_fp16 = const()[name = tensor("norm22_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(562251584)))]; + tensor norm22_2_bias_to_fp16 = const()[name = tensor("norm22_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(562253696)))]; + tensor var_5612_to_fp16 = const()[name = tensor("op_5612_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_273_cast_fp16 = layer_norm(axes = input_273_axes_0, beta = norm22_2_bias_to_fp16, epsilon = var_5612_to_fp16, gamma = norm22_2_weight_to_fp16, x = input_271_cast_fp16)[name = tensor("input_273_cast_fp16")]; + tensor linear22_1_weight_to_fp16 = const()[name = tensor("linear22_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(562255808)))]; + tensor linear_90_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear22_1_weight_to_fp16, x = input_273_cast_fp16)[name = tensor("linear_90_cast_fp16")]; + tensor input_275_mode_0 = const()[name = tensor("input_275_mode_0"), val = tensor("EXACT")]; + tensor input_275_cast_fp16 = gelu(mode = input_275_mode_0, x = linear_90_cast_fp16)[name = tensor("input_275_cast_fp16")]; + tensor linear22_2_weight_to_fp16 = const()[name = tensor("linear22_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(570644480)))]; + tensor linear_91_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear22_2_weight_to_fp16, x = input_275_cast_fp16)[name = tensor("linear_91_cast_fp16")]; + tensor input_277_cast_fp16 = add(x = input_271_cast_fp16, y = linear_91_cast_fp16)[name = tensor("input_277_cast_fp16")]; + tensor input_279_axes_0 = const()[name = tensor("input_279_axes_0"), val = tensor([-1])]; + tensor norm23_1_weight_to_fp16 = const()[name = tensor("norm23_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(579033152)))]; + tensor norm23_1_bias_to_fp16 = const()[name = tensor("norm23_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(579035264)))]; + tensor var_5629_to_fp16 = const()[name = tensor("op_5629_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_279_cast_fp16 = layer_norm(axes = input_279_axes_0, beta = norm23_1_bias_to_fp16, epsilon = var_5629_to_fp16, gamma = norm23_1_weight_to_fp16, x = input_277_cast_fp16)[name = tensor("input_279_cast_fp16")]; + tensor attn23_in_proj_weight_to_fp16 = const()[name = tensor("attn23_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(579037376)))]; + tensor linear_92_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn23_in_proj_weight_to_fp16, x = input_279_cast_fp16)[name = tensor("linear_92_cast_fp16")]; + tensor var_5643 = const()[name = tensor("op_5643"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_cast_fp16 = reshape(shape = var_5643, x = linear_92_cast_fp16)[name = tensor("qkv_cast_fp16")]; + tensor x_begin_0 = const()[name = tensor("x_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor x_end_0 = const()[name = tensor("x_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor x_end_mask_0 = const()[name = tensor("x_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor x_squeeze_mask_0 = const()[name = tensor("x_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor x_cast_fp16 = slice_by_index(begin = x_begin_0, end = x_end_0, end_mask = x_end_mask_0, squeeze_mask = x_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("x_cast_fp16")]; + tensor v_begin_0 = const()[name = tensor("v_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_end_0 = const()[name = tensor("v_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_end_mask_0 = const()[name = tensor("v_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_squeeze_mask_0 = const()[name = tensor("v_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_cast_fp16 = slice_by_index(begin = v_begin_0, end = v_end_0, end_mask = v_end_mask_0, squeeze_mask = v_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("v_cast_fp16")]; + tensor var_5675 = const()[name = tensor("op_5675"), val = tensor([256, 16, 32, 2])]; + tensor pairs_cast_fp16 = reshape(shape = var_5675, x = x_cast_fp16)[name = tensor("pairs_cast_fp16")]; + tensor xr_begin_0 = const()[name = tensor("xr_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_end_0 = const()[name = tensor("xr_end_0"), val = tensor([256, 16, 32, 1])]; + tensor xr_end_mask_0 = const()[name = tensor("xr_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_squeeze_mask_0 = const()[name = tensor("xr_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_cast_fp16 = slice_by_index(begin = xr_begin_0, end = xr_end_0, end_mask = xr_end_mask_0, squeeze_mask = xr_squeeze_mask_0, x = pairs_cast_fp16)[name = tensor("xr_cast_fp16")]; + tensor xi_begin_0 = const()[name = tensor("xi_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_end_0 = const()[name = tensor("xi_end_0"), val = tensor([256, 16, 32, 2])]; + tensor xi_end_mask_0 = const()[name = tensor("xi_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_squeeze_mask_0 = const()[name = tensor("xi_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_cast_fp16 = slice_by_index(begin = xi_begin_0, end = xi_end_0, end_mask = xi_end_mask_0, squeeze_mask = xi_squeeze_mask_0, x = pairs_cast_fp16)[name = tensor("xi_cast_fp16")]; + tensor var_5683_cast_fp16 = mul(x = xr_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5683_cast_fp16")]; + tensor var_5684_cast_fp16 = mul(x = xi_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5684_cast_fp16")]; + tensor xor__cast_fp16 = sub(x = var_5683_cast_fp16, y = var_5684_cast_fp16)[name = tensor("xor__cast_fp16")]; + tensor var_5687_cast_fp16 = mul(x = xr_cast_fp16, y = roti_cast_fp16)[name = tensor("op_5687_cast_fp16")]; + tensor var_5688_cast_fp16 = mul(x = xi_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_5688_cast_fp16")]; + tensor xoi_cast_fp16 = add(x = var_5687_cast_fp16, y = var_5688_cast_fp16)[name = tensor("xoi_cast_fp16")]; + tensor out_axis_0 = const()[name = tensor("out_axis_0"), val = tensor(-1)]; + tensor out_cast_fp16 = stack(axis = out_axis_0, values = (xor__cast_fp16, xoi_cast_fp16))[name = tensor("out_cast_fp16")]; + tensor var_5705 = const()[name = tensor("op_5705"), val = tensor([256, 1024])]; + tensor var_5706_cast_fp16 = reshape(shape = var_5705, x = out_cast_fp16)[name = tensor("op_5706_cast_fp16")]; + tensor written_k_transpose_x_1 = const()[name = tensor("written_k_transpose_x_1"), val = tensor(true)]; + tensor written_k_transpose_y_1 = const()[name = tensor("written_k_transpose_y_1"), val = tensor(false)]; + tensor written_k_cast_fp16 = matmul(transpose_x = written_k_transpose_x_1, transpose_y = written_k_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_5706_cast_fp16)[name = tensor("written_k_cast_fp16")]; + tensor var_5713 = const()[name = tensor("op_5713"), val = tensor([256, 1024])]; + tensor var_5714_cast_fp16 = reshape(shape = var_5713, x = v_cast_fp16)[name = tensor("op_5714_cast_fp16")]; + tensor written_v_transpose_x_1 = const()[name = tensor("written_v_transpose_x_1"), val = tensor(true)]; + tensor written_v_transpose_y_1 = const()[name = tensor("written_v_transpose_y_1"), val = tensor(false)]; + tensor written_v_cast_fp16 = matmul(transpose_x = written_v_transpose_x_1, transpose_y = written_v_transpose_y_1, x = var_268_cast_fp16_to_fp16, y = var_5714_cast_fp16)[name = tensor("written_v_cast_fp16")]; + tensor k_cache23_to_fp16_dtype_0 = const()[name = tensor("k_cache23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache23_to_fp16 = cast(dtype = k_cache23_to_fp16_dtype_0, x = k_cache23)[name = tensor("cast_327")]; + tensor var_5725_cast_fp16 = mul(x = k_cache23_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_5725_cast_fp16")]; + tensor var_5730 = const()[name = tensor("op_5730"), val = tensor([1, 512, 16, 64])]; + tensor var_5731_cast_fp16 = reshape(shape = var_5730, x = written_k_cast_fp16)[name = tensor("op_5731_cast_fp16")]; + tensor new_k_cache_cast_fp16 = add(x = var_5725_cast_fp16, y = var_5731_cast_fp16)[name = tensor("new_k_cache_cast_fp16")]; + tensor new_k_cache_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache23_to_fp16_dtype_0 = const()[name = tensor("v_cache23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache23_to_fp16 = cast(dtype = v_cache23_to_fp16_dtype_0, x = v_cache23)[name = tensor("cast_325")]; + tensor var_5734_cast_fp16 = mul(x = v_cache23_to_fp16, y = keep_1_cast_fp16)[name = tensor("op_5734_cast_fp16")]; + tensor var_5739 = const()[name = tensor("op_5739"), val = tensor([1, 512, 16, 64])]; + tensor var_5740_cast_fp16 = reshape(shape = var_5739, x = written_v_cast_fp16)[name = tensor("op_5740_cast_fp16")]; + tensor new_v_cache_cast_fp16 = add(x = var_5734_cast_fp16, y = var_5740_cast_fp16)[name = tensor("new_v_cache_cast_fp16")]; + tensor new_v_cache_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor position23_to_fp16_dtype_0 = const()[name = tensor("position23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position23_to_fp16 = cast(dtype = position23_to_fp16_dtype_0, x = position23)[name = tensor("cast_323")]; + tensor var_5747_cast_fp16 = add(x = position23_to_fp16, y = valid_len_to_fp16)[name = tensor("op_5747_cast_fp16")]; + tensor var_5747_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5747_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor new_position23 = cast(dtype = var_5747_cast_fp16_to_fp32_dtype_0, x = var_5747_cast_fp16)[name = tensor("cast_322")]; + tensor new_v_cache23 = cast(dtype = new_v_cache_cast_fp16_to_fp32_dtype_0, x = new_v_cache_cast_fp16)[name = tensor("cast_324")]; + tensor new_k_cache23 = cast(dtype = new_k_cache_cast_fp16_to_fp32_dtype_0, x = new_k_cache_cast_fp16)[name = tensor("cast_326")]; + tensor new_position22 = cast(dtype = var_5608_cast_fp16_to_fp32_dtype_0, x = var_5608_cast_fp16)[name = tensor("cast_328")]; + tensor new_v_cache22 = cast(dtype = new_v_cache_45_cast_fp16_to_fp32_dtype_0, x = new_v_cache_45_cast_fp16)[name = tensor("cast_330")]; + tensor new_k_cache22 = cast(dtype = new_k_cache_45_cast_fp16_to_fp32_dtype_0, x = new_k_cache_45_cast_fp16)[name = tensor("cast_332")]; + tensor new_position21 = cast(dtype = var_5377_cast_fp16_to_fp32_dtype_0, x = var_5377_cast_fp16)[name = tensor("cast_334")]; + tensor new_v_cache21 = cast(dtype = new_v_cache_43_cast_fp16_to_fp32_dtype_0, x = new_v_cache_43_cast_fp16)[name = tensor("cast_336")]; + tensor new_k_cache21 = cast(dtype = new_k_cache_43_cast_fp16_to_fp32_dtype_0, x = new_k_cache_43_cast_fp16)[name = tensor("cast_338")]; + tensor new_position20 = cast(dtype = var_5146_cast_fp16_to_fp32_dtype_0, x = var_5146_cast_fp16)[name = tensor("cast_340")]; + tensor new_v_cache20 = cast(dtype = new_v_cache_41_cast_fp16_to_fp32_dtype_0, x = new_v_cache_41_cast_fp16)[name = tensor("cast_342")]; + tensor new_k_cache20 = cast(dtype = new_k_cache_41_cast_fp16_to_fp32_dtype_0, x = new_k_cache_41_cast_fp16)[name = tensor("cast_344")]; + tensor new_position19 = cast(dtype = var_4915_cast_fp16_to_fp32_dtype_0, x = var_4915_cast_fp16)[name = tensor("cast_346")]; + tensor new_v_cache19 = cast(dtype = new_v_cache_39_cast_fp16_to_fp32_dtype_0, x = new_v_cache_39_cast_fp16)[name = tensor("cast_348")]; + tensor new_k_cache19 = cast(dtype = new_k_cache_39_cast_fp16_to_fp32_dtype_0, x = new_k_cache_39_cast_fp16)[name = tensor("cast_350")]; + tensor new_position18 = cast(dtype = var_4684_cast_fp16_to_fp32_dtype_0, x = var_4684_cast_fp16)[name = tensor("cast_352")]; + tensor new_v_cache18 = cast(dtype = new_v_cache_37_cast_fp16_to_fp32_dtype_0, x = new_v_cache_37_cast_fp16)[name = tensor("cast_354")]; + tensor new_k_cache18 = cast(dtype = new_k_cache_37_cast_fp16_to_fp32_dtype_0, x = new_k_cache_37_cast_fp16)[name = tensor("cast_356")]; + tensor new_position17 = cast(dtype = var_4453_cast_fp16_to_fp32_dtype_0, x = var_4453_cast_fp16)[name = tensor("cast_358")]; + tensor new_v_cache17 = cast(dtype = new_v_cache_35_cast_fp16_to_fp32_dtype_0, x = new_v_cache_35_cast_fp16)[name = tensor("cast_360")]; + tensor new_k_cache17 = cast(dtype = new_k_cache_35_cast_fp16_to_fp32_dtype_0, x = new_k_cache_35_cast_fp16)[name = tensor("cast_362")]; + tensor new_position16 = cast(dtype = var_4222_cast_fp16_to_fp32_dtype_0, x = var_4222_cast_fp16)[name = tensor("cast_364")]; + tensor new_v_cache16 = cast(dtype = new_v_cache_33_cast_fp16_to_fp32_dtype_0, x = new_v_cache_33_cast_fp16)[name = tensor("cast_366")]; + tensor new_k_cache16 = cast(dtype = new_k_cache_33_cast_fp16_to_fp32_dtype_0, x = new_k_cache_33_cast_fp16)[name = tensor("cast_368")]; + tensor new_position15 = cast(dtype = var_3991_cast_fp16_to_fp32_dtype_0, x = var_3991_cast_fp16)[name = tensor("cast_370")]; + tensor new_v_cache15 = cast(dtype = new_v_cache_31_cast_fp16_to_fp32_dtype_0, x = new_v_cache_31_cast_fp16)[name = tensor("cast_372")]; + tensor new_k_cache15 = cast(dtype = new_k_cache_31_cast_fp16_to_fp32_dtype_0, x = new_k_cache_31_cast_fp16)[name = tensor("cast_374")]; + tensor new_position14 = cast(dtype = var_3760_cast_fp16_to_fp32_dtype_0, x = var_3760_cast_fp16)[name = tensor("cast_376")]; + tensor new_v_cache14 = cast(dtype = new_v_cache_29_cast_fp16_to_fp32_dtype_0, x = new_v_cache_29_cast_fp16)[name = tensor("cast_378")]; + tensor new_k_cache14 = cast(dtype = new_k_cache_29_cast_fp16_to_fp32_dtype_0, x = new_k_cache_29_cast_fp16)[name = tensor("cast_380")]; + tensor new_position13 = cast(dtype = var_3529_cast_fp16_to_fp32_dtype_0, x = var_3529_cast_fp16)[name = tensor("cast_382")]; + tensor new_v_cache13 = cast(dtype = new_v_cache_27_cast_fp16_to_fp32_dtype_0, x = new_v_cache_27_cast_fp16)[name = tensor("cast_384")]; + tensor new_k_cache13 = cast(dtype = new_k_cache_27_cast_fp16_to_fp32_dtype_0, x = new_k_cache_27_cast_fp16)[name = tensor("cast_386")]; + tensor new_position12 = cast(dtype = var_3298_cast_fp16_to_fp32_dtype_0, x = var_3298_cast_fp16)[name = tensor("cast_388")]; + tensor new_v_cache12 = cast(dtype = new_v_cache_25_cast_fp16_to_fp32_dtype_0, x = new_v_cache_25_cast_fp16)[name = tensor("cast_390")]; + tensor new_k_cache12 = cast(dtype = new_k_cache_25_cast_fp16_to_fp32_dtype_0, x = new_k_cache_25_cast_fp16)[name = tensor("cast_392")]; + tensor new_position11 = cast(dtype = var_3067_cast_fp16_to_fp32_dtype_0, x = var_3067_cast_fp16)[name = tensor("cast_394")]; + tensor new_v_cache11 = cast(dtype = new_v_cache_23_cast_fp16_to_fp32_dtype_0, x = new_v_cache_23_cast_fp16)[name = tensor("cast_396")]; + tensor new_k_cache11 = cast(dtype = new_k_cache_23_cast_fp16_to_fp32_dtype_0, x = new_k_cache_23_cast_fp16)[name = tensor("cast_398")]; + tensor new_position10 = cast(dtype = var_2836_cast_fp16_to_fp32_dtype_0, x = var_2836_cast_fp16)[name = tensor("cast_400")]; + tensor new_v_cache10 = cast(dtype = new_v_cache_21_cast_fp16_to_fp32_dtype_0, x = new_v_cache_21_cast_fp16)[name = tensor("cast_402")]; + tensor new_k_cache10 = cast(dtype = new_k_cache_21_cast_fp16_to_fp32_dtype_0, x = new_k_cache_21_cast_fp16)[name = tensor("cast_404")]; + tensor new_position9 = cast(dtype = var_2605_cast_fp16_to_fp32_dtype_0, x = var_2605_cast_fp16)[name = tensor("cast_406")]; + tensor new_v_cache9 = cast(dtype = new_v_cache_19_cast_fp16_to_fp32_dtype_0, x = new_v_cache_19_cast_fp16)[name = tensor("cast_408")]; + tensor new_k_cache9 = cast(dtype = new_k_cache_19_cast_fp16_to_fp32_dtype_0, x = new_k_cache_19_cast_fp16)[name = tensor("cast_410")]; + tensor new_position8 = cast(dtype = var_2374_cast_fp16_to_fp32_dtype_0, x = var_2374_cast_fp16)[name = tensor("cast_412")]; + tensor new_v_cache8 = cast(dtype = new_v_cache_17_cast_fp16_to_fp32_dtype_0, x = new_v_cache_17_cast_fp16)[name = tensor("cast_414")]; + tensor new_k_cache8 = cast(dtype = new_k_cache_17_cast_fp16_to_fp32_dtype_0, x = new_k_cache_17_cast_fp16)[name = tensor("cast_416")]; + tensor new_position7 = cast(dtype = var_2143_cast_fp16_to_fp32_dtype_0, x = var_2143_cast_fp16)[name = tensor("cast_418")]; + tensor new_v_cache7 = cast(dtype = new_v_cache_15_cast_fp16_to_fp32_dtype_0, x = new_v_cache_15_cast_fp16)[name = tensor("cast_420")]; + tensor new_k_cache7 = cast(dtype = new_k_cache_15_cast_fp16_to_fp32_dtype_0, x = new_k_cache_15_cast_fp16)[name = tensor("cast_422")]; + tensor new_position6 = cast(dtype = var_1912_cast_fp16_to_fp32_dtype_0, x = var_1912_cast_fp16)[name = tensor("cast_424")]; + tensor new_v_cache6 = cast(dtype = new_v_cache_13_cast_fp16_to_fp32_dtype_0, x = new_v_cache_13_cast_fp16)[name = tensor("cast_426")]; + tensor new_k_cache6 = cast(dtype = new_k_cache_13_cast_fp16_to_fp32_dtype_0, x = new_k_cache_13_cast_fp16)[name = tensor("cast_428")]; + tensor new_position5 = cast(dtype = var_1681_cast_fp16_to_fp32_dtype_0, x = var_1681_cast_fp16)[name = tensor("cast_430")]; + tensor new_v_cache5 = cast(dtype = new_v_cache_11_cast_fp16_to_fp32_dtype_0, x = new_v_cache_11_cast_fp16)[name = tensor("cast_432")]; + tensor new_k_cache5 = cast(dtype = new_k_cache_11_cast_fp16_to_fp32_dtype_0, x = new_k_cache_11_cast_fp16)[name = tensor("cast_434")]; + tensor new_position4 = cast(dtype = var_1450_cast_fp16_to_fp32_dtype_0, x = var_1450_cast_fp16)[name = tensor("cast_436")]; + tensor new_v_cache4 = cast(dtype = new_v_cache_9_cast_fp16_to_fp32_dtype_0, x = new_v_cache_9_cast_fp16)[name = tensor("cast_438")]; + tensor new_k_cache4 = cast(dtype = new_k_cache_9_cast_fp16_to_fp32_dtype_0, x = new_k_cache_9_cast_fp16)[name = tensor("cast_440")]; + tensor new_position3 = cast(dtype = var_1219_cast_fp16_to_fp32_dtype_0, x = var_1219_cast_fp16)[name = tensor("cast_442")]; + tensor new_v_cache3 = cast(dtype = new_v_cache_7_cast_fp16_to_fp32_dtype_0, x = new_v_cache_7_cast_fp16)[name = tensor("cast_444")]; + tensor new_k_cache3 = cast(dtype = new_k_cache_7_cast_fp16_to_fp32_dtype_0, x = new_k_cache_7_cast_fp16)[name = tensor("cast_446")]; + tensor new_position2 = cast(dtype = var_988_cast_fp16_to_fp32_dtype_0, x = var_988_cast_fp16)[name = tensor("cast_448")]; + tensor new_v_cache2 = cast(dtype = new_v_cache_5_cast_fp16_to_fp32_dtype_0, x = new_v_cache_5_cast_fp16)[name = tensor("cast_450")]; + tensor new_k_cache2 = cast(dtype = new_k_cache_5_cast_fp16_to_fp32_dtype_0, x = new_k_cache_5_cast_fp16)[name = tensor("cast_452")]; + tensor new_position1 = cast(dtype = var_757_cast_fp16_to_fp32_dtype_0, x = var_757_cast_fp16)[name = tensor("cast_454")]; + tensor new_v_cache1 = cast(dtype = new_v_cache_3_cast_fp16_to_fp32_dtype_0, x = new_v_cache_3_cast_fp16)[name = tensor("cast_456")]; + tensor new_k_cache1 = cast(dtype = new_k_cache_3_cast_fp16_to_fp32_dtype_0, x = new_k_cache_3_cast_fp16)[name = tensor("cast_458")]; + tensor new_position0 = cast(dtype = var_526_cast_fp16_to_fp32_dtype_0, x = var_297_cast_fp16)[name = tensor("cast_460")]; + tensor new_v_cache0 = cast(dtype = new_v_cache_1_cast_fp16_to_fp32_dtype_0, x = new_v_cache_1_cast_fp16)[name = tensor("cast_461")]; + tensor new_k_cache0 = cast(dtype = new_k_cache_1_cast_fp16_to_fp32_dtype_0, x = new_k_cache_1_cast_fp16)[name = tensor("cast_463")]; + } -> (new_k_cache0, new_v_cache0, new_position0, new_k_cache1, new_v_cache1, new_position1, new_k_cache2, new_v_cache2, new_position2, new_k_cache3, new_v_cache3, new_position3, new_k_cache4, new_v_cache4, new_position4, new_k_cache5, new_v_cache5, new_position5, new_k_cache6, new_v_cache6, new_position6, new_k_cache7, new_v_cache7, new_position7, new_k_cache8, new_v_cache8, new_position8, new_k_cache9, new_v_cache9, new_position9, new_k_cache10, new_v_cache10, new_position10, new_k_cache11, new_v_cache11, new_position11, new_k_cache12, new_v_cache12, new_position12, new_k_cache13, new_v_cache13, new_position13, new_k_cache14, new_v_cache14, new_position14, new_k_cache15, new_v_cache15, new_position15, new_k_cache16, new_v_cache16, new_position16, new_k_cache17, new_v_cache17, new_position17, new_k_cache18, new_v_cache18, new_position18, new_k_cache19, new_v_cache19, new_position19, new_k_cache20, new_v_cache20, new_position20, new_k_cache21, new_v_cache21, new_position21, new_k_cache22, new_v_cache22, new_position22, new_k_cache23, new_v_cache23, new_position23); +} \ No newline at end of file diff --git a/v2.1/spanish_24l/cond_prefill_ane.mlmodelc/weights/weight.bin b/v2.1/spanish_24l/cond_prefill_ane.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..c81dd765be39dcc0bee08d025289ee4ac59c5573 --- /dev/null +++ b/v2.1/spanish_24l/cond_prefill_ane.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ac6c67e9965ee2008e783b3bcec65ed880688e5124a81a9b4de03ee86c393272 +size 585328896 diff --git a/v2.1/spanish_24l/cond_prefill_ane.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/spanish_24l/cond_prefill_ane.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..4d555cf18a5a848178c413eeab365e9a8a4b275c --- /dev/null +++ b/v2.1/spanish_24l/cond_prefill_ane.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:943f4e1037ba5c9d1cb987344c33765107acdd80a22cd8f98dff317bc0a2dc13 +size 483717 diff --git a/v2.1/spanish_24l/cond_prefill_ane.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/spanish_24l/cond_prefill_ane.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..c81dd765be39dcc0bee08d025289ee4ac59c5573 --- /dev/null +++ b/v2.1/spanish_24l/cond_prefill_ane.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ac6c67e9965ee2008e783b3bcec65ed880688e5124a81a9b4de03ee86c393272 +size 585328896 diff --git a/v2.1/spanish_24l/cond_prefill_ane.mlpackage/Manifest.json b/v2.1/spanish_24l/cond_prefill_ane.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..1f6d9ee69cce834636a4302819d1a4f4e6a0663c --- /dev/null +++ b/v2.1/spanish_24l/cond_prefill_ane.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "219A2294-4D11-4BDD-9D63-1F1C60921830": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "E40BAC90-AF7D-4AC7-A017-FE2E32C42A9C": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "219A2294-4D11-4BDD-9D63-1F1C60921830" +} diff --git a/v2.1/spanish_24l/flowlm_step_ane.mlmodelc/analytics/coremldata.bin b/v2.1/spanish_24l/flowlm_step_ane.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..096d644c6181bc02777b78d0d13ee8239345e5d0 --- /dev/null +++ b/v2.1/spanish_24l/flowlm_step_ane.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a85ca25745cd6ebb968b1ed0f34e482d95e02e3939077ba4aa8fafade8f953b1 +size 243 diff --git a/v2.1/spanish_24l/flowlm_step_ane.mlmodelc/coremldata.bin b/v2.1/spanish_24l/flowlm_step_ane.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..4a8840ec36945a06c186518f4def322fa257a522 --- /dev/null +++ b/v2.1/spanish_24l/flowlm_step_ane.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5e86b2bd9f2f24deb72d21c67c187b0894cf0b531176a12063860015af0c2656 +size 4522 diff --git a/v2.1/spanish_24l/flowlm_step_ane.mlmodelc/model.mil b/v2.1/spanish_24l/flowlm_step_ane.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..7d9834ef71775a1a27dc2e99881550438732aad7 --- /dev/null +++ b/v2.1/spanish_24l/flowlm_step_ane.mlmodelc/model.mil @@ -0,0 +1,3699 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.9.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor k_cache0, tensor k_cache1, tensor k_cache10, tensor k_cache11, tensor k_cache12, tensor k_cache13, tensor k_cache14, tensor k_cache15, tensor k_cache16, tensor k_cache17, tensor k_cache18, tensor k_cache19, tensor k_cache2, tensor k_cache20, tensor k_cache21, tensor k_cache22, tensor k_cache23, tensor k_cache3, tensor k_cache4, tensor k_cache5, tensor k_cache6, tensor k_cache7, tensor k_cache8, tensor k_cache9, tensor position0, tensor position1, tensor position10, tensor position11, tensor position12, tensor position13, tensor position14, tensor position15, tensor position16, tensor position17, tensor position18, tensor position19, tensor position2, tensor position20, tensor position21, tensor position22, tensor position23, tensor position3, tensor position4, tensor position5, tensor position6, tensor position7, tensor position8, tensor position9, tensor sequence, tensor v_cache0, tensor v_cache1, tensor v_cache10, tensor v_cache11, tensor v_cache12, tensor v_cache13, tensor v_cache14, tensor v_cache15, tensor v_cache16, tensor v_cache17, tensor v_cache18, tensor v_cache19, tensor v_cache2, tensor v_cache20, tensor v_cache21, tensor v_cache22, tensor v_cache23, tensor v_cache3, tensor v_cache4, tensor v_cache5, tensor v_cache6, tensor v_cache7, tensor v_cache8, tensor v_cache9) { + tensor sequence_to_fp16_dtype_0 = const()[name = tensor("sequence_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor input_linear_weight_to_fp16 = const()[name = tensor("input_linear_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(65664)))]; + tensor sequence_to_fp16 = cast(dtype = sequence_to_fp16_dtype_0, x = sequence)[name = tensor("cast_578")]; + tensor linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = input_linear_weight_to_fp16, x = sequence_to_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; + tensor norm0_1_weight_to_fp16 = const()[name = tensor("norm0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67776)))]; + tensor norm0_1_bias_to_fp16 = const()[name = tensor("norm0_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(69888)))]; + tensor var_227_to_fp16 = const()[name = tensor("op_227_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_5_cast_fp16 = layer_norm(axes = input_5_axes_0, beta = norm0_1_bias_to_fp16, epsilon = var_227_to_fp16, gamma = norm0_1_weight_to_fp16, x = linear_0_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor attn0_in_proj_weight_to_fp16 = const()[name = tensor("attn0_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(72000)))]; + tensor linear_1_bias_0_to_fp16 = const()[name = tensor("linear_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6363520)))]; + tensor linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn0_in_proj_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("linear_1_cast_fp16")]; + tensor var_240 = const()[name = tensor("op_240"), val = tensor([1, 3, 16, 64])]; + tensor qkv_1_cast_fp16 = reshape(shape = var_240, x = linear_1_cast_fp16)[name = tensor("qkv_1_cast_fp16")]; + tensor x_1_begin_0 = const()[name = tensor("x_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_1_end_0 = const()[name = tensor("x_1_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_1_end_mask_0 = const()[name = tensor("x_1_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_1_squeeze_mask_0 = const()[name = tensor("x_1_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_1_cast_fp16 = slice_by_index(begin = x_1_begin_0, end = x_1_end_0, end_mask = x_1_end_mask_0, squeeze_mask = x_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("x_1_cast_fp16")]; + tensor x_3_begin_0 = const()[name = tensor("x_3_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_3_end_0 = const()[name = tensor("x_3_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_3_end_mask_0 = const()[name = tensor("x_3_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_3_squeeze_mask_0 = const()[name = tensor("x_3_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_3_cast_fp16 = slice_by_index(begin = x_3_begin_0, end = x_3_end_0, end_mask = x_3_end_mask_0, squeeze_mask = x_3_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("x_3_cast_fp16")]; + tensor v_1_begin_0 = const()[name = tensor("v_1_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_1_end_0 = const()[name = tensor("v_1_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_1_end_mask_0 = const()[name = tensor("v_1_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_1_squeeze_mask_0 = const()[name = tensor("v_1_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_1_cast_fp16 = slice_by_index(begin = v_1_begin_0, end = v_1_end_0, end_mask = v_1_end_mask_0, squeeze_mask = v_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("v_1_cast_fp16")]; + tensor var_269 = const()[name = tensor("op_269"), val = tensor([1, 1, 1])]; + tensor position0_to_fp16_dtype_0 = const()[name = tensor("position0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position0_to_fp16 = cast(dtype = position0_to_fp16_dtype_0, x = position0)[name = tensor("cast_577")]; + tensor var_270_cast_fp16 = reshape(shape = var_269, x = position0_to_fp16)[name = tensor("op_270_cast_fp16")]; + tensor rope_freqs_to_fp16 = const()[name = tensor("rope_freqs_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6369728)))]; + tensor angles_1_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_270_cast_fp16)[name = tensor("angles_1_cast_fp16")]; + tensor rotr_1_cast_fp16 = cos(x = angles_1_cast_fp16)[name = tensor("rotr_1_cast_fp16")]; + tensor roti_1_cast_fp16 = sin(x = angles_1_cast_fp16)[name = tensor("roti_1_cast_fp16")]; + tensor var_278 = const()[name = tensor("op_278"), val = tensor([1, 16, 32, 2])]; + tensor pairs_1_cast_fp16 = reshape(shape = var_278, x = x_1_cast_fp16)[name = tensor("pairs_1_cast_fp16")]; + tensor xr_1_begin_0 = const()[name = tensor("xr_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_1_end_0 = const()[name = tensor("xr_1_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_1_end_mask_0 = const()[name = tensor("xr_1_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_1_squeeze_mask_0 = const()[name = tensor("xr_1_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_1_cast_fp16 = slice_by_index(begin = xr_1_begin_0, end = xr_1_end_0, end_mask = xr_1_end_mask_0, squeeze_mask = xr_1_squeeze_mask_0, x = pairs_1_cast_fp16)[name = tensor("xr_1_cast_fp16")]; + tensor xi_1_begin_0 = const()[name = tensor("xi_1_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_1_end_0 = const()[name = tensor("xi_1_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_1_end_mask_0 = const()[name = tensor("xi_1_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_1_squeeze_mask_0 = const()[name = tensor("xi_1_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_1_cast_fp16 = slice_by_index(begin = xi_1_begin_0, end = xi_1_end_0, end_mask = xi_1_end_mask_0, squeeze_mask = xi_1_squeeze_mask_0, x = pairs_1_cast_fp16)[name = tensor("xi_1_cast_fp16")]; + tensor var_286_cast_fp16 = mul(x = xr_1_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_286_cast_fp16")]; + tensor var_287_cast_fp16 = mul(x = xi_1_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_287_cast_fp16")]; + tensor xor__1_cast_fp16 = sub(x = var_286_cast_fp16, y = var_287_cast_fp16)[name = tensor("xor__1_cast_fp16")]; + tensor var_290_cast_fp16 = mul(x = xr_1_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_290_cast_fp16")]; + tensor var_291_cast_fp16 = mul(x = xi_1_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_291_cast_fp16")]; + tensor xoi_1_cast_fp16 = add(x = var_290_cast_fp16, y = var_291_cast_fp16)[name = tensor("xoi_1_cast_fp16")]; + tensor out_1_axis_0 = const()[name = tensor("out_1_axis_0"), val = tensor(-1)]; + tensor out_1_cast_fp16 = stack(axis = out_1_axis_0, values = (xor__1_cast_fp16, xoi_1_cast_fp16))[name = tensor("out_1_cast_fp16")]; + tensor var_314 = const()[name = tensor("op_314"), val = tensor([1, 16, 32, 2])]; + tensor pairs_3_cast_fp16 = reshape(shape = var_314, x = x_3_cast_fp16)[name = tensor("pairs_3_cast_fp16")]; + tensor xr_3_begin_0 = const()[name = tensor("xr_3_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_3_end_0 = const()[name = tensor("xr_3_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_3_end_mask_0 = const()[name = tensor("xr_3_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_3_squeeze_mask_0 = const()[name = tensor("xr_3_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_3_cast_fp16 = slice_by_index(begin = xr_3_begin_0, end = xr_3_end_0, end_mask = xr_3_end_mask_0, squeeze_mask = xr_3_squeeze_mask_0, x = pairs_3_cast_fp16)[name = tensor("xr_3_cast_fp16")]; + tensor xi_3_begin_0 = const()[name = tensor("xi_3_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_3_end_0 = const()[name = tensor("xi_3_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_3_end_mask_0 = const()[name = tensor("xi_3_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_3_squeeze_mask_0 = const()[name = tensor("xi_3_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_3_cast_fp16 = slice_by_index(begin = xi_3_begin_0, end = xi_3_end_0, end_mask = xi_3_end_mask_0, squeeze_mask = xi_3_squeeze_mask_0, x = pairs_3_cast_fp16)[name = tensor("xi_3_cast_fp16")]; + tensor var_322_cast_fp16 = mul(x = xr_3_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_322_cast_fp16")]; + tensor var_323_cast_fp16 = mul(x = xi_3_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_323_cast_fp16")]; + tensor xor__3_cast_fp16 = sub(x = var_322_cast_fp16, y = var_323_cast_fp16)[name = tensor("xor__3_cast_fp16")]; + tensor var_326_cast_fp16 = mul(x = xr_3_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_326_cast_fp16")]; + tensor var_327_cast_fp16 = mul(x = xi_3_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_327_cast_fp16")]; + tensor xoi_3_cast_fp16 = add(x = var_326_cast_fp16, y = var_327_cast_fp16)[name = tensor("xoi_3_cast_fp16")]; + tensor out_3_axis_0 = const()[name = tensor("out_3_axis_0"), val = tensor(-1)]; + tensor out_3_cast_fp16 = stack(axis = out_3_axis_0, values = (xor__3_cast_fp16, xoi_3_cast_fp16))[name = tensor("out_3_cast_fp16")]; + tensor _inversed_339_y_0_to_fp16 = const()[name = tensor("_inversed_339_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_339_cast_fp16 = mul(x = position0_to_fp16, y = _inversed_339_y_0_to_fp16)[name = tensor("_inversed_339_cast_fp16")]; + tensor var_340_cast_fp16 = floor(x = _inversed_339_cast_fp16)[name = tensor("op_340_cast_fp16")]; + tensor var_341_to_fp16 = const()[name = tensor("op_341_to_fp16"), val = tensor(0x1p+9)]; + tensor var_342_cast_fp16 = mul(x = var_340_cast_fp16, y = var_341_to_fp16)[name = tensor("op_342_cast_fp16")]; + tensor write_idx_1_cast_fp16 = sub(x = position0_to_fp16, y = var_342_cast_fp16)[name = tensor("write_idx_1_cast_fp16")]; + tensor slot_indices_to_fp16 = const()[name = tensor("slot_indices_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6369856)))]; + tensor var_345_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_1_cast_fp16)[name = tensor("op_345_cast_fp16")]; + tensor var_355 = const()[name = tensor("op_355"), val = tensor([1, 512, 1, 1])]; + tensor onehot_1_to_fp16_dtype_0 = const()[name = tensor("onehot_1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_345_cast_fp16_to_fp16 = cast(dtype = onehot_1_to_fp16_dtype_0, x = var_345_cast_fp16)[name = tensor("cast_576")]; + tensor onehot_3_cast_fp16 = reshape(shape = var_355, x = var_345_cast_fp16_to_fp16)[name = tensor("onehot_3_cast_fp16")]; + tensor var_357_to_fp16 = const()[name = tensor("op_357_to_fp16"), val = tensor(0x1p+0)]; + tensor var_359_cast_fp16 = sub(x = var_357_to_fp16, y = onehot_3_cast_fp16)[name = tensor("op_359_cast_fp16")]; + tensor k_cache0_to_fp16_dtype_0 = const()[name = tensor("k_cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache0_to_fp16 = cast(dtype = k_cache0_to_fp16_dtype_0, x = k_cache0)[name = tensor("cast_575")]; + tensor var_360_cast_fp16 = mul(x = k_cache0_to_fp16, y = var_359_cast_fp16)[name = tensor("op_360_cast_fp16")]; + tensor var_365 = const()[name = tensor("op_365"), val = tensor([1, 1, 16, 64])]; + tensor var_366_cast_fp16 = reshape(shape = var_365, x = out_3_cast_fp16)[name = tensor("op_366_cast_fp16")]; + tensor var_367_cast_fp16 = mul(x = var_366_cast_fp16, y = onehot_3_cast_fp16)[name = tensor("op_367_cast_fp16")]; + tensor new_k_cache_1_cast_fp16 = add(x = var_360_cast_fp16, y = var_367_cast_fp16)[name = tensor("new_k_cache_1_cast_fp16")]; + tensor new_k_cache_1_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_1_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache0_to_fp16_dtype_0 = const()[name = tensor("v_cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache0_to_fp16 = cast(dtype = v_cache0_to_fp16_dtype_0, x = v_cache0)[name = tensor("cast_573")]; + tensor var_373_cast_fp16 = mul(x = v_cache0_to_fp16, y = var_359_cast_fp16)[name = tensor("op_373_cast_fp16")]; + tensor var_378 = const()[name = tensor("op_378"), val = tensor([1, 1, 16, 64])]; + tensor var_379_cast_fp16 = reshape(shape = var_378, x = v_1_cast_fp16)[name = tensor("op_379_cast_fp16")]; + tensor var_380_cast_fp16 = mul(x = var_379_cast_fp16, y = onehot_3_cast_fp16)[name = tensor("op_380_cast_fp16")]; + tensor new_v_cache_1_cast_fp16 = add(x = var_373_cast_fp16, y = var_380_cast_fp16)[name = tensor("new_v_cache_1_cast_fp16")]; + tensor new_v_cache_1_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_1_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_0_perm_0 = const()[name = tensor("transpose_0_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_393 = const()[name = tensor("op_393"), val = tensor([0, 2, 1, 3])]; + tensor var_399 = const()[name = tensor("op_399"), val = tensor([1, 16, 1, 64])]; + tensor q4_1_cast_fp16 = reshape(shape = var_399, x = out_1_cast_fp16)[name = tensor("q4_1_cast_fp16")]; + tensor var_404_transpose_x_0 = const()[name = tensor("op_404_transpose_x_0"), val = tensor(false)]; + tensor var_404_transpose_y_0 = const()[name = tensor("op_404_transpose_y_0"), val = tensor(false)]; + tensor transpose_0_cast_fp16 = transpose(perm = transpose_0_perm_0, x = new_k_cache_1_cast_fp16)[name = tensor("transpose_95")]; + tensor var_404_cast_fp16 = matmul(transpose_x = var_404_transpose_x_0, transpose_y = var_404_transpose_y_0, x = q4_1_cast_fp16, y = transpose_0_cast_fp16)[name = tensor("op_404_cast_fp16")]; + tensor var_405_to_fp16 = const()[name = tensor("op_405_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_1_cast_fp16 = mul(x = var_404_cast_fp16, y = var_405_to_fp16)[name = tensor("scores_1_cast_fp16")]; + tensor var_409 = const()[name = tensor("op_409"), val = tensor([1, 1])]; + tensor var_410_cast_fp16 = reshape(shape = var_409, x = position0_to_fp16)[name = tensor("op_410_cast_fp16")]; + tensor var_411_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_410_cast_fp16)[name = tensor("op_411_cast_fp16")]; + tensor mask_1_to_fp16_dtype_0 = const()[name = tensor("mask_1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_418_to_fp16 = const()[name = tensor("op_418_to_fp16"), val = tensor(0x1p+0)]; + tensor var_411_cast_fp16_to_fp16 = cast(dtype = mask_1_to_fp16_dtype_0, x = var_411_cast_fp16)[name = tensor("cast_571")]; + tensor var_419_cast_fp16 = sub(x = var_411_cast_fp16_to_fp16, y = var_418_to_fp16)[name = tensor("op_419_cast_fp16")]; + tensor var_424 = const()[name = tensor("op_424"), val = tensor([1, 1, 1, 512])]; + tensor var_425_cast_fp16 = reshape(shape = var_424, x = var_419_cast_fp16)[name = tensor("op_425_cast_fp16")]; + tensor var_426_to_fp16 = const()[name = tensor("op_426_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_427_cast_fp16 = mul(x = var_425_cast_fp16, y = var_426_to_fp16)[name = tensor("op_427_cast_fp16")]; + tensor scores_3_cast_fp16 = add(x = scores_1_cast_fp16, y = var_427_cast_fp16)[name = tensor("scores_3_cast_fp16")]; + tensor var_430 = const()[name = tensor("op_430"), val = tensor(-1)]; + tensor weights_1_cast_fp16 = softmax(axis = var_430, x = scores_3_cast_fp16)[name = tensor("weights_1_cast_fp16")]; + tensor attn_1_transpose_x_0 = const()[name = tensor("attn_1_transpose_x_0"), val = tensor(false)]; + tensor attn_1_transpose_y_0 = const()[name = tensor("attn_1_transpose_y_0"), val = tensor(false)]; + tensor values_1_cast_fp16 = transpose(perm = var_393, x = new_v_cache_1_cast_fp16)[name = tensor("transpose_94")]; + tensor attn_1_cast_fp16 = matmul(transpose_x = attn_1_transpose_x_0, transpose_y = attn_1_transpose_y_0, x = weights_1_cast_fp16, y = values_1_cast_fp16)[name = tensor("attn_1_cast_fp16")]; + tensor var_438 = const()[name = tensor("op_438"), val = tensor([0, 2, 1, 3])]; + tensor var_443 = const()[name = tensor("op_443"), val = tensor([1, 1, 1024])]; + tensor var_439_cast_fp16 = transpose(perm = var_438, x = attn_1_cast_fp16)[name = tensor("transpose_93")]; + tensor input_7_cast_fp16 = reshape(shape = var_443, x = var_439_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor attn0_out_proj_weight_to_fp16 = const()[name = tensor("attn0_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6370944)))]; + tensor linear_2_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn0_out_proj_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor var_449_to_fp16 = const()[name = tensor("op_449_to_fp16"), val = tensor(0x1p+0)]; + tensor var_450_cast_fp16 = add(x = position0_to_fp16, y = var_449_to_fp16)[name = tensor("op_450_cast_fp16")]; + tensor var_450_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_450_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_9_cast_fp16 = add(x = linear_0_cast_fp16, y = linear_2_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor input_11_axes_0 = const()[name = tensor("input_11_axes_0"), val = tensor([-1])]; + tensor norm0_2_weight_to_fp16 = const()[name = tensor("norm0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8468160)))]; + tensor norm0_2_bias_to_fp16 = const()[name = tensor("norm0_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8470272)))]; + tensor var_454_to_fp16 = const()[name = tensor("op_454_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_11_cast_fp16 = layer_norm(axes = input_11_axes_0, beta = norm0_2_bias_to_fp16, epsilon = var_454_to_fp16, gamma = norm0_2_weight_to_fp16, x = input_9_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor linear0_1_weight_to_fp16 = const()[name = tensor("linear0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8472384)))]; + tensor linear_3_bias_0_to_fp16 = const()[name = tensor("linear_3_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16861056)))]; + tensor linear_3_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear0_1_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("linear_3_cast_fp16")]; + tensor input_13_mode_0 = const()[name = tensor("input_13_mode_0"), val = tensor("EXACT")]; + tensor input_13_cast_fp16 = gelu(mode = input_13_mode_0, x = linear_3_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor linear0_2_weight_to_fp16 = const()[name = tensor("linear0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16869312)))]; + tensor linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear0_2_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor input_15_cast_fp16 = add(x = input_9_cast_fp16, y = linear_4_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor input_17_axes_0 = const()[name = tensor("input_17_axes_0"), val = tensor([-1])]; + tensor norm1_1_weight_to_fp16 = const()[name = tensor("norm1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25257984)))]; + tensor norm1_1_bias_to_fp16 = const()[name = tensor("norm1_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25260096)))]; + tensor var_471_to_fp16 = const()[name = tensor("op_471_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_17_cast_fp16 = layer_norm(axes = input_17_axes_0, beta = norm1_1_bias_to_fp16, epsilon = var_471_to_fp16, gamma = norm1_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor attn1_in_proj_weight_to_fp16 = const()[name = tensor("attn1_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25262208)))]; + tensor linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn1_in_proj_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor var_484 = const()[name = tensor("op_484"), val = tensor([1, 3, 16, 64])]; + tensor qkv_3_cast_fp16 = reshape(shape = var_484, x = linear_5_cast_fp16)[name = tensor("qkv_3_cast_fp16")]; + tensor x_5_begin_0 = const()[name = tensor("x_5_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_5_end_0 = const()[name = tensor("x_5_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_5_end_mask_0 = const()[name = tensor("x_5_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_5_squeeze_mask_0 = const()[name = tensor("x_5_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_5_cast_fp16 = slice_by_index(begin = x_5_begin_0, end = x_5_end_0, end_mask = x_5_end_mask_0, squeeze_mask = x_5_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("x_5_cast_fp16")]; + tensor x_7_begin_0 = const()[name = tensor("x_7_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_7_end_0 = const()[name = tensor("x_7_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_7_end_mask_0 = const()[name = tensor("x_7_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_7_squeeze_mask_0 = const()[name = tensor("x_7_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_7_cast_fp16 = slice_by_index(begin = x_7_begin_0, end = x_7_end_0, end_mask = x_7_end_mask_0, squeeze_mask = x_7_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("x_7_cast_fp16")]; + tensor v_3_begin_0 = const()[name = tensor("v_3_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_3_end_0 = const()[name = tensor("v_3_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_3_end_mask_0 = const()[name = tensor("v_3_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_3_squeeze_mask_0 = const()[name = tensor("v_3_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_3_cast_fp16 = slice_by_index(begin = v_3_begin_0, end = v_3_end_0, end_mask = v_3_end_mask_0, squeeze_mask = v_3_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("v_3_cast_fp16")]; + tensor var_513 = const()[name = tensor("op_513"), val = tensor([1, 1, 1])]; + tensor position1_to_fp16_dtype_0 = const()[name = tensor("position1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position1_to_fp16 = cast(dtype = position1_to_fp16_dtype_0, x = position1)[name = tensor("cast_569")]; + tensor var_514_cast_fp16 = reshape(shape = var_513, x = position1_to_fp16)[name = tensor("op_514_cast_fp16")]; + tensor angles_5_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_514_cast_fp16)[name = tensor("angles_5_cast_fp16")]; + tensor rotr_5_cast_fp16 = cos(x = angles_5_cast_fp16)[name = tensor("rotr_5_cast_fp16")]; + tensor roti_5_cast_fp16 = sin(x = angles_5_cast_fp16)[name = tensor("roti_5_cast_fp16")]; + tensor var_522 = const()[name = tensor("op_522"), val = tensor([1, 16, 32, 2])]; + tensor pairs_5_cast_fp16 = reshape(shape = var_522, x = x_5_cast_fp16)[name = tensor("pairs_5_cast_fp16")]; + tensor xr_5_begin_0 = const()[name = tensor("xr_5_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_5_end_0 = const()[name = tensor("xr_5_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_5_end_mask_0 = const()[name = tensor("xr_5_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_5_squeeze_mask_0 = const()[name = tensor("xr_5_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_5_cast_fp16 = slice_by_index(begin = xr_5_begin_0, end = xr_5_end_0, end_mask = xr_5_end_mask_0, squeeze_mask = xr_5_squeeze_mask_0, x = pairs_5_cast_fp16)[name = tensor("xr_5_cast_fp16")]; + tensor xi_5_begin_0 = const()[name = tensor("xi_5_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_5_end_0 = const()[name = tensor("xi_5_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_5_end_mask_0 = const()[name = tensor("xi_5_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_5_squeeze_mask_0 = const()[name = tensor("xi_5_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_5_cast_fp16 = slice_by_index(begin = xi_5_begin_0, end = xi_5_end_0, end_mask = xi_5_end_mask_0, squeeze_mask = xi_5_squeeze_mask_0, x = pairs_5_cast_fp16)[name = tensor("xi_5_cast_fp16")]; + tensor var_530_cast_fp16 = mul(x = xr_5_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_530_cast_fp16")]; + tensor var_531_cast_fp16 = mul(x = xi_5_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_531_cast_fp16")]; + tensor xor__5_cast_fp16 = sub(x = var_530_cast_fp16, y = var_531_cast_fp16)[name = tensor("xor__5_cast_fp16")]; + tensor var_534_cast_fp16 = mul(x = xr_5_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_534_cast_fp16")]; + tensor var_535_cast_fp16 = mul(x = xi_5_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_535_cast_fp16")]; + tensor xoi_5_cast_fp16 = add(x = var_534_cast_fp16, y = var_535_cast_fp16)[name = tensor("xoi_5_cast_fp16")]; + tensor out_5_axis_0 = const()[name = tensor("out_5_axis_0"), val = tensor(-1)]; + tensor out_5_cast_fp16 = stack(axis = out_5_axis_0, values = (xor__5_cast_fp16, xoi_5_cast_fp16))[name = tensor("out_5_cast_fp16")]; + tensor var_558 = const()[name = tensor("op_558"), val = tensor([1, 16, 32, 2])]; + tensor pairs_7_cast_fp16 = reshape(shape = var_558, x = x_7_cast_fp16)[name = tensor("pairs_7_cast_fp16")]; + tensor xr_7_begin_0 = const()[name = tensor("xr_7_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_7_end_0 = const()[name = tensor("xr_7_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_7_end_mask_0 = const()[name = tensor("xr_7_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_7_squeeze_mask_0 = const()[name = tensor("xr_7_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_7_cast_fp16 = slice_by_index(begin = xr_7_begin_0, end = xr_7_end_0, end_mask = xr_7_end_mask_0, squeeze_mask = xr_7_squeeze_mask_0, x = pairs_7_cast_fp16)[name = tensor("xr_7_cast_fp16")]; + tensor xi_7_begin_0 = const()[name = tensor("xi_7_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_7_end_0 = const()[name = tensor("xi_7_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_7_end_mask_0 = const()[name = tensor("xi_7_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_7_squeeze_mask_0 = const()[name = tensor("xi_7_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_7_cast_fp16 = slice_by_index(begin = xi_7_begin_0, end = xi_7_end_0, end_mask = xi_7_end_mask_0, squeeze_mask = xi_7_squeeze_mask_0, x = pairs_7_cast_fp16)[name = tensor("xi_7_cast_fp16")]; + tensor var_566_cast_fp16 = mul(x = xr_7_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_566_cast_fp16")]; + tensor var_567_cast_fp16 = mul(x = xi_7_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_567_cast_fp16")]; + tensor xor__7_cast_fp16 = sub(x = var_566_cast_fp16, y = var_567_cast_fp16)[name = tensor("xor__7_cast_fp16")]; + tensor var_570_cast_fp16 = mul(x = xr_7_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_570_cast_fp16")]; + tensor var_571_cast_fp16 = mul(x = xi_7_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_571_cast_fp16")]; + tensor xoi_7_cast_fp16 = add(x = var_570_cast_fp16, y = var_571_cast_fp16)[name = tensor("xoi_7_cast_fp16")]; + tensor out_7_axis_0 = const()[name = tensor("out_7_axis_0"), val = tensor(-1)]; + tensor out_7_cast_fp16 = stack(axis = out_7_axis_0, values = (xor__7_cast_fp16, xoi_7_cast_fp16))[name = tensor("out_7_cast_fp16")]; + tensor _inversed_583_y_0_to_fp16 = const()[name = tensor("_inversed_583_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_583_cast_fp16 = mul(x = position1_to_fp16, y = _inversed_583_y_0_to_fp16)[name = tensor("_inversed_583_cast_fp16")]; + tensor var_584_cast_fp16 = floor(x = _inversed_583_cast_fp16)[name = tensor("op_584_cast_fp16")]; + tensor var_585_to_fp16 = const()[name = tensor("op_585_to_fp16"), val = tensor(0x1p+9)]; + tensor var_586_cast_fp16 = mul(x = var_584_cast_fp16, y = var_585_to_fp16)[name = tensor("op_586_cast_fp16")]; + tensor write_idx_3_cast_fp16 = sub(x = position1_to_fp16, y = var_586_cast_fp16)[name = tensor("write_idx_3_cast_fp16")]; + tensor var_589_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_3_cast_fp16)[name = tensor("op_589_cast_fp16")]; + tensor var_599 = const()[name = tensor("op_599"), val = tensor([1, 512, 1, 1])]; + tensor onehot_5_to_fp16_dtype_0 = const()[name = tensor("onehot_5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_589_cast_fp16_to_fp16 = cast(dtype = onehot_5_to_fp16_dtype_0, x = var_589_cast_fp16)[name = tensor("cast_568")]; + tensor onehot_7_cast_fp16 = reshape(shape = var_599, x = var_589_cast_fp16_to_fp16)[name = tensor("onehot_7_cast_fp16")]; + tensor var_601_to_fp16 = const()[name = tensor("op_601_to_fp16"), val = tensor(0x1p+0)]; + tensor var_603_cast_fp16 = sub(x = var_601_to_fp16, y = onehot_7_cast_fp16)[name = tensor("op_603_cast_fp16")]; + tensor k_cache1_to_fp16_dtype_0 = const()[name = tensor("k_cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache1_to_fp16 = cast(dtype = k_cache1_to_fp16_dtype_0, x = k_cache1)[name = tensor("cast_567")]; + tensor var_604_cast_fp16 = mul(x = k_cache1_to_fp16, y = var_603_cast_fp16)[name = tensor("op_604_cast_fp16")]; + tensor var_609 = const()[name = tensor("op_609"), val = tensor([1, 1, 16, 64])]; + tensor var_610_cast_fp16 = reshape(shape = var_609, x = out_7_cast_fp16)[name = tensor("op_610_cast_fp16")]; + tensor var_611_cast_fp16 = mul(x = var_610_cast_fp16, y = onehot_7_cast_fp16)[name = tensor("op_611_cast_fp16")]; + tensor new_k_cache_3_cast_fp16 = add(x = var_604_cast_fp16, y = var_611_cast_fp16)[name = tensor("new_k_cache_3_cast_fp16")]; + tensor new_k_cache_3_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_3_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache1_to_fp16_dtype_0 = const()[name = tensor("v_cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache1_to_fp16 = cast(dtype = v_cache1_to_fp16_dtype_0, x = v_cache1)[name = tensor("cast_565")]; + tensor var_617_cast_fp16 = mul(x = v_cache1_to_fp16, y = var_603_cast_fp16)[name = tensor("op_617_cast_fp16")]; + tensor var_622 = const()[name = tensor("op_622"), val = tensor([1, 1, 16, 64])]; + tensor var_623_cast_fp16 = reshape(shape = var_622, x = v_3_cast_fp16)[name = tensor("op_623_cast_fp16")]; + tensor var_624_cast_fp16 = mul(x = var_623_cast_fp16, y = onehot_7_cast_fp16)[name = tensor("op_624_cast_fp16")]; + tensor new_v_cache_3_cast_fp16 = add(x = var_617_cast_fp16, y = var_624_cast_fp16)[name = tensor("new_v_cache_3_cast_fp16")]; + tensor new_v_cache_3_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_3_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_1_perm_0 = const()[name = tensor("transpose_1_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_637 = const()[name = tensor("op_637"), val = tensor([0, 2, 1, 3])]; + tensor var_643 = const()[name = tensor("op_643"), val = tensor([1, 16, 1, 64])]; + tensor q4_3_cast_fp16 = reshape(shape = var_643, x = out_5_cast_fp16)[name = tensor("q4_3_cast_fp16")]; + tensor var_648_transpose_x_0 = const()[name = tensor("op_648_transpose_x_0"), val = tensor(false)]; + tensor var_648_transpose_y_0 = const()[name = tensor("op_648_transpose_y_0"), val = tensor(false)]; + tensor transpose_1_cast_fp16 = transpose(perm = transpose_1_perm_0, x = new_k_cache_3_cast_fp16)[name = tensor("transpose_92")]; + tensor var_648_cast_fp16 = matmul(transpose_x = var_648_transpose_x_0, transpose_y = var_648_transpose_y_0, x = q4_3_cast_fp16, y = transpose_1_cast_fp16)[name = tensor("op_648_cast_fp16")]; + tensor var_649_to_fp16 = const()[name = tensor("op_649_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_5_cast_fp16 = mul(x = var_648_cast_fp16, y = var_649_to_fp16)[name = tensor("scores_5_cast_fp16")]; + tensor var_653 = const()[name = tensor("op_653"), val = tensor([1, 1])]; + tensor var_654_cast_fp16 = reshape(shape = var_653, x = position1_to_fp16)[name = tensor("op_654_cast_fp16")]; + tensor var_655_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_654_cast_fp16)[name = tensor("op_655_cast_fp16")]; + tensor mask_3_to_fp16_dtype_0 = const()[name = tensor("mask_3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_662_to_fp16 = const()[name = tensor("op_662_to_fp16"), val = tensor(0x1p+0)]; + tensor var_655_cast_fp16_to_fp16 = cast(dtype = mask_3_to_fp16_dtype_0, x = var_655_cast_fp16)[name = tensor("cast_563")]; + tensor var_663_cast_fp16 = sub(x = var_655_cast_fp16_to_fp16, y = var_662_to_fp16)[name = tensor("op_663_cast_fp16")]; + tensor var_668 = const()[name = tensor("op_668"), val = tensor([1, 1, 1, 512])]; + tensor var_669_cast_fp16 = reshape(shape = var_668, x = var_663_cast_fp16)[name = tensor("op_669_cast_fp16")]; + tensor var_670_to_fp16 = const()[name = tensor("op_670_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_671_cast_fp16 = mul(x = var_669_cast_fp16, y = var_670_to_fp16)[name = tensor("op_671_cast_fp16")]; + tensor scores_7_cast_fp16 = add(x = scores_5_cast_fp16, y = var_671_cast_fp16)[name = tensor("scores_7_cast_fp16")]; + tensor var_674 = const()[name = tensor("op_674"), val = tensor(-1)]; + tensor weights_3_cast_fp16 = softmax(axis = var_674, x = scores_7_cast_fp16)[name = tensor("weights_3_cast_fp16")]; + tensor attn_3_transpose_x_0 = const()[name = tensor("attn_3_transpose_x_0"), val = tensor(false)]; + tensor attn_3_transpose_y_0 = const()[name = tensor("attn_3_transpose_y_0"), val = tensor(false)]; + tensor values_3_cast_fp16 = transpose(perm = var_637, x = new_v_cache_3_cast_fp16)[name = tensor("transpose_91")]; + tensor attn_3_cast_fp16 = matmul(transpose_x = attn_3_transpose_x_0, transpose_y = attn_3_transpose_y_0, x = weights_3_cast_fp16, y = values_3_cast_fp16)[name = tensor("attn_3_cast_fp16")]; + tensor var_682 = const()[name = tensor("op_682"), val = tensor([0, 2, 1, 3])]; + tensor var_687 = const()[name = tensor("op_687"), val = tensor([1, 1, 1024])]; + tensor var_683_cast_fp16 = transpose(perm = var_682, x = attn_3_cast_fp16)[name = tensor("transpose_90")]; + tensor input_19_cast_fp16 = reshape(shape = var_687, x = var_683_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor attn1_out_proj_weight_to_fp16 = const()[name = tensor("attn1_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31553728)))]; + tensor linear_6_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn1_out_proj_weight_to_fp16, x = input_19_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor var_693_to_fp16 = const()[name = tensor("op_693_to_fp16"), val = tensor(0x1p+0)]; + tensor var_694_cast_fp16 = add(x = position1_to_fp16, y = var_693_to_fp16)[name = tensor("op_694_cast_fp16")]; + tensor var_694_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_694_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_21_cast_fp16 = add(x = input_15_cast_fp16, y = linear_6_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor input_23_axes_0 = const()[name = tensor("input_23_axes_0"), val = tensor([-1])]; + tensor norm1_2_weight_to_fp16 = const()[name = tensor("norm1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33650944)))]; + tensor norm1_2_bias_to_fp16 = const()[name = tensor("norm1_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33653056)))]; + tensor var_698_to_fp16 = const()[name = tensor("op_698_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_23_cast_fp16 = layer_norm(axes = input_23_axes_0, beta = norm1_2_bias_to_fp16, epsilon = var_698_to_fp16, gamma = norm1_2_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor linear1_1_weight_to_fp16 = const()[name = tensor("linear1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33655168)))]; + tensor linear_7_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear1_1_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_25_mode_0 = const()[name = tensor("input_25_mode_0"), val = tensor("EXACT")]; + tensor input_25_cast_fp16 = gelu(mode = input_25_mode_0, x = linear_7_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor linear1_2_weight_to_fp16 = const()[name = tensor("linear1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42043840)))]; + tensor linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear1_2_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor input_27_cast_fp16 = add(x = input_21_cast_fp16, y = linear_8_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor input_29_axes_0 = const()[name = tensor("input_29_axes_0"), val = tensor([-1])]; + tensor norm2_1_weight_to_fp16 = const()[name = tensor("norm2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50432512)))]; + tensor norm2_1_bias_to_fp16 = const()[name = tensor("norm2_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50434624)))]; + tensor var_715_to_fp16 = const()[name = tensor("op_715_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_29_cast_fp16 = layer_norm(axes = input_29_axes_0, beta = norm2_1_bias_to_fp16, epsilon = var_715_to_fp16, gamma = norm2_1_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor attn2_in_proj_weight_to_fp16 = const()[name = tensor("attn2_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50436736)))]; + tensor linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn2_in_proj_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor var_728 = const()[name = tensor("op_728"), val = tensor([1, 3, 16, 64])]; + tensor qkv_5_cast_fp16 = reshape(shape = var_728, x = linear_9_cast_fp16)[name = tensor("qkv_5_cast_fp16")]; + tensor x_9_begin_0 = const()[name = tensor("x_9_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_9_end_0 = const()[name = tensor("x_9_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_9_end_mask_0 = const()[name = tensor("x_9_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_9_squeeze_mask_0 = const()[name = tensor("x_9_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_9_cast_fp16 = slice_by_index(begin = x_9_begin_0, end = x_9_end_0, end_mask = x_9_end_mask_0, squeeze_mask = x_9_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("x_9_cast_fp16")]; + tensor x_11_begin_0 = const()[name = tensor("x_11_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_11_end_0 = const()[name = tensor("x_11_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_11_end_mask_0 = const()[name = tensor("x_11_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_11_squeeze_mask_0 = const()[name = tensor("x_11_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_11_cast_fp16 = slice_by_index(begin = x_11_begin_0, end = x_11_end_0, end_mask = x_11_end_mask_0, squeeze_mask = x_11_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("x_11_cast_fp16")]; + tensor v_5_begin_0 = const()[name = tensor("v_5_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_5_end_0 = const()[name = tensor("v_5_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_5_end_mask_0 = const()[name = tensor("v_5_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_5_squeeze_mask_0 = const()[name = tensor("v_5_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_5_cast_fp16 = slice_by_index(begin = v_5_begin_0, end = v_5_end_0, end_mask = v_5_end_mask_0, squeeze_mask = v_5_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("v_5_cast_fp16")]; + tensor var_757 = const()[name = tensor("op_757"), val = tensor([1, 1, 1])]; + tensor position2_to_fp16_dtype_0 = const()[name = tensor("position2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position2_to_fp16 = cast(dtype = position2_to_fp16_dtype_0, x = position2)[name = tensor("cast_561")]; + tensor var_758_cast_fp16 = reshape(shape = var_757, x = position2_to_fp16)[name = tensor("op_758_cast_fp16")]; + tensor angles_9_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_758_cast_fp16)[name = tensor("angles_9_cast_fp16")]; + tensor rotr_9_cast_fp16 = cos(x = angles_9_cast_fp16)[name = tensor("rotr_9_cast_fp16")]; + tensor roti_9_cast_fp16 = sin(x = angles_9_cast_fp16)[name = tensor("roti_9_cast_fp16")]; + tensor var_766 = const()[name = tensor("op_766"), val = tensor([1, 16, 32, 2])]; + tensor pairs_9_cast_fp16 = reshape(shape = var_766, x = x_9_cast_fp16)[name = tensor("pairs_9_cast_fp16")]; + tensor xr_9_begin_0 = const()[name = tensor("xr_9_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_9_end_0 = const()[name = tensor("xr_9_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_9_end_mask_0 = const()[name = tensor("xr_9_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_9_squeeze_mask_0 = const()[name = tensor("xr_9_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_9_cast_fp16 = slice_by_index(begin = xr_9_begin_0, end = xr_9_end_0, end_mask = xr_9_end_mask_0, squeeze_mask = xr_9_squeeze_mask_0, x = pairs_9_cast_fp16)[name = tensor("xr_9_cast_fp16")]; + tensor xi_9_begin_0 = const()[name = tensor("xi_9_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_9_end_0 = const()[name = tensor("xi_9_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_9_end_mask_0 = const()[name = tensor("xi_9_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_9_squeeze_mask_0 = const()[name = tensor("xi_9_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_9_cast_fp16 = slice_by_index(begin = xi_9_begin_0, end = xi_9_end_0, end_mask = xi_9_end_mask_0, squeeze_mask = xi_9_squeeze_mask_0, x = pairs_9_cast_fp16)[name = tensor("xi_9_cast_fp16")]; + tensor var_774_cast_fp16 = mul(x = xr_9_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_774_cast_fp16")]; + tensor var_775_cast_fp16 = mul(x = xi_9_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_775_cast_fp16")]; + tensor xor__9_cast_fp16 = sub(x = var_774_cast_fp16, y = var_775_cast_fp16)[name = tensor("xor__9_cast_fp16")]; + tensor var_778_cast_fp16 = mul(x = xr_9_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_778_cast_fp16")]; + tensor var_779_cast_fp16 = mul(x = xi_9_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_779_cast_fp16")]; + tensor xoi_9_cast_fp16 = add(x = var_778_cast_fp16, y = var_779_cast_fp16)[name = tensor("xoi_9_cast_fp16")]; + tensor out_9_axis_0 = const()[name = tensor("out_9_axis_0"), val = tensor(-1)]; + tensor out_9_cast_fp16 = stack(axis = out_9_axis_0, values = (xor__9_cast_fp16, xoi_9_cast_fp16))[name = tensor("out_9_cast_fp16")]; + tensor var_802 = const()[name = tensor("op_802"), val = tensor([1, 16, 32, 2])]; + tensor pairs_11_cast_fp16 = reshape(shape = var_802, x = x_11_cast_fp16)[name = tensor("pairs_11_cast_fp16")]; + tensor xr_11_begin_0 = const()[name = tensor("xr_11_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_11_end_0 = const()[name = tensor("xr_11_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_11_end_mask_0 = const()[name = tensor("xr_11_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_11_squeeze_mask_0 = const()[name = tensor("xr_11_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_11_cast_fp16 = slice_by_index(begin = xr_11_begin_0, end = xr_11_end_0, end_mask = xr_11_end_mask_0, squeeze_mask = xr_11_squeeze_mask_0, x = pairs_11_cast_fp16)[name = tensor("xr_11_cast_fp16")]; + tensor xi_11_begin_0 = const()[name = tensor("xi_11_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_11_end_0 = const()[name = tensor("xi_11_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_11_end_mask_0 = const()[name = tensor("xi_11_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_11_squeeze_mask_0 = const()[name = tensor("xi_11_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_11_cast_fp16 = slice_by_index(begin = xi_11_begin_0, end = xi_11_end_0, end_mask = xi_11_end_mask_0, squeeze_mask = xi_11_squeeze_mask_0, x = pairs_11_cast_fp16)[name = tensor("xi_11_cast_fp16")]; + tensor var_810_cast_fp16 = mul(x = xr_11_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_810_cast_fp16")]; + tensor var_811_cast_fp16 = mul(x = xi_11_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_811_cast_fp16")]; + tensor xor__11_cast_fp16 = sub(x = var_810_cast_fp16, y = var_811_cast_fp16)[name = tensor("xor__11_cast_fp16")]; + tensor var_814_cast_fp16 = mul(x = xr_11_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_814_cast_fp16")]; + tensor var_815_cast_fp16 = mul(x = xi_11_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_815_cast_fp16")]; + tensor xoi_11_cast_fp16 = add(x = var_814_cast_fp16, y = var_815_cast_fp16)[name = tensor("xoi_11_cast_fp16")]; + tensor out_11_axis_0 = const()[name = tensor("out_11_axis_0"), val = tensor(-1)]; + tensor out_11_cast_fp16 = stack(axis = out_11_axis_0, values = (xor__11_cast_fp16, xoi_11_cast_fp16))[name = tensor("out_11_cast_fp16")]; + tensor _inversed_827_y_0_to_fp16 = const()[name = tensor("_inversed_827_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_827_cast_fp16 = mul(x = position2_to_fp16, y = _inversed_827_y_0_to_fp16)[name = tensor("_inversed_827_cast_fp16")]; + tensor var_828_cast_fp16 = floor(x = _inversed_827_cast_fp16)[name = tensor("op_828_cast_fp16")]; + tensor var_829_to_fp16 = const()[name = tensor("op_829_to_fp16"), val = tensor(0x1p+9)]; + tensor var_830_cast_fp16 = mul(x = var_828_cast_fp16, y = var_829_to_fp16)[name = tensor("op_830_cast_fp16")]; + tensor write_idx_5_cast_fp16 = sub(x = position2_to_fp16, y = var_830_cast_fp16)[name = tensor("write_idx_5_cast_fp16")]; + tensor var_833_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_5_cast_fp16)[name = tensor("op_833_cast_fp16")]; + tensor var_843 = const()[name = tensor("op_843"), val = tensor([1, 512, 1, 1])]; + tensor onehot_9_to_fp16_dtype_0 = const()[name = tensor("onehot_9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_833_cast_fp16_to_fp16 = cast(dtype = onehot_9_to_fp16_dtype_0, x = var_833_cast_fp16)[name = tensor("cast_560")]; + tensor onehot_11_cast_fp16 = reshape(shape = var_843, x = var_833_cast_fp16_to_fp16)[name = tensor("onehot_11_cast_fp16")]; + tensor var_845_to_fp16 = const()[name = tensor("op_845_to_fp16"), val = tensor(0x1p+0)]; + tensor var_847_cast_fp16 = sub(x = var_845_to_fp16, y = onehot_11_cast_fp16)[name = tensor("op_847_cast_fp16")]; + tensor k_cache2_to_fp16_dtype_0 = const()[name = tensor("k_cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache2_to_fp16 = cast(dtype = k_cache2_to_fp16_dtype_0, x = k_cache2)[name = tensor("cast_559")]; + tensor var_848_cast_fp16 = mul(x = k_cache2_to_fp16, y = var_847_cast_fp16)[name = tensor("op_848_cast_fp16")]; + tensor var_853 = const()[name = tensor("op_853"), val = tensor([1, 1, 16, 64])]; + tensor var_854_cast_fp16 = reshape(shape = var_853, x = out_11_cast_fp16)[name = tensor("op_854_cast_fp16")]; + tensor var_855_cast_fp16 = mul(x = var_854_cast_fp16, y = onehot_11_cast_fp16)[name = tensor("op_855_cast_fp16")]; + tensor new_k_cache_5_cast_fp16 = add(x = var_848_cast_fp16, y = var_855_cast_fp16)[name = tensor("new_k_cache_5_cast_fp16")]; + tensor new_k_cache_5_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_5_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache2_to_fp16_dtype_0 = const()[name = tensor("v_cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache2_to_fp16 = cast(dtype = v_cache2_to_fp16_dtype_0, x = v_cache2)[name = tensor("cast_557")]; + tensor var_861_cast_fp16 = mul(x = v_cache2_to_fp16, y = var_847_cast_fp16)[name = tensor("op_861_cast_fp16")]; + tensor var_866 = const()[name = tensor("op_866"), val = tensor([1, 1, 16, 64])]; + tensor var_867_cast_fp16 = reshape(shape = var_866, x = v_5_cast_fp16)[name = tensor("op_867_cast_fp16")]; + tensor var_868_cast_fp16 = mul(x = var_867_cast_fp16, y = onehot_11_cast_fp16)[name = tensor("op_868_cast_fp16")]; + tensor new_v_cache_5_cast_fp16 = add(x = var_861_cast_fp16, y = var_868_cast_fp16)[name = tensor("new_v_cache_5_cast_fp16")]; + tensor new_v_cache_5_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_5_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_2_perm_0 = const()[name = tensor("transpose_2_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_881 = const()[name = tensor("op_881"), val = tensor([0, 2, 1, 3])]; + tensor var_887 = const()[name = tensor("op_887"), val = tensor([1, 16, 1, 64])]; + tensor q4_5_cast_fp16 = reshape(shape = var_887, x = out_9_cast_fp16)[name = tensor("q4_5_cast_fp16")]; + tensor var_892_transpose_x_0 = const()[name = tensor("op_892_transpose_x_0"), val = tensor(false)]; + tensor var_892_transpose_y_0 = const()[name = tensor("op_892_transpose_y_0"), val = tensor(false)]; + tensor transpose_2_cast_fp16 = transpose(perm = transpose_2_perm_0, x = new_k_cache_5_cast_fp16)[name = tensor("transpose_89")]; + tensor var_892_cast_fp16 = matmul(transpose_x = var_892_transpose_x_0, transpose_y = var_892_transpose_y_0, x = q4_5_cast_fp16, y = transpose_2_cast_fp16)[name = tensor("op_892_cast_fp16")]; + tensor var_893_to_fp16 = const()[name = tensor("op_893_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_9_cast_fp16 = mul(x = var_892_cast_fp16, y = var_893_to_fp16)[name = tensor("scores_9_cast_fp16")]; + tensor var_897 = const()[name = tensor("op_897"), val = tensor([1, 1])]; + tensor var_898_cast_fp16 = reshape(shape = var_897, x = position2_to_fp16)[name = tensor("op_898_cast_fp16")]; + tensor var_899_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_898_cast_fp16)[name = tensor("op_899_cast_fp16")]; + tensor mask_5_to_fp16_dtype_0 = const()[name = tensor("mask_5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_906_to_fp16 = const()[name = tensor("op_906_to_fp16"), val = tensor(0x1p+0)]; + tensor var_899_cast_fp16_to_fp16 = cast(dtype = mask_5_to_fp16_dtype_0, x = var_899_cast_fp16)[name = tensor("cast_555")]; + tensor var_907_cast_fp16 = sub(x = var_899_cast_fp16_to_fp16, y = var_906_to_fp16)[name = tensor("op_907_cast_fp16")]; + tensor var_912 = const()[name = tensor("op_912"), val = tensor([1, 1, 1, 512])]; + tensor var_913_cast_fp16 = reshape(shape = var_912, x = var_907_cast_fp16)[name = tensor("op_913_cast_fp16")]; + tensor var_914_to_fp16 = const()[name = tensor("op_914_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_915_cast_fp16 = mul(x = var_913_cast_fp16, y = var_914_to_fp16)[name = tensor("op_915_cast_fp16")]; + tensor scores_11_cast_fp16 = add(x = scores_9_cast_fp16, y = var_915_cast_fp16)[name = tensor("scores_11_cast_fp16")]; + tensor var_918 = const()[name = tensor("op_918"), val = tensor(-1)]; + tensor weights_5_cast_fp16 = softmax(axis = var_918, x = scores_11_cast_fp16)[name = tensor("weights_5_cast_fp16")]; + tensor attn_5_transpose_x_0 = const()[name = tensor("attn_5_transpose_x_0"), val = tensor(false)]; + tensor attn_5_transpose_y_0 = const()[name = tensor("attn_5_transpose_y_0"), val = tensor(false)]; + tensor values_5_cast_fp16 = transpose(perm = var_881, x = new_v_cache_5_cast_fp16)[name = tensor("transpose_88")]; + tensor attn_5_cast_fp16 = matmul(transpose_x = attn_5_transpose_x_0, transpose_y = attn_5_transpose_y_0, x = weights_5_cast_fp16, y = values_5_cast_fp16)[name = tensor("attn_5_cast_fp16")]; + tensor var_926 = const()[name = tensor("op_926"), val = tensor([0, 2, 1, 3])]; + tensor var_931 = const()[name = tensor("op_931"), val = tensor([1, 1, 1024])]; + tensor var_927_cast_fp16 = transpose(perm = var_926, x = attn_5_cast_fp16)[name = tensor("transpose_87")]; + tensor input_31_cast_fp16 = reshape(shape = var_931, x = var_927_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor attn2_out_proj_weight_to_fp16 = const()[name = tensor("attn2_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(56728256)))]; + tensor linear_10_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn2_out_proj_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor var_937_to_fp16 = const()[name = tensor("op_937_to_fp16"), val = tensor(0x1p+0)]; + tensor var_938_cast_fp16 = add(x = position2_to_fp16, y = var_937_to_fp16)[name = tensor("op_938_cast_fp16")]; + tensor var_938_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_938_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_33_cast_fp16 = add(x = input_27_cast_fp16, y = linear_10_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor input_35_axes_0 = const()[name = tensor("input_35_axes_0"), val = tensor([-1])]; + tensor norm2_2_weight_to_fp16 = const()[name = tensor("norm2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58825472)))]; + tensor norm2_2_bias_to_fp16 = const()[name = tensor("norm2_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58827584)))]; + tensor var_942_to_fp16 = const()[name = tensor("op_942_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_35_cast_fp16 = layer_norm(axes = input_35_axes_0, beta = norm2_2_bias_to_fp16, epsilon = var_942_to_fp16, gamma = norm2_2_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor linear2_1_weight_to_fp16 = const()[name = tensor("linear2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58829696)))]; + tensor linear_11_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear2_1_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor input_37_mode_0 = const()[name = tensor("input_37_mode_0"), val = tensor("EXACT")]; + tensor input_37_cast_fp16 = gelu(mode = input_37_mode_0, x = linear_11_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor linear2_2_weight_to_fp16 = const()[name = tensor("linear2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67218368)))]; + tensor linear_12_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear2_2_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor input_39_cast_fp16 = add(x = input_33_cast_fp16, y = linear_12_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor input_41_axes_0 = const()[name = tensor("input_41_axes_0"), val = tensor([-1])]; + tensor norm3_1_weight_to_fp16 = const()[name = tensor("norm3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75607040)))]; + tensor norm3_1_bias_to_fp16 = const()[name = tensor("norm3_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75609152)))]; + tensor var_959_to_fp16 = const()[name = tensor("op_959_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_41_cast_fp16 = layer_norm(axes = input_41_axes_0, beta = norm3_1_bias_to_fp16, epsilon = var_959_to_fp16, gamma = norm3_1_weight_to_fp16, x = input_39_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor attn3_in_proj_weight_to_fp16 = const()[name = tensor("attn3_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75611264)))]; + tensor linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn3_in_proj_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor var_972 = const()[name = tensor("op_972"), val = tensor([1, 3, 16, 64])]; + tensor qkv_7_cast_fp16 = reshape(shape = var_972, x = linear_13_cast_fp16)[name = tensor("qkv_7_cast_fp16")]; + tensor x_13_begin_0 = const()[name = tensor("x_13_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_13_end_0 = const()[name = tensor("x_13_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_13_end_mask_0 = const()[name = tensor("x_13_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_13_squeeze_mask_0 = const()[name = tensor("x_13_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_13_cast_fp16 = slice_by_index(begin = x_13_begin_0, end = x_13_end_0, end_mask = x_13_end_mask_0, squeeze_mask = x_13_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("x_13_cast_fp16")]; + tensor x_15_begin_0 = const()[name = tensor("x_15_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_15_end_0 = const()[name = tensor("x_15_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_15_end_mask_0 = const()[name = tensor("x_15_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_15_squeeze_mask_0 = const()[name = tensor("x_15_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_15_cast_fp16 = slice_by_index(begin = x_15_begin_0, end = x_15_end_0, end_mask = x_15_end_mask_0, squeeze_mask = x_15_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("x_15_cast_fp16")]; + tensor v_7_begin_0 = const()[name = tensor("v_7_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_7_end_0 = const()[name = tensor("v_7_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_7_end_mask_0 = const()[name = tensor("v_7_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_7_squeeze_mask_0 = const()[name = tensor("v_7_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_7_cast_fp16 = slice_by_index(begin = v_7_begin_0, end = v_7_end_0, end_mask = v_7_end_mask_0, squeeze_mask = v_7_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("v_7_cast_fp16")]; + tensor var_1001 = const()[name = tensor("op_1001"), val = tensor([1, 1, 1])]; + tensor position3_to_fp16_dtype_0 = const()[name = tensor("position3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position3_to_fp16 = cast(dtype = position3_to_fp16_dtype_0, x = position3)[name = tensor("cast_553")]; + tensor var_1002_cast_fp16 = reshape(shape = var_1001, x = position3_to_fp16)[name = tensor("op_1002_cast_fp16")]; + tensor angles_13_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_1002_cast_fp16)[name = tensor("angles_13_cast_fp16")]; + tensor rotr_13_cast_fp16 = cos(x = angles_13_cast_fp16)[name = tensor("rotr_13_cast_fp16")]; + tensor roti_13_cast_fp16 = sin(x = angles_13_cast_fp16)[name = tensor("roti_13_cast_fp16")]; + tensor var_1010 = const()[name = tensor("op_1010"), val = tensor([1, 16, 32, 2])]; + tensor pairs_13_cast_fp16 = reshape(shape = var_1010, x = x_13_cast_fp16)[name = tensor("pairs_13_cast_fp16")]; + tensor xr_13_begin_0 = const()[name = tensor("xr_13_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_13_end_0 = const()[name = tensor("xr_13_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_13_end_mask_0 = const()[name = tensor("xr_13_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_13_squeeze_mask_0 = const()[name = tensor("xr_13_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_13_cast_fp16 = slice_by_index(begin = xr_13_begin_0, end = xr_13_end_0, end_mask = xr_13_end_mask_0, squeeze_mask = xr_13_squeeze_mask_0, x = pairs_13_cast_fp16)[name = tensor("xr_13_cast_fp16")]; + tensor xi_13_begin_0 = const()[name = tensor("xi_13_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_13_end_0 = const()[name = tensor("xi_13_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_13_end_mask_0 = const()[name = tensor("xi_13_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_13_squeeze_mask_0 = const()[name = tensor("xi_13_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_13_cast_fp16 = slice_by_index(begin = xi_13_begin_0, end = xi_13_end_0, end_mask = xi_13_end_mask_0, squeeze_mask = xi_13_squeeze_mask_0, x = pairs_13_cast_fp16)[name = tensor("xi_13_cast_fp16")]; + tensor var_1018_cast_fp16 = mul(x = xr_13_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_1018_cast_fp16")]; + tensor var_1019_cast_fp16 = mul(x = xi_13_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_1019_cast_fp16")]; + tensor xor__13_cast_fp16 = sub(x = var_1018_cast_fp16, y = var_1019_cast_fp16)[name = tensor("xor__13_cast_fp16")]; + tensor var_1022_cast_fp16 = mul(x = xr_13_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_1022_cast_fp16")]; + tensor var_1023_cast_fp16 = mul(x = xi_13_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_1023_cast_fp16")]; + tensor xoi_13_cast_fp16 = add(x = var_1022_cast_fp16, y = var_1023_cast_fp16)[name = tensor("xoi_13_cast_fp16")]; + tensor out_13_axis_0 = const()[name = tensor("out_13_axis_0"), val = tensor(-1)]; + tensor out_13_cast_fp16 = stack(axis = out_13_axis_0, values = (xor__13_cast_fp16, xoi_13_cast_fp16))[name = tensor("out_13_cast_fp16")]; + tensor var_1046 = const()[name = tensor("op_1046"), val = tensor([1, 16, 32, 2])]; + tensor pairs_15_cast_fp16 = reshape(shape = var_1046, x = x_15_cast_fp16)[name = tensor("pairs_15_cast_fp16")]; + tensor xr_15_begin_0 = const()[name = tensor("xr_15_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_15_end_0 = const()[name = tensor("xr_15_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_15_end_mask_0 = const()[name = tensor("xr_15_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_15_squeeze_mask_0 = const()[name = tensor("xr_15_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_15_cast_fp16 = slice_by_index(begin = xr_15_begin_0, end = xr_15_end_0, end_mask = xr_15_end_mask_0, squeeze_mask = xr_15_squeeze_mask_0, x = pairs_15_cast_fp16)[name = tensor("xr_15_cast_fp16")]; + tensor xi_15_begin_0 = const()[name = tensor("xi_15_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_15_end_0 = const()[name = tensor("xi_15_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_15_end_mask_0 = const()[name = tensor("xi_15_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_15_squeeze_mask_0 = const()[name = tensor("xi_15_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_15_cast_fp16 = slice_by_index(begin = xi_15_begin_0, end = xi_15_end_0, end_mask = xi_15_end_mask_0, squeeze_mask = xi_15_squeeze_mask_0, x = pairs_15_cast_fp16)[name = tensor("xi_15_cast_fp16")]; + tensor var_1054_cast_fp16 = mul(x = xr_15_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_1054_cast_fp16")]; + tensor var_1055_cast_fp16 = mul(x = xi_15_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_1055_cast_fp16")]; + tensor xor__15_cast_fp16 = sub(x = var_1054_cast_fp16, y = var_1055_cast_fp16)[name = tensor("xor__15_cast_fp16")]; + tensor var_1058_cast_fp16 = mul(x = xr_15_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_1058_cast_fp16")]; + tensor var_1059_cast_fp16 = mul(x = xi_15_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_1059_cast_fp16")]; + tensor xoi_15_cast_fp16 = add(x = var_1058_cast_fp16, y = var_1059_cast_fp16)[name = tensor("xoi_15_cast_fp16")]; + tensor out_15_axis_0 = const()[name = tensor("out_15_axis_0"), val = tensor(-1)]; + tensor out_15_cast_fp16 = stack(axis = out_15_axis_0, values = (xor__15_cast_fp16, xoi_15_cast_fp16))[name = tensor("out_15_cast_fp16")]; + tensor _inversed_1071_y_0_to_fp16 = const()[name = tensor("_inversed_1071_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1071_cast_fp16 = mul(x = position3_to_fp16, y = _inversed_1071_y_0_to_fp16)[name = tensor("_inversed_1071_cast_fp16")]; + tensor var_1072_cast_fp16 = floor(x = _inversed_1071_cast_fp16)[name = tensor("op_1072_cast_fp16")]; + tensor var_1073_to_fp16 = const()[name = tensor("op_1073_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1074_cast_fp16 = mul(x = var_1072_cast_fp16, y = var_1073_to_fp16)[name = tensor("op_1074_cast_fp16")]; + tensor write_idx_7_cast_fp16 = sub(x = position3_to_fp16, y = var_1074_cast_fp16)[name = tensor("write_idx_7_cast_fp16")]; + tensor var_1077_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_7_cast_fp16)[name = tensor("op_1077_cast_fp16")]; + tensor var_1087 = const()[name = tensor("op_1087"), val = tensor([1, 512, 1, 1])]; + tensor onehot_13_to_fp16_dtype_0 = const()[name = tensor("onehot_13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_1077_cast_fp16_to_fp16 = cast(dtype = onehot_13_to_fp16_dtype_0, x = var_1077_cast_fp16)[name = tensor("cast_552")]; + tensor onehot_15_cast_fp16 = reshape(shape = var_1087, x = var_1077_cast_fp16_to_fp16)[name = tensor("onehot_15_cast_fp16")]; + tensor var_1089_to_fp16 = const()[name = tensor("op_1089_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1091_cast_fp16 = sub(x = var_1089_to_fp16, y = onehot_15_cast_fp16)[name = tensor("op_1091_cast_fp16")]; + tensor k_cache3_to_fp16_dtype_0 = const()[name = tensor("k_cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache3_to_fp16 = cast(dtype = k_cache3_to_fp16_dtype_0, x = k_cache3)[name = tensor("cast_551")]; + tensor var_1092_cast_fp16 = mul(x = k_cache3_to_fp16, y = var_1091_cast_fp16)[name = tensor("op_1092_cast_fp16")]; + tensor var_1097 = const()[name = tensor("op_1097"), val = tensor([1, 1, 16, 64])]; + tensor var_1098_cast_fp16 = reshape(shape = var_1097, x = out_15_cast_fp16)[name = tensor("op_1098_cast_fp16")]; + tensor var_1099_cast_fp16 = mul(x = var_1098_cast_fp16, y = onehot_15_cast_fp16)[name = tensor("op_1099_cast_fp16")]; + tensor new_k_cache_7_cast_fp16 = add(x = var_1092_cast_fp16, y = var_1099_cast_fp16)[name = tensor("new_k_cache_7_cast_fp16")]; + tensor new_k_cache_7_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_7_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache3_to_fp16_dtype_0 = const()[name = tensor("v_cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache3_to_fp16 = cast(dtype = v_cache3_to_fp16_dtype_0, x = v_cache3)[name = tensor("cast_549")]; + tensor var_1105_cast_fp16 = mul(x = v_cache3_to_fp16, y = var_1091_cast_fp16)[name = tensor("op_1105_cast_fp16")]; + tensor var_1110 = const()[name = tensor("op_1110"), val = tensor([1, 1, 16, 64])]; + tensor var_1111_cast_fp16 = reshape(shape = var_1110, x = v_7_cast_fp16)[name = tensor("op_1111_cast_fp16")]; + tensor var_1112_cast_fp16 = mul(x = var_1111_cast_fp16, y = onehot_15_cast_fp16)[name = tensor("op_1112_cast_fp16")]; + tensor new_v_cache_7_cast_fp16 = add(x = var_1105_cast_fp16, y = var_1112_cast_fp16)[name = tensor("new_v_cache_7_cast_fp16")]; + tensor new_v_cache_7_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_7_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_3_perm_0 = const()[name = tensor("transpose_3_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_1125 = const()[name = tensor("op_1125"), val = tensor([0, 2, 1, 3])]; + tensor var_1131 = const()[name = tensor("op_1131"), val = tensor([1, 16, 1, 64])]; + tensor q4_7_cast_fp16 = reshape(shape = var_1131, x = out_13_cast_fp16)[name = tensor("q4_7_cast_fp16")]; + tensor var_1136_transpose_x_0 = const()[name = tensor("op_1136_transpose_x_0"), val = tensor(false)]; + tensor var_1136_transpose_y_0 = const()[name = tensor("op_1136_transpose_y_0"), val = tensor(false)]; + tensor transpose_3_cast_fp16 = transpose(perm = transpose_3_perm_0, x = new_k_cache_7_cast_fp16)[name = tensor("transpose_86")]; + tensor var_1136_cast_fp16 = matmul(transpose_x = var_1136_transpose_x_0, transpose_y = var_1136_transpose_y_0, x = q4_7_cast_fp16, y = transpose_3_cast_fp16)[name = tensor("op_1136_cast_fp16")]; + tensor var_1137_to_fp16 = const()[name = tensor("op_1137_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_13_cast_fp16 = mul(x = var_1136_cast_fp16, y = var_1137_to_fp16)[name = tensor("scores_13_cast_fp16")]; + tensor var_1141 = const()[name = tensor("op_1141"), val = tensor([1, 1])]; + tensor var_1142_cast_fp16 = reshape(shape = var_1141, x = position3_to_fp16)[name = tensor("op_1142_cast_fp16")]; + tensor var_1143_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_1142_cast_fp16)[name = tensor("op_1143_cast_fp16")]; + tensor mask_7_to_fp16_dtype_0 = const()[name = tensor("mask_7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_1150_to_fp16 = const()[name = tensor("op_1150_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1143_cast_fp16_to_fp16 = cast(dtype = mask_7_to_fp16_dtype_0, x = var_1143_cast_fp16)[name = tensor("cast_547")]; + tensor var_1151_cast_fp16 = sub(x = var_1143_cast_fp16_to_fp16, y = var_1150_to_fp16)[name = tensor("op_1151_cast_fp16")]; + tensor var_1156 = const()[name = tensor("op_1156"), val = tensor([1, 1, 1, 512])]; + tensor var_1157_cast_fp16 = reshape(shape = var_1156, x = var_1151_cast_fp16)[name = tensor("op_1157_cast_fp16")]; + tensor var_1158_to_fp16 = const()[name = tensor("op_1158_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_1159_cast_fp16 = mul(x = var_1157_cast_fp16, y = var_1158_to_fp16)[name = tensor("op_1159_cast_fp16")]; + tensor scores_15_cast_fp16 = add(x = scores_13_cast_fp16, y = var_1159_cast_fp16)[name = tensor("scores_15_cast_fp16")]; + tensor var_1162 = const()[name = tensor("op_1162"), val = tensor(-1)]; + tensor weights_7_cast_fp16 = softmax(axis = var_1162, x = scores_15_cast_fp16)[name = tensor("weights_7_cast_fp16")]; + tensor attn_7_transpose_x_0 = const()[name = tensor("attn_7_transpose_x_0"), val = tensor(false)]; + tensor attn_7_transpose_y_0 = const()[name = tensor("attn_7_transpose_y_0"), val = tensor(false)]; + tensor values_7_cast_fp16 = transpose(perm = var_1125, x = new_v_cache_7_cast_fp16)[name = tensor("transpose_85")]; + tensor attn_7_cast_fp16 = matmul(transpose_x = attn_7_transpose_x_0, transpose_y = attn_7_transpose_y_0, x = weights_7_cast_fp16, y = values_7_cast_fp16)[name = tensor("attn_7_cast_fp16")]; + tensor var_1170 = const()[name = tensor("op_1170"), val = tensor([0, 2, 1, 3])]; + tensor var_1175 = const()[name = tensor("op_1175"), val = tensor([1, 1, 1024])]; + tensor var_1171_cast_fp16 = transpose(perm = var_1170, x = attn_7_cast_fp16)[name = tensor("transpose_84")]; + tensor input_43_cast_fp16 = reshape(shape = var_1175, x = var_1171_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor attn3_out_proj_weight_to_fp16 = const()[name = tensor("attn3_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(81902784)))]; + tensor linear_14_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn3_out_proj_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor var_1181_to_fp16 = const()[name = tensor("op_1181_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1182_cast_fp16 = add(x = position3_to_fp16, y = var_1181_to_fp16)[name = tensor("op_1182_cast_fp16")]; + tensor var_1182_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1182_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_45_cast_fp16 = add(x = input_39_cast_fp16, y = linear_14_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor input_47_axes_0 = const()[name = tensor("input_47_axes_0"), val = tensor([-1])]; + tensor norm3_2_weight_to_fp16 = const()[name = tensor("norm3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84000000)))]; + tensor norm3_2_bias_to_fp16 = const()[name = tensor("norm3_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84002112)))]; + tensor var_1186_to_fp16 = const()[name = tensor("op_1186_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_47_cast_fp16 = layer_norm(axes = input_47_axes_0, beta = norm3_2_bias_to_fp16, epsilon = var_1186_to_fp16, gamma = norm3_2_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor linear3_1_weight_to_fp16 = const()[name = tensor("linear3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84004224)))]; + tensor linear_15_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear3_1_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor input_49_mode_0 = const()[name = tensor("input_49_mode_0"), val = tensor("EXACT")]; + tensor input_49_cast_fp16 = gelu(mode = input_49_mode_0, x = linear_15_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor linear3_2_weight_to_fp16 = const()[name = tensor("linear3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(92392896)))]; + tensor linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear3_2_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor input_51_cast_fp16 = add(x = input_45_cast_fp16, y = linear_16_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor input_53_axes_0 = const()[name = tensor("input_53_axes_0"), val = tensor([-1])]; + tensor norm4_1_weight_to_fp16 = const()[name = tensor("norm4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100781568)))]; + tensor norm4_1_bias_to_fp16 = const()[name = tensor("norm4_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100783680)))]; + tensor var_1203_to_fp16 = const()[name = tensor("op_1203_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_53_cast_fp16 = layer_norm(axes = input_53_axes_0, beta = norm4_1_bias_to_fp16, epsilon = var_1203_to_fp16, gamma = norm4_1_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor attn4_in_proj_weight_to_fp16 = const()[name = tensor("attn4_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100785792)))]; + tensor linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn4_in_proj_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor var_1216 = const()[name = tensor("op_1216"), val = tensor([1, 3, 16, 64])]; + tensor qkv_9_cast_fp16 = reshape(shape = var_1216, x = linear_17_cast_fp16)[name = tensor("qkv_9_cast_fp16")]; + tensor x_17_begin_0 = const()[name = tensor("x_17_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_17_end_0 = const()[name = tensor("x_17_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_17_end_mask_0 = const()[name = tensor("x_17_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_17_squeeze_mask_0 = const()[name = tensor("x_17_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_17_cast_fp16 = slice_by_index(begin = x_17_begin_0, end = x_17_end_0, end_mask = x_17_end_mask_0, squeeze_mask = x_17_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("x_17_cast_fp16")]; + tensor x_19_begin_0 = const()[name = tensor("x_19_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_19_end_0 = const()[name = tensor("x_19_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_19_end_mask_0 = const()[name = tensor("x_19_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_19_squeeze_mask_0 = const()[name = tensor("x_19_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_19_cast_fp16 = slice_by_index(begin = x_19_begin_0, end = x_19_end_0, end_mask = x_19_end_mask_0, squeeze_mask = x_19_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("x_19_cast_fp16")]; + tensor v_9_begin_0 = const()[name = tensor("v_9_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_9_end_0 = const()[name = tensor("v_9_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_9_end_mask_0 = const()[name = tensor("v_9_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_9_squeeze_mask_0 = const()[name = tensor("v_9_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_9_cast_fp16 = slice_by_index(begin = v_9_begin_0, end = v_9_end_0, end_mask = v_9_end_mask_0, squeeze_mask = v_9_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("v_9_cast_fp16")]; + tensor var_1245 = const()[name = tensor("op_1245"), val = tensor([1, 1, 1])]; + tensor position4_to_fp16_dtype_0 = const()[name = tensor("position4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position4_to_fp16 = cast(dtype = position4_to_fp16_dtype_0, x = position4)[name = tensor("cast_545")]; + tensor var_1246_cast_fp16 = reshape(shape = var_1245, x = position4_to_fp16)[name = tensor("op_1246_cast_fp16")]; + tensor angles_17_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_1246_cast_fp16)[name = tensor("angles_17_cast_fp16")]; + tensor rotr_17_cast_fp16 = cos(x = angles_17_cast_fp16)[name = tensor("rotr_17_cast_fp16")]; + tensor roti_17_cast_fp16 = sin(x = angles_17_cast_fp16)[name = tensor("roti_17_cast_fp16")]; + tensor var_1254 = const()[name = tensor("op_1254"), val = tensor([1, 16, 32, 2])]; + tensor pairs_17_cast_fp16 = reshape(shape = var_1254, x = x_17_cast_fp16)[name = tensor("pairs_17_cast_fp16")]; + tensor xr_17_begin_0 = const()[name = tensor("xr_17_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_17_end_0 = const()[name = tensor("xr_17_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_17_end_mask_0 = const()[name = tensor("xr_17_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_17_squeeze_mask_0 = const()[name = tensor("xr_17_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_17_cast_fp16 = slice_by_index(begin = xr_17_begin_0, end = xr_17_end_0, end_mask = xr_17_end_mask_0, squeeze_mask = xr_17_squeeze_mask_0, x = pairs_17_cast_fp16)[name = tensor("xr_17_cast_fp16")]; + tensor xi_17_begin_0 = const()[name = tensor("xi_17_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_17_end_0 = const()[name = tensor("xi_17_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_17_end_mask_0 = const()[name = tensor("xi_17_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_17_squeeze_mask_0 = const()[name = tensor("xi_17_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_17_cast_fp16 = slice_by_index(begin = xi_17_begin_0, end = xi_17_end_0, end_mask = xi_17_end_mask_0, squeeze_mask = xi_17_squeeze_mask_0, x = pairs_17_cast_fp16)[name = tensor("xi_17_cast_fp16")]; + tensor var_1262_cast_fp16 = mul(x = xr_17_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_1262_cast_fp16")]; + tensor var_1263_cast_fp16 = mul(x = xi_17_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_1263_cast_fp16")]; + tensor xor__17_cast_fp16 = sub(x = var_1262_cast_fp16, y = var_1263_cast_fp16)[name = tensor("xor__17_cast_fp16")]; + tensor var_1266_cast_fp16 = mul(x = xr_17_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_1266_cast_fp16")]; + tensor var_1267_cast_fp16 = mul(x = xi_17_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_1267_cast_fp16")]; + tensor xoi_17_cast_fp16 = add(x = var_1266_cast_fp16, y = var_1267_cast_fp16)[name = tensor("xoi_17_cast_fp16")]; + tensor out_17_axis_0 = const()[name = tensor("out_17_axis_0"), val = tensor(-1)]; + tensor out_17_cast_fp16 = stack(axis = out_17_axis_0, values = (xor__17_cast_fp16, xoi_17_cast_fp16))[name = tensor("out_17_cast_fp16")]; + tensor var_1290 = const()[name = tensor("op_1290"), val = tensor([1, 16, 32, 2])]; + tensor pairs_19_cast_fp16 = reshape(shape = var_1290, x = x_19_cast_fp16)[name = tensor("pairs_19_cast_fp16")]; + tensor xr_19_begin_0 = const()[name = tensor("xr_19_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_19_end_0 = const()[name = tensor("xr_19_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_19_end_mask_0 = const()[name = tensor("xr_19_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_19_squeeze_mask_0 = const()[name = tensor("xr_19_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_19_cast_fp16 = slice_by_index(begin = xr_19_begin_0, end = xr_19_end_0, end_mask = xr_19_end_mask_0, squeeze_mask = xr_19_squeeze_mask_0, x = pairs_19_cast_fp16)[name = tensor("xr_19_cast_fp16")]; + tensor xi_19_begin_0 = const()[name = tensor("xi_19_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_19_end_0 = const()[name = tensor("xi_19_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_19_end_mask_0 = const()[name = tensor("xi_19_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_19_squeeze_mask_0 = const()[name = tensor("xi_19_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_19_cast_fp16 = slice_by_index(begin = xi_19_begin_0, end = xi_19_end_0, end_mask = xi_19_end_mask_0, squeeze_mask = xi_19_squeeze_mask_0, x = pairs_19_cast_fp16)[name = tensor("xi_19_cast_fp16")]; + tensor var_1298_cast_fp16 = mul(x = xr_19_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_1298_cast_fp16")]; + tensor var_1299_cast_fp16 = mul(x = xi_19_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_1299_cast_fp16")]; + tensor xor__19_cast_fp16 = sub(x = var_1298_cast_fp16, y = var_1299_cast_fp16)[name = tensor("xor__19_cast_fp16")]; + tensor var_1302_cast_fp16 = mul(x = xr_19_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_1302_cast_fp16")]; + tensor var_1303_cast_fp16 = mul(x = xi_19_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_1303_cast_fp16")]; + tensor xoi_19_cast_fp16 = add(x = var_1302_cast_fp16, y = var_1303_cast_fp16)[name = tensor("xoi_19_cast_fp16")]; + tensor out_19_axis_0 = const()[name = tensor("out_19_axis_0"), val = tensor(-1)]; + tensor out_19_cast_fp16 = stack(axis = out_19_axis_0, values = (xor__19_cast_fp16, xoi_19_cast_fp16))[name = tensor("out_19_cast_fp16")]; + tensor _inversed_1315_y_0_to_fp16 = const()[name = tensor("_inversed_1315_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1315_cast_fp16 = mul(x = position4_to_fp16, y = _inversed_1315_y_0_to_fp16)[name = tensor("_inversed_1315_cast_fp16")]; + tensor var_1316_cast_fp16 = floor(x = _inversed_1315_cast_fp16)[name = tensor("op_1316_cast_fp16")]; + tensor var_1317_to_fp16 = const()[name = tensor("op_1317_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1318_cast_fp16 = mul(x = var_1316_cast_fp16, y = var_1317_to_fp16)[name = tensor("op_1318_cast_fp16")]; + tensor write_idx_9_cast_fp16 = sub(x = position4_to_fp16, y = var_1318_cast_fp16)[name = tensor("write_idx_9_cast_fp16")]; + tensor var_1321_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_9_cast_fp16)[name = tensor("op_1321_cast_fp16")]; + tensor var_1331 = const()[name = tensor("op_1331"), val = tensor([1, 512, 1, 1])]; + tensor onehot_17_to_fp16_dtype_0 = const()[name = tensor("onehot_17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_1321_cast_fp16_to_fp16 = cast(dtype = onehot_17_to_fp16_dtype_0, x = var_1321_cast_fp16)[name = tensor("cast_544")]; + tensor onehot_19_cast_fp16 = reshape(shape = var_1331, x = var_1321_cast_fp16_to_fp16)[name = tensor("onehot_19_cast_fp16")]; + tensor var_1333_to_fp16 = const()[name = tensor("op_1333_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1335_cast_fp16 = sub(x = var_1333_to_fp16, y = onehot_19_cast_fp16)[name = tensor("op_1335_cast_fp16")]; + tensor k_cache4_to_fp16_dtype_0 = const()[name = tensor("k_cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache4_to_fp16 = cast(dtype = k_cache4_to_fp16_dtype_0, x = k_cache4)[name = tensor("cast_543")]; + tensor var_1336_cast_fp16 = mul(x = k_cache4_to_fp16, y = var_1335_cast_fp16)[name = tensor("op_1336_cast_fp16")]; + tensor var_1341 = const()[name = tensor("op_1341"), val = tensor([1, 1, 16, 64])]; + tensor var_1342_cast_fp16 = reshape(shape = var_1341, x = out_19_cast_fp16)[name = tensor("op_1342_cast_fp16")]; + tensor var_1343_cast_fp16 = mul(x = var_1342_cast_fp16, y = onehot_19_cast_fp16)[name = tensor("op_1343_cast_fp16")]; + tensor new_k_cache_9_cast_fp16 = add(x = var_1336_cast_fp16, y = var_1343_cast_fp16)[name = tensor("new_k_cache_9_cast_fp16")]; + tensor new_k_cache_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache4_to_fp16_dtype_0 = const()[name = tensor("v_cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache4_to_fp16 = cast(dtype = v_cache4_to_fp16_dtype_0, x = v_cache4)[name = tensor("cast_541")]; + tensor var_1349_cast_fp16 = mul(x = v_cache4_to_fp16, y = var_1335_cast_fp16)[name = tensor("op_1349_cast_fp16")]; + tensor var_1354 = const()[name = tensor("op_1354"), val = tensor([1, 1, 16, 64])]; + tensor var_1355_cast_fp16 = reshape(shape = var_1354, x = v_9_cast_fp16)[name = tensor("op_1355_cast_fp16")]; + tensor var_1356_cast_fp16 = mul(x = var_1355_cast_fp16, y = onehot_19_cast_fp16)[name = tensor("op_1356_cast_fp16")]; + tensor new_v_cache_9_cast_fp16 = add(x = var_1349_cast_fp16, y = var_1356_cast_fp16)[name = tensor("new_v_cache_9_cast_fp16")]; + tensor new_v_cache_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_4_perm_0 = const()[name = tensor("transpose_4_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_1369 = const()[name = tensor("op_1369"), val = tensor([0, 2, 1, 3])]; + tensor var_1375 = const()[name = tensor("op_1375"), val = tensor([1, 16, 1, 64])]; + tensor q4_9_cast_fp16 = reshape(shape = var_1375, x = out_17_cast_fp16)[name = tensor("q4_9_cast_fp16")]; + tensor var_1380_transpose_x_0 = const()[name = tensor("op_1380_transpose_x_0"), val = tensor(false)]; + tensor var_1380_transpose_y_0 = const()[name = tensor("op_1380_transpose_y_0"), val = tensor(false)]; + tensor transpose_4_cast_fp16 = transpose(perm = transpose_4_perm_0, x = new_k_cache_9_cast_fp16)[name = tensor("transpose_83")]; + tensor var_1380_cast_fp16 = matmul(transpose_x = var_1380_transpose_x_0, transpose_y = var_1380_transpose_y_0, x = q4_9_cast_fp16, y = transpose_4_cast_fp16)[name = tensor("op_1380_cast_fp16")]; + tensor var_1381_to_fp16 = const()[name = tensor("op_1381_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_17_cast_fp16 = mul(x = var_1380_cast_fp16, y = var_1381_to_fp16)[name = tensor("scores_17_cast_fp16")]; + tensor var_1385 = const()[name = tensor("op_1385"), val = tensor([1, 1])]; + tensor var_1386_cast_fp16 = reshape(shape = var_1385, x = position4_to_fp16)[name = tensor("op_1386_cast_fp16")]; + tensor var_1387_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_1386_cast_fp16)[name = tensor("op_1387_cast_fp16")]; + tensor mask_9_to_fp16_dtype_0 = const()[name = tensor("mask_9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_1394_to_fp16 = const()[name = tensor("op_1394_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1387_cast_fp16_to_fp16 = cast(dtype = mask_9_to_fp16_dtype_0, x = var_1387_cast_fp16)[name = tensor("cast_539")]; + tensor var_1395_cast_fp16 = sub(x = var_1387_cast_fp16_to_fp16, y = var_1394_to_fp16)[name = tensor("op_1395_cast_fp16")]; + tensor var_1400 = const()[name = tensor("op_1400"), val = tensor([1, 1, 1, 512])]; + tensor var_1401_cast_fp16 = reshape(shape = var_1400, x = var_1395_cast_fp16)[name = tensor("op_1401_cast_fp16")]; + tensor var_1402_to_fp16 = const()[name = tensor("op_1402_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_1403_cast_fp16 = mul(x = var_1401_cast_fp16, y = var_1402_to_fp16)[name = tensor("op_1403_cast_fp16")]; + tensor scores_19_cast_fp16 = add(x = scores_17_cast_fp16, y = var_1403_cast_fp16)[name = tensor("scores_19_cast_fp16")]; + tensor var_1406 = const()[name = tensor("op_1406"), val = tensor(-1)]; + tensor weights_9_cast_fp16 = softmax(axis = var_1406, x = scores_19_cast_fp16)[name = tensor("weights_9_cast_fp16")]; + tensor attn_9_transpose_x_0 = const()[name = tensor("attn_9_transpose_x_0"), val = tensor(false)]; + tensor attn_9_transpose_y_0 = const()[name = tensor("attn_9_transpose_y_0"), val = tensor(false)]; + tensor values_9_cast_fp16 = transpose(perm = var_1369, x = new_v_cache_9_cast_fp16)[name = tensor("transpose_82")]; + tensor attn_9_cast_fp16 = matmul(transpose_x = attn_9_transpose_x_0, transpose_y = attn_9_transpose_y_0, x = weights_9_cast_fp16, y = values_9_cast_fp16)[name = tensor("attn_9_cast_fp16")]; + tensor var_1414 = const()[name = tensor("op_1414"), val = tensor([0, 2, 1, 3])]; + tensor var_1419 = const()[name = tensor("op_1419"), val = tensor([1, 1, 1024])]; + tensor var_1415_cast_fp16 = transpose(perm = var_1414, x = attn_9_cast_fp16)[name = tensor("transpose_81")]; + tensor input_55_cast_fp16 = reshape(shape = var_1419, x = var_1415_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor attn4_out_proj_weight_to_fp16 = const()[name = tensor("attn4_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(107077312)))]; + tensor linear_18_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn4_out_proj_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor var_1425_to_fp16 = const()[name = tensor("op_1425_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1426_cast_fp16 = add(x = position4_to_fp16, y = var_1425_to_fp16)[name = tensor("op_1426_cast_fp16")]; + tensor var_1426_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1426_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_57_cast_fp16 = add(x = input_51_cast_fp16, y = linear_18_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor input_59_axes_0 = const()[name = tensor("input_59_axes_0"), val = tensor([-1])]; + tensor norm4_2_weight_to_fp16 = const()[name = tensor("norm4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109174528)))]; + tensor norm4_2_bias_to_fp16 = const()[name = tensor("norm4_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109176640)))]; + tensor var_1430_to_fp16 = const()[name = tensor("op_1430_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_59_cast_fp16 = layer_norm(axes = input_59_axes_0, beta = norm4_2_bias_to_fp16, epsilon = var_1430_to_fp16, gamma = norm4_2_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor linear4_1_weight_to_fp16 = const()[name = tensor("linear4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109178752)))]; + tensor linear_19_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear4_1_weight_to_fp16, x = input_59_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_61_mode_0 = const()[name = tensor("input_61_mode_0"), val = tensor("EXACT")]; + tensor input_61_cast_fp16 = gelu(mode = input_61_mode_0, x = linear_19_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor linear4_2_weight_to_fp16 = const()[name = tensor("linear4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(117567424)))]; + tensor linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear4_2_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor input_63_cast_fp16 = add(x = input_57_cast_fp16, y = linear_20_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor input_65_axes_0 = const()[name = tensor("input_65_axes_0"), val = tensor([-1])]; + tensor norm5_1_weight_to_fp16 = const()[name = tensor("norm5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125956096)))]; + tensor norm5_1_bias_to_fp16 = const()[name = tensor("norm5_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125958208)))]; + tensor var_1447_to_fp16 = const()[name = tensor("op_1447_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_65_cast_fp16 = layer_norm(axes = input_65_axes_0, beta = norm5_1_bias_to_fp16, epsilon = var_1447_to_fp16, gamma = norm5_1_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("input_65_cast_fp16")]; + tensor attn5_in_proj_weight_to_fp16 = const()[name = tensor("attn5_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125960320)))]; + tensor linear_21_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn5_in_proj_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("linear_21_cast_fp16")]; + tensor var_1460 = const()[name = tensor("op_1460"), val = tensor([1, 3, 16, 64])]; + tensor qkv_11_cast_fp16 = reshape(shape = var_1460, x = linear_21_cast_fp16)[name = tensor("qkv_11_cast_fp16")]; + tensor x_21_begin_0 = const()[name = tensor("x_21_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_21_end_0 = const()[name = tensor("x_21_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_21_end_mask_0 = const()[name = tensor("x_21_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_21_squeeze_mask_0 = const()[name = tensor("x_21_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_21_cast_fp16 = slice_by_index(begin = x_21_begin_0, end = x_21_end_0, end_mask = x_21_end_mask_0, squeeze_mask = x_21_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("x_21_cast_fp16")]; + tensor x_23_begin_0 = const()[name = tensor("x_23_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_23_end_0 = const()[name = tensor("x_23_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_23_end_mask_0 = const()[name = tensor("x_23_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_23_squeeze_mask_0 = const()[name = tensor("x_23_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_23_cast_fp16 = slice_by_index(begin = x_23_begin_0, end = x_23_end_0, end_mask = x_23_end_mask_0, squeeze_mask = x_23_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("x_23_cast_fp16")]; + tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_11_cast_fp16 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("v_11_cast_fp16")]; + tensor var_1489 = const()[name = tensor("op_1489"), val = tensor([1, 1, 1])]; + tensor position5_to_fp16_dtype_0 = const()[name = tensor("position5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position5_to_fp16 = cast(dtype = position5_to_fp16_dtype_0, x = position5)[name = tensor("cast_537")]; + tensor var_1490_cast_fp16 = reshape(shape = var_1489, x = position5_to_fp16)[name = tensor("op_1490_cast_fp16")]; + tensor angles_21_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_1490_cast_fp16)[name = tensor("angles_21_cast_fp16")]; + tensor rotr_21_cast_fp16 = cos(x = angles_21_cast_fp16)[name = tensor("rotr_21_cast_fp16")]; + tensor roti_21_cast_fp16 = sin(x = angles_21_cast_fp16)[name = tensor("roti_21_cast_fp16")]; + tensor var_1498 = const()[name = tensor("op_1498"), val = tensor([1, 16, 32, 2])]; + tensor pairs_21_cast_fp16 = reshape(shape = var_1498, x = x_21_cast_fp16)[name = tensor("pairs_21_cast_fp16")]; + tensor xr_21_begin_0 = const()[name = tensor("xr_21_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_21_end_0 = const()[name = tensor("xr_21_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_21_end_mask_0 = const()[name = tensor("xr_21_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_21_squeeze_mask_0 = const()[name = tensor("xr_21_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_21_cast_fp16 = slice_by_index(begin = xr_21_begin_0, end = xr_21_end_0, end_mask = xr_21_end_mask_0, squeeze_mask = xr_21_squeeze_mask_0, x = pairs_21_cast_fp16)[name = tensor("xr_21_cast_fp16")]; + tensor xi_21_begin_0 = const()[name = tensor("xi_21_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_21_end_0 = const()[name = tensor("xi_21_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_21_end_mask_0 = const()[name = tensor("xi_21_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_21_squeeze_mask_0 = const()[name = tensor("xi_21_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_21_cast_fp16 = slice_by_index(begin = xi_21_begin_0, end = xi_21_end_0, end_mask = xi_21_end_mask_0, squeeze_mask = xi_21_squeeze_mask_0, x = pairs_21_cast_fp16)[name = tensor("xi_21_cast_fp16")]; + tensor var_1506_cast_fp16 = mul(x = xr_21_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_1506_cast_fp16")]; + tensor var_1507_cast_fp16 = mul(x = xi_21_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_1507_cast_fp16")]; + tensor xor__21_cast_fp16 = sub(x = var_1506_cast_fp16, y = var_1507_cast_fp16)[name = tensor("xor__21_cast_fp16")]; + tensor var_1510_cast_fp16 = mul(x = xr_21_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_1510_cast_fp16")]; + tensor var_1511_cast_fp16 = mul(x = xi_21_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_1511_cast_fp16")]; + tensor xoi_21_cast_fp16 = add(x = var_1510_cast_fp16, y = var_1511_cast_fp16)[name = tensor("xoi_21_cast_fp16")]; + tensor out_21_axis_0 = const()[name = tensor("out_21_axis_0"), val = tensor(-1)]; + tensor out_21_cast_fp16 = stack(axis = out_21_axis_0, values = (xor__21_cast_fp16, xoi_21_cast_fp16))[name = tensor("out_21_cast_fp16")]; + tensor var_1534 = const()[name = tensor("op_1534"), val = tensor([1, 16, 32, 2])]; + tensor pairs_23_cast_fp16 = reshape(shape = var_1534, x = x_23_cast_fp16)[name = tensor("pairs_23_cast_fp16")]; + tensor xr_23_begin_0 = const()[name = tensor("xr_23_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_23_end_0 = const()[name = tensor("xr_23_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_23_end_mask_0 = const()[name = tensor("xr_23_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_23_squeeze_mask_0 = const()[name = tensor("xr_23_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_23_cast_fp16 = slice_by_index(begin = xr_23_begin_0, end = xr_23_end_0, end_mask = xr_23_end_mask_0, squeeze_mask = xr_23_squeeze_mask_0, x = pairs_23_cast_fp16)[name = tensor("xr_23_cast_fp16")]; + tensor xi_23_begin_0 = const()[name = tensor("xi_23_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_23_end_0 = const()[name = tensor("xi_23_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_23_end_mask_0 = const()[name = tensor("xi_23_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_23_squeeze_mask_0 = const()[name = tensor("xi_23_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_23_cast_fp16 = slice_by_index(begin = xi_23_begin_0, end = xi_23_end_0, end_mask = xi_23_end_mask_0, squeeze_mask = xi_23_squeeze_mask_0, x = pairs_23_cast_fp16)[name = tensor("xi_23_cast_fp16")]; + tensor var_1542_cast_fp16 = mul(x = xr_23_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_1542_cast_fp16")]; + tensor var_1543_cast_fp16 = mul(x = xi_23_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_1543_cast_fp16")]; + tensor xor__23_cast_fp16 = sub(x = var_1542_cast_fp16, y = var_1543_cast_fp16)[name = tensor("xor__23_cast_fp16")]; + tensor var_1546_cast_fp16 = mul(x = xr_23_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_1546_cast_fp16")]; + tensor var_1547_cast_fp16 = mul(x = xi_23_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_1547_cast_fp16")]; + tensor xoi_23_cast_fp16 = add(x = var_1546_cast_fp16, y = var_1547_cast_fp16)[name = tensor("xoi_23_cast_fp16")]; + tensor out_23_axis_0 = const()[name = tensor("out_23_axis_0"), val = tensor(-1)]; + tensor out_23_cast_fp16 = stack(axis = out_23_axis_0, values = (xor__23_cast_fp16, xoi_23_cast_fp16))[name = tensor("out_23_cast_fp16")]; + tensor _inversed_1559_y_0_to_fp16 = const()[name = tensor("_inversed_1559_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1559_cast_fp16 = mul(x = position5_to_fp16, y = _inversed_1559_y_0_to_fp16)[name = tensor("_inversed_1559_cast_fp16")]; + tensor var_1560_cast_fp16 = floor(x = _inversed_1559_cast_fp16)[name = tensor("op_1560_cast_fp16")]; + tensor var_1561_to_fp16 = const()[name = tensor("op_1561_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1562_cast_fp16 = mul(x = var_1560_cast_fp16, y = var_1561_to_fp16)[name = tensor("op_1562_cast_fp16")]; + tensor write_idx_11_cast_fp16 = sub(x = position5_to_fp16, y = var_1562_cast_fp16)[name = tensor("write_idx_11_cast_fp16")]; + tensor var_1565_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_11_cast_fp16)[name = tensor("op_1565_cast_fp16")]; + tensor var_1575 = const()[name = tensor("op_1575"), val = tensor([1, 512, 1, 1])]; + tensor onehot_21_to_fp16_dtype_0 = const()[name = tensor("onehot_21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_1565_cast_fp16_to_fp16 = cast(dtype = onehot_21_to_fp16_dtype_0, x = var_1565_cast_fp16)[name = tensor("cast_536")]; + tensor onehot_23_cast_fp16 = reshape(shape = var_1575, x = var_1565_cast_fp16_to_fp16)[name = tensor("onehot_23_cast_fp16")]; + tensor var_1577_to_fp16 = const()[name = tensor("op_1577_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1579_cast_fp16 = sub(x = var_1577_to_fp16, y = onehot_23_cast_fp16)[name = tensor("op_1579_cast_fp16")]; + tensor k_cache5_to_fp16_dtype_0 = const()[name = tensor("k_cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache5_to_fp16 = cast(dtype = k_cache5_to_fp16_dtype_0, x = k_cache5)[name = tensor("cast_535")]; + tensor var_1580_cast_fp16 = mul(x = k_cache5_to_fp16, y = var_1579_cast_fp16)[name = tensor("op_1580_cast_fp16")]; + tensor var_1585 = const()[name = tensor("op_1585"), val = tensor([1, 1, 16, 64])]; + tensor var_1586_cast_fp16 = reshape(shape = var_1585, x = out_23_cast_fp16)[name = tensor("op_1586_cast_fp16")]; + tensor var_1587_cast_fp16 = mul(x = var_1586_cast_fp16, y = onehot_23_cast_fp16)[name = tensor("op_1587_cast_fp16")]; + tensor new_k_cache_11_cast_fp16 = add(x = var_1580_cast_fp16, y = var_1587_cast_fp16)[name = tensor("new_k_cache_11_cast_fp16")]; + tensor new_k_cache_11_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_11_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache5_to_fp16_dtype_0 = const()[name = tensor("v_cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache5_to_fp16 = cast(dtype = v_cache5_to_fp16_dtype_0, x = v_cache5)[name = tensor("cast_533")]; + tensor var_1593_cast_fp16 = mul(x = v_cache5_to_fp16, y = var_1579_cast_fp16)[name = tensor("op_1593_cast_fp16")]; + tensor var_1598 = const()[name = tensor("op_1598"), val = tensor([1, 1, 16, 64])]; + tensor var_1599_cast_fp16 = reshape(shape = var_1598, x = v_11_cast_fp16)[name = tensor("op_1599_cast_fp16")]; + tensor var_1600_cast_fp16 = mul(x = var_1599_cast_fp16, y = onehot_23_cast_fp16)[name = tensor("op_1600_cast_fp16")]; + tensor new_v_cache_11_cast_fp16 = add(x = var_1593_cast_fp16, y = var_1600_cast_fp16)[name = tensor("new_v_cache_11_cast_fp16")]; + tensor new_v_cache_11_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_11_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_5_perm_0 = const()[name = tensor("transpose_5_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_1613 = const()[name = tensor("op_1613"), val = tensor([0, 2, 1, 3])]; + tensor var_1619 = const()[name = tensor("op_1619"), val = tensor([1, 16, 1, 64])]; + tensor q4_11_cast_fp16 = reshape(shape = var_1619, x = out_21_cast_fp16)[name = tensor("q4_11_cast_fp16")]; + tensor var_1624_transpose_x_0 = const()[name = tensor("op_1624_transpose_x_0"), val = tensor(false)]; + tensor var_1624_transpose_y_0 = const()[name = tensor("op_1624_transpose_y_0"), val = tensor(false)]; + tensor transpose_5_cast_fp16 = transpose(perm = transpose_5_perm_0, x = new_k_cache_11_cast_fp16)[name = tensor("transpose_80")]; + tensor var_1624_cast_fp16 = matmul(transpose_x = var_1624_transpose_x_0, transpose_y = var_1624_transpose_y_0, x = q4_11_cast_fp16, y = transpose_5_cast_fp16)[name = tensor("op_1624_cast_fp16")]; + tensor var_1625_to_fp16 = const()[name = tensor("op_1625_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_21_cast_fp16 = mul(x = var_1624_cast_fp16, y = var_1625_to_fp16)[name = tensor("scores_21_cast_fp16")]; + tensor var_1629 = const()[name = tensor("op_1629"), val = tensor([1, 1])]; + tensor var_1630_cast_fp16 = reshape(shape = var_1629, x = position5_to_fp16)[name = tensor("op_1630_cast_fp16")]; + tensor var_1631_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_1630_cast_fp16)[name = tensor("op_1631_cast_fp16")]; + tensor mask_11_to_fp16_dtype_0 = const()[name = tensor("mask_11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_1638_to_fp16 = const()[name = tensor("op_1638_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1631_cast_fp16_to_fp16 = cast(dtype = mask_11_to_fp16_dtype_0, x = var_1631_cast_fp16)[name = tensor("cast_531")]; + tensor var_1639_cast_fp16 = sub(x = var_1631_cast_fp16_to_fp16, y = var_1638_to_fp16)[name = tensor("op_1639_cast_fp16")]; + tensor var_1644 = const()[name = tensor("op_1644"), val = tensor([1, 1, 1, 512])]; + tensor var_1645_cast_fp16 = reshape(shape = var_1644, x = var_1639_cast_fp16)[name = tensor("op_1645_cast_fp16")]; + tensor var_1646_to_fp16 = const()[name = tensor("op_1646_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_1647_cast_fp16 = mul(x = var_1645_cast_fp16, y = var_1646_to_fp16)[name = tensor("op_1647_cast_fp16")]; + tensor scores_23_cast_fp16 = add(x = scores_21_cast_fp16, y = var_1647_cast_fp16)[name = tensor("scores_23_cast_fp16")]; + tensor var_1650 = const()[name = tensor("op_1650"), val = tensor(-1)]; + tensor weights_11_cast_fp16 = softmax(axis = var_1650, x = scores_23_cast_fp16)[name = tensor("weights_11_cast_fp16")]; + tensor attn_11_transpose_x_0 = const()[name = tensor("attn_11_transpose_x_0"), val = tensor(false)]; + tensor attn_11_transpose_y_0 = const()[name = tensor("attn_11_transpose_y_0"), val = tensor(false)]; + tensor values_11_cast_fp16 = transpose(perm = var_1613, x = new_v_cache_11_cast_fp16)[name = tensor("transpose_79")]; + tensor attn_11_cast_fp16 = matmul(transpose_x = attn_11_transpose_x_0, transpose_y = attn_11_transpose_y_0, x = weights_11_cast_fp16, y = values_11_cast_fp16)[name = tensor("attn_11_cast_fp16")]; + tensor var_1658 = const()[name = tensor("op_1658"), val = tensor([0, 2, 1, 3])]; + tensor var_1663 = const()[name = tensor("op_1663"), val = tensor([1, 1, 1024])]; + tensor var_1659_cast_fp16 = transpose(perm = var_1658, x = attn_11_cast_fp16)[name = tensor("transpose_78")]; + tensor input_67_cast_fp16 = reshape(shape = var_1663, x = var_1659_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor attn5_out_proj_weight_to_fp16 = const()[name = tensor("attn5_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132251840)))]; + tensor linear_22_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn5_out_proj_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("linear_22_cast_fp16")]; + tensor var_1669_to_fp16 = const()[name = tensor("op_1669_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1670_cast_fp16 = add(x = position5_to_fp16, y = var_1669_to_fp16)[name = tensor("op_1670_cast_fp16")]; + tensor var_1670_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1670_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_69_cast_fp16 = add(x = input_63_cast_fp16, y = linear_22_cast_fp16)[name = tensor("input_69_cast_fp16")]; + tensor input_71_axes_0 = const()[name = tensor("input_71_axes_0"), val = tensor([-1])]; + tensor norm5_2_weight_to_fp16 = const()[name = tensor("norm5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134349056)))]; + tensor norm5_2_bias_to_fp16 = const()[name = tensor("norm5_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134351168)))]; + tensor var_1674_to_fp16 = const()[name = tensor("op_1674_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_71_cast_fp16 = layer_norm(axes = input_71_axes_0, beta = norm5_2_bias_to_fp16, epsilon = var_1674_to_fp16, gamma = norm5_2_weight_to_fp16, x = input_69_cast_fp16)[name = tensor("input_71_cast_fp16")]; + tensor linear5_1_weight_to_fp16 = const()[name = tensor("linear5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134353280)))]; + tensor linear_23_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear5_1_weight_to_fp16, x = input_71_cast_fp16)[name = tensor("linear_23_cast_fp16")]; + tensor input_73_mode_0 = const()[name = tensor("input_73_mode_0"), val = tensor("EXACT")]; + tensor input_73_cast_fp16 = gelu(mode = input_73_mode_0, x = linear_23_cast_fp16)[name = tensor("input_73_cast_fp16")]; + tensor linear5_2_weight_to_fp16 = const()[name = tensor("linear5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142741952)))]; + tensor linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear5_2_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("linear_24_cast_fp16")]; + tensor input_75_cast_fp16 = add(x = input_69_cast_fp16, y = linear_24_cast_fp16)[name = tensor("input_75_cast_fp16")]; + tensor input_77_axes_0 = const()[name = tensor("input_77_axes_0"), val = tensor([-1])]; + tensor norm6_1_weight_to_fp16 = const()[name = tensor("norm6_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151130624)))]; + tensor norm6_1_bias_to_fp16 = const()[name = tensor("norm6_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151132736)))]; + tensor var_1691_to_fp16 = const()[name = tensor("op_1691_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_77_cast_fp16 = layer_norm(axes = input_77_axes_0, beta = norm6_1_bias_to_fp16, epsilon = var_1691_to_fp16, gamma = norm6_1_weight_to_fp16, x = input_75_cast_fp16)[name = tensor("input_77_cast_fp16")]; + tensor attn6_in_proj_weight_to_fp16 = const()[name = tensor("attn6_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151134848)))]; + tensor linear_25_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn6_in_proj_weight_to_fp16, x = input_77_cast_fp16)[name = tensor("linear_25_cast_fp16")]; + tensor var_1704 = const()[name = tensor("op_1704"), val = tensor([1, 3, 16, 64])]; + tensor qkv_13_cast_fp16 = reshape(shape = var_1704, x = linear_25_cast_fp16)[name = tensor("qkv_13_cast_fp16")]; + tensor x_25_begin_0 = const()[name = tensor("x_25_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_25_end_0 = const()[name = tensor("x_25_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_25_end_mask_0 = const()[name = tensor("x_25_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_25_squeeze_mask_0 = const()[name = tensor("x_25_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_25_cast_fp16 = slice_by_index(begin = x_25_begin_0, end = x_25_end_0, end_mask = x_25_end_mask_0, squeeze_mask = x_25_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("x_25_cast_fp16")]; + tensor x_27_begin_0 = const()[name = tensor("x_27_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_27_end_0 = const()[name = tensor("x_27_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_27_end_mask_0 = const()[name = tensor("x_27_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_27_squeeze_mask_0 = const()[name = tensor("x_27_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_27_cast_fp16 = slice_by_index(begin = x_27_begin_0, end = x_27_end_0, end_mask = x_27_end_mask_0, squeeze_mask = x_27_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("x_27_cast_fp16")]; + tensor v_13_begin_0 = const()[name = tensor("v_13_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_13_end_0 = const()[name = tensor("v_13_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_13_end_mask_0 = const()[name = tensor("v_13_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_13_squeeze_mask_0 = const()[name = tensor("v_13_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_13_cast_fp16 = slice_by_index(begin = v_13_begin_0, end = v_13_end_0, end_mask = v_13_end_mask_0, squeeze_mask = v_13_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("v_13_cast_fp16")]; + tensor var_1733 = const()[name = tensor("op_1733"), val = tensor([1, 1, 1])]; + tensor position6_to_fp16_dtype_0 = const()[name = tensor("position6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position6_to_fp16 = cast(dtype = position6_to_fp16_dtype_0, x = position6)[name = tensor("cast_529")]; + tensor var_1734_cast_fp16 = reshape(shape = var_1733, x = position6_to_fp16)[name = tensor("op_1734_cast_fp16")]; + tensor angles_25_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_1734_cast_fp16)[name = tensor("angles_25_cast_fp16")]; + tensor rotr_25_cast_fp16 = cos(x = angles_25_cast_fp16)[name = tensor("rotr_25_cast_fp16")]; + tensor roti_25_cast_fp16 = sin(x = angles_25_cast_fp16)[name = tensor("roti_25_cast_fp16")]; + tensor var_1742 = const()[name = tensor("op_1742"), val = tensor([1, 16, 32, 2])]; + tensor pairs_25_cast_fp16 = reshape(shape = var_1742, x = x_25_cast_fp16)[name = tensor("pairs_25_cast_fp16")]; + tensor xr_25_begin_0 = const()[name = tensor("xr_25_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_25_end_0 = const()[name = tensor("xr_25_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_25_end_mask_0 = const()[name = tensor("xr_25_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_25_squeeze_mask_0 = const()[name = tensor("xr_25_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_25_cast_fp16 = slice_by_index(begin = xr_25_begin_0, end = xr_25_end_0, end_mask = xr_25_end_mask_0, squeeze_mask = xr_25_squeeze_mask_0, x = pairs_25_cast_fp16)[name = tensor("xr_25_cast_fp16")]; + tensor xi_25_begin_0 = const()[name = tensor("xi_25_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_25_end_0 = const()[name = tensor("xi_25_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_25_end_mask_0 = const()[name = tensor("xi_25_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_25_squeeze_mask_0 = const()[name = tensor("xi_25_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_25_cast_fp16 = slice_by_index(begin = xi_25_begin_0, end = xi_25_end_0, end_mask = xi_25_end_mask_0, squeeze_mask = xi_25_squeeze_mask_0, x = pairs_25_cast_fp16)[name = tensor("xi_25_cast_fp16")]; + tensor var_1750_cast_fp16 = mul(x = xr_25_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_1750_cast_fp16")]; + tensor var_1751_cast_fp16 = mul(x = xi_25_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_1751_cast_fp16")]; + tensor xor__25_cast_fp16 = sub(x = var_1750_cast_fp16, y = var_1751_cast_fp16)[name = tensor("xor__25_cast_fp16")]; + tensor var_1754_cast_fp16 = mul(x = xr_25_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_1754_cast_fp16")]; + tensor var_1755_cast_fp16 = mul(x = xi_25_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_1755_cast_fp16")]; + tensor xoi_25_cast_fp16 = add(x = var_1754_cast_fp16, y = var_1755_cast_fp16)[name = tensor("xoi_25_cast_fp16")]; + tensor out_25_axis_0 = const()[name = tensor("out_25_axis_0"), val = tensor(-1)]; + tensor out_25_cast_fp16 = stack(axis = out_25_axis_0, values = (xor__25_cast_fp16, xoi_25_cast_fp16))[name = tensor("out_25_cast_fp16")]; + tensor var_1778 = const()[name = tensor("op_1778"), val = tensor([1, 16, 32, 2])]; + tensor pairs_27_cast_fp16 = reshape(shape = var_1778, x = x_27_cast_fp16)[name = tensor("pairs_27_cast_fp16")]; + tensor xr_27_begin_0 = const()[name = tensor("xr_27_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_27_end_0 = const()[name = tensor("xr_27_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_27_end_mask_0 = const()[name = tensor("xr_27_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_27_squeeze_mask_0 = const()[name = tensor("xr_27_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_27_cast_fp16 = slice_by_index(begin = xr_27_begin_0, end = xr_27_end_0, end_mask = xr_27_end_mask_0, squeeze_mask = xr_27_squeeze_mask_0, x = pairs_27_cast_fp16)[name = tensor("xr_27_cast_fp16")]; + tensor xi_27_begin_0 = const()[name = tensor("xi_27_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_27_end_0 = const()[name = tensor("xi_27_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_27_end_mask_0 = const()[name = tensor("xi_27_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_27_squeeze_mask_0 = const()[name = tensor("xi_27_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_27_cast_fp16 = slice_by_index(begin = xi_27_begin_0, end = xi_27_end_0, end_mask = xi_27_end_mask_0, squeeze_mask = xi_27_squeeze_mask_0, x = pairs_27_cast_fp16)[name = tensor("xi_27_cast_fp16")]; + tensor var_1786_cast_fp16 = mul(x = xr_27_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_1786_cast_fp16")]; + tensor var_1787_cast_fp16 = mul(x = xi_27_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_1787_cast_fp16")]; + tensor xor__27_cast_fp16 = sub(x = var_1786_cast_fp16, y = var_1787_cast_fp16)[name = tensor("xor__27_cast_fp16")]; + tensor var_1790_cast_fp16 = mul(x = xr_27_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_1790_cast_fp16")]; + tensor var_1791_cast_fp16 = mul(x = xi_27_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_1791_cast_fp16")]; + tensor xoi_27_cast_fp16 = add(x = var_1790_cast_fp16, y = var_1791_cast_fp16)[name = tensor("xoi_27_cast_fp16")]; + tensor out_27_axis_0 = const()[name = tensor("out_27_axis_0"), val = tensor(-1)]; + tensor out_27_cast_fp16 = stack(axis = out_27_axis_0, values = (xor__27_cast_fp16, xoi_27_cast_fp16))[name = tensor("out_27_cast_fp16")]; + tensor _inversed_1803_y_0_to_fp16 = const()[name = tensor("_inversed_1803_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1803_cast_fp16 = mul(x = position6_to_fp16, y = _inversed_1803_y_0_to_fp16)[name = tensor("_inversed_1803_cast_fp16")]; + tensor var_1804_cast_fp16 = floor(x = _inversed_1803_cast_fp16)[name = tensor("op_1804_cast_fp16")]; + tensor var_1805_to_fp16 = const()[name = tensor("op_1805_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1806_cast_fp16 = mul(x = var_1804_cast_fp16, y = var_1805_to_fp16)[name = tensor("op_1806_cast_fp16")]; + tensor write_idx_13_cast_fp16 = sub(x = position6_to_fp16, y = var_1806_cast_fp16)[name = tensor("write_idx_13_cast_fp16")]; + tensor var_1809_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_13_cast_fp16)[name = tensor("op_1809_cast_fp16")]; + tensor var_1819 = const()[name = tensor("op_1819"), val = tensor([1, 512, 1, 1])]; + tensor onehot_25_to_fp16_dtype_0 = const()[name = tensor("onehot_25_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_1809_cast_fp16_to_fp16 = cast(dtype = onehot_25_to_fp16_dtype_0, x = var_1809_cast_fp16)[name = tensor("cast_528")]; + tensor onehot_27_cast_fp16 = reshape(shape = var_1819, x = var_1809_cast_fp16_to_fp16)[name = tensor("onehot_27_cast_fp16")]; + tensor var_1821_to_fp16 = const()[name = tensor("op_1821_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1823_cast_fp16 = sub(x = var_1821_to_fp16, y = onehot_27_cast_fp16)[name = tensor("op_1823_cast_fp16")]; + tensor k_cache6_to_fp16_dtype_0 = const()[name = tensor("k_cache6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache6_to_fp16 = cast(dtype = k_cache6_to_fp16_dtype_0, x = k_cache6)[name = tensor("cast_527")]; + tensor var_1824_cast_fp16 = mul(x = k_cache6_to_fp16, y = var_1823_cast_fp16)[name = tensor("op_1824_cast_fp16")]; + tensor var_1829 = const()[name = tensor("op_1829"), val = tensor([1, 1, 16, 64])]; + tensor var_1830_cast_fp16 = reshape(shape = var_1829, x = out_27_cast_fp16)[name = tensor("op_1830_cast_fp16")]; + tensor var_1831_cast_fp16 = mul(x = var_1830_cast_fp16, y = onehot_27_cast_fp16)[name = tensor("op_1831_cast_fp16")]; + tensor new_k_cache_13_cast_fp16 = add(x = var_1824_cast_fp16, y = var_1831_cast_fp16)[name = tensor("new_k_cache_13_cast_fp16")]; + tensor new_k_cache_13_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_13_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache6_to_fp16_dtype_0 = const()[name = tensor("v_cache6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache6_to_fp16 = cast(dtype = v_cache6_to_fp16_dtype_0, x = v_cache6)[name = tensor("cast_525")]; + tensor var_1837_cast_fp16 = mul(x = v_cache6_to_fp16, y = var_1823_cast_fp16)[name = tensor("op_1837_cast_fp16")]; + tensor var_1842 = const()[name = tensor("op_1842"), val = tensor([1, 1, 16, 64])]; + tensor var_1843_cast_fp16 = reshape(shape = var_1842, x = v_13_cast_fp16)[name = tensor("op_1843_cast_fp16")]; + tensor var_1844_cast_fp16 = mul(x = var_1843_cast_fp16, y = onehot_27_cast_fp16)[name = tensor("op_1844_cast_fp16")]; + tensor new_v_cache_13_cast_fp16 = add(x = var_1837_cast_fp16, y = var_1844_cast_fp16)[name = tensor("new_v_cache_13_cast_fp16")]; + tensor new_v_cache_13_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_13_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_6_perm_0 = const()[name = tensor("transpose_6_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_1857 = const()[name = tensor("op_1857"), val = tensor([0, 2, 1, 3])]; + tensor var_1863 = const()[name = tensor("op_1863"), val = tensor([1, 16, 1, 64])]; + tensor q4_13_cast_fp16 = reshape(shape = var_1863, x = out_25_cast_fp16)[name = tensor("q4_13_cast_fp16")]; + tensor var_1868_transpose_x_0 = const()[name = tensor("op_1868_transpose_x_0"), val = tensor(false)]; + tensor var_1868_transpose_y_0 = const()[name = tensor("op_1868_transpose_y_0"), val = tensor(false)]; + tensor transpose_6_cast_fp16 = transpose(perm = transpose_6_perm_0, x = new_k_cache_13_cast_fp16)[name = tensor("transpose_77")]; + tensor var_1868_cast_fp16 = matmul(transpose_x = var_1868_transpose_x_0, transpose_y = var_1868_transpose_y_0, x = q4_13_cast_fp16, y = transpose_6_cast_fp16)[name = tensor("op_1868_cast_fp16")]; + tensor var_1869_to_fp16 = const()[name = tensor("op_1869_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_25_cast_fp16 = mul(x = var_1868_cast_fp16, y = var_1869_to_fp16)[name = tensor("scores_25_cast_fp16")]; + tensor var_1873 = const()[name = tensor("op_1873"), val = tensor([1, 1])]; + tensor var_1874_cast_fp16 = reshape(shape = var_1873, x = position6_to_fp16)[name = tensor("op_1874_cast_fp16")]; + tensor var_1875_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_1874_cast_fp16)[name = tensor("op_1875_cast_fp16")]; + tensor mask_13_to_fp16_dtype_0 = const()[name = tensor("mask_13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_1882_to_fp16 = const()[name = tensor("op_1882_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1875_cast_fp16_to_fp16 = cast(dtype = mask_13_to_fp16_dtype_0, x = var_1875_cast_fp16)[name = tensor("cast_523")]; + tensor var_1883_cast_fp16 = sub(x = var_1875_cast_fp16_to_fp16, y = var_1882_to_fp16)[name = tensor("op_1883_cast_fp16")]; + tensor var_1888 = const()[name = tensor("op_1888"), val = tensor([1, 1, 1, 512])]; + tensor var_1889_cast_fp16 = reshape(shape = var_1888, x = var_1883_cast_fp16)[name = tensor("op_1889_cast_fp16")]; + tensor var_1890_to_fp16 = const()[name = tensor("op_1890_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_1891_cast_fp16 = mul(x = var_1889_cast_fp16, y = var_1890_to_fp16)[name = tensor("op_1891_cast_fp16")]; + tensor scores_27_cast_fp16 = add(x = scores_25_cast_fp16, y = var_1891_cast_fp16)[name = tensor("scores_27_cast_fp16")]; + tensor var_1894 = const()[name = tensor("op_1894"), val = tensor(-1)]; + tensor weights_13_cast_fp16 = softmax(axis = var_1894, x = scores_27_cast_fp16)[name = tensor("weights_13_cast_fp16")]; + tensor attn_13_transpose_x_0 = const()[name = tensor("attn_13_transpose_x_0"), val = tensor(false)]; + tensor attn_13_transpose_y_0 = const()[name = tensor("attn_13_transpose_y_0"), val = tensor(false)]; + tensor values_13_cast_fp16 = transpose(perm = var_1857, x = new_v_cache_13_cast_fp16)[name = tensor("transpose_76")]; + tensor attn_13_cast_fp16 = matmul(transpose_x = attn_13_transpose_x_0, transpose_y = attn_13_transpose_y_0, x = weights_13_cast_fp16, y = values_13_cast_fp16)[name = tensor("attn_13_cast_fp16")]; + tensor var_1902 = const()[name = tensor("op_1902"), val = tensor([0, 2, 1, 3])]; + tensor var_1907 = const()[name = tensor("op_1907"), val = tensor([1, 1, 1024])]; + tensor var_1903_cast_fp16 = transpose(perm = var_1902, x = attn_13_cast_fp16)[name = tensor("transpose_75")]; + tensor input_79_cast_fp16 = reshape(shape = var_1907, x = var_1903_cast_fp16)[name = tensor("input_79_cast_fp16")]; + tensor attn6_out_proj_weight_to_fp16 = const()[name = tensor("attn6_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(157426368)))]; + tensor linear_26_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn6_out_proj_weight_to_fp16, x = input_79_cast_fp16)[name = tensor("linear_26_cast_fp16")]; + tensor var_1913_to_fp16 = const()[name = tensor("op_1913_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1914_cast_fp16 = add(x = position6_to_fp16, y = var_1913_to_fp16)[name = tensor("op_1914_cast_fp16")]; + tensor var_1914_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1914_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_81_cast_fp16 = add(x = input_75_cast_fp16, y = linear_26_cast_fp16)[name = tensor("input_81_cast_fp16")]; + tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([-1])]; + tensor norm6_2_weight_to_fp16 = const()[name = tensor("norm6_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159523584)))]; + tensor norm6_2_bias_to_fp16 = const()[name = tensor("norm6_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159525696)))]; + tensor var_1918_to_fp16 = const()[name = tensor("op_1918_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_83_cast_fp16 = layer_norm(axes = input_83_axes_0, beta = norm6_2_bias_to_fp16, epsilon = var_1918_to_fp16, gamma = norm6_2_weight_to_fp16, x = input_81_cast_fp16)[name = tensor("input_83_cast_fp16")]; + tensor linear6_1_weight_to_fp16 = const()[name = tensor("linear6_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159527808)))]; + tensor linear_27_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear6_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_27_cast_fp16")]; + tensor input_85_mode_0 = const()[name = tensor("input_85_mode_0"), val = tensor("EXACT")]; + tensor input_85_cast_fp16 = gelu(mode = input_85_mode_0, x = linear_27_cast_fp16)[name = tensor("input_85_cast_fp16")]; + tensor linear6_2_weight_to_fp16 = const()[name = tensor("linear6_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(167916480)))]; + tensor linear_28_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear6_2_weight_to_fp16, x = input_85_cast_fp16)[name = tensor("linear_28_cast_fp16")]; + tensor input_87_cast_fp16 = add(x = input_81_cast_fp16, y = linear_28_cast_fp16)[name = tensor("input_87_cast_fp16")]; + tensor input_89_axes_0 = const()[name = tensor("input_89_axes_0"), val = tensor([-1])]; + tensor norm7_1_weight_to_fp16 = const()[name = tensor("norm7_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176305152)))]; + tensor norm7_1_bias_to_fp16 = const()[name = tensor("norm7_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176307264)))]; + tensor var_1935_to_fp16 = const()[name = tensor("op_1935_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_89_cast_fp16 = layer_norm(axes = input_89_axes_0, beta = norm7_1_bias_to_fp16, epsilon = var_1935_to_fp16, gamma = norm7_1_weight_to_fp16, x = input_87_cast_fp16)[name = tensor("input_89_cast_fp16")]; + tensor attn7_in_proj_weight_to_fp16 = const()[name = tensor("attn7_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176309376)))]; + tensor linear_29_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn7_in_proj_weight_to_fp16, x = input_89_cast_fp16)[name = tensor("linear_29_cast_fp16")]; + tensor var_1948 = const()[name = tensor("op_1948"), val = tensor([1, 3, 16, 64])]; + tensor qkv_15_cast_fp16 = reshape(shape = var_1948, x = linear_29_cast_fp16)[name = tensor("qkv_15_cast_fp16")]; + tensor x_29_begin_0 = const()[name = tensor("x_29_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_29_end_0 = const()[name = tensor("x_29_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_29_end_mask_0 = const()[name = tensor("x_29_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_29_squeeze_mask_0 = const()[name = tensor("x_29_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_29_cast_fp16 = slice_by_index(begin = x_29_begin_0, end = x_29_end_0, end_mask = x_29_end_mask_0, squeeze_mask = x_29_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("x_29_cast_fp16")]; + tensor x_31_begin_0 = const()[name = tensor("x_31_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_31_end_0 = const()[name = tensor("x_31_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_31_end_mask_0 = const()[name = tensor("x_31_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_31_squeeze_mask_0 = const()[name = tensor("x_31_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_31_cast_fp16 = slice_by_index(begin = x_31_begin_0, end = x_31_end_0, end_mask = x_31_end_mask_0, squeeze_mask = x_31_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("x_31_cast_fp16")]; + tensor v_15_begin_0 = const()[name = tensor("v_15_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_15_end_0 = const()[name = tensor("v_15_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_15_end_mask_0 = const()[name = tensor("v_15_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_15_squeeze_mask_0 = const()[name = tensor("v_15_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_15_cast_fp16 = slice_by_index(begin = v_15_begin_0, end = v_15_end_0, end_mask = v_15_end_mask_0, squeeze_mask = v_15_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("v_15_cast_fp16")]; + tensor var_1977 = const()[name = tensor("op_1977"), val = tensor([1, 1, 1])]; + tensor position7_to_fp16_dtype_0 = const()[name = tensor("position7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position7_to_fp16 = cast(dtype = position7_to_fp16_dtype_0, x = position7)[name = tensor("cast_521")]; + tensor var_1978_cast_fp16 = reshape(shape = var_1977, x = position7_to_fp16)[name = tensor("op_1978_cast_fp16")]; + tensor angles_29_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_1978_cast_fp16)[name = tensor("angles_29_cast_fp16")]; + tensor rotr_29_cast_fp16 = cos(x = angles_29_cast_fp16)[name = tensor("rotr_29_cast_fp16")]; + tensor roti_29_cast_fp16 = sin(x = angles_29_cast_fp16)[name = tensor("roti_29_cast_fp16")]; + tensor var_1986 = const()[name = tensor("op_1986"), val = tensor([1, 16, 32, 2])]; + tensor pairs_29_cast_fp16 = reshape(shape = var_1986, x = x_29_cast_fp16)[name = tensor("pairs_29_cast_fp16")]; + tensor xr_29_begin_0 = const()[name = tensor("xr_29_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_29_end_0 = const()[name = tensor("xr_29_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_29_end_mask_0 = const()[name = tensor("xr_29_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_29_squeeze_mask_0 = const()[name = tensor("xr_29_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_29_cast_fp16 = slice_by_index(begin = xr_29_begin_0, end = xr_29_end_0, end_mask = xr_29_end_mask_0, squeeze_mask = xr_29_squeeze_mask_0, x = pairs_29_cast_fp16)[name = tensor("xr_29_cast_fp16")]; + tensor xi_29_begin_0 = const()[name = tensor("xi_29_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_29_end_0 = const()[name = tensor("xi_29_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_29_end_mask_0 = const()[name = tensor("xi_29_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_29_squeeze_mask_0 = const()[name = tensor("xi_29_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_29_cast_fp16 = slice_by_index(begin = xi_29_begin_0, end = xi_29_end_0, end_mask = xi_29_end_mask_0, squeeze_mask = xi_29_squeeze_mask_0, x = pairs_29_cast_fp16)[name = tensor("xi_29_cast_fp16")]; + tensor var_1994_cast_fp16 = mul(x = xr_29_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_1994_cast_fp16")]; + tensor var_1995_cast_fp16 = mul(x = xi_29_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_1995_cast_fp16")]; + tensor xor__29_cast_fp16 = sub(x = var_1994_cast_fp16, y = var_1995_cast_fp16)[name = tensor("xor__29_cast_fp16")]; + tensor var_1998_cast_fp16 = mul(x = xr_29_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_1998_cast_fp16")]; + tensor var_1999_cast_fp16 = mul(x = xi_29_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_1999_cast_fp16")]; + tensor xoi_29_cast_fp16 = add(x = var_1998_cast_fp16, y = var_1999_cast_fp16)[name = tensor("xoi_29_cast_fp16")]; + tensor out_29_axis_0 = const()[name = tensor("out_29_axis_0"), val = tensor(-1)]; + tensor out_29_cast_fp16 = stack(axis = out_29_axis_0, values = (xor__29_cast_fp16, xoi_29_cast_fp16))[name = tensor("out_29_cast_fp16")]; + tensor var_2022 = const()[name = tensor("op_2022"), val = tensor([1, 16, 32, 2])]; + tensor pairs_31_cast_fp16 = reshape(shape = var_2022, x = x_31_cast_fp16)[name = tensor("pairs_31_cast_fp16")]; + tensor xr_31_begin_0 = const()[name = tensor("xr_31_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_31_end_0 = const()[name = tensor("xr_31_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_31_end_mask_0 = const()[name = tensor("xr_31_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_31_squeeze_mask_0 = const()[name = tensor("xr_31_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_31_cast_fp16 = slice_by_index(begin = xr_31_begin_0, end = xr_31_end_0, end_mask = xr_31_end_mask_0, squeeze_mask = xr_31_squeeze_mask_0, x = pairs_31_cast_fp16)[name = tensor("xr_31_cast_fp16")]; + tensor xi_31_begin_0 = const()[name = tensor("xi_31_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_31_end_0 = const()[name = tensor("xi_31_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_31_end_mask_0 = const()[name = tensor("xi_31_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_31_squeeze_mask_0 = const()[name = tensor("xi_31_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_31_cast_fp16 = slice_by_index(begin = xi_31_begin_0, end = xi_31_end_0, end_mask = xi_31_end_mask_0, squeeze_mask = xi_31_squeeze_mask_0, x = pairs_31_cast_fp16)[name = tensor("xi_31_cast_fp16")]; + tensor var_2030_cast_fp16 = mul(x = xr_31_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_2030_cast_fp16")]; + tensor var_2031_cast_fp16 = mul(x = xi_31_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_2031_cast_fp16")]; + tensor xor__31_cast_fp16 = sub(x = var_2030_cast_fp16, y = var_2031_cast_fp16)[name = tensor("xor__31_cast_fp16")]; + tensor var_2034_cast_fp16 = mul(x = xr_31_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_2034_cast_fp16")]; + tensor var_2035_cast_fp16 = mul(x = xi_31_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_2035_cast_fp16")]; + tensor xoi_31_cast_fp16 = add(x = var_2034_cast_fp16, y = var_2035_cast_fp16)[name = tensor("xoi_31_cast_fp16")]; + tensor out_31_axis_0 = const()[name = tensor("out_31_axis_0"), val = tensor(-1)]; + tensor out_31_cast_fp16 = stack(axis = out_31_axis_0, values = (xor__31_cast_fp16, xoi_31_cast_fp16))[name = tensor("out_31_cast_fp16")]; + tensor _inversed_2047_y_0_to_fp16 = const()[name = tensor("_inversed_2047_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2047_cast_fp16 = mul(x = position7_to_fp16, y = _inversed_2047_y_0_to_fp16)[name = tensor("_inversed_2047_cast_fp16")]; + tensor var_2048_cast_fp16 = floor(x = _inversed_2047_cast_fp16)[name = tensor("op_2048_cast_fp16")]; + tensor var_2049_to_fp16 = const()[name = tensor("op_2049_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2050_cast_fp16 = mul(x = var_2048_cast_fp16, y = var_2049_to_fp16)[name = tensor("op_2050_cast_fp16")]; + tensor write_idx_15_cast_fp16 = sub(x = position7_to_fp16, y = var_2050_cast_fp16)[name = tensor("write_idx_15_cast_fp16")]; + tensor var_2053_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_15_cast_fp16)[name = tensor("op_2053_cast_fp16")]; + tensor var_2063 = const()[name = tensor("op_2063"), val = tensor([1, 512, 1, 1])]; + tensor onehot_29_to_fp16_dtype_0 = const()[name = tensor("onehot_29_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_2053_cast_fp16_to_fp16 = cast(dtype = onehot_29_to_fp16_dtype_0, x = var_2053_cast_fp16)[name = tensor("cast_520")]; + tensor onehot_31_cast_fp16 = reshape(shape = var_2063, x = var_2053_cast_fp16_to_fp16)[name = tensor("onehot_31_cast_fp16")]; + tensor var_2065_to_fp16 = const()[name = tensor("op_2065_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2067_cast_fp16 = sub(x = var_2065_to_fp16, y = onehot_31_cast_fp16)[name = tensor("op_2067_cast_fp16")]; + tensor k_cache7_to_fp16_dtype_0 = const()[name = tensor("k_cache7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache7_to_fp16 = cast(dtype = k_cache7_to_fp16_dtype_0, x = k_cache7)[name = tensor("cast_519")]; + tensor var_2068_cast_fp16 = mul(x = k_cache7_to_fp16, y = var_2067_cast_fp16)[name = tensor("op_2068_cast_fp16")]; + tensor var_2073 = const()[name = tensor("op_2073"), val = tensor([1, 1, 16, 64])]; + tensor var_2074_cast_fp16 = reshape(shape = var_2073, x = out_31_cast_fp16)[name = tensor("op_2074_cast_fp16")]; + tensor var_2075_cast_fp16 = mul(x = var_2074_cast_fp16, y = onehot_31_cast_fp16)[name = tensor("op_2075_cast_fp16")]; + tensor new_k_cache_15_cast_fp16 = add(x = var_2068_cast_fp16, y = var_2075_cast_fp16)[name = tensor("new_k_cache_15_cast_fp16")]; + tensor new_k_cache_15_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_15_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache7_to_fp16_dtype_0 = const()[name = tensor("v_cache7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache7_to_fp16 = cast(dtype = v_cache7_to_fp16_dtype_0, x = v_cache7)[name = tensor("cast_517")]; + tensor var_2081_cast_fp16 = mul(x = v_cache7_to_fp16, y = var_2067_cast_fp16)[name = tensor("op_2081_cast_fp16")]; + tensor var_2086 = const()[name = tensor("op_2086"), val = tensor([1, 1, 16, 64])]; + tensor var_2087_cast_fp16 = reshape(shape = var_2086, x = v_15_cast_fp16)[name = tensor("op_2087_cast_fp16")]; + tensor var_2088_cast_fp16 = mul(x = var_2087_cast_fp16, y = onehot_31_cast_fp16)[name = tensor("op_2088_cast_fp16")]; + tensor new_v_cache_15_cast_fp16 = add(x = var_2081_cast_fp16, y = var_2088_cast_fp16)[name = tensor("new_v_cache_15_cast_fp16")]; + tensor new_v_cache_15_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_15_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_7_perm_0 = const()[name = tensor("transpose_7_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_2101 = const()[name = tensor("op_2101"), val = tensor([0, 2, 1, 3])]; + tensor var_2107 = const()[name = tensor("op_2107"), val = tensor([1, 16, 1, 64])]; + tensor q4_15_cast_fp16 = reshape(shape = var_2107, x = out_29_cast_fp16)[name = tensor("q4_15_cast_fp16")]; + tensor var_2112_transpose_x_0 = const()[name = tensor("op_2112_transpose_x_0"), val = tensor(false)]; + tensor var_2112_transpose_y_0 = const()[name = tensor("op_2112_transpose_y_0"), val = tensor(false)]; + tensor transpose_7_cast_fp16 = transpose(perm = transpose_7_perm_0, x = new_k_cache_15_cast_fp16)[name = tensor("transpose_74")]; + tensor var_2112_cast_fp16 = matmul(transpose_x = var_2112_transpose_x_0, transpose_y = var_2112_transpose_y_0, x = q4_15_cast_fp16, y = transpose_7_cast_fp16)[name = tensor("op_2112_cast_fp16")]; + tensor var_2113_to_fp16 = const()[name = tensor("op_2113_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_29_cast_fp16 = mul(x = var_2112_cast_fp16, y = var_2113_to_fp16)[name = tensor("scores_29_cast_fp16")]; + tensor var_2117 = const()[name = tensor("op_2117"), val = tensor([1, 1])]; + tensor var_2118_cast_fp16 = reshape(shape = var_2117, x = position7_to_fp16)[name = tensor("op_2118_cast_fp16")]; + tensor var_2119_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_2118_cast_fp16)[name = tensor("op_2119_cast_fp16")]; + tensor mask_15_to_fp16_dtype_0 = const()[name = tensor("mask_15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_2126_to_fp16 = const()[name = tensor("op_2126_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2119_cast_fp16_to_fp16 = cast(dtype = mask_15_to_fp16_dtype_0, x = var_2119_cast_fp16)[name = tensor("cast_515")]; + tensor var_2127_cast_fp16 = sub(x = var_2119_cast_fp16_to_fp16, y = var_2126_to_fp16)[name = tensor("op_2127_cast_fp16")]; + tensor var_2132 = const()[name = tensor("op_2132"), val = tensor([1, 1, 1, 512])]; + tensor var_2133_cast_fp16 = reshape(shape = var_2132, x = var_2127_cast_fp16)[name = tensor("op_2133_cast_fp16")]; + tensor var_2134_to_fp16 = const()[name = tensor("op_2134_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_2135_cast_fp16 = mul(x = var_2133_cast_fp16, y = var_2134_to_fp16)[name = tensor("op_2135_cast_fp16")]; + tensor scores_31_cast_fp16 = add(x = scores_29_cast_fp16, y = var_2135_cast_fp16)[name = tensor("scores_31_cast_fp16")]; + tensor var_2138 = const()[name = tensor("op_2138"), val = tensor(-1)]; + tensor weights_15_cast_fp16 = softmax(axis = var_2138, x = scores_31_cast_fp16)[name = tensor("weights_15_cast_fp16")]; + tensor attn_15_transpose_x_0 = const()[name = tensor("attn_15_transpose_x_0"), val = tensor(false)]; + tensor attn_15_transpose_y_0 = const()[name = tensor("attn_15_transpose_y_0"), val = tensor(false)]; + tensor values_15_cast_fp16 = transpose(perm = var_2101, x = new_v_cache_15_cast_fp16)[name = tensor("transpose_73")]; + tensor attn_15_cast_fp16 = matmul(transpose_x = attn_15_transpose_x_0, transpose_y = attn_15_transpose_y_0, x = weights_15_cast_fp16, y = values_15_cast_fp16)[name = tensor("attn_15_cast_fp16")]; + tensor var_2146 = const()[name = tensor("op_2146"), val = tensor([0, 2, 1, 3])]; + tensor var_2151 = const()[name = tensor("op_2151"), val = tensor([1, 1, 1024])]; + tensor var_2147_cast_fp16 = transpose(perm = var_2146, x = attn_15_cast_fp16)[name = tensor("transpose_72")]; + tensor input_91_cast_fp16 = reshape(shape = var_2151, x = var_2147_cast_fp16)[name = tensor("input_91_cast_fp16")]; + tensor attn7_out_proj_weight_to_fp16 = const()[name = tensor("attn7_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(182600896)))]; + tensor linear_30_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn7_out_proj_weight_to_fp16, x = input_91_cast_fp16)[name = tensor("linear_30_cast_fp16")]; + tensor var_2157_to_fp16 = const()[name = tensor("op_2157_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2158_cast_fp16 = add(x = position7_to_fp16, y = var_2157_to_fp16)[name = tensor("op_2158_cast_fp16")]; + tensor var_2158_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2158_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_93_cast_fp16 = add(x = input_87_cast_fp16, y = linear_30_cast_fp16)[name = tensor("input_93_cast_fp16")]; + tensor input_95_axes_0 = const()[name = tensor("input_95_axes_0"), val = tensor([-1])]; + tensor norm7_2_weight_to_fp16 = const()[name = tensor("norm7_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184698112)))]; + tensor norm7_2_bias_to_fp16 = const()[name = tensor("norm7_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184700224)))]; + tensor var_2162_to_fp16 = const()[name = tensor("op_2162_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_95_cast_fp16 = layer_norm(axes = input_95_axes_0, beta = norm7_2_bias_to_fp16, epsilon = var_2162_to_fp16, gamma = norm7_2_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("input_95_cast_fp16")]; + tensor linear7_1_weight_to_fp16 = const()[name = tensor("linear7_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184702336)))]; + tensor linear_31_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear7_1_weight_to_fp16, x = input_95_cast_fp16)[name = tensor("linear_31_cast_fp16")]; + tensor input_97_mode_0 = const()[name = tensor("input_97_mode_0"), val = tensor("EXACT")]; + tensor input_97_cast_fp16 = gelu(mode = input_97_mode_0, x = linear_31_cast_fp16)[name = tensor("input_97_cast_fp16")]; + tensor linear7_2_weight_to_fp16 = const()[name = tensor("linear7_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(193091008)))]; + tensor linear_32_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear7_2_weight_to_fp16, x = input_97_cast_fp16)[name = tensor("linear_32_cast_fp16")]; + tensor input_99_cast_fp16 = add(x = input_93_cast_fp16, y = linear_32_cast_fp16)[name = tensor("input_99_cast_fp16")]; + tensor input_101_axes_0 = const()[name = tensor("input_101_axes_0"), val = tensor([-1])]; + tensor norm8_1_weight_to_fp16 = const()[name = tensor("norm8_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201479680)))]; + tensor norm8_1_bias_to_fp16 = const()[name = tensor("norm8_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201481792)))]; + tensor var_2179_to_fp16 = const()[name = tensor("op_2179_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_101_cast_fp16 = layer_norm(axes = input_101_axes_0, beta = norm8_1_bias_to_fp16, epsilon = var_2179_to_fp16, gamma = norm8_1_weight_to_fp16, x = input_99_cast_fp16)[name = tensor("input_101_cast_fp16")]; + tensor attn8_in_proj_weight_to_fp16 = const()[name = tensor("attn8_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201483904)))]; + tensor linear_33_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn8_in_proj_weight_to_fp16, x = input_101_cast_fp16)[name = tensor("linear_33_cast_fp16")]; + tensor var_2192 = const()[name = tensor("op_2192"), val = tensor([1, 3, 16, 64])]; + tensor qkv_17_cast_fp16 = reshape(shape = var_2192, x = linear_33_cast_fp16)[name = tensor("qkv_17_cast_fp16")]; + tensor x_33_begin_0 = const()[name = tensor("x_33_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_33_end_0 = const()[name = tensor("x_33_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_33_end_mask_0 = const()[name = tensor("x_33_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_33_squeeze_mask_0 = const()[name = tensor("x_33_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_33_cast_fp16 = slice_by_index(begin = x_33_begin_0, end = x_33_end_0, end_mask = x_33_end_mask_0, squeeze_mask = x_33_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("x_33_cast_fp16")]; + tensor x_35_begin_0 = const()[name = tensor("x_35_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_35_end_0 = const()[name = tensor("x_35_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_35_end_mask_0 = const()[name = tensor("x_35_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_35_squeeze_mask_0 = const()[name = tensor("x_35_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_35_cast_fp16 = slice_by_index(begin = x_35_begin_0, end = x_35_end_0, end_mask = x_35_end_mask_0, squeeze_mask = x_35_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("x_35_cast_fp16")]; + tensor v_17_begin_0 = const()[name = tensor("v_17_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_17_end_0 = const()[name = tensor("v_17_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_17_end_mask_0 = const()[name = tensor("v_17_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_17_squeeze_mask_0 = const()[name = tensor("v_17_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_17_cast_fp16 = slice_by_index(begin = v_17_begin_0, end = v_17_end_0, end_mask = v_17_end_mask_0, squeeze_mask = v_17_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("v_17_cast_fp16")]; + tensor var_2221 = const()[name = tensor("op_2221"), val = tensor([1, 1, 1])]; + tensor position8_to_fp16_dtype_0 = const()[name = tensor("position8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position8_to_fp16 = cast(dtype = position8_to_fp16_dtype_0, x = position8)[name = tensor("cast_513")]; + tensor var_2222_cast_fp16 = reshape(shape = var_2221, x = position8_to_fp16)[name = tensor("op_2222_cast_fp16")]; + tensor angles_33_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_2222_cast_fp16)[name = tensor("angles_33_cast_fp16")]; + tensor rotr_33_cast_fp16 = cos(x = angles_33_cast_fp16)[name = tensor("rotr_33_cast_fp16")]; + tensor roti_33_cast_fp16 = sin(x = angles_33_cast_fp16)[name = tensor("roti_33_cast_fp16")]; + tensor var_2230 = const()[name = tensor("op_2230"), val = tensor([1, 16, 32, 2])]; + tensor pairs_33_cast_fp16 = reshape(shape = var_2230, x = x_33_cast_fp16)[name = tensor("pairs_33_cast_fp16")]; + tensor xr_33_begin_0 = const()[name = tensor("xr_33_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_33_end_0 = const()[name = tensor("xr_33_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_33_end_mask_0 = const()[name = tensor("xr_33_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_33_squeeze_mask_0 = const()[name = tensor("xr_33_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_33_cast_fp16 = slice_by_index(begin = xr_33_begin_0, end = xr_33_end_0, end_mask = xr_33_end_mask_0, squeeze_mask = xr_33_squeeze_mask_0, x = pairs_33_cast_fp16)[name = tensor("xr_33_cast_fp16")]; + tensor xi_33_begin_0 = const()[name = tensor("xi_33_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_33_end_0 = const()[name = tensor("xi_33_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_33_end_mask_0 = const()[name = tensor("xi_33_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_33_squeeze_mask_0 = const()[name = tensor("xi_33_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_33_cast_fp16 = slice_by_index(begin = xi_33_begin_0, end = xi_33_end_0, end_mask = xi_33_end_mask_0, squeeze_mask = xi_33_squeeze_mask_0, x = pairs_33_cast_fp16)[name = tensor("xi_33_cast_fp16")]; + tensor var_2238_cast_fp16 = mul(x = xr_33_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_2238_cast_fp16")]; + tensor var_2239_cast_fp16 = mul(x = xi_33_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_2239_cast_fp16")]; + tensor xor__33_cast_fp16 = sub(x = var_2238_cast_fp16, y = var_2239_cast_fp16)[name = tensor("xor__33_cast_fp16")]; + tensor var_2242_cast_fp16 = mul(x = xr_33_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_2242_cast_fp16")]; + tensor var_2243_cast_fp16 = mul(x = xi_33_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_2243_cast_fp16")]; + tensor xoi_33_cast_fp16 = add(x = var_2242_cast_fp16, y = var_2243_cast_fp16)[name = tensor("xoi_33_cast_fp16")]; + tensor out_33_axis_0 = const()[name = tensor("out_33_axis_0"), val = tensor(-1)]; + tensor out_33_cast_fp16 = stack(axis = out_33_axis_0, values = (xor__33_cast_fp16, xoi_33_cast_fp16))[name = tensor("out_33_cast_fp16")]; + tensor var_2266 = const()[name = tensor("op_2266"), val = tensor([1, 16, 32, 2])]; + tensor pairs_35_cast_fp16 = reshape(shape = var_2266, x = x_35_cast_fp16)[name = tensor("pairs_35_cast_fp16")]; + tensor xr_35_begin_0 = const()[name = tensor("xr_35_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_35_end_0 = const()[name = tensor("xr_35_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_35_end_mask_0 = const()[name = tensor("xr_35_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_35_squeeze_mask_0 = const()[name = tensor("xr_35_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_35_cast_fp16 = slice_by_index(begin = xr_35_begin_0, end = xr_35_end_0, end_mask = xr_35_end_mask_0, squeeze_mask = xr_35_squeeze_mask_0, x = pairs_35_cast_fp16)[name = tensor("xr_35_cast_fp16")]; + tensor xi_35_begin_0 = const()[name = tensor("xi_35_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_35_end_0 = const()[name = tensor("xi_35_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_35_end_mask_0 = const()[name = tensor("xi_35_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_35_squeeze_mask_0 = const()[name = tensor("xi_35_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_35_cast_fp16 = slice_by_index(begin = xi_35_begin_0, end = xi_35_end_0, end_mask = xi_35_end_mask_0, squeeze_mask = xi_35_squeeze_mask_0, x = pairs_35_cast_fp16)[name = tensor("xi_35_cast_fp16")]; + tensor var_2274_cast_fp16 = mul(x = xr_35_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_2274_cast_fp16")]; + tensor var_2275_cast_fp16 = mul(x = xi_35_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_2275_cast_fp16")]; + tensor xor__35_cast_fp16 = sub(x = var_2274_cast_fp16, y = var_2275_cast_fp16)[name = tensor("xor__35_cast_fp16")]; + tensor var_2278_cast_fp16 = mul(x = xr_35_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_2278_cast_fp16")]; + tensor var_2279_cast_fp16 = mul(x = xi_35_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_2279_cast_fp16")]; + tensor xoi_35_cast_fp16 = add(x = var_2278_cast_fp16, y = var_2279_cast_fp16)[name = tensor("xoi_35_cast_fp16")]; + tensor out_35_axis_0 = const()[name = tensor("out_35_axis_0"), val = tensor(-1)]; + tensor out_35_cast_fp16 = stack(axis = out_35_axis_0, values = (xor__35_cast_fp16, xoi_35_cast_fp16))[name = tensor("out_35_cast_fp16")]; + tensor _inversed_2291_y_0_to_fp16 = const()[name = tensor("_inversed_2291_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2291_cast_fp16 = mul(x = position8_to_fp16, y = _inversed_2291_y_0_to_fp16)[name = tensor("_inversed_2291_cast_fp16")]; + tensor var_2292_cast_fp16 = floor(x = _inversed_2291_cast_fp16)[name = tensor("op_2292_cast_fp16")]; + tensor var_2293_to_fp16 = const()[name = tensor("op_2293_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2294_cast_fp16 = mul(x = var_2292_cast_fp16, y = var_2293_to_fp16)[name = tensor("op_2294_cast_fp16")]; + tensor write_idx_17_cast_fp16 = sub(x = position8_to_fp16, y = var_2294_cast_fp16)[name = tensor("write_idx_17_cast_fp16")]; + tensor var_2297_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_17_cast_fp16)[name = tensor("op_2297_cast_fp16")]; + tensor var_2307 = const()[name = tensor("op_2307"), val = tensor([1, 512, 1, 1])]; + tensor onehot_33_to_fp16_dtype_0 = const()[name = tensor("onehot_33_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_2297_cast_fp16_to_fp16 = cast(dtype = onehot_33_to_fp16_dtype_0, x = var_2297_cast_fp16)[name = tensor("cast_512")]; + tensor onehot_35_cast_fp16 = reshape(shape = var_2307, x = var_2297_cast_fp16_to_fp16)[name = tensor("onehot_35_cast_fp16")]; + tensor var_2309_to_fp16 = const()[name = tensor("op_2309_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2311_cast_fp16 = sub(x = var_2309_to_fp16, y = onehot_35_cast_fp16)[name = tensor("op_2311_cast_fp16")]; + tensor k_cache8_to_fp16_dtype_0 = const()[name = tensor("k_cache8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache8_to_fp16 = cast(dtype = k_cache8_to_fp16_dtype_0, x = k_cache8)[name = tensor("cast_511")]; + tensor var_2312_cast_fp16 = mul(x = k_cache8_to_fp16, y = var_2311_cast_fp16)[name = tensor("op_2312_cast_fp16")]; + tensor var_2317 = const()[name = tensor("op_2317"), val = tensor([1, 1, 16, 64])]; + tensor var_2318_cast_fp16 = reshape(shape = var_2317, x = out_35_cast_fp16)[name = tensor("op_2318_cast_fp16")]; + tensor var_2319_cast_fp16 = mul(x = var_2318_cast_fp16, y = onehot_35_cast_fp16)[name = tensor("op_2319_cast_fp16")]; + tensor new_k_cache_17_cast_fp16 = add(x = var_2312_cast_fp16, y = var_2319_cast_fp16)[name = tensor("new_k_cache_17_cast_fp16")]; + tensor new_k_cache_17_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_17_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache8_to_fp16_dtype_0 = const()[name = tensor("v_cache8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache8_to_fp16 = cast(dtype = v_cache8_to_fp16_dtype_0, x = v_cache8)[name = tensor("cast_509")]; + tensor var_2325_cast_fp16 = mul(x = v_cache8_to_fp16, y = var_2311_cast_fp16)[name = tensor("op_2325_cast_fp16")]; + tensor var_2330 = const()[name = tensor("op_2330"), val = tensor([1, 1, 16, 64])]; + tensor var_2331_cast_fp16 = reshape(shape = var_2330, x = v_17_cast_fp16)[name = tensor("op_2331_cast_fp16")]; + tensor var_2332_cast_fp16 = mul(x = var_2331_cast_fp16, y = onehot_35_cast_fp16)[name = tensor("op_2332_cast_fp16")]; + tensor new_v_cache_17_cast_fp16 = add(x = var_2325_cast_fp16, y = var_2332_cast_fp16)[name = tensor("new_v_cache_17_cast_fp16")]; + tensor new_v_cache_17_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_17_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_8_perm_0 = const()[name = tensor("transpose_8_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_2345 = const()[name = tensor("op_2345"), val = tensor([0, 2, 1, 3])]; + tensor var_2351 = const()[name = tensor("op_2351"), val = tensor([1, 16, 1, 64])]; + tensor q4_17_cast_fp16 = reshape(shape = var_2351, x = out_33_cast_fp16)[name = tensor("q4_17_cast_fp16")]; + tensor var_2356_transpose_x_0 = const()[name = tensor("op_2356_transpose_x_0"), val = tensor(false)]; + tensor var_2356_transpose_y_0 = const()[name = tensor("op_2356_transpose_y_0"), val = tensor(false)]; + tensor transpose_8_cast_fp16 = transpose(perm = transpose_8_perm_0, x = new_k_cache_17_cast_fp16)[name = tensor("transpose_71")]; + tensor var_2356_cast_fp16 = matmul(transpose_x = var_2356_transpose_x_0, transpose_y = var_2356_transpose_y_0, x = q4_17_cast_fp16, y = transpose_8_cast_fp16)[name = tensor("op_2356_cast_fp16")]; + tensor var_2357_to_fp16 = const()[name = tensor("op_2357_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_33_cast_fp16 = mul(x = var_2356_cast_fp16, y = var_2357_to_fp16)[name = tensor("scores_33_cast_fp16")]; + tensor var_2361 = const()[name = tensor("op_2361"), val = tensor([1, 1])]; + tensor var_2362_cast_fp16 = reshape(shape = var_2361, x = position8_to_fp16)[name = tensor("op_2362_cast_fp16")]; + tensor var_2363_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_2362_cast_fp16)[name = tensor("op_2363_cast_fp16")]; + tensor mask_17_to_fp16_dtype_0 = const()[name = tensor("mask_17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_2370_to_fp16 = const()[name = tensor("op_2370_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2363_cast_fp16_to_fp16 = cast(dtype = mask_17_to_fp16_dtype_0, x = var_2363_cast_fp16)[name = tensor("cast_507")]; + tensor var_2371_cast_fp16 = sub(x = var_2363_cast_fp16_to_fp16, y = var_2370_to_fp16)[name = tensor("op_2371_cast_fp16")]; + tensor var_2376 = const()[name = tensor("op_2376"), val = tensor([1, 1, 1, 512])]; + tensor var_2377_cast_fp16 = reshape(shape = var_2376, x = var_2371_cast_fp16)[name = tensor("op_2377_cast_fp16")]; + tensor var_2378_to_fp16 = const()[name = tensor("op_2378_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_2379_cast_fp16 = mul(x = var_2377_cast_fp16, y = var_2378_to_fp16)[name = tensor("op_2379_cast_fp16")]; + tensor scores_35_cast_fp16 = add(x = scores_33_cast_fp16, y = var_2379_cast_fp16)[name = tensor("scores_35_cast_fp16")]; + tensor var_2382 = const()[name = tensor("op_2382"), val = tensor(-1)]; + tensor weights_17_cast_fp16 = softmax(axis = var_2382, x = scores_35_cast_fp16)[name = tensor("weights_17_cast_fp16")]; + tensor attn_17_transpose_x_0 = const()[name = tensor("attn_17_transpose_x_0"), val = tensor(false)]; + tensor attn_17_transpose_y_0 = const()[name = tensor("attn_17_transpose_y_0"), val = tensor(false)]; + tensor values_17_cast_fp16 = transpose(perm = var_2345, x = new_v_cache_17_cast_fp16)[name = tensor("transpose_70")]; + tensor attn_17_cast_fp16 = matmul(transpose_x = attn_17_transpose_x_0, transpose_y = attn_17_transpose_y_0, x = weights_17_cast_fp16, y = values_17_cast_fp16)[name = tensor("attn_17_cast_fp16")]; + tensor var_2390 = const()[name = tensor("op_2390"), val = tensor([0, 2, 1, 3])]; + tensor var_2395 = const()[name = tensor("op_2395"), val = tensor([1, 1, 1024])]; + tensor var_2391_cast_fp16 = transpose(perm = var_2390, x = attn_17_cast_fp16)[name = tensor("transpose_69")]; + tensor input_103_cast_fp16 = reshape(shape = var_2395, x = var_2391_cast_fp16)[name = tensor("input_103_cast_fp16")]; + tensor attn8_out_proj_weight_to_fp16 = const()[name = tensor("attn8_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(207775424)))]; + tensor linear_34_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn8_out_proj_weight_to_fp16, x = input_103_cast_fp16)[name = tensor("linear_34_cast_fp16")]; + tensor var_2401_to_fp16 = const()[name = tensor("op_2401_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2402_cast_fp16 = add(x = position8_to_fp16, y = var_2401_to_fp16)[name = tensor("op_2402_cast_fp16")]; + tensor var_2402_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2402_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_105_cast_fp16 = add(x = input_99_cast_fp16, y = linear_34_cast_fp16)[name = tensor("input_105_cast_fp16")]; + tensor input_107_axes_0 = const()[name = tensor("input_107_axes_0"), val = tensor([-1])]; + tensor norm8_2_weight_to_fp16 = const()[name = tensor("norm8_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209872640)))]; + tensor norm8_2_bias_to_fp16 = const()[name = tensor("norm8_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209874752)))]; + tensor var_2406_to_fp16 = const()[name = tensor("op_2406_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_107_cast_fp16 = layer_norm(axes = input_107_axes_0, beta = norm8_2_bias_to_fp16, epsilon = var_2406_to_fp16, gamma = norm8_2_weight_to_fp16, x = input_105_cast_fp16)[name = tensor("input_107_cast_fp16")]; + tensor linear8_1_weight_to_fp16 = const()[name = tensor("linear8_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(209876864)))]; + tensor linear_35_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear8_1_weight_to_fp16, x = input_107_cast_fp16)[name = tensor("linear_35_cast_fp16")]; + tensor input_109_mode_0 = const()[name = tensor("input_109_mode_0"), val = tensor("EXACT")]; + tensor input_109_cast_fp16 = gelu(mode = input_109_mode_0, x = linear_35_cast_fp16)[name = tensor("input_109_cast_fp16")]; + tensor linear8_2_weight_to_fp16 = const()[name = tensor("linear8_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(218265536)))]; + tensor linear_36_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear8_2_weight_to_fp16, x = input_109_cast_fp16)[name = tensor("linear_36_cast_fp16")]; + tensor input_111_cast_fp16 = add(x = input_105_cast_fp16, y = linear_36_cast_fp16)[name = tensor("input_111_cast_fp16")]; + tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([-1])]; + tensor norm9_1_weight_to_fp16 = const()[name = tensor("norm9_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226654208)))]; + tensor norm9_1_bias_to_fp16 = const()[name = tensor("norm9_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226656320)))]; + tensor var_2423_to_fp16 = const()[name = tensor("op_2423_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_113_cast_fp16 = layer_norm(axes = input_113_axes_0, beta = norm9_1_bias_to_fp16, epsilon = var_2423_to_fp16, gamma = norm9_1_weight_to_fp16, x = input_111_cast_fp16)[name = tensor("input_113_cast_fp16")]; + tensor attn9_in_proj_weight_to_fp16 = const()[name = tensor("attn9_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226658432)))]; + tensor linear_37_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn9_in_proj_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("linear_37_cast_fp16")]; + tensor var_2436 = const()[name = tensor("op_2436"), val = tensor([1, 3, 16, 64])]; + tensor qkv_19_cast_fp16 = reshape(shape = var_2436, x = linear_37_cast_fp16)[name = tensor("qkv_19_cast_fp16")]; + tensor x_37_begin_0 = const()[name = tensor("x_37_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_37_end_0 = const()[name = tensor("x_37_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_37_end_mask_0 = const()[name = tensor("x_37_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_37_squeeze_mask_0 = const()[name = tensor("x_37_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_37_cast_fp16 = slice_by_index(begin = x_37_begin_0, end = x_37_end_0, end_mask = x_37_end_mask_0, squeeze_mask = x_37_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("x_37_cast_fp16")]; + tensor x_39_begin_0 = const()[name = tensor("x_39_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_39_end_0 = const()[name = tensor("x_39_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_39_end_mask_0 = const()[name = tensor("x_39_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_39_squeeze_mask_0 = const()[name = tensor("x_39_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_39_cast_fp16 = slice_by_index(begin = x_39_begin_0, end = x_39_end_0, end_mask = x_39_end_mask_0, squeeze_mask = x_39_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("x_39_cast_fp16")]; + tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_19_cast_fp16 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("v_19_cast_fp16")]; + tensor var_2465 = const()[name = tensor("op_2465"), val = tensor([1, 1, 1])]; + tensor position9_to_fp16_dtype_0 = const()[name = tensor("position9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position9_to_fp16 = cast(dtype = position9_to_fp16_dtype_0, x = position9)[name = tensor("cast_505")]; + tensor var_2466_cast_fp16 = reshape(shape = var_2465, x = position9_to_fp16)[name = tensor("op_2466_cast_fp16")]; + tensor angles_37_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_2466_cast_fp16)[name = tensor("angles_37_cast_fp16")]; + tensor rotr_37_cast_fp16 = cos(x = angles_37_cast_fp16)[name = tensor("rotr_37_cast_fp16")]; + tensor roti_37_cast_fp16 = sin(x = angles_37_cast_fp16)[name = tensor("roti_37_cast_fp16")]; + tensor var_2474 = const()[name = tensor("op_2474"), val = tensor([1, 16, 32, 2])]; + tensor pairs_37_cast_fp16 = reshape(shape = var_2474, x = x_37_cast_fp16)[name = tensor("pairs_37_cast_fp16")]; + tensor xr_37_begin_0 = const()[name = tensor("xr_37_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_37_end_0 = const()[name = tensor("xr_37_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_37_end_mask_0 = const()[name = tensor("xr_37_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_37_squeeze_mask_0 = const()[name = tensor("xr_37_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_37_cast_fp16 = slice_by_index(begin = xr_37_begin_0, end = xr_37_end_0, end_mask = xr_37_end_mask_0, squeeze_mask = xr_37_squeeze_mask_0, x = pairs_37_cast_fp16)[name = tensor("xr_37_cast_fp16")]; + tensor xi_37_begin_0 = const()[name = tensor("xi_37_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_37_end_0 = const()[name = tensor("xi_37_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_37_end_mask_0 = const()[name = tensor("xi_37_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_37_squeeze_mask_0 = const()[name = tensor("xi_37_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_37_cast_fp16 = slice_by_index(begin = xi_37_begin_0, end = xi_37_end_0, end_mask = xi_37_end_mask_0, squeeze_mask = xi_37_squeeze_mask_0, x = pairs_37_cast_fp16)[name = tensor("xi_37_cast_fp16")]; + tensor var_2482_cast_fp16 = mul(x = xr_37_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_2482_cast_fp16")]; + tensor var_2483_cast_fp16 = mul(x = xi_37_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_2483_cast_fp16")]; + tensor xor__37_cast_fp16 = sub(x = var_2482_cast_fp16, y = var_2483_cast_fp16)[name = tensor("xor__37_cast_fp16")]; + tensor var_2486_cast_fp16 = mul(x = xr_37_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_2486_cast_fp16")]; + tensor var_2487_cast_fp16 = mul(x = xi_37_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_2487_cast_fp16")]; + tensor xoi_37_cast_fp16 = add(x = var_2486_cast_fp16, y = var_2487_cast_fp16)[name = tensor("xoi_37_cast_fp16")]; + tensor out_37_axis_0 = const()[name = tensor("out_37_axis_0"), val = tensor(-1)]; + tensor out_37_cast_fp16 = stack(axis = out_37_axis_0, values = (xor__37_cast_fp16, xoi_37_cast_fp16))[name = tensor("out_37_cast_fp16")]; + tensor var_2510 = const()[name = tensor("op_2510"), val = tensor([1, 16, 32, 2])]; + tensor pairs_39_cast_fp16 = reshape(shape = var_2510, x = x_39_cast_fp16)[name = tensor("pairs_39_cast_fp16")]; + tensor xr_39_begin_0 = const()[name = tensor("xr_39_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_39_end_0 = const()[name = tensor("xr_39_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_39_end_mask_0 = const()[name = tensor("xr_39_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_39_squeeze_mask_0 = const()[name = tensor("xr_39_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_39_cast_fp16 = slice_by_index(begin = xr_39_begin_0, end = xr_39_end_0, end_mask = xr_39_end_mask_0, squeeze_mask = xr_39_squeeze_mask_0, x = pairs_39_cast_fp16)[name = tensor("xr_39_cast_fp16")]; + tensor xi_39_begin_0 = const()[name = tensor("xi_39_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_39_end_0 = const()[name = tensor("xi_39_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_39_end_mask_0 = const()[name = tensor("xi_39_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_39_squeeze_mask_0 = const()[name = tensor("xi_39_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_39_cast_fp16 = slice_by_index(begin = xi_39_begin_0, end = xi_39_end_0, end_mask = xi_39_end_mask_0, squeeze_mask = xi_39_squeeze_mask_0, x = pairs_39_cast_fp16)[name = tensor("xi_39_cast_fp16")]; + tensor var_2518_cast_fp16 = mul(x = xr_39_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_2518_cast_fp16")]; + tensor var_2519_cast_fp16 = mul(x = xi_39_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_2519_cast_fp16")]; + tensor xor__39_cast_fp16 = sub(x = var_2518_cast_fp16, y = var_2519_cast_fp16)[name = tensor("xor__39_cast_fp16")]; + tensor var_2522_cast_fp16 = mul(x = xr_39_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_2522_cast_fp16")]; + tensor var_2523_cast_fp16 = mul(x = xi_39_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_2523_cast_fp16")]; + tensor xoi_39_cast_fp16 = add(x = var_2522_cast_fp16, y = var_2523_cast_fp16)[name = tensor("xoi_39_cast_fp16")]; + tensor out_39_axis_0 = const()[name = tensor("out_39_axis_0"), val = tensor(-1)]; + tensor out_39_cast_fp16 = stack(axis = out_39_axis_0, values = (xor__39_cast_fp16, xoi_39_cast_fp16))[name = tensor("out_39_cast_fp16")]; + tensor _inversed_2535_y_0_to_fp16 = const()[name = tensor("_inversed_2535_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2535_cast_fp16 = mul(x = position9_to_fp16, y = _inversed_2535_y_0_to_fp16)[name = tensor("_inversed_2535_cast_fp16")]; + tensor var_2536_cast_fp16 = floor(x = _inversed_2535_cast_fp16)[name = tensor("op_2536_cast_fp16")]; + tensor var_2537_to_fp16 = const()[name = tensor("op_2537_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2538_cast_fp16 = mul(x = var_2536_cast_fp16, y = var_2537_to_fp16)[name = tensor("op_2538_cast_fp16")]; + tensor write_idx_19_cast_fp16 = sub(x = position9_to_fp16, y = var_2538_cast_fp16)[name = tensor("write_idx_19_cast_fp16")]; + tensor var_2541_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_19_cast_fp16)[name = tensor("op_2541_cast_fp16")]; + tensor var_2551 = const()[name = tensor("op_2551"), val = tensor([1, 512, 1, 1])]; + tensor onehot_37_to_fp16_dtype_0 = const()[name = tensor("onehot_37_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_2541_cast_fp16_to_fp16 = cast(dtype = onehot_37_to_fp16_dtype_0, x = var_2541_cast_fp16)[name = tensor("cast_504")]; + tensor onehot_39_cast_fp16 = reshape(shape = var_2551, x = var_2541_cast_fp16_to_fp16)[name = tensor("onehot_39_cast_fp16")]; + tensor var_2553_to_fp16 = const()[name = tensor("op_2553_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2555_cast_fp16 = sub(x = var_2553_to_fp16, y = onehot_39_cast_fp16)[name = tensor("op_2555_cast_fp16")]; + tensor k_cache9_to_fp16_dtype_0 = const()[name = tensor("k_cache9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache9_to_fp16 = cast(dtype = k_cache9_to_fp16_dtype_0, x = k_cache9)[name = tensor("cast_503")]; + tensor var_2556_cast_fp16 = mul(x = k_cache9_to_fp16, y = var_2555_cast_fp16)[name = tensor("op_2556_cast_fp16")]; + tensor var_2561 = const()[name = tensor("op_2561"), val = tensor([1, 1, 16, 64])]; + tensor var_2562_cast_fp16 = reshape(shape = var_2561, x = out_39_cast_fp16)[name = tensor("op_2562_cast_fp16")]; + tensor var_2563_cast_fp16 = mul(x = var_2562_cast_fp16, y = onehot_39_cast_fp16)[name = tensor("op_2563_cast_fp16")]; + tensor new_k_cache_19_cast_fp16 = add(x = var_2556_cast_fp16, y = var_2563_cast_fp16)[name = tensor("new_k_cache_19_cast_fp16")]; + tensor new_k_cache_19_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_19_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache9_to_fp16_dtype_0 = const()[name = tensor("v_cache9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache9_to_fp16 = cast(dtype = v_cache9_to_fp16_dtype_0, x = v_cache9)[name = tensor("cast_501")]; + tensor var_2569_cast_fp16 = mul(x = v_cache9_to_fp16, y = var_2555_cast_fp16)[name = tensor("op_2569_cast_fp16")]; + tensor var_2574 = const()[name = tensor("op_2574"), val = tensor([1, 1, 16, 64])]; + tensor var_2575_cast_fp16 = reshape(shape = var_2574, x = v_19_cast_fp16)[name = tensor("op_2575_cast_fp16")]; + tensor var_2576_cast_fp16 = mul(x = var_2575_cast_fp16, y = onehot_39_cast_fp16)[name = tensor("op_2576_cast_fp16")]; + tensor new_v_cache_19_cast_fp16 = add(x = var_2569_cast_fp16, y = var_2576_cast_fp16)[name = tensor("new_v_cache_19_cast_fp16")]; + tensor new_v_cache_19_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_19_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_9_perm_0 = const()[name = tensor("transpose_9_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_2589 = const()[name = tensor("op_2589"), val = tensor([0, 2, 1, 3])]; + tensor var_2595 = const()[name = tensor("op_2595"), val = tensor([1, 16, 1, 64])]; + tensor q4_19_cast_fp16 = reshape(shape = var_2595, x = out_37_cast_fp16)[name = tensor("q4_19_cast_fp16")]; + tensor var_2600_transpose_x_0 = const()[name = tensor("op_2600_transpose_x_0"), val = tensor(false)]; + tensor var_2600_transpose_y_0 = const()[name = tensor("op_2600_transpose_y_0"), val = tensor(false)]; + tensor transpose_9_cast_fp16 = transpose(perm = transpose_9_perm_0, x = new_k_cache_19_cast_fp16)[name = tensor("transpose_68")]; + tensor var_2600_cast_fp16 = matmul(transpose_x = var_2600_transpose_x_0, transpose_y = var_2600_transpose_y_0, x = q4_19_cast_fp16, y = transpose_9_cast_fp16)[name = tensor("op_2600_cast_fp16")]; + tensor var_2601_to_fp16 = const()[name = tensor("op_2601_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_37_cast_fp16 = mul(x = var_2600_cast_fp16, y = var_2601_to_fp16)[name = tensor("scores_37_cast_fp16")]; + tensor var_2605 = const()[name = tensor("op_2605"), val = tensor([1, 1])]; + tensor var_2606_cast_fp16 = reshape(shape = var_2605, x = position9_to_fp16)[name = tensor("op_2606_cast_fp16")]; + tensor var_2607_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_2606_cast_fp16)[name = tensor("op_2607_cast_fp16")]; + tensor mask_19_to_fp16_dtype_0 = const()[name = tensor("mask_19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_2614_to_fp16 = const()[name = tensor("op_2614_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2607_cast_fp16_to_fp16 = cast(dtype = mask_19_to_fp16_dtype_0, x = var_2607_cast_fp16)[name = tensor("cast_499")]; + tensor var_2615_cast_fp16 = sub(x = var_2607_cast_fp16_to_fp16, y = var_2614_to_fp16)[name = tensor("op_2615_cast_fp16")]; + tensor var_2620 = const()[name = tensor("op_2620"), val = tensor([1, 1, 1, 512])]; + tensor var_2621_cast_fp16 = reshape(shape = var_2620, x = var_2615_cast_fp16)[name = tensor("op_2621_cast_fp16")]; + tensor var_2622_to_fp16 = const()[name = tensor("op_2622_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_2623_cast_fp16 = mul(x = var_2621_cast_fp16, y = var_2622_to_fp16)[name = tensor("op_2623_cast_fp16")]; + tensor scores_39_cast_fp16 = add(x = scores_37_cast_fp16, y = var_2623_cast_fp16)[name = tensor("scores_39_cast_fp16")]; + tensor var_2626 = const()[name = tensor("op_2626"), val = tensor(-1)]; + tensor weights_19_cast_fp16 = softmax(axis = var_2626, x = scores_39_cast_fp16)[name = tensor("weights_19_cast_fp16")]; + tensor attn_19_transpose_x_0 = const()[name = tensor("attn_19_transpose_x_0"), val = tensor(false)]; + tensor attn_19_transpose_y_0 = const()[name = tensor("attn_19_transpose_y_0"), val = tensor(false)]; + tensor values_19_cast_fp16 = transpose(perm = var_2589, x = new_v_cache_19_cast_fp16)[name = tensor("transpose_67")]; + tensor attn_19_cast_fp16 = matmul(transpose_x = attn_19_transpose_x_0, transpose_y = attn_19_transpose_y_0, x = weights_19_cast_fp16, y = values_19_cast_fp16)[name = tensor("attn_19_cast_fp16")]; + tensor var_2634 = const()[name = tensor("op_2634"), val = tensor([0, 2, 1, 3])]; + tensor var_2639 = const()[name = tensor("op_2639"), val = tensor([1, 1, 1024])]; + tensor var_2635_cast_fp16 = transpose(perm = var_2634, x = attn_19_cast_fp16)[name = tensor("transpose_66")]; + tensor input_115_cast_fp16 = reshape(shape = var_2639, x = var_2635_cast_fp16)[name = tensor("input_115_cast_fp16")]; + tensor attn9_out_proj_weight_to_fp16 = const()[name = tensor("attn9_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(232949952)))]; + tensor linear_38_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn9_out_proj_weight_to_fp16, x = input_115_cast_fp16)[name = tensor("linear_38_cast_fp16")]; + tensor var_2645_to_fp16 = const()[name = tensor("op_2645_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2646_cast_fp16 = add(x = position9_to_fp16, y = var_2645_to_fp16)[name = tensor("op_2646_cast_fp16")]; + tensor var_2646_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2646_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_117_cast_fp16 = add(x = input_111_cast_fp16, y = linear_38_cast_fp16)[name = tensor("input_117_cast_fp16")]; + tensor input_119_axes_0 = const()[name = tensor("input_119_axes_0"), val = tensor([-1])]; + tensor norm9_2_weight_to_fp16 = const()[name = tensor("norm9_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(235047168)))]; + tensor norm9_2_bias_to_fp16 = const()[name = tensor("norm9_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(235049280)))]; + tensor var_2650_to_fp16 = const()[name = tensor("op_2650_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_119_cast_fp16 = layer_norm(axes = input_119_axes_0, beta = norm9_2_bias_to_fp16, epsilon = var_2650_to_fp16, gamma = norm9_2_weight_to_fp16, x = input_117_cast_fp16)[name = tensor("input_119_cast_fp16")]; + tensor linear9_1_weight_to_fp16 = const()[name = tensor("linear9_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(235051392)))]; + tensor linear_39_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear9_1_weight_to_fp16, x = input_119_cast_fp16)[name = tensor("linear_39_cast_fp16")]; + tensor input_121_mode_0 = const()[name = tensor("input_121_mode_0"), val = tensor("EXACT")]; + tensor input_121_cast_fp16 = gelu(mode = input_121_mode_0, x = linear_39_cast_fp16)[name = tensor("input_121_cast_fp16")]; + tensor linear9_2_weight_to_fp16 = const()[name = tensor("linear9_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(243440064)))]; + tensor linear_40_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear9_2_weight_to_fp16, x = input_121_cast_fp16)[name = tensor("linear_40_cast_fp16")]; + tensor input_123_cast_fp16 = add(x = input_117_cast_fp16, y = linear_40_cast_fp16)[name = tensor("input_123_cast_fp16")]; + tensor input_125_axes_0 = const()[name = tensor("input_125_axes_0"), val = tensor([-1])]; + tensor norm10_1_weight_to_fp16 = const()[name = tensor("norm10_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(251828736)))]; + tensor norm10_1_bias_to_fp16 = const()[name = tensor("norm10_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(251830848)))]; + tensor var_2667_to_fp16 = const()[name = tensor("op_2667_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_125_cast_fp16 = layer_norm(axes = input_125_axes_0, beta = norm10_1_bias_to_fp16, epsilon = var_2667_to_fp16, gamma = norm10_1_weight_to_fp16, x = input_123_cast_fp16)[name = tensor("input_125_cast_fp16")]; + tensor attn10_in_proj_weight_to_fp16 = const()[name = tensor("attn10_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(251832960)))]; + tensor linear_41_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn10_in_proj_weight_to_fp16, x = input_125_cast_fp16)[name = tensor("linear_41_cast_fp16")]; + tensor var_2680 = const()[name = tensor("op_2680"), val = tensor([1, 3, 16, 64])]; + tensor qkv_21_cast_fp16 = reshape(shape = var_2680, x = linear_41_cast_fp16)[name = tensor("qkv_21_cast_fp16")]; + tensor x_41_begin_0 = const()[name = tensor("x_41_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_41_end_0 = const()[name = tensor("x_41_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_41_end_mask_0 = const()[name = tensor("x_41_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_41_squeeze_mask_0 = const()[name = tensor("x_41_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_41_cast_fp16 = slice_by_index(begin = x_41_begin_0, end = x_41_end_0, end_mask = x_41_end_mask_0, squeeze_mask = x_41_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("x_41_cast_fp16")]; + tensor x_43_begin_0 = const()[name = tensor("x_43_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_43_end_0 = const()[name = tensor("x_43_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_43_end_mask_0 = const()[name = tensor("x_43_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_43_squeeze_mask_0 = const()[name = tensor("x_43_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_43_cast_fp16 = slice_by_index(begin = x_43_begin_0, end = x_43_end_0, end_mask = x_43_end_mask_0, squeeze_mask = x_43_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("x_43_cast_fp16")]; + tensor v_21_begin_0 = const()[name = tensor("v_21_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_21_end_0 = const()[name = tensor("v_21_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_21_end_mask_0 = const()[name = tensor("v_21_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_21_squeeze_mask_0 = const()[name = tensor("v_21_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_21_cast_fp16 = slice_by_index(begin = v_21_begin_0, end = v_21_end_0, end_mask = v_21_end_mask_0, squeeze_mask = v_21_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("v_21_cast_fp16")]; + tensor var_2709 = const()[name = tensor("op_2709"), val = tensor([1, 1, 1])]; + tensor position10_to_fp16_dtype_0 = const()[name = tensor("position10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position10_to_fp16 = cast(dtype = position10_to_fp16_dtype_0, x = position10)[name = tensor("cast_497")]; + tensor var_2710_cast_fp16 = reshape(shape = var_2709, x = position10_to_fp16)[name = tensor("op_2710_cast_fp16")]; + tensor angles_41_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_2710_cast_fp16)[name = tensor("angles_41_cast_fp16")]; + tensor rotr_41_cast_fp16 = cos(x = angles_41_cast_fp16)[name = tensor("rotr_41_cast_fp16")]; + tensor roti_41_cast_fp16 = sin(x = angles_41_cast_fp16)[name = tensor("roti_41_cast_fp16")]; + tensor var_2718 = const()[name = tensor("op_2718"), val = tensor([1, 16, 32, 2])]; + tensor pairs_41_cast_fp16 = reshape(shape = var_2718, x = x_41_cast_fp16)[name = tensor("pairs_41_cast_fp16")]; + tensor xr_41_begin_0 = const()[name = tensor("xr_41_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_41_end_0 = const()[name = tensor("xr_41_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_41_end_mask_0 = const()[name = tensor("xr_41_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_41_squeeze_mask_0 = const()[name = tensor("xr_41_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_41_cast_fp16 = slice_by_index(begin = xr_41_begin_0, end = xr_41_end_0, end_mask = xr_41_end_mask_0, squeeze_mask = xr_41_squeeze_mask_0, x = pairs_41_cast_fp16)[name = tensor("xr_41_cast_fp16")]; + tensor xi_41_begin_0 = const()[name = tensor("xi_41_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_41_end_0 = const()[name = tensor("xi_41_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_41_end_mask_0 = const()[name = tensor("xi_41_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_41_squeeze_mask_0 = const()[name = tensor("xi_41_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_41_cast_fp16 = slice_by_index(begin = xi_41_begin_0, end = xi_41_end_0, end_mask = xi_41_end_mask_0, squeeze_mask = xi_41_squeeze_mask_0, x = pairs_41_cast_fp16)[name = tensor("xi_41_cast_fp16")]; + tensor var_2726_cast_fp16 = mul(x = xr_41_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_2726_cast_fp16")]; + tensor var_2727_cast_fp16 = mul(x = xi_41_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_2727_cast_fp16")]; + tensor xor__41_cast_fp16 = sub(x = var_2726_cast_fp16, y = var_2727_cast_fp16)[name = tensor("xor__41_cast_fp16")]; + tensor var_2730_cast_fp16 = mul(x = xr_41_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_2730_cast_fp16")]; + tensor var_2731_cast_fp16 = mul(x = xi_41_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_2731_cast_fp16")]; + tensor xoi_41_cast_fp16 = add(x = var_2730_cast_fp16, y = var_2731_cast_fp16)[name = tensor("xoi_41_cast_fp16")]; + tensor out_41_axis_0 = const()[name = tensor("out_41_axis_0"), val = tensor(-1)]; + tensor out_41_cast_fp16 = stack(axis = out_41_axis_0, values = (xor__41_cast_fp16, xoi_41_cast_fp16))[name = tensor("out_41_cast_fp16")]; + tensor var_2754 = const()[name = tensor("op_2754"), val = tensor([1, 16, 32, 2])]; + tensor pairs_43_cast_fp16 = reshape(shape = var_2754, x = x_43_cast_fp16)[name = tensor("pairs_43_cast_fp16")]; + tensor xr_43_begin_0 = const()[name = tensor("xr_43_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_43_end_0 = const()[name = tensor("xr_43_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_43_end_mask_0 = const()[name = tensor("xr_43_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_43_squeeze_mask_0 = const()[name = tensor("xr_43_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_43_cast_fp16 = slice_by_index(begin = xr_43_begin_0, end = xr_43_end_0, end_mask = xr_43_end_mask_0, squeeze_mask = xr_43_squeeze_mask_0, x = pairs_43_cast_fp16)[name = tensor("xr_43_cast_fp16")]; + tensor xi_43_begin_0 = const()[name = tensor("xi_43_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_43_end_0 = const()[name = tensor("xi_43_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_43_end_mask_0 = const()[name = tensor("xi_43_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_43_squeeze_mask_0 = const()[name = tensor("xi_43_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_43_cast_fp16 = slice_by_index(begin = xi_43_begin_0, end = xi_43_end_0, end_mask = xi_43_end_mask_0, squeeze_mask = xi_43_squeeze_mask_0, x = pairs_43_cast_fp16)[name = tensor("xi_43_cast_fp16")]; + tensor var_2762_cast_fp16 = mul(x = xr_43_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_2762_cast_fp16")]; + tensor var_2763_cast_fp16 = mul(x = xi_43_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_2763_cast_fp16")]; + tensor xor__43_cast_fp16 = sub(x = var_2762_cast_fp16, y = var_2763_cast_fp16)[name = tensor("xor__43_cast_fp16")]; + tensor var_2766_cast_fp16 = mul(x = xr_43_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_2766_cast_fp16")]; + tensor var_2767_cast_fp16 = mul(x = xi_43_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_2767_cast_fp16")]; + tensor xoi_43_cast_fp16 = add(x = var_2766_cast_fp16, y = var_2767_cast_fp16)[name = tensor("xoi_43_cast_fp16")]; + tensor out_43_axis_0 = const()[name = tensor("out_43_axis_0"), val = tensor(-1)]; + tensor out_43_cast_fp16 = stack(axis = out_43_axis_0, values = (xor__43_cast_fp16, xoi_43_cast_fp16))[name = tensor("out_43_cast_fp16")]; + tensor _inversed_2779_y_0_to_fp16 = const()[name = tensor("_inversed_2779_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2779_cast_fp16 = mul(x = position10_to_fp16, y = _inversed_2779_y_0_to_fp16)[name = tensor("_inversed_2779_cast_fp16")]; + tensor var_2780_cast_fp16 = floor(x = _inversed_2779_cast_fp16)[name = tensor("op_2780_cast_fp16")]; + tensor var_2781_to_fp16 = const()[name = tensor("op_2781_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2782_cast_fp16 = mul(x = var_2780_cast_fp16, y = var_2781_to_fp16)[name = tensor("op_2782_cast_fp16")]; + tensor write_idx_21_cast_fp16 = sub(x = position10_to_fp16, y = var_2782_cast_fp16)[name = tensor("write_idx_21_cast_fp16")]; + tensor var_2785_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_21_cast_fp16)[name = tensor("op_2785_cast_fp16")]; + tensor var_2795 = const()[name = tensor("op_2795"), val = tensor([1, 512, 1, 1])]; + tensor onehot_41_to_fp16_dtype_0 = const()[name = tensor("onehot_41_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_2785_cast_fp16_to_fp16 = cast(dtype = onehot_41_to_fp16_dtype_0, x = var_2785_cast_fp16)[name = tensor("cast_496")]; + tensor onehot_43_cast_fp16 = reshape(shape = var_2795, x = var_2785_cast_fp16_to_fp16)[name = tensor("onehot_43_cast_fp16")]; + tensor var_2797_to_fp16 = const()[name = tensor("op_2797_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2799_cast_fp16 = sub(x = var_2797_to_fp16, y = onehot_43_cast_fp16)[name = tensor("op_2799_cast_fp16")]; + tensor k_cache10_to_fp16_dtype_0 = const()[name = tensor("k_cache10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache10_to_fp16 = cast(dtype = k_cache10_to_fp16_dtype_0, x = k_cache10)[name = tensor("cast_495")]; + tensor var_2800_cast_fp16 = mul(x = k_cache10_to_fp16, y = var_2799_cast_fp16)[name = tensor("op_2800_cast_fp16")]; + tensor var_2805 = const()[name = tensor("op_2805"), val = tensor([1, 1, 16, 64])]; + tensor var_2806_cast_fp16 = reshape(shape = var_2805, x = out_43_cast_fp16)[name = tensor("op_2806_cast_fp16")]; + tensor var_2807_cast_fp16 = mul(x = var_2806_cast_fp16, y = onehot_43_cast_fp16)[name = tensor("op_2807_cast_fp16")]; + tensor new_k_cache_21_cast_fp16 = add(x = var_2800_cast_fp16, y = var_2807_cast_fp16)[name = tensor("new_k_cache_21_cast_fp16")]; + tensor new_k_cache_21_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_21_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache10_to_fp16_dtype_0 = const()[name = tensor("v_cache10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache10_to_fp16 = cast(dtype = v_cache10_to_fp16_dtype_0, x = v_cache10)[name = tensor("cast_493")]; + tensor var_2813_cast_fp16 = mul(x = v_cache10_to_fp16, y = var_2799_cast_fp16)[name = tensor("op_2813_cast_fp16")]; + tensor var_2818 = const()[name = tensor("op_2818"), val = tensor([1, 1, 16, 64])]; + tensor var_2819_cast_fp16 = reshape(shape = var_2818, x = v_21_cast_fp16)[name = tensor("op_2819_cast_fp16")]; + tensor var_2820_cast_fp16 = mul(x = var_2819_cast_fp16, y = onehot_43_cast_fp16)[name = tensor("op_2820_cast_fp16")]; + tensor new_v_cache_21_cast_fp16 = add(x = var_2813_cast_fp16, y = var_2820_cast_fp16)[name = tensor("new_v_cache_21_cast_fp16")]; + tensor new_v_cache_21_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_21_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_10_perm_0 = const()[name = tensor("transpose_10_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_2833 = const()[name = tensor("op_2833"), val = tensor([0, 2, 1, 3])]; + tensor var_2839 = const()[name = tensor("op_2839"), val = tensor([1, 16, 1, 64])]; + tensor q4_21_cast_fp16 = reshape(shape = var_2839, x = out_41_cast_fp16)[name = tensor("q4_21_cast_fp16")]; + tensor var_2844_transpose_x_0 = const()[name = tensor("op_2844_transpose_x_0"), val = tensor(false)]; + tensor var_2844_transpose_y_0 = const()[name = tensor("op_2844_transpose_y_0"), val = tensor(false)]; + tensor transpose_10_cast_fp16 = transpose(perm = transpose_10_perm_0, x = new_k_cache_21_cast_fp16)[name = tensor("transpose_65")]; + tensor var_2844_cast_fp16 = matmul(transpose_x = var_2844_transpose_x_0, transpose_y = var_2844_transpose_y_0, x = q4_21_cast_fp16, y = transpose_10_cast_fp16)[name = tensor("op_2844_cast_fp16")]; + tensor var_2845_to_fp16 = const()[name = tensor("op_2845_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_41_cast_fp16 = mul(x = var_2844_cast_fp16, y = var_2845_to_fp16)[name = tensor("scores_41_cast_fp16")]; + tensor var_2849 = const()[name = tensor("op_2849"), val = tensor([1, 1])]; + tensor var_2850_cast_fp16 = reshape(shape = var_2849, x = position10_to_fp16)[name = tensor("op_2850_cast_fp16")]; + tensor var_2851_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_2850_cast_fp16)[name = tensor("op_2851_cast_fp16")]; + tensor mask_21_to_fp16_dtype_0 = const()[name = tensor("mask_21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_2858_to_fp16 = const()[name = tensor("op_2858_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2851_cast_fp16_to_fp16 = cast(dtype = mask_21_to_fp16_dtype_0, x = var_2851_cast_fp16)[name = tensor("cast_491")]; + tensor var_2859_cast_fp16 = sub(x = var_2851_cast_fp16_to_fp16, y = var_2858_to_fp16)[name = tensor("op_2859_cast_fp16")]; + tensor var_2864 = const()[name = tensor("op_2864"), val = tensor([1, 1, 1, 512])]; + tensor var_2865_cast_fp16 = reshape(shape = var_2864, x = var_2859_cast_fp16)[name = tensor("op_2865_cast_fp16")]; + tensor var_2866_to_fp16 = const()[name = tensor("op_2866_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_2867_cast_fp16 = mul(x = var_2865_cast_fp16, y = var_2866_to_fp16)[name = tensor("op_2867_cast_fp16")]; + tensor scores_43_cast_fp16 = add(x = scores_41_cast_fp16, y = var_2867_cast_fp16)[name = tensor("scores_43_cast_fp16")]; + tensor var_2870 = const()[name = tensor("op_2870"), val = tensor(-1)]; + tensor weights_21_cast_fp16 = softmax(axis = var_2870, x = scores_43_cast_fp16)[name = tensor("weights_21_cast_fp16")]; + tensor attn_21_transpose_x_0 = const()[name = tensor("attn_21_transpose_x_0"), val = tensor(false)]; + tensor attn_21_transpose_y_0 = const()[name = tensor("attn_21_transpose_y_0"), val = tensor(false)]; + tensor values_21_cast_fp16 = transpose(perm = var_2833, x = new_v_cache_21_cast_fp16)[name = tensor("transpose_64")]; + tensor attn_21_cast_fp16 = matmul(transpose_x = attn_21_transpose_x_0, transpose_y = attn_21_transpose_y_0, x = weights_21_cast_fp16, y = values_21_cast_fp16)[name = tensor("attn_21_cast_fp16")]; + tensor var_2878 = const()[name = tensor("op_2878"), val = tensor([0, 2, 1, 3])]; + tensor var_2883 = const()[name = tensor("op_2883"), val = tensor([1, 1, 1024])]; + tensor var_2879_cast_fp16 = transpose(perm = var_2878, x = attn_21_cast_fp16)[name = tensor("transpose_63")]; + tensor input_127_cast_fp16 = reshape(shape = var_2883, x = var_2879_cast_fp16)[name = tensor("input_127_cast_fp16")]; + tensor attn10_out_proj_weight_to_fp16 = const()[name = tensor("attn10_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(258124480)))]; + tensor linear_42_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn10_out_proj_weight_to_fp16, x = input_127_cast_fp16)[name = tensor("linear_42_cast_fp16")]; + tensor var_2889_to_fp16 = const()[name = tensor("op_2889_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2890_cast_fp16 = add(x = position10_to_fp16, y = var_2889_to_fp16)[name = tensor("op_2890_cast_fp16")]; + tensor var_2890_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2890_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_129_cast_fp16 = add(x = input_123_cast_fp16, y = linear_42_cast_fp16)[name = tensor("input_129_cast_fp16")]; + tensor input_131_axes_0 = const()[name = tensor("input_131_axes_0"), val = tensor([-1])]; + tensor norm10_2_weight_to_fp16 = const()[name = tensor("norm10_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260221696)))]; + tensor norm10_2_bias_to_fp16 = const()[name = tensor("norm10_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260223808)))]; + tensor var_2894_to_fp16 = const()[name = tensor("op_2894_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_131_cast_fp16 = layer_norm(axes = input_131_axes_0, beta = norm10_2_bias_to_fp16, epsilon = var_2894_to_fp16, gamma = norm10_2_weight_to_fp16, x = input_129_cast_fp16)[name = tensor("input_131_cast_fp16")]; + tensor linear10_1_weight_to_fp16 = const()[name = tensor("linear10_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260225920)))]; + tensor linear_43_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear10_1_weight_to_fp16, x = input_131_cast_fp16)[name = tensor("linear_43_cast_fp16")]; + tensor input_133_mode_0 = const()[name = tensor("input_133_mode_0"), val = tensor("EXACT")]; + tensor input_133_cast_fp16 = gelu(mode = input_133_mode_0, x = linear_43_cast_fp16)[name = tensor("input_133_cast_fp16")]; + tensor linear10_2_weight_to_fp16 = const()[name = tensor("linear10_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(268614592)))]; + tensor linear_44_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear10_2_weight_to_fp16, x = input_133_cast_fp16)[name = tensor("linear_44_cast_fp16")]; + tensor input_135_cast_fp16 = add(x = input_129_cast_fp16, y = linear_44_cast_fp16)[name = tensor("input_135_cast_fp16")]; + tensor input_137_axes_0 = const()[name = tensor("input_137_axes_0"), val = tensor([-1])]; + tensor norm11_1_weight_to_fp16 = const()[name = tensor("norm11_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277003264)))]; + tensor norm11_1_bias_to_fp16 = const()[name = tensor("norm11_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277005376)))]; + tensor var_2911_to_fp16 = const()[name = tensor("op_2911_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_137_cast_fp16 = layer_norm(axes = input_137_axes_0, beta = norm11_1_bias_to_fp16, epsilon = var_2911_to_fp16, gamma = norm11_1_weight_to_fp16, x = input_135_cast_fp16)[name = tensor("input_137_cast_fp16")]; + tensor attn11_in_proj_weight_to_fp16 = const()[name = tensor("attn11_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277007488)))]; + tensor linear_45_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn11_in_proj_weight_to_fp16, x = input_137_cast_fp16)[name = tensor("linear_45_cast_fp16")]; + tensor var_2924 = const()[name = tensor("op_2924"), val = tensor([1, 3, 16, 64])]; + tensor qkv_23_cast_fp16 = reshape(shape = var_2924, x = linear_45_cast_fp16)[name = tensor("qkv_23_cast_fp16")]; + tensor x_45_begin_0 = const()[name = tensor("x_45_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_45_end_0 = const()[name = tensor("x_45_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_45_end_mask_0 = const()[name = tensor("x_45_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_45_squeeze_mask_0 = const()[name = tensor("x_45_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_45_cast_fp16 = slice_by_index(begin = x_45_begin_0, end = x_45_end_0, end_mask = x_45_end_mask_0, squeeze_mask = x_45_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("x_45_cast_fp16")]; + tensor x_47_begin_0 = const()[name = tensor("x_47_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_47_end_0 = const()[name = tensor("x_47_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_47_end_mask_0 = const()[name = tensor("x_47_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_47_squeeze_mask_0 = const()[name = tensor("x_47_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_47_cast_fp16 = slice_by_index(begin = x_47_begin_0, end = x_47_end_0, end_mask = x_47_end_mask_0, squeeze_mask = x_47_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("x_47_cast_fp16")]; + tensor v_23_begin_0 = const()[name = tensor("v_23_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_23_end_0 = const()[name = tensor("v_23_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_23_end_mask_0 = const()[name = tensor("v_23_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_23_squeeze_mask_0 = const()[name = tensor("v_23_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_23_cast_fp16 = slice_by_index(begin = v_23_begin_0, end = v_23_end_0, end_mask = v_23_end_mask_0, squeeze_mask = v_23_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("v_23_cast_fp16")]; + tensor var_2953 = const()[name = tensor("op_2953"), val = tensor([1, 1, 1])]; + tensor position11_to_fp16_dtype_0 = const()[name = tensor("position11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position11_to_fp16 = cast(dtype = position11_to_fp16_dtype_0, x = position11)[name = tensor("cast_489")]; + tensor var_2954_cast_fp16 = reshape(shape = var_2953, x = position11_to_fp16)[name = tensor("op_2954_cast_fp16")]; + tensor angles_45_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_2954_cast_fp16)[name = tensor("angles_45_cast_fp16")]; + tensor rotr_45_cast_fp16 = cos(x = angles_45_cast_fp16)[name = tensor("rotr_45_cast_fp16")]; + tensor roti_45_cast_fp16 = sin(x = angles_45_cast_fp16)[name = tensor("roti_45_cast_fp16")]; + tensor var_2962 = const()[name = tensor("op_2962"), val = tensor([1, 16, 32, 2])]; + tensor pairs_45_cast_fp16 = reshape(shape = var_2962, x = x_45_cast_fp16)[name = tensor("pairs_45_cast_fp16")]; + tensor xr_45_begin_0 = const()[name = tensor("xr_45_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_45_end_0 = const()[name = tensor("xr_45_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_45_end_mask_0 = const()[name = tensor("xr_45_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_45_squeeze_mask_0 = const()[name = tensor("xr_45_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_45_cast_fp16 = slice_by_index(begin = xr_45_begin_0, end = xr_45_end_0, end_mask = xr_45_end_mask_0, squeeze_mask = xr_45_squeeze_mask_0, x = pairs_45_cast_fp16)[name = tensor("xr_45_cast_fp16")]; + tensor xi_45_begin_0 = const()[name = tensor("xi_45_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_45_end_0 = const()[name = tensor("xi_45_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_45_end_mask_0 = const()[name = tensor("xi_45_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_45_squeeze_mask_0 = const()[name = tensor("xi_45_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_45_cast_fp16 = slice_by_index(begin = xi_45_begin_0, end = xi_45_end_0, end_mask = xi_45_end_mask_0, squeeze_mask = xi_45_squeeze_mask_0, x = pairs_45_cast_fp16)[name = tensor("xi_45_cast_fp16")]; + tensor var_2970_cast_fp16 = mul(x = xr_45_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_2970_cast_fp16")]; + tensor var_2971_cast_fp16 = mul(x = xi_45_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_2971_cast_fp16")]; + tensor xor__45_cast_fp16 = sub(x = var_2970_cast_fp16, y = var_2971_cast_fp16)[name = tensor("xor__45_cast_fp16")]; + tensor var_2974_cast_fp16 = mul(x = xr_45_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_2974_cast_fp16")]; + tensor var_2975_cast_fp16 = mul(x = xi_45_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_2975_cast_fp16")]; + tensor xoi_45_cast_fp16 = add(x = var_2974_cast_fp16, y = var_2975_cast_fp16)[name = tensor("xoi_45_cast_fp16")]; + tensor out_45_axis_0 = const()[name = tensor("out_45_axis_0"), val = tensor(-1)]; + tensor out_45_cast_fp16 = stack(axis = out_45_axis_0, values = (xor__45_cast_fp16, xoi_45_cast_fp16))[name = tensor("out_45_cast_fp16")]; + tensor var_2998 = const()[name = tensor("op_2998"), val = tensor([1, 16, 32, 2])]; + tensor pairs_47_cast_fp16 = reshape(shape = var_2998, x = x_47_cast_fp16)[name = tensor("pairs_47_cast_fp16")]; + tensor xr_47_begin_0 = const()[name = tensor("xr_47_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_47_end_0 = const()[name = tensor("xr_47_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_47_end_mask_0 = const()[name = tensor("xr_47_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_47_squeeze_mask_0 = const()[name = tensor("xr_47_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_47_cast_fp16 = slice_by_index(begin = xr_47_begin_0, end = xr_47_end_0, end_mask = xr_47_end_mask_0, squeeze_mask = xr_47_squeeze_mask_0, x = pairs_47_cast_fp16)[name = tensor("xr_47_cast_fp16")]; + tensor xi_47_begin_0 = const()[name = tensor("xi_47_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_47_end_0 = const()[name = tensor("xi_47_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_47_end_mask_0 = const()[name = tensor("xi_47_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_47_squeeze_mask_0 = const()[name = tensor("xi_47_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_47_cast_fp16 = slice_by_index(begin = xi_47_begin_0, end = xi_47_end_0, end_mask = xi_47_end_mask_0, squeeze_mask = xi_47_squeeze_mask_0, x = pairs_47_cast_fp16)[name = tensor("xi_47_cast_fp16")]; + tensor var_3006_cast_fp16 = mul(x = xr_47_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_3006_cast_fp16")]; + tensor var_3007_cast_fp16 = mul(x = xi_47_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_3007_cast_fp16")]; + tensor xor__47_cast_fp16 = sub(x = var_3006_cast_fp16, y = var_3007_cast_fp16)[name = tensor("xor__47_cast_fp16")]; + tensor var_3010_cast_fp16 = mul(x = xr_47_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_3010_cast_fp16")]; + tensor var_3011_cast_fp16 = mul(x = xi_47_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_3011_cast_fp16")]; + tensor xoi_47_cast_fp16 = add(x = var_3010_cast_fp16, y = var_3011_cast_fp16)[name = tensor("xoi_47_cast_fp16")]; + tensor out_47_axis_0 = const()[name = tensor("out_47_axis_0"), val = tensor(-1)]; + tensor out_47_cast_fp16 = stack(axis = out_47_axis_0, values = (xor__47_cast_fp16, xoi_47_cast_fp16))[name = tensor("out_47_cast_fp16")]; + tensor _inversed_3023_y_0_to_fp16 = const()[name = tensor("_inversed_3023_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3023_cast_fp16 = mul(x = position11_to_fp16, y = _inversed_3023_y_0_to_fp16)[name = tensor("_inversed_3023_cast_fp16")]; + tensor var_3024_cast_fp16 = floor(x = _inversed_3023_cast_fp16)[name = tensor("op_3024_cast_fp16")]; + tensor var_3025_to_fp16 = const()[name = tensor("op_3025_to_fp16"), val = tensor(0x1p+9)]; + tensor var_3026_cast_fp16 = mul(x = var_3024_cast_fp16, y = var_3025_to_fp16)[name = tensor("op_3026_cast_fp16")]; + tensor write_idx_23_cast_fp16 = sub(x = position11_to_fp16, y = var_3026_cast_fp16)[name = tensor("write_idx_23_cast_fp16")]; + tensor var_3029_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_23_cast_fp16)[name = tensor("op_3029_cast_fp16")]; + tensor var_3039 = const()[name = tensor("op_3039"), val = tensor([1, 512, 1, 1])]; + tensor onehot_45_to_fp16_dtype_0 = const()[name = tensor("onehot_45_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_3029_cast_fp16_to_fp16 = cast(dtype = onehot_45_to_fp16_dtype_0, x = var_3029_cast_fp16)[name = tensor("cast_488")]; + tensor onehot_47_cast_fp16 = reshape(shape = var_3039, x = var_3029_cast_fp16_to_fp16)[name = tensor("onehot_47_cast_fp16")]; + tensor var_3041_to_fp16 = const()[name = tensor("op_3041_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3043_cast_fp16 = sub(x = var_3041_to_fp16, y = onehot_47_cast_fp16)[name = tensor("op_3043_cast_fp16")]; + tensor k_cache11_to_fp16_dtype_0 = const()[name = tensor("k_cache11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache11_to_fp16 = cast(dtype = k_cache11_to_fp16_dtype_0, x = k_cache11)[name = tensor("cast_487")]; + tensor var_3044_cast_fp16 = mul(x = k_cache11_to_fp16, y = var_3043_cast_fp16)[name = tensor("op_3044_cast_fp16")]; + tensor var_3049 = const()[name = tensor("op_3049"), val = tensor([1, 1, 16, 64])]; + tensor var_3050_cast_fp16 = reshape(shape = var_3049, x = out_47_cast_fp16)[name = tensor("op_3050_cast_fp16")]; + tensor var_3051_cast_fp16 = mul(x = var_3050_cast_fp16, y = onehot_47_cast_fp16)[name = tensor("op_3051_cast_fp16")]; + tensor new_k_cache_23_cast_fp16 = add(x = var_3044_cast_fp16, y = var_3051_cast_fp16)[name = tensor("new_k_cache_23_cast_fp16")]; + tensor new_k_cache_23_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_23_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache11_to_fp16_dtype_0 = const()[name = tensor("v_cache11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache11_to_fp16 = cast(dtype = v_cache11_to_fp16_dtype_0, x = v_cache11)[name = tensor("cast_485")]; + tensor var_3057_cast_fp16 = mul(x = v_cache11_to_fp16, y = var_3043_cast_fp16)[name = tensor("op_3057_cast_fp16")]; + tensor var_3062 = const()[name = tensor("op_3062"), val = tensor([1, 1, 16, 64])]; + tensor var_3063_cast_fp16 = reshape(shape = var_3062, x = v_23_cast_fp16)[name = tensor("op_3063_cast_fp16")]; + tensor var_3064_cast_fp16 = mul(x = var_3063_cast_fp16, y = onehot_47_cast_fp16)[name = tensor("op_3064_cast_fp16")]; + tensor new_v_cache_23_cast_fp16 = add(x = var_3057_cast_fp16, y = var_3064_cast_fp16)[name = tensor("new_v_cache_23_cast_fp16")]; + tensor new_v_cache_23_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_23_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_11_perm_0 = const()[name = tensor("transpose_11_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_3077 = const()[name = tensor("op_3077"), val = tensor([0, 2, 1, 3])]; + tensor var_3083 = const()[name = tensor("op_3083"), val = tensor([1, 16, 1, 64])]; + tensor q4_23_cast_fp16 = reshape(shape = var_3083, x = out_45_cast_fp16)[name = tensor("q4_23_cast_fp16")]; + tensor var_3088_transpose_x_0 = const()[name = tensor("op_3088_transpose_x_0"), val = tensor(false)]; + tensor var_3088_transpose_y_0 = const()[name = tensor("op_3088_transpose_y_0"), val = tensor(false)]; + tensor transpose_11_cast_fp16 = transpose(perm = transpose_11_perm_0, x = new_k_cache_23_cast_fp16)[name = tensor("transpose_62")]; + tensor var_3088_cast_fp16 = matmul(transpose_x = var_3088_transpose_x_0, transpose_y = var_3088_transpose_y_0, x = q4_23_cast_fp16, y = transpose_11_cast_fp16)[name = tensor("op_3088_cast_fp16")]; + tensor var_3089_to_fp16 = const()[name = tensor("op_3089_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_45_cast_fp16 = mul(x = var_3088_cast_fp16, y = var_3089_to_fp16)[name = tensor("scores_45_cast_fp16")]; + tensor var_3093 = const()[name = tensor("op_3093"), val = tensor([1, 1])]; + tensor var_3094_cast_fp16 = reshape(shape = var_3093, x = position11_to_fp16)[name = tensor("op_3094_cast_fp16")]; + tensor var_3095_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_3094_cast_fp16)[name = tensor("op_3095_cast_fp16")]; + tensor mask_23_to_fp16_dtype_0 = const()[name = tensor("mask_23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_3102_to_fp16 = const()[name = tensor("op_3102_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3095_cast_fp16_to_fp16 = cast(dtype = mask_23_to_fp16_dtype_0, x = var_3095_cast_fp16)[name = tensor("cast_483")]; + tensor var_3103_cast_fp16 = sub(x = var_3095_cast_fp16_to_fp16, y = var_3102_to_fp16)[name = tensor("op_3103_cast_fp16")]; + tensor var_3108 = const()[name = tensor("op_3108"), val = tensor([1, 1, 1, 512])]; + tensor var_3109_cast_fp16 = reshape(shape = var_3108, x = var_3103_cast_fp16)[name = tensor("op_3109_cast_fp16")]; + tensor var_3110_to_fp16 = const()[name = tensor("op_3110_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_3111_cast_fp16 = mul(x = var_3109_cast_fp16, y = var_3110_to_fp16)[name = tensor("op_3111_cast_fp16")]; + tensor scores_47_cast_fp16 = add(x = scores_45_cast_fp16, y = var_3111_cast_fp16)[name = tensor("scores_47_cast_fp16")]; + tensor var_3114 = const()[name = tensor("op_3114"), val = tensor(-1)]; + tensor weights_23_cast_fp16 = softmax(axis = var_3114, x = scores_47_cast_fp16)[name = tensor("weights_23_cast_fp16")]; + tensor attn_23_transpose_x_0 = const()[name = tensor("attn_23_transpose_x_0"), val = tensor(false)]; + tensor attn_23_transpose_y_0 = const()[name = tensor("attn_23_transpose_y_0"), val = tensor(false)]; + tensor values_23_cast_fp16 = transpose(perm = var_3077, x = new_v_cache_23_cast_fp16)[name = tensor("transpose_61")]; + tensor attn_23_cast_fp16 = matmul(transpose_x = attn_23_transpose_x_0, transpose_y = attn_23_transpose_y_0, x = weights_23_cast_fp16, y = values_23_cast_fp16)[name = tensor("attn_23_cast_fp16")]; + tensor var_3122 = const()[name = tensor("op_3122"), val = tensor([0, 2, 1, 3])]; + tensor var_3127 = const()[name = tensor("op_3127"), val = tensor([1, 1, 1024])]; + tensor var_3123_cast_fp16 = transpose(perm = var_3122, x = attn_23_cast_fp16)[name = tensor("transpose_60")]; + tensor input_139_cast_fp16 = reshape(shape = var_3127, x = var_3123_cast_fp16)[name = tensor("input_139_cast_fp16")]; + tensor attn11_out_proj_weight_to_fp16 = const()[name = tensor("attn11_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(283299008)))]; + tensor linear_46_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn11_out_proj_weight_to_fp16, x = input_139_cast_fp16)[name = tensor("linear_46_cast_fp16")]; + tensor var_3133_to_fp16 = const()[name = tensor("op_3133_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3134_cast_fp16 = add(x = position11_to_fp16, y = var_3133_to_fp16)[name = tensor("op_3134_cast_fp16")]; + tensor var_3134_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3134_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_141_cast_fp16 = add(x = input_135_cast_fp16, y = linear_46_cast_fp16)[name = tensor("input_141_cast_fp16")]; + tensor input_143_axes_0 = const()[name = tensor("input_143_axes_0"), val = tensor([-1])]; + tensor norm11_2_weight_to_fp16 = const()[name = tensor("norm11_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285396224)))]; + tensor norm11_2_bias_to_fp16 = const()[name = tensor("norm11_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285398336)))]; + tensor var_3138_to_fp16 = const()[name = tensor("op_3138_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_143_cast_fp16 = layer_norm(axes = input_143_axes_0, beta = norm11_2_bias_to_fp16, epsilon = var_3138_to_fp16, gamma = norm11_2_weight_to_fp16, x = input_141_cast_fp16)[name = tensor("input_143_cast_fp16")]; + tensor linear11_1_weight_to_fp16 = const()[name = tensor("linear11_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285400448)))]; + tensor linear_47_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear11_1_weight_to_fp16, x = input_143_cast_fp16)[name = tensor("linear_47_cast_fp16")]; + tensor input_145_mode_0 = const()[name = tensor("input_145_mode_0"), val = tensor("EXACT")]; + tensor input_145_cast_fp16 = gelu(mode = input_145_mode_0, x = linear_47_cast_fp16)[name = tensor("input_145_cast_fp16")]; + tensor linear11_2_weight_to_fp16 = const()[name = tensor("linear11_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293789120)))]; + tensor linear_48_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear11_2_weight_to_fp16, x = input_145_cast_fp16)[name = tensor("linear_48_cast_fp16")]; + tensor input_147_cast_fp16 = add(x = input_141_cast_fp16, y = linear_48_cast_fp16)[name = tensor("input_147_cast_fp16")]; + tensor input_149_axes_0 = const()[name = tensor("input_149_axes_0"), val = tensor([-1])]; + tensor norm12_1_weight_to_fp16 = const()[name = tensor("norm12_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302177792)))]; + tensor norm12_1_bias_to_fp16 = const()[name = tensor("norm12_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302179904)))]; + tensor var_3155_to_fp16 = const()[name = tensor("op_3155_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_149_cast_fp16 = layer_norm(axes = input_149_axes_0, beta = norm12_1_bias_to_fp16, epsilon = var_3155_to_fp16, gamma = norm12_1_weight_to_fp16, x = input_147_cast_fp16)[name = tensor("input_149_cast_fp16")]; + tensor attn12_in_proj_weight_to_fp16 = const()[name = tensor("attn12_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302182016)))]; + tensor linear_49_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn12_in_proj_weight_to_fp16, x = input_149_cast_fp16)[name = tensor("linear_49_cast_fp16")]; + tensor var_3168 = const()[name = tensor("op_3168"), val = tensor([1, 3, 16, 64])]; + tensor qkv_25_cast_fp16 = reshape(shape = var_3168, x = linear_49_cast_fp16)[name = tensor("qkv_25_cast_fp16")]; + tensor x_49_begin_0 = const()[name = tensor("x_49_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_49_end_0 = const()[name = tensor("x_49_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_49_end_mask_0 = const()[name = tensor("x_49_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_49_squeeze_mask_0 = const()[name = tensor("x_49_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_49_cast_fp16 = slice_by_index(begin = x_49_begin_0, end = x_49_end_0, end_mask = x_49_end_mask_0, squeeze_mask = x_49_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("x_49_cast_fp16")]; + tensor x_51_begin_0 = const()[name = tensor("x_51_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_51_end_0 = const()[name = tensor("x_51_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_51_end_mask_0 = const()[name = tensor("x_51_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_51_squeeze_mask_0 = const()[name = tensor("x_51_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_51_cast_fp16 = slice_by_index(begin = x_51_begin_0, end = x_51_end_0, end_mask = x_51_end_mask_0, squeeze_mask = x_51_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("x_51_cast_fp16")]; + tensor v_25_begin_0 = const()[name = tensor("v_25_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_25_end_0 = const()[name = tensor("v_25_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_25_end_mask_0 = const()[name = tensor("v_25_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_25_squeeze_mask_0 = const()[name = tensor("v_25_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_25_cast_fp16 = slice_by_index(begin = v_25_begin_0, end = v_25_end_0, end_mask = v_25_end_mask_0, squeeze_mask = v_25_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("v_25_cast_fp16")]; + tensor var_3197 = const()[name = tensor("op_3197"), val = tensor([1, 1, 1])]; + tensor position12_to_fp16_dtype_0 = const()[name = tensor("position12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position12_to_fp16 = cast(dtype = position12_to_fp16_dtype_0, x = position12)[name = tensor("cast_481")]; + tensor var_3198_cast_fp16 = reshape(shape = var_3197, x = position12_to_fp16)[name = tensor("op_3198_cast_fp16")]; + tensor angles_49_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_3198_cast_fp16)[name = tensor("angles_49_cast_fp16")]; + tensor rotr_49_cast_fp16 = cos(x = angles_49_cast_fp16)[name = tensor("rotr_49_cast_fp16")]; + tensor roti_49_cast_fp16 = sin(x = angles_49_cast_fp16)[name = tensor("roti_49_cast_fp16")]; + tensor var_3206 = const()[name = tensor("op_3206"), val = tensor([1, 16, 32, 2])]; + tensor pairs_49_cast_fp16 = reshape(shape = var_3206, x = x_49_cast_fp16)[name = tensor("pairs_49_cast_fp16")]; + tensor xr_49_begin_0 = const()[name = tensor("xr_49_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_49_end_0 = const()[name = tensor("xr_49_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_49_end_mask_0 = const()[name = tensor("xr_49_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_49_squeeze_mask_0 = const()[name = tensor("xr_49_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_49_cast_fp16 = slice_by_index(begin = xr_49_begin_0, end = xr_49_end_0, end_mask = xr_49_end_mask_0, squeeze_mask = xr_49_squeeze_mask_0, x = pairs_49_cast_fp16)[name = tensor("xr_49_cast_fp16")]; + tensor xi_49_begin_0 = const()[name = tensor("xi_49_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_49_end_0 = const()[name = tensor("xi_49_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_49_end_mask_0 = const()[name = tensor("xi_49_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_49_squeeze_mask_0 = const()[name = tensor("xi_49_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_49_cast_fp16 = slice_by_index(begin = xi_49_begin_0, end = xi_49_end_0, end_mask = xi_49_end_mask_0, squeeze_mask = xi_49_squeeze_mask_0, x = pairs_49_cast_fp16)[name = tensor("xi_49_cast_fp16")]; + tensor var_3214_cast_fp16 = mul(x = xr_49_cast_fp16, y = rotr_49_cast_fp16)[name = tensor("op_3214_cast_fp16")]; + tensor var_3215_cast_fp16 = mul(x = xi_49_cast_fp16, y = roti_49_cast_fp16)[name = tensor("op_3215_cast_fp16")]; + tensor xor__49_cast_fp16 = sub(x = var_3214_cast_fp16, y = var_3215_cast_fp16)[name = tensor("xor__49_cast_fp16")]; + tensor var_3218_cast_fp16 = mul(x = xr_49_cast_fp16, y = roti_49_cast_fp16)[name = tensor("op_3218_cast_fp16")]; + tensor var_3219_cast_fp16 = mul(x = xi_49_cast_fp16, y = rotr_49_cast_fp16)[name = tensor("op_3219_cast_fp16")]; + tensor xoi_49_cast_fp16 = add(x = var_3218_cast_fp16, y = var_3219_cast_fp16)[name = tensor("xoi_49_cast_fp16")]; + tensor out_49_axis_0 = const()[name = tensor("out_49_axis_0"), val = tensor(-1)]; + tensor out_49_cast_fp16 = stack(axis = out_49_axis_0, values = (xor__49_cast_fp16, xoi_49_cast_fp16))[name = tensor("out_49_cast_fp16")]; + tensor var_3242 = const()[name = tensor("op_3242"), val = tensor([1, 16, 32, 2])]; + tensor pairs_51_cast_fp16 = reshape(shape = var_3242, x = x_51_cast_fp16)[name = tensor("pairs_51_cast_fp16")]; + tensor xr_51_begin_0 = const()[name = tensor("xr_51_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_51_end_0 = const()[name = tensor("xr_51_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_51_end_mask_0 = const()[name = tensor("xr_51_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_51_squeeze_mask_0 = const()[name = tensor("xr_51_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_51_cast_fp16 = slice_by_index(begin = xr_51_begin_0, end = xr_51_end_0, end_mask = xr_51_end_mask_0, squeeze_mask = xr_51_squeeze_mask_0, x = pairs_51_cast_fp16)[name = tensor("xr_51_cast_fp16")]; + tensor xi_51_begin_0 = const()[name = tensor("xi_51_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_51_end_0 = const()[name = tensor("xi_51_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_51_end_mask_0 = const()[name = tensor("xi_51_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_51_squeeze_mask_0 = const()[name = tensor("xi_51_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_51_cast_fp16 = slice_by_index(begin = xi_51_begin_0, end = xi_51_end_0, end_mask = xi_51_end_mask_0, squeeze_mask = xi_51_squeeze_mask_0, x = pairs_51_cast_fp16)[name = tensor("xi_51_cast_fp16")]; + tensor var_3250_cast_fp16 = mul(x = xr_51_cast_fp16, y = rotr_49_cast_fp16)[name = tensor("op_3250_cast_fp16")]; + tensor var_3251_cast_fp16 = mul(x = xi_51_cast_fp16, y = roti_49_cast_fp16)[name = tensor("op_3251_cast_fp16")]; + tensor xor__51_cast_fp16 = sub(x = var_3250_cast_fp16, y = var_3251_cast_fp16)[name = tensor("xor__51_cast_fp16")]; + tensor var_3254_cast_fp16 = mul(x = xr_51_cast_fp16, y = roti_49_cast_fp16)[name = tensor("op_3254_cast_fp16")]; + tensor var_3255_cast_fp16 = mul(x = xi_51_cast_fp16, y = rotr_49_cast_fp16)[name = tensor("op_3255_cast_fp16")]; + tensor xoi_51_cast_fp16 = add(x = var_3254_cast_fp16, y = var_3255_cast_fp16)[name = tensor("xoi_51_cast_fp16")]; + tensor out_51_axis_0 = const()[name = tensor("out_51_axis_0"), val = tensor(-1)]; + tensor out_51_cast_fp16 = stack(axis = out_51_axis_0, values = (xor__51_cast_fp16, xoi_51_cast_fp16))[name = tensor("out_51_cast_fp16")]; + tensor _inversed_3267_y_0_to_fp16 = const()[name = tensor("_inversed_3267_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3267_cast_fp16 = mul(x = position12_to_fp16, y = _inversed_3267_y_0_to_fp16)[name = tensor("_inversed_3267_cast_fp16")]; + tensor var_3268_cast_fp16 = floor(x = _inversed_3267_cast_fp16)[name = tensor("op_3268_cast_fp16")]; + tensor var_3269_to_fp16 = const()[name = tensor("op_3269_to_fp16"), val = tensor(0x1p+9)]; + tensor var_3270_cast_fp16 = mul(x = var_3268_cast_fp16, y = var_3269_to_fp16)[name = tensor("op_3270_cast_fp16")]; + tensor write_idx_25_cast_fp16 = sub(x = position12_to_fp16, y = var_3270_cast_fp16)[name = tensor("write_idx_25_cast_fp16")]; + tensor var_3273_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_25_cast_fp16)[name = tensor("op_3273_cast_fp16")]; + tensor var_3283 = const()[name = tensor("op_3283"), val = tensor([1, 512, 1, 1])]; + tensor onehot_49_to_fp16_dtype_0 = const()[name = tensor("onehot_49_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_3273_cast_fp16_to_fp16 = cast(dtype = onehot_49_to_fp16_dtype_0, x = var_3273_cast_fp16)[name = tensor("cast_480")]; + tensor onehot_51_cast_fp16 = reshape(shape = var_3283, x = var_3273_cast_fp16_to_fp16)[name = tensor("onehot_51_cast_fp16")]; + tensor var_3285_to_fp16 = const()[name = tensor("op_3285_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3287_cast_fp16 = sub(x = var_3285_to_fp16, y = onehot_51_cast_fp16)[name = tensor("op_3287_cast_fp16")]; + tensor k_cache12_to_fp16_dtype_0 = const()[name = tensor("k_cache12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache12_to_fp16 = cast(dtype = k_cache12_to_fp16_dtype_0, x = k_cache12)[name = tensor("cast_479")]; + tensor var_3288_cast_fp16 = mul(x = k_cache12_to_fp16, y = var_3287_cast_fp16)[name = tensor("op_3288_cast_fp16")]; + tensor var_3293 = const()[name = tensor("op_3293"), val = tensor([1, 1, 16, 64])]; + tensor var_3294_cast_fp16 = reshape(shape = var_3293, x = out_51_cast_fp16)[name = tensor("op_3294_cast_fp16")]; + tensor var_3295_cast_fp16 = mul(x = var_3294_cast_fp16, y = onehot_51_cast_fp16)[name = tensor("op_3295_cast_fp16")]; + tensor new_k_cache_25_cast_fp16 = add(x = var_3288_cast_fp16, y = var_3295_cast_fp16)[name = tensor("new_k_cache_25_cast_fp16")]; + tensor new_k_cache_25_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_25_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache12_to_fp16_dtype_0 = const()[name = tensor("v_cache12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache12_to_fp16 = cast(dtype = v_cache12_to_fp16_dtype_0, x = v_cache12)[name = tensor("cast_477")]; + tensor var_3301_cast_fp16 = mul(x = v_cache12_to_fp16, y = var_3287_cast_fp16)[name = tensor("op_3301_cast_fp16")]; + tensor var_3306 = const()[name = tensor("op_3306"), val = tensor([1, 1, 16, 64])]; + tensor var_3307_cast_fp16 = reshape(shape = var_3306, x = v_25_cast_fp16)[name = tensor("op_3307_cast_fp16")]; + tensor var_3308_cast_fp16 = mul(x = var_3307_cast_fp16, y = onehot_51_cast_fp16)[name = tensor("op_3308_cast_fp16")]; + tensor new_v_cache_25_cast_fp16 = add(x = var_3301_cast_fp16, y = var_3308_cast_fp16)[name = tensor("new_v_cache_25_cast_fp16")]; + tensor new_v_cache_25_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_25_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_12_perm_0 = const()[name = tensor("transpose_12_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_3321 = const()[name = tensor("op_3321"), val = tensor([0, 2, 1, 3])]; + tensor var_3327 = const()[name = tensor("op_3327"), val = tensor([1, 16, 1, 64])]; + tensor q4_25_cast_fp16 = reshape(shape = var_3327, x = out_49_cast_fp16)[name = tensor("q4_25_cast_fp16")]; + tensor var_3332_transpose_x_0 = const()[name = tensor("op_3332_transpose_x_0"), val = tensor(false)]; + tensor var_3332_transpose_y_0 = const()[name = tensor("op_3332_transpose_y_0"), val = tensor(false)]; + tensor transpose_12_cast_fp16 = transpose(perm = transpose_12_perm_0, x = new_k_cache_25_cast_fp16)[name = tensor("transpose_59")]; + tensor var_3332_cast_fp16 = matmul(transpose_x = var_3332_transpose_x_0, transpose_y = var_3332_transpose_y_0, x = q4_25_cast_fp16, y = transpose_12_cast_fp16)[name = tensor("op_3332_cast_fp16")]; + tensor var_3333_to_fp16 = const()[name = tensor("op_3333_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_49_cast_fp16 = mul(x = var_3332_cast_fp16, y = var_3333_to_fp16)[name = tensor("scores_49_cast_fp16")]; + tensor var_3337 = const()[name = tensor("op_3337"), val = tensor([1, 1])]; + tensor var_3338_cast_fp16 = reshape(shape = var_3337, x = position12_to_fp16)[name = tensor("op_3338_cast_fp16")]; + tensor var_3339_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_3338_cast_fp16)[name = tensor("op_3339_cast_fp16")]; + tensor mask_25_to_fp16_dtype_0 = const()[name = tensor("mask_25_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_3346_to_fp16 = const()[name = tensor("op_3346_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3339_cast_fp16_to_fp16 = cast(dtype = mask_25_to_fp16_dtype_0, x = var_3339_cast_fp16)[name = tensor("cast_475")]; + tensor var_3347_cast_fp16 = sub(x = var_3339_cast_fp16_to_fp16, y = var_3346_to_fp16)[name = tensor("op_3347_cast_fp16")]; + tensor var_3352 = const()[name = tensor("op_3352"), val = tensor([1, 1, 1, 512])]; + tensor var_3353_cast_fp16 = reshape(shape = var_3352, x = var_3347_cast_fp16)[name = tensor("op_3353_cast_fp16")]; + tensor var_3354_to_fp16 = const()[name = tensor("op_3354_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_3355_cast_fp16 = mul(x = var_3353_cast_fp16, y = var_3354_to_fp16)[name = tensor("op_3355_cast_fp16")]; + tensor scores_51_cast_fp16 = add(x = scores_49_cast_fp16, y = var_3355_cast_fp16)[name = tensor("scores_51_cast_fp16")]; + tensor var_3358 = const()[name = tensor("op_3358"), val = tensor(-1)]; + tensor weights_25_cast_fp16 = softmax(axis = var_3358, x = scores_51_cast_fp16)[name = tensor("weights_25_cast_fp16")]; + tensor attn_25_transpose_x_0 = const()[name = tensor("attn_25_transpose_x_0"), val = tensor(false)]; + tensor attn_25_transpose_y_0 = const()[name = tensor("attn_25_transpose_y_0"), val = tensor(false)]; + tensor values_25_cast_fp16 = transpose(perm = var_3321, x = new_v_cache_25_cast_fp16)[name = tensor("transpose_58")]; + tensor attn_25_cast_fp16 = matmul(transpose_x = attn_25_transpose_x_0, transpose_y = attn_25_transpose_y_0, x = weights_25_cast_fp16, y = values_25_cast_fp16)[name = tensor("attn_25_cast_fp16")]; + tensor var_3366 = const()[name = tensor("op_3366"), val = tensor([0, 2, 1, 3])]; + tensor var_3371 = const()[name = tensor("op_3371"), val = tensor([1, 1, 1024])]; + tensor var_3367_cast_fp16 = transpose(perm = var_3366, x = attn_25_cast_fp16)[name = tensor("transpose_57")]; + tensor input_151_cast_fp16 = reshape(shape = var_3371, x = var_3367_cast_fp16)[name = tensor("input_151_cast_fp16")]; + tensor attn12_out_proj_weight_to_fp16 = const()[name = tensor("attn12_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(308473536)))]; + tensor linear_50_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn12_out_proj_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_50_cast_fp16")]; + tensor var_3377_to_fp16 = const()[name = tensor("op_3377_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3378_cast_fp16 = add(x = position12_to_fp16, y = var_3377_to_fp16)[name = tensor("op_3378_cast_fp16")]; + tensor var_3378_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3378_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_153_cast_fp16 = add(x = input_147_cast_fp16, y = linear_50_cast_fp16)[name = tensor("input_153_cast_fp16")]; + tensor input_155_axes_0 = const()[name = tensor("input_155_axes_0"), val = tensor([-1])]; + tensor norm12_2_weight_to_fp16 = const()[name = tensor("norm12_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(310570752)))]; + tensor norm12_2_bias_to_fp16 = const()[name = tensor("norm12_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(310572864)))]; + tensor var_3382_to_fp16 = const()[name = tensor("op_3382_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_155_cast_fp16 = layer_norm(axes = input_155_axes_0, beta = norm12_2_bias_to_fp16, epsilon = var_3382_to_fp16, gamma = norm12_2_weight_to_fp16, x = input_153_cast_fp16)[name = tensor("input_155_cast_fp16")]; + tensor linear12_1_weight_to_fp16 = const()[name = tensor("linear12_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(310574976)))]; + tensor linear_51_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear12_1_weight_to_fp16, x = input_155_cast_fp16)[name = tensor("linear_51_cast_fp16")]; + tensor input_157_mode_0 = const()[name = tensor("input_157_mode_0"), val = tensor("EXACT")]; + tensor input_157_cast_fp16 = gelu(mode = input_157_mode_0, x = linear_51_cast_fp16)[name = tensor("input_157_cast_fp16")]; + tensor linear12_2_weight_to_fp16 = const()[name = tensor("linear12_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(318963648)))]; + tensor linear_52_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear12_2_weight_to_fp16, x = input_157_cast_fp16)[name = tensor("linear_52_cast_fp16")]; + tensor input_159_cast_fp16 = add(x = input_153_cast_fp16, y = linear_52_cast_fp16)[name = tensor("input_159_cast_fp16")]; + tensor input_161_axes_0 = const()[name = tensor("input_161_axes_0"), val = tensor([-1])]; + tensor norm13_1_weight_to_fp16 = const()[name = tensor("norm13_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(327352320)))]; + tensor norm13_1_bias_to_fp16 = const()[name = tensor("norm13_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(327354432)))]; + tensor var_3399_to_fp16 = const()[name = tensor("op_3399_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_161_cast_fp16 = layer_norm(axes = input_161_axes_0, beta = norm13_1_bias_to_fp16, epsilon = var_3399_to_fp16, gamma = norm13_1_weight_to_fp16, x = input_159_cast_fp16)[name = tensor("input_161_cast_fp16")]; + tensor attn13_in_proj_weight_to_fp16 = const()[name = tensor("attn13_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(327356544)))]; + tensor linear_53_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn13_in_proj_weight_to_fp16, x = input_161_cast_fp16)[name = tensor("linear_53_cast_fp16")]; + tensor var_3412 = const()[name = tensor("op_3412"), val = tensor([1, 3, 16, 64])]; + tensor qkv_27_cast_fp16 = reshape(shape = var_3412, x = linear_53_cast_fp16)[name = tensor("qkv_27_cast_fp16")]; + tensor x_53_begin_0 = const()[name = tensor("x_53_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_53_end_0 = const()[name = tensor("x_53_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_53_end_mask_0 = const()[name = tensor("x_53_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_53_squeeze_mask_0 = const()[name = tensor("x_53_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_53_cast_fp16 = slice_by_index(begin = x_53_begin_0, end = x_53_end_0, end_mask = x_53_end_mask_0, squeeze_mask = x_53_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("x_53_cast_fp16")]; + tensor x_55_begin_0 = const()[name = tensor("x_55_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_55_end_0 = const()[name = tensor("x_55_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_55_end_mask_0 = const()[name = tensor("x_55_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_55_squeeze_mask_0 = const()[name = tensor("x_55_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_55_cast_fp16 = slice_by_index(begin = x_55_begin_0, end = x_55_end_0, end_mask = x_55_end_mask_0, squeeze_mask = x_55_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("x_55_cast_fp16")]; + tensor v_27_begin_0 = const()[name = tensor("v_27_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_27_end_0 = const()[name = tensor("v_27_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_27_end_mask_0 = const()[name = tensor("v_27_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_27_squeeze_mask_0 = const()[name = tensor("v_27_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_27_cast_fp16 = slice_by_index(begin = v_27_begin_0, end = v_27_end_0, end_mask = v_27_end_mask_0, squeeze_mask = v_27_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("v_27_cast_fp16")]; + tensor var_3441 = const()[name = tensor("op_3441"), val = tensor([1, 1, 1])]; + tensor position13_to_fp16_dtype_0 = const()[name = tensor("position13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position13_to_fp16 = cast(dtype = position13_to_fp16_dtype_0, x = position13)[name = tensor("cast_473")]; + tensor var_3442_cast_fp16 = reshape(shape = var_3441, x = position13_to_fp16)[name = tensor("op_3442_cast_fp16")]; + tensor angles_53_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_3442_cast_fp16)[name = tensor("angles_53_cast_fp16")]; + tensor rotr_53_cast_fp16 = cos(x = angles_53_cast_fp16)[name = tensor("rotr_53_cast_fp16")]; + tensor roti_53_cast_fp16 = sin(x = angles_53_cast_fp16)[name = tensor("roti_53_cast_fp16")]; + tensor var_3450 = const()[name = tensor("op_3450"), val = tensor([1, 16, 32, 2])]; + tensor pairs_53_cast_fp16 = reshape(shape = var_3450, x = x_53_cast_fp16)[name = tensor("pairs_53_cast_fp16")]; + tensor xr_53_begin_0 = const()[name = tensor("xr_53_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_53_end_0 = const()[name = tensor("xr_53_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_53_end_mask_0 = const()[name = tensor("xr_53_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_53_squeeze_mask_0 = const()[name = tensor("xr_53_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_53_cast_fp16 = slice_by_index(begin = xr_53_begin_0, end = xr_53_end_0, end_mask = xr_53_end_mask_0, squeeze_mask = xr_53_squeeze_mask_0, x = pairs_53_cast_fp16)[name = tensor("xr_53_cast_fp16")]; + tensor xi_53_begin_0 = const()[name = tensor("xi_53_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_53_end_0 = const()[name = tensor("xi_53_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_53_end_mask_0 = const()[name = tensor("xi_53_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_53_squeeze_mask_0 = const()[name = tensor("xi_53_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_53_cast_fp16 = slice_by_index(begin = xi_53_begin_0, end = xi_53_end_0, end_mask = xi_53_end_mask_0, squeeze_mask = xi_53_squeeze_mask_0, x = pairs_53_cast_fp16)[name = tensor("xi_53_cast_fp16")]; + tensor var_3458_cast_fp16 = mul(x = xr_53_cast_fp16, y = rotr_53_cast_fp16)[name = tensor("op_3458_cast_fp16")]; + tensor var_3459_cast_fp16 = mul(x = xi_53_cast_fp16, y = roti_53_cast_fp16)[name = tensor("op_3459_cast_fp16")]; + tensor xor__53_cast_fp16 = sub(x = var_3458_cast_fp16, y = var_3459_cast_fp16)[name = tensor("xor__53_cast_fp16")]; + tensor var_3462_cast_fp16 = mul(x = xr_53_cast_fp16, y = roti_53_cast_fp16)[name = tensor("op_3462_cast_fp16")]; + tensor var_3463_cast_fp16 = mul(x = xi_53_cast_fp16, y = rotr_53_cast_fp16)[name = tensor("op_3463_cast_fp16")]; + tensor xoi_53_cast_fp16 = add(x = var_3462_cast_fp16, y = var_3463_cast_fp16)[name = tensor("xoi_53_cast_fp16")]; + tensor out_53_axis_0 = const()[name = tensor("out_53_axis_0"), val = tensor(-1)]; + tensor out_53_cast_fp16 = stack(axis = out_53_axis_0, values = (xor__53_cast_fp16, xoi_53_cast_fp16))[name = tensor("out_53_cast_fp16")]; + tensor var_3486 = const()[name = tensor("op_3486"), val = tensor([1, 16, 32, 2])]; + tensor pairs_55_cast_fp16 = reshape(shape = var_3486, x = x_55_cast_fp16)[name = tensor("pairs_55_cast_fp16")]; + tensor xr_55_begin_0 = const()[name = tensor("xr_55_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_55_end_0 = const()[name = tensor("xr_55_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_55_end_mask_0 = const()[name = tensor("xr_55_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_55_squeeze_mask_0 = const()[name = tensor("xr_55_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_55_cast_fp16 = slice_by_index(begin = xr_55_begin_0, end = xr_55_end_0, end_mask = xr_55_end_mask_0, squeeze_mask = xr_55_squeeze_mask_0, x = pairs_55_cast_fp16)[name = tensor("xr_55_cast_fp16")]; + tensor xi_55_begin_0 = const()[name = tensor("xi_55_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_55_end_0 = const()[name = tensor("xi_55_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_55_end_mask_0 = const()[name = tensor("xi_55_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_55_squeeze_mask_0 = const()[name = tensor("xi_55_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_55_cast_fp16 = slice_by_index(begin = xi_55_begin_0, end = xi_55_end_0, end_mask = xi_55_end_mask_0, squeeze_mask = xi_55_squeeze_mask_0, x = pairs_55_cast_fp16)[name = tensor("xi_55_cast_fp16")]; + tensor var_3494_cast_fp16 = mul(x = xr_55_cast_fp16, y = rotr_53_cast_fp16)[name = tensor("op_3494_cast_fp16")]; + tensor var_3495_cast_fp16 = mul(x = xi_55_cast_fp16, y = roti_53_cast_fp16)[name = tensor("op_3495_cast_fp16")]; + tensor xor__55_cast_fp16 = sub(x = var_3494_cast_fp16, y = var_3495_cast_fp16)[name = tensor("xor__55_cast_fp16")]; + tensor var_3498_cast_fp16 = mul(x = xr_55_cast_fp16, y = roti_53_cast_fp16)[name = tensor("op_3498_cast_fp16")]; + tensor var_3499_cast_fp16 = mul(x = xi_55_cast_fp16, y = rotr_53_cast_fp16)[name = tensor("op_3499_cast_fp16")]; + tensor xoi_55_cast_fp16 = add(x = var_3498_cast_fp16, y = var_3499_cast_fp16)[name = tensor("xoi_55_cast_fp16")]; + tensor out_55_axis_0 = const()[name = tensor("out_55_axis_0"), val = tensor(-1)]; + tensor out_55_cast_fp16 = stack(axis = out_55_axis_0, values = (xor__55_cast_fp16, xoi_55_cast_fp16))[name = tensor("out_55_cast_fp16")]; + tensor _inversed_3511_y_0_to_fp16 = const()[name = tensor("_inversed_3511_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3511_cast_fp16 = mul(x = position13_to_fp16, y = _inversed_3511_y_0_to_fp16)[name = tensor("_inversed_3511_cast_fp16")]; + tensor var_3512_cast_fp16 = floor(x = _inversed_3511_cast_fp16)[name = tensor("op_3512_cast_fp16")]; + tensor var_3513_to_fp16 = const()[name = tensor("op_3513_to_fp16"), val = tensor(0x1p+9)]; + tensor var_3514_cast_fp16 = mul(x = var_3512_cast_fp16, y = var_3513_to_fp16)[name = tensor("op_3514_cast_fp16")]; + tensor write_idx_27_cast_fp16 = sub(x = position13_to_fp16, y = var_3514_cast_fp16)[name = tensor("write_idx_27_cast_fp16")]; + tensor var_3517_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_27_cast_fp16)[name = tensor("op_3517_cast_fp16")]; + tensor var_3527 = const()[name = tensor("op_3527"), val = tensor([1, 512, 1, 1])]; + tensor onehot_53_to_fp16_dtype_0 = const()[name = tensor("onehot_53_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_3517_cast_fp16_to_fp16 = cast(dtype = onehot_53_to_fp16_dtype_0, x = var_3517_cast_fp16)[name = tensor("cast_472")]; + tensor onehot_55_cast_fp16 = reshape(shape = var_3527, x = var_3517_cast_fp16_to_fp16)[name = tensor("onehot_55_cast_fp16")]; + tensor var_3529_to_fp16 = const()[name = tensor("op_3529_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3531_cast_fp16 = sub(x = var_3529_to_fp16, y = onehot_55_cast_fp16)[name = tensor("op_3531_cast_fp16")]; + tensor k_cache13_to_fp16_dtype_0 = const()[name = tensor("k_cache13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache13_to_fp16 = cast(dtype = k_cache13_to_fp16_dtype_0, x = k_cache13)[name = tensor("cast_471")]; + tensor var_3532_cast_fp16 = mul(x = k_cache13_to_fp16, y = var_3531_cast_fp16)[name = tensor("op_3532_cast_fp16")]; + tensor var_3537 = const()[name = tensor("op_3537"), val = tensor([1, 1, 16, 64])]; + tensor var_3538_cast_fp16 = reshape(shape = var_3537, x = out_55_cast_fp16)[name = tensor("op_3538_cast_fp16")]; + tensor var_3539_cast_fp16 = mul(x = var_3538_cast_fp16, y = onehot_55_cast_fp16)[name = tensor("op_3539_cast_fp16")]; + tensor new_k_cache_27_cast_fp16 = add(x = var_3532_cast_fp16, y = var_3539_cast_fp16)[name = tensor("new_k_cache_27_cast_fp16")]; + tensor new_k_cache_27_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_27_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache13_to_fp16_dtype_0 = const()[name = tensor("v_cache13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache13_to_fp16 = cast(dtype = v_cache13_to_fp16_dtype_0, x = v_cache13)[name = tensor("cast_469")]; + tensor var_3545_cast_fp16 = mul(x = v_cache13_to_fp16, y = var_3531_cast_fp16)[name = tensor("op_3545_cast_fp16")]; + tensor var_3550 = const()[name = tensor("op_3550"), val = tensor([1, 1, 16, 64])]; + tensor var_3551_cast_fp16 = reshape(shape = var_3550, x = v_27_cast_fp16)[name = tensor("op_3551_cast_fp16")]; + tensor var_3552_cast_fp16 = mul(x = var_3551_cast_fp16, y = onehot_55_cast_fp16)[name = tensor("op_3552_cast_fp16")]; + tensor new_v_cache_27_cast_fp16 = add(x = var_3545_cast_fp16, y = var_3552_cast_fp16)[name = tensor("new_v_cache_27_cast_fp16")]; + tensor new_v_cache_27_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_27_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_13_perm_0 = const()[name = tensor("transpose_13_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_3565 = const()[name = tensor("op_3565"), val = tensor([0, 2, 1, 3])]; + tensor var_3571 = const()[name = tensor("op_3571"), val = tensor([1, 16, 1, 64])]; + tensor q4_27_cast_fp16 = reshape(shape = var_3571, x = out_53_cast_fp16)[name = tensor("q4_27_cast_fp16")]; + tensor var_3576_transpose_x_0 = const()[name = tensor("op_3576_transpose_x_0"), val = tensor(false)]; + tensor var_3576_transpose_y_0 = const()[name = tensor("op_3576_transpose_y_0"), val = tensor(false)]; + tensor transpose_13_cast_fp16 = transpose(perm = transpose_13_perm_0, x = new_k_cache_27_cast_fp16)[name = tensor("transpose_56")]; + tensor var_3576_cast_fp16 = matmul(transpose_x = var_3576_transpose_x_0, transpose_y = var_3576_transpose_y_0, x = q4_27_cast_fp16, y = transpose_13_cast_fp16)[name = tensor("op_3576_cast_fp16")]; + tensor var_3577_to_fp16 = const()[name = tensor("op_3577_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_53_cast_fp16 = mul(x = var_3576_cast_fp16, y = var_3577_to_fp16)[name = tensor("scores_53_cast_fp16")]; + tensor var_3581 = const()[name = tensor("op_3581"), val = tensor([1, 1])]; + tensor var_3582_cast_fp16 = reshape(shape = var_3581, x = position13_to_fp16)[name = tensor("op_3582_cast_fp16")]; + tensor var_3583_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_3582_cast_fp16)[name = tensor("op_3583_cast_fp16")]; + tensor mask_27_to_fp16_dtype_0 = const()[name = tensor("mask_27_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_3590_to_fp16 = const()[name = tensor("op_3590_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3583_cast_fp16_to_fp16 = cast(dtype = mask_27_to_fp16_dtype_0, x = var_3583_cast_fp16)[name = tensor("cast_467")]; + tensor var_3591_cast_fp16 = sub(x = var_3583_cast_fp16_to_fp16, y = var_3590_to_fp16)[name = tensor("op_3591_cast_fp16")]; + tensor var_3596 = const()[name = tensor("op_3596"), val = tensor([1, 1, 1, 512])]; + tensor var_3597_cast_fp16 = reshape(shape = var_3596, x = var_3591_cast_fp16)[name = tensor("op_3597_cast_fp16")]; + tensor var_3598_to_fp16 = const()[name = tensor("op_3598_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_3599_cast_fp16 = mul(x = var_3597_cast_fp16, y = var_3598_to_fp16)[name = tensor("op_3599_cast_fp16")]; + tensor scores_55_cast_fp16 = add(x = scores_53_cast_fp16, y = var_3599_cast_fp16)[name = tensor("scores_55_cast_fp16")]; + tensor var_3602 = const()[name = tensor("op_3602"), val = tensor(-1)]; + tensor weights_27_cast_fp16 = softmax(axis = var_3602, x = scores_55_cast_fp16)[name = tensor("weights_27_cast_fp16")]; + tensor attn_27_transpose_x_0 = const()[name = tensor("attn_27_transpose_x_0"), val = tensor(false)]; + tensor attn_27_transpose_y_0 = const()[name = tensor("attn_27_transpose_y_0"), val = tensor(false)]; + tensor values_27_cast_fp16 = transpose(perm = var_3565, x = new_v_cache_27_cast_fp16)[name = tensor("transpose_55")]; + tensor attn_27_cast_fp16 = matmul(transpose_x = attn_27_transpose_x_0, transpose_y = attn_27_transpose_y_0, x = weights_27_cast_fp16, y = values_27_cast_fp16)[name = tensor("attn_27_cast_fp16")]; + tensor var_3610 = const()[name = tensor("op_3610"), val = tensor([0, 2, 1, 3])]; + tensor var_3615 = const()[name = tensor("op_3615"), val = tensor([1, 1, 1024])]; + tensor var_3611_cast_fp16 = transpose(perm = var_3610, x = attn_27_cast_fp16)[name = tensor("transpose_54")]; + tensor input_163_cast_fp16 = reshape(shape = var_3615, x = var_3611_cast_fp16)[name = tensor("input_163_cast_fp16")]; + tensor attn13_out_proj_weight_to_fp16 = const()[name = tensor("attn13_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(333648064)))]; + tensor linear_54_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn13_out_proj_weight_to_fp16, x = input_163_cast_fp16)[name = tensor("linear_54_cast_fp16")]; + tensor var_3621_to_fp16 = const()[name = tensor("op_3621_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3622_cast_fp16 = add(x = position13_to_fp16, y = var_3621_to_fp16)[name = tensor("op_3622_cast_fp16")]; + tensor var_3622_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3622_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_165_cast_fp16 = add(x = input_159_cast_fp16, y = linear_54_cast_fp16)[name = tensor("input_165_cast_fp16")]; + tensor input_167_axes_0 = const()[name = tensor("input_167_axes_0"), val = tensor([-1])]; + tensor norm13_2_weight_to_fp16 = const()[name = tensor("norm13_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335745280)))]; + tensor norm13_2_bias_to_fp16 = const()[name = tensor("norm13_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335747392)))]; + tensor var_3626_to_fp16 = const()[name = tensor("op_3626_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_167_cast_fp16 = layer_norm(axes = input_167_axes_0, beta = norm13_2_bias_to_fp16, epsilon = var_3626_to_fp16, gamma = norm13_2_weight_to_fp16, x = input_165_cast_fp16)[name = tensor("input_167_cast_fp16")]; + tensor linear13_1_weight_to_fp16 = const()[name = tensor("linear13_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335749504)))]; + tensor linear_55_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear13_1_weight_to_fp16, x = input_167_cast_fp16)[name = tensor("linear_55_cast_fp16")]; + tensor input_169_mode_0 = const()[name = tensor("input_169_mode_0"), val = tensor("EXACT")]; + tensor input_169_cast_fp16 = gelu(mode = input_169_mode_0, x = linear_55_cast_fp16)[name = tensor("input_169_cast_fp16")]; + tensor linear13_2_weight_to_fp16 = const()[name = tensor("linear13_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(344138176)))]; + tensor linear_56_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear13_2_weight_to_fp16, x = input_169_cast_fp16)[name = tensor("linear_56_cast_fp16")]; + tensor input_171_cast_fp16 = add(x = input_165_cast_fp16, y = linear_56_cast_fp16)[name = tensor("input_171_cast_fp16")]; + tensor input_173_axes_0 = const()[name = tensor("input_173_axes_0"), val = tensor([-1])]; + tensor norm14_1_weight_to_fp16 = const()[name = tensor("norm14_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(352526848)))]; + tensor norm14_1_bias_to_fp16 = const()[name = tensor("norm14_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(352528960)))]; + tensor var_3643_to_fp16 = const()[name = tensor("op_3643_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_173_cast_fp16 = layer_norm(axes = input_173_axes_0, beta = norm14_1_bias_to_fp16, epsilon = var_3643_to_fp16, gamma = norm14_1_weight_to_fp16, x = input_171_cast_fp16)[name = tensor("input_173_cast_fp16")]; + tensor attn14_in_proj_weight_to_fp16 = const()[name = tensor("attn14_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(352531072)))]; + tensor linear_57_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn14_in_proj_weight_to_fp16, x = input_173_cast_fp16)[name = tensor("linear_57_cast_fp16")]; + tensor var_3656 = const()[name = tensor("op_3656"), val = tensor([1, 3, 16, 64])]; + tensor qkv_29_cast_fp16 = reshape(shape = var_3656, x = linear_57_cast_fp16)[name = tensor("qkv_29_cast_fp16")]; + tensor x_57_begin_0 = const()[name = tensor("x_57_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_57_end_0 = const()[name = tensor("x_57_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_57_end_mask_0 = const()[name = tensor("x_57_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_57_squeeze_mask_0 = const()[name = tensor("x_57_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_57_cast_fp16 = slice_by_index(begin = x_57_begin_0, end = x_57_end_0, end_mask = x_57_end_mask_0, squeeze_mask = x_57_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("x_57_cast_fp16")]; + tensor x_59_begin_0 = const()[name = tensor("x_59_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_59_end_0 = const()[name = tensor("x_59_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_59_end_mask_0 = const()[name = tensor("x_59_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_59_squeeze_mask_0 = const()[name = tensor("x_59_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_59_cast_fp16 = slice_by_index(begin = x_59_begin_0, end = x_59_end_0, end_mask = x_59_end_mask_0, squeeze_mask = x_59_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("x_59_cast_fp16")]; + tensor v_29_begin_0 = const()[name = tensor("v_29_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_29_end_0 = const()[name = tensor("v_29_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_29_end_mask_0 = const()[name = tensor("v_29_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_29_squeeze_mask_0 = const()[name = tensor("v_29_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_29_cast_fp16 = slice_by_index(begin = v_29_begin_0, end = v_29_end_0, end_mask = v_29_end_mask_0, squeeze_mask = v_29_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("v_29_cast_fp16")]; + tensor var_3685 = const()[name = tensor("op_3685"), val = tensor([1, 1, 1])]; + tensor position14_to_fp16_dtype_0 = const()[name = tensor("position14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position14_to_fp16 = cast(dtype = position14_to_fp16_dtype_0, x = position14)[name = tensor("cast_465")]; + tensor var_3686_cast_fp16 = reshape(shape = var_3685, x = position14_to_fp16)[name = tensor("op_3686_cast_fp16")]; + tensor angles_57_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_3686_cast_fp16)[name = tensor("angles_57_cast_fp16")]; + tensor rotr_57_cast_fp16 = cos(x = angles_57_cast_fp16)[name = tensor("rotr_57_cast_fp16")]; + tensor roti_57_cast_fp16 = sin(x = angles_57_cast_fp16)[name = tensor("roti_57_cast_fp16")]; + tensor var_3694 = const()[name = tensor("op_3694"), val = tensor([1, 16, 32, 2])]; + tensor pairs_57_cast_fp16 = reshape(shape = var_3694, x = x_57_cast_fp16)[name = tensor("pairs_57_cast_fp16")]; + tensor xr_57_begin_0 = const()[name = tensor("xr_57_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_57_end_0 = const()[name = tensor("xr_57_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_57_end_mask_0 = const()[name = tensor("xr_57_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_57_squeeze_mask_0 = const()[name = tensor("xr_57_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_57_cast_fp16 = slice_by_index(begin = xr_57_begin_0, end = xr_57_end_0, end_mask = xr_57_end_mask_0, squeeze_mask = xr_57_squeeze_mask_0, x = pairs_57_cast_fp16)[name = tensor("xr_57_cast_fp16")]; + tensor xi_57_begin_0 = const()[name = tensor("xi_57_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_57_end_0 = const()[name = tensor("xi_57_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_57_end_mask_0 = const()[name = tensor("xi_57_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_57_squeeze_mask_0 = const()[name = tensor("xi_57_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_57_cast_fp16 = slice_by_index(begin = xi_57_begin_0, end = xi_57_end_0, end_mask = xi_57_end_mask_0, squeeze_mask = xi_57_squeeze_mask_0, x = pairs_57_cast_fp16)[name = tensor("xi_57_cast_fp16")]; + tensor var_3702_cast_fp16 = mul(x = xr_57_cast_fp16, y = rotr_57_cast_fp16)[name = tensor("op_3702_cast_fp16")]; + tensor var_3703_cast_fp16 = mul(x = xi_57_cast_fp16, y = roti_57_cast_fp16)[name = tensor("op_3703_cast_fp16")]; + tensor xor__57_cast_fp16 = sub(x = var_3702_cast_fp16, y = var_3703_cast_fp16)[name = tensor("xor__57_cast_fp16")]; + tensor var_3706_cast_fp16 = mul(x = xr_57_cast_fp16, y = roti_57_cast_fp16)[name = tensor("op_3706_cast_fp16")]; + tensor var_3707_cast_fp16 = mul(x = xi_57_cast_fp16, y = rotr_57_cast_fp16)[name = tensor("op_3707_cast_fp16")]; + tensor xoi_57_cast_fp16 = add(x = var_3706_cast_fp16, y = var_3707_cast_fp16)[name = tensor("xoi_57_cast_fp16")]; + tensor out_57_axis_0 = const()[name = tensor("out_57_axis_0"), val = tensor(-1)]; + tensor out_57_cast_fp16 = stack(axis = out_57_axis_0, values = (xor__57_cast_fp16, xoi_57_cast_fp16))[name = tensor("out_57_cast_fp16")]; + tensor var_3730 = const()[name = tensor("op_3730"), val = tensor([1, 16, 32, 2])]; + tensor pairs_59_cast_fp16 = reshape(shape = var_3730, x = x_59_cast_fp16)[name = tensor("pairs_59_cast_fp16")]; + tensor xr_59_begin_0 = const()[name = tensor("xr_59_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_59_end_0 = const()[name = tensor("xr_59_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_59_end_mask_0 = const()[name = tensor("xr_59_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_59_squeeze_mask_0 = const()[name = tensor("xr_59_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_59_cast_fp16 = slice_by_index(begin = xr_59_begin_0, end = xr_59_end_0, end_mask = xr_59_end_mask_0, squeeze_mask = xr_59_squeeze_mask_0, x = pairs_59_cast_fp16)[name = tensor("xr_59_cast_fp16")]; + tensor xi_59_begin_0 = const()[name = tensor("xi_59_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_59_end_0 = const()[name = tensor("xi_59_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_59_end_mask_0 = const()[name = tensor("xi_59_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_59_squeeze_mask_0 = const()[name = tensor("xi_59_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_59_cast_fp16 = slice_by_index(begin = xi_59_begin_0, end = xi_59_end_0, end_mask = xi_59_end_mask_0, squeeze_mask = xi_59_squeeze_mask_0, x = pairs_59_cast_fp16)[name = tensor("xi_59_cast_fp16")]; + tensor var_3738_cast_fp16 = mul(x = xr_59_cast_fp16, y = rotr_57_cast_fp16)[name = tensor("op_3738_cast_fp16")]; + tensor var_3739_cast_fp16 = mul(x = xi_59_cast_fp16, y = roti_57_cast_fp16)[name = tensor("op_3739_cast_fp16")]; + tensor xor__59_cast_fp16 = sub(x = var_3738_cast_fp16, y = var_3739_cast_fp16)[name = tensor("xor__59_cast_fp16")]; + tensor var_3742_cast_fp16 = mul(x = xr_59_cast_fp16, y = roti_57_cast_fp16)[name = tensor("op_3742_cast_fp16")]; + tensor var_3743_cast_fp16 = mul(x = xi_59_cast_fp16, y = rotr_57_cast_fp16)[name = tensor("op_3743_cast_fp16")]; + tensor xoi_59_cast_fp16 = add(x = var_3742_cast_fp16, y = var_3743_cast_fp16)[name = tensor("xoi_59_cast_fp16")]; + tensor out_59_axis_0 = const()[name = tensor("out_59_axis_0"), val = tensor(-1)]; + tensor out_59_cast_fp16 = stack(axis = out_59_axis_0, values = (xor__59_cast_fp16, xoi_59_cast_fp16))[name = tensor("out_59_cast_fp16")]; + tensor _inversed_3755_y_0_to_fp16 = const()[name = tensor("_inversed_3755_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3755_cast_fp16 = mul(x = position14_to_fp16, y = _inversed_3755_y_0_to_fp16)[name = tensor("_inversed_3755_cast_fp16")]; + tensor var_3756_cast_fp16 = floor(x = _inversed_3755_cast_fp16)[name = tensor("op_3756_cast_fp16")]; + tensor var_3757_to_fp16 = const()[name = tensor("op_3757_to_fp16"), val = tensor(0x1p+9)]; + tensor var_3758_cast_fp16 = mul(x = var_3756_cast_fp16, y = var_3757_to_fp16)[name = tensor("op_3758_cast_fp16")]; + tensor write_idx_29_cast_fp16 = sub(x = position14_to_fp16, y = var_3758_cast_fp16)[name = tensor("write_idx_29_cast_fp16")]; + tensor var_3761_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_29_cast_fp16)[name = tensor("op_3761_cast_fp16")]; + tensor var_3771 = const()[name = tensor("op_3771"), val = tensor([1, 512, 1, 1])]; + tensor onehot_57_to_fp16_dtype_0 = const()[name = tensor("onehot_57_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_3761_cast_fp16_to_fp16 = cast(dtype = onehot_57_to_fp16_dtype_0, x = var_3761_cast_fp16)[name = tensor("cast_464")]; + tensor onehot_59_cast_fp16 = reshape(shape = var_3771, x = var_3761_cast_fp16_to_fp16)[name = tensor("onehot_59_cast_fp16")]; + tensor var_3773_to_fp16 = const()[name = tensor("op_3773_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3775_cast_fp16 = sub(x = var_3773_to_fp16, y = onehot_59_cast_fp16)[name = tensor("op_3775_cast_fp16")]; + tensor k_cache14_to_fp16_dtype_0 = const()[name = tensor("k_cache14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache14_to_fp16 = cast(dtype = k_cache14_to_fp16_dtype_0, x = k_cache14)[name = tensor("cast_463")]; + tensor var_3776_cast_fp16 = mul(x = k_cache14_to_fp16, y = var_3775_cast_fp16)[name = tensor("op_3776_cast_fp16")]; + tensor var_3781 = const()[name = tensor("op_3781"), val = tensor([1, 1, 16, 64])]; + tensor var_3782_cast_fp16 = reshape(shape = var_3781, x = out_59_cast_fp16)[name = tensor("op_3782_cast_fp16")]; + tensor var_3783_cast_fp16 = mul(x = var_3782_cast_fp16, y = onehot_59_cast_fp16)[name = tensor("op_3783_cast_fp16")]; + tensor new_k_cache_29_cast_fp16 = add(x = var_3776_cast_fp16, y = var_3783_cast_fp16)[name = tensor("new_k_cache_29_cast_fp16")]; + tensor new_k_cache_29_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_29_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache14_to_fp16_dtype_0 = const()[name = tensor("v_cache14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache14_to_fp16 = cast(dtype = v_cache14_to_fp16_dtype_0, x = v_cache14)[name = tensor("cast_461")]; + tensor var_3789_cast_fp16 = mul(x = v_cache14_to_fp16, y = var_3775_cast_fp16)[name = tensor("op_3789_cast_fp16")]; + tensor var_3794 = const()[name = tensor("op_3794"), val = tensor([1, 1, 16, 64])]; + tensor var_3795_cast_fp16 = reshape(shape = var_3794, x = v_29_cast_fp16)[name = tensor("op_3795_cast_fp16")]; + tensor var_3796_cast_fp16 = mul(x = var_3795_cast_fp16, y = onehot_59_cast_fp16)[name = tensor("op_3796_cast_fp16")]; + tensor new_v_cache_29_cast_fp16 = add(x = var_3789_cast_fp16, y = var_3796_cast_fp16)[name = tensor("new_v_cache_29_cast_fp16")]; + tensor new_v_cache_29_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_29_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_14_perm_0 = const()[name = tensor("transpose_14_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_3809 = const()[name = tensor("op_3809"), val = tensor([0, 2, 1, 3])]; + tensor var_3815 = const()[name = tensor("op_3815"), val = tensor([1, 16, 1, 64])]; + tensor q4_29_cast_fp16 = reshape(shape = var_3815, x = out_57_cast_fp16)[name = tensor("q4_29_cast_fp16")]; + tensor var_3820_transpose_x_0 = const()[name = tensor("op_3820_transpose_x_0"), val = tensor(false)]; + tensor var_3820_transpose_y_0 = const()[name = tensor("op_3820_transpose_y_0"), val = tensor(false)]; + tensor transpose_14_cast_fp16 = transpose(perm = transpose_14_perm_0, x = new_k_cache_29_cast_fp16)[name = tensor("transpose_53")]; + tensor var_3820_cast_fp16 = matmul(transpose_x = var_3820_transpose_x_0, transpose_y = var_3820_transpose_y_0, x = q4_29_cast_fp16, y = transpose_14_cast_fp16)[name = tensor("op_3820_cast_fp16")]; + tensor var_3821_to_fp16 = const()[name = tensor("op_3821_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_57_cast_fp16 = mul(x = var_3820_cast_fp16, y = var_3821_to_fp16)[name = tensor("scores_57_cast_fp16")]; + tensor var_3825 = const()[name = tensor("op_3825"), val = tensor([1, 1])]; + tensor var_3826_cast_fp16 = reshape(shape = var_3825, x = position14_to_fp16)[name = tensor("op_3826_cast_fp16")]; + tensor var_3827_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_3826_cast_fp16)[name = tensor("op_3827_cast_fp16")]; + tensor mask_29_to_fp16_dtype_0 = const()[name = tensor("mask_29_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_3834_to_fp16 = const()[name = tensor("op_3834_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3827_cast_fp16_to_fp16 = cast(dtype = mask_29_to_fp16_dtype_0, x = var_3827_cast_fp16)[name = tensor("cast_459")]; + tensor var_3835_cast_fp16 = sub(x = var_3827_cast_fp16_to_fp16, y = var_3834_to_fp16)[name = tensor("op_3835_cast_fp16")]; + tensor var_3840 = const()[name = tensor("op_3840"), val = tensor([1, 1, 1, 512])]; + tensor var_3841_cast_fp16 = reshape(shape = var_3840, x = var_3835_cast_fp16)[name = tensor("op_3841_cast_fp16")]; + tensor var_3842_to_fp16 = const()[name = tensor("op_3842_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_3843_cast_fp16 = mul(x = var_3841_cast_fp16, y = var_3842_to_fp16)[name = tensor("op_3843_cast_fp16")]; + tensor scores_59_cast_fp16 = add(x = scores_57_cast_fp16, y = var_3843_cast_fp16)[name = tensor("scores_59_cast_fp16")]; + tensor var_3846 = const()[name = tensor("op_3846"), val = tensor(-1)]; + tensor weights_29_cast_fp16 = softmax(axis = var_3846, x = scores_59_cast_fp16)[name = tensor("weights_29_cast_fp16")]; + tensor attn_29_transpose_x_0 = const()[name = tensor("attn_29_transpose_x_0"), val = tensor(false)]; + tensor attn_29_transpose_y_0 = const()[name = tensor("attn_29_transpose_y_0"), val = tensor(false)]; + tensor values_29_cast_fp16 = transpose(perm = var_3809, x = new_v_cache_29_cast_fp16)[name = tensor("transpose_52")]; + tensor attn_29_cast_fp16 = matmul(transpose_x = attn_29_transpose_x_0, transpose_y = attn_29_transpose_y_0, x = weights_29_cast_fp16, y = values_29_cast_fp16)[name = tensor("attn_29_cast_fp16")]; + tensor var_3854 = const()[name = tensor("op_3854"), val = tensor([0, 2, 1, 3])]; + tensor var_3859 = const()[name = tensor("op_3859"), val = tensor([1, 1, 1024])]; + tensor var_3855_cast_fp16 = transpose(perm = var_3854, x = attn_29_cast_fp16)[name = tensor("transpose_51")]; + tensor input_175_cast_fp16 = reshape(shape = var_3859, x = var_3855_cast_fp16)[name = tensor("input_175_cast_fp16")]; + tensor attn14_out_proj_weight_to_fp16 = const()[name = tensor("attn14_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(358822592)))]; + tensor linear_58_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn14_out_proj_weight_to_fp16, x = input_175_cast_fp16)[name = tensor("linear_58_cast_fp16")]; + tensor var_3865_to_fp16 = const()[name = tensor("op_3865_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3866_cast_fp16 = add(x = position14_to_fp16, y = var_3865_to_fp16)[name = tensor("op_3866_cast_fp16")]; + tensor var_3866_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3866_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_177_cast_fp16 = add(x = input_171_cast_fp16, y = linear_58_cast_fp16)[name = tensor("input_177_cast_fp16")]; + tensor input_179_axes_0 = const()[name = tensor("input_179_axes_0"), val = tensor([-1])]; + tensor norm14_2_weight_to_fp16 = const()[name = tensor("norm14_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(360919808)))]; + tensor norm14_2_bias_to_fp16 = const()[name = tensor("norm14_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(360921920)))]; + tensor var_3870_to_fp16 = const()[name = tensor("op_3870_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_179_cast_fp16 = layer_norm(axes = input_179_axes_0, beta = norm14_2_bias_to_fp16, epsilon = var_3870_to_fp16, gamma = norm14_2_weight_to_fp16, x = input_177_cast_fp16)[name = tensor("input_179_cast_fp16")]; + tensor linear14_1_weight_to_fp16 = const()[name = tensor("linear14_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(360924032)))]; + tensor linear_59_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear14_1_weight_to_fp16, x = input_179_cast_fp16)[name = tensor("linear_59_cast_fp16")]; + tensor input_181_mode_0 = const()[name = tensor("input_181_mode_0"), val = tensor("EXACT")]; + tensor input_181_cast_fp16 = gelu(mode = input_181_mode_0, x = linear_59_cast_fp16)[name = tensor("input_181_cast_fp16")]; + tensor linear14_2_weight_to_fp16 = const()[name = tensor("linear14_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(369312704)))]; + tensor linear_60_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear14_2_weight_to_fp16, x = input_181_cast_fp16)[name = tensor("linear_60_cast_fp16")]; + tensor input_183_cast_fp16 = add(x = input_177_cast_fp16, y = linear_60_cast_fp16)[name = tensor("input_183_cast_fp16")]; + tensor input_185_axes_0 = const()[name = tensor("input_185_axes_0"), val = tensor([-1])]; + tensor norm15_1_weight_to_fp16 = const()[name = tensor("norm15_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(377701376)))]; + tensor norm15_1_bias_to_fp16 = const()[name = tensor("norm15_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(377703488)))]; + tensor var_3887_to_fp16 = const()[name = tensor("op_3887_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_185_cast_fp16 = layer_norm(axes = input_185_axes_0, beta = norm15_1_bias_to_fp16, epsilon = var_3887_to_fp16, gamma = norm15_1_weight_to_fp16, x = input_183_cast_fp16)[name = tensor("input_185_cast_fp16")]; + tensor attn15_in_proj_weight_to_fp16 = const()[name = tensor("attn15_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(377705600)))]; + tensor linear_61_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn15_in_proj_weight_to_fp16, x = input_185_cast_fp16)[name = tensor("linear_61_cast_fp16")]; + tensor var_3900 = const()[name = tensor("op_3900"), val = tensor([1, 3, 16, 64])]; + tensor qkv_31_cast_fp16 = reshape(shape = var_3900, x = linear_61_cast_fp16)[name = tensor("qkv_31_cast_fp16")]; + tensor x_61_begin_0 = const()[name = tensor("x_61_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_61_end_0 = const()[name = tensor("x_61_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_61_end_mask_0 = const()[name = tensor("x_61_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_61_squeeze_mask_0 = const()[name = tensor("x_61_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_61_cast_fp16 = slice_by_index(begin = x_61_begin_0, end = x_61_end_0, end_mask = x_61_end_mask_0, squeeze_mask = x_61_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("x_61_cast_fp16")]; + tensor x_63_begin_0 = const()[name = tensor("x_63_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_63_end_0 = const()[name = tensor("x_63_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_63_end_mask_0 = const()[name = tensor("x_63_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_63_squeeze_mask_0 = const()[name = tensor("x_63_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_63_cast_fp16 = slice_by_index(begin = x_63_begin_0, end = x_63_end_0, end_mask = x_63_end_mask_0, squeeze_mask = x_63_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("x_63_cast_fp16")]; + tensor v_31_begin_0 = const()[name = tensor("v_31_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_31_end_0 = const()[name = tensor("v_31_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_31_end_mask_0 = const()[name = tensor("v_31_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_31_squeeze_mask_0 = const()[name = tensor("v_31_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_31_cast_fp16 = slice_by_index(begin = v_31_begin_0, end = v_31_end_0, end_mask = v_31_end_mask_0, squeeze_mask = v_31_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("v_31_cast_fp16")]; + tensor var_3929 = const()[name = tensor("op_3929"), val = tensor([1, 1, 1])]; + tensor position15_to_fp16_dtype_0 = const()[name = tensor("position15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position15_to_fp16 = cast(dtype = position15_to_fp16_dtype_0, x = position15)[name = tensor("cast_457")]; + tensor var_3930_cast_fp16 = reshape(shape = var_3929, x = position15_to_fp16)[name = tensor("op_3930_cast_fp16")]; + tensor angles_61_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_3930_cast_fp16)[name = tensor("angles_61_cast_fp16")]; + tensor rotr_61_cast_fp16 = cos(x = angles_61_cast_fp16)[name = tensor("rotr_61_cast_fp16")]; + tensor roti_61_cast_fp16 = sin(x = angles_61_cast_fp16)[name = tensor("roti_61_cast_fp16")]; + tensor var_3938 = const()[name = tensor("op_3938"), val = tensor([1, 16, 32, 2])]; + tensor pairs_61_cast_fp16 = reshape(shape = var_3938, x = x_61_cast_fp16)[name = tensor("pairs_61_cast_fp16")]; + tensor xr_61_begin_0 = const()[name = tensor("xr_61_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_61_end_0 = const()[name = tensor("xr_61_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_61_end_mask_0 = const()[name = tensor("xr_61_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_61_squeeze_mask_0 = const()[name = tensor("xr_61_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_61_cast_fp16 = slice_by_index(begin = xr_61_begin_0, end = xr_61_end_0, end_mask = xr_61_end_mask_0, squeeze_mask = xr_61_squeeze_mask_0, x = pairs_61_cast_fp16)[name = tensor("xr_61_cast_fp16")]; + tensor xi_61_begin_0 = const()[name = tensor("xi_61_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_61_end_0 = const()[name = tensor("xi_61_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_61_end_mask_0 = const()[name = tensor("xi_61_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_61_squeeze_mask_0 = const()[name = tensor("xi_61_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_61_cast_fp16 = slice_by_index(begin = xi_61_begin_0, end = xi_61_end_0, end_mask = xi_61_end_mask_0, squeeze_mask = xi_61_squeeze_mask_0, x = pairs_61_cast_fp16)[name = tensor("xi_61_cast_fp16")]; + tensor var_3946_cast_fp16 = mul(x = xr_61_cast_fp16, y = rotr_61_cast_fp16)[name = tensor("op_3946_cast_fp16")]; + tensor var_3947_cast_fp16 = mul(x = xi_61_cast_fp16, y = roti_61_cast_fp16)[name = tensor("op_3947_cast_fp16")]; + tensor xor__61_cast_fp16 = sub(x = var_3946_cast_fp16, y = var_3947_cast_fp16)[name = tensor("xor__61_cast_fp16")]; + tensor var_3950_cast_fp16 = mul(x = xr_61_cast_fp16, y = roti_61_cast_fp16)[name = tensor("op_3950_cast_fp16")]; + tensor var_3951_cast_fp16 = mul(x = xi_61_cast_fp16, y = rotr_61_cast_fp16)[name = tensor("op_3951_cast_fp16")]; + tensor xoi_61_cast_fp16 = add(x = var_3950_cast_fp16, y = var_3951_cast_fp16)[name = tensor("xoi_61_cast_fp16")]; + tensor out_61_axis_0 = const()[name = tensor("out_61_axis_0"), val = tensor(-1)]; + tensor out_61_cast_fp16 = stack(axis = out_61_axis_0, values = (xor__61_cast_fp16, xoi_61_cast_fp16))[name = tensor("out_61_cast_fp16")]; + tensor var_3974 = const()[name = tensor("op_3974"), val = tensor([1, 16, 32, 2])]; + tensor pairs_63_cast_fp16 = reshape(shape = var_3974, x = x_63_cast_fp16)[name = tensor("pairs_63_cast_fp16")]; + tensor xr_63_begin_0 = const()[name = tensor("xr_63_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_63_end_0 = const()[name = tensor("xr_63_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_63_end_mask_0 = const()[name = tensor("xr_63_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_63_squeeze_mask_0 = const()[name = tensor("xr_63_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_63_cast_fp16 = slice_by_index(begin = xr_63_begin_0, end = xr_63_end_0, end_mask = xr_63_end_mask_0, squeeze_mask = xr_63_squeeze_mask_0, x = pairs_63_cast_fp16)[name = tensor("xr_63_cast_fp16")]; + tensor xi_63_begin_0 = const()[name = tensor("xi_63_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_63_end_0 = const()[name = tensor("xi_63_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_63_end_mask_0 = const()[name = tensor("xi_63_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_63_squeeze_mask_0 = const()[name = tensor("xi_63_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_63_cast_fp16 = slice_by_index(begin = xi_63_begin_0, end = xi_63_end_0, end_mask = xi_63_end_mask_0, squeeze_mask = xi_63_squeeze_mask_0, x = pairs_63_cast_fp16)[name = tensor("xi_63_cast_fp16")]; + tensor var_3982_cast_fp16 = mul(x = xr_63_cast_fp16, y = rotr_61_cast_fp16)[name = tensor("op_3982_cast_fp16")]; + tensor var_3983_cast_fp16 = mul(x = xi_63_cast_fp16, y = roti_61_cast_fp16)[name = tensor("op_3983_cast_fp16")]; + tensor xor__63_cast_fp16 = sub(x = var_3982_cast_fp16, y = var_3983_cast_fp16)[name = tensor("xor__63_cast_fp16")]; + tensor var_3986_cast_fp16 = mul(x = xr_63_cast_fp16, y = roti_61_cast_fp16)[name = tensor("op_3986_cast_fp16")]; + tensor var_3987_cast_fp16 = mul(x = xi_63_cast_fp16, y = rotr_61_cast_fp16)[name = tensor("op_3987_cast_fp16")]; + tensor xoi_63_cast_fp16 = add(x = var_3986_cast_fp16, y = var_3987_cast_fp16)[name = tensor("xoi_63_cast_fp16")]; + tensor out_63_axis_0 = const()[name = tensor("out_63_axis_0"), val = tensor(-1)]; + tensor out_63_cast_fp16 = stack(axis = out_63_axis_0, values = (xor__63_cast_fp16, xoi_63_cast_fp16))[name = tensor("out_63_cast_fp16")]; + tensor _inversed_3999_y_0_to_fp16 = const()[name = tensor("_inversed_3999_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3999_cast_fp16 = mul(x = position15_to_fp16, y = _inversed_3999_y_0_to_fp16)[name = tensor("_inversed_3999_cast_fp16")]; + tensor var_4000_cast_fp16 = floor(x = _inversed_3999_cast_fp16)[name = tensor("op_4000_cast_fp16")]; + tensor var_4001_to_fp16 = const()[name = tensor("op_4001_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4002_cast_fp16 = mul(x = var_4000_cast_fp16, y = var_4001_to_fp16)[name = tensor("op_4002_cast_fp16")]; + tensor write_idx_31_cast_fp16 = sub(x = position15_to_fp16, y = var_4002_cast_fp16)[name = tensor("write_idx_31_cast_fp16")]; + tensor var_4005_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_31_cast_fp16)[name = tensor("op_4005_cast_fp16")]; + tensor var_4015 = const()[name = tensor("op_4015"), val = tensor([1, 512, 1, 1])]; + tensor onehot_61_to_fp16_dtype_0 = const()[name = tensor("onehot_61_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_4005_cast_fp16_to_fp16 = cast(dtype = onehot_61_to_fp16_dtype_0, x = var_4005_cast_fp16)[name = tensor("cast_456")]; + tensor onehot_63_cast_fp16 = reshape(shape = var_4015, x = var_4005_cast_fp16_to_fp16)[name = tensor("onehot_63_cast_fp16")]; + tensor var_4017_to_fp16 = const()[name = tensor("op_4017_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4019_cast_fp16 = sub(x = var_4017_to_fp16, y = onehot_63_cast_fp16)[name = tensor("op_4019_cast_fp16")]; + tensor k_cache15_to_fp16_dtype_0 = const()[name = tensor("k_cache15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache15_to_fp16 = cast(dtype = k_cache15_to_fp16_dtype_0, x = k_cache15)[name = tensor("cast_455")]; + tensor var_4020_cast_fp16 = mul(x = k_cache15_to_fp16, y = var_4019_cast_fp16)[name = tensor("op_4020_cast_fp16")]; + tensor var_4025 = const()[name = tensor("op_4025"), val = tensor([1, 1, 16, 64])]; + tensor var_4026_cast_fp16 = reshape(shape = var_4025, x = out_63_cast_fp16)[name = tensor("op_4026_cast_fp16")]; + tensor var_4027_cast_fp16 = mul(x = var_4026_cast_fp16, y = onehot_63_cast_fp16)[name = tensor("op_4027_cast_fp16")]; + tensor new_k_cache_31_cast_fp16 = add(x = var_4020_cast_fp16, y = var_4027_cast_fp16)[name = tensor("new_k_cache_31_cast_fp16")]; + tensor new_k_cache_31_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_31_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache15_to_fp16_dtype_0 = const()[name = tensor("v_cache15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache15_to_fp16 = cast(dtype = v_cache15_to_fp16_dtype_0, x = v_cache15)[name = tensor("cast_453")]; + tensor var_4033_cast_fp16 = mul(x = v_cache15_to_fp16, y = var_4019_cast_fp16)[name = tensor("op_4033_cast_fp16")]; + tensor var_4038 = const()[name = tensor("op_4038"), val = tensor([1, 1, 16, 64])]; + tensor var_4039_cast_fp16 = reshape(shape = var_4038, x = v_31_cast_fp16)[name = tensor("op_4039_cast_fp16")]; + tensor var_4040_cast_fp16 = mul(x = var_4039_cast_fp16, y = onehot_63_cast_fp16)[name = tensor("op_4040_cast_fp16")]; + tensor new_v_cache_31_cast_fp16 = add(x = var_4033_cast_fp16, y = var_4040_cast_fp16)[name = tensor("new_v_cache_31_cast_fp16")]; + tensor new_v_cache_31_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_31_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_15_perm_0 = const()[name = tensor("transpose_15_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_4053 = const()[name = tensor("op_4053"), val = tensor([0, 2, 1, 3])]; + tensor var_4059 = const()[name = tensor("op_4059"), val = tensor([1, 16, 1, 64])]; + tensor q4_31_cast_fp16 = reshape(shape = var_4059, x = out_61_cast_fp16)[name = tensor("q4_31_cast_fp16")]; + tensor var_4064_transpose_x_0 = const()[name = tensor("op_4064_transpose_x_0"), val = tensor(false)]; + tensor var_4064_transpose_y_0 = const()[name = tensor("op_4064_transpose_y_0"), val = tensor(false)]; + tensor transpose_15_cast_fp16 = transpose(perm = transpose_15_perm_0, x = new_k_cache_31_cast_fp16)[name = tensor("transpose_50")]; + tensor var_4064_cast_fp16 = matmul(transpose_x = var_4064_transpose_x_0, transpose_y = var_4064_transpose_y_0, x = q4_31_cast_fp16, y = transpose_15_cast_fp16)[name = tensor("op_4064_cast_fp16")]; + tensor var_4065_to_fp16 = const()[name = tensor("op_4065_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_61_cast_fp16 = mul(x = var_4064_cast_fp16, y = var_4065_to_fp16)[name = tensor("scores_61_cast_fp16")]; + tensor var_4069 = const()[name = tensor("op_4069"), val = tensor([1, 1])]; + tensor var_4070_cast_fp16 = reshape(shape = var_4069, x = position15_to_fp16)[name = tensor("op_4070_cast_fp16")]; + tensor var_4071_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_4070_cast_fp16)[name = tensor("op_4071_cast_fp16")]; + tensor mask_31_to_fp16_dtype_0 = const()[name = tensor("mask_31_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_4078_to_fp16 = const()[name = tensor("op_4078_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4071_cast_fp16_to_fp16 = cast(dtype = mask_31_to_fp16_dtype_0, x = var_4071_cast_fp16)[name = tensor("cast_451")]; + tensor var_4079_cast_fp16 = sub(x = var_4071_cast_fp16_to_fp16, y = var_4078_to_fp16)[name = tensor("op_4079_cast_fp16")]; + tensor var_4084 = const()[name = tensor("op_4084"), val = tensor([1, 1, 1, 512])]; + tensor var_4085_cast_fp16 = reshape(shape = var_4084, x = var_4079_cast_fp16)[name = tensor("op_4085_cast_fp16")]; + tensor var_4086_to_fp16 = const()[name = tensor("op_4086_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_4087_cast_fp16 = mul(x = var_4085_cast_fp16, y = var_4086_to_fp16)[name = tensor("op_4087_cast_fp16")]; + tensor scores_63_cast_fp16 = add(x = scores_61_cast_fp16, y = var_4087_cast_fp16)[name = tensor("scores_63_cast_fp16")]; + tensor var_4090 = const()[name = tensor("op_4090"), val = tensor(-1)]; + tensor weights_31_cast_fp16 = softmax(axis = var_4090, x = scores_63_cast_fp16)[name = tensor("weights_31_cast_fp16")]; + tensor attn_31_transpose_x_0 = const()[name = tensor("attn_31_transpose_x_0"), val = tensor(false)]; + tensor attn_31_transpose_y_0 = const()[name = tensor("attn_31_transpose_y_0"), val = tensor(false)]; + tensor values_31_cast_fp16 = transpose(perm = var_4053, x = new_v_cache_31_cast_fp16)[name = tensor("transpose_49")]; + tensor attn_31_cast_fp16 = matmul(transpose_x = attn_31_transpose_x_0, transpose_y = attn_31_transpose_y_0, x = weights_31_cast_fp16, y = values_31_cast_fp16)[name = tensor("attn_31_cast_fp16")]; + tensor var_4098 = const()[name = tensor("op_4098"), val = tensor([0, 2, 1, 3])]; + tensor var_4103 = const()[name = tensor("op_4103"), val = tensor([1, 1, 1024])]; + tensor var_4099_cast_fp16 = transpose(perm = var_4098, x = attn_31_cast_fp16)[name = tensor("transpose_48")]; + tensor input_187_cast_fp16 = reshape(shape = var_4103, x = var_4099_cast_fp16)[name = tensor("input_187_cast_fp16")]; + tensor attn15_out_proj_weight_to_fp16 = const()[name = tensor("attn15_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(383997120)))]; + tensor linear_62_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn15_out_proj_weight_to_fp16, x = input_187_cast_fp16)[name = tensor("linear_62_cast_fp16")]; + tensor var_4109_to_fp16 = const()[name = tensor("op_4109_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4110_cast_fp16 = add(x = position15_to_fp16, y = var_4109_to_fp16)[name = tensor("op_4110_cast_fp16")]; + tensor var_4110_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4110_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_189_cast_fp16 = add(x = input_183_cast_fp16, y = linear_62_cast_fp16)[name = tensor("input_189_cast_fp16")]; + tensor input_191_axes_0 = const()[name = tensor("input_191_axes_0"), val = tensor([-1])]; + tensor norm15_2_weight_to_fp16 = const()[name = tensor("norm15_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(386094336)))]; + tensor norm15_2_bias_to_fp16 = const()[name = tensor("norm15_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(386096448)))]; + tensor var_4114_to_fp16 = const()[name = tensor("op_4114_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_191_cast_fp16 = layer_norm(axes = input_191_axes_0, beta = norm15_2_bias_to_fp16, epsilon = var_4114_to_fp16, gamma = norm15_2_weight_to_fp16, x = input_189_cast_fp16)[name = tensor("input_191_cast_fp16")]; + tensor linear15_1_weight_to_fp16 = const()[name = tensor("linear15_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(386098560)))]; + tensor linear_63_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear15_1_weight_to_fp16, x = input_191_cast_fp16)[name = tensor("linear_63_cast_fp16")]; + tensor input_193_mode_0 = const()[name = tensor("input_193_mode_0"), val = tensor("EXACT")]; + tensor input_193_cast_fp16 = gelu(mode = input_193_mode_0, x = linear_63_cast_fp16)[name = tensor("input_193_cast_fp16")]; + tensor linear15_2_weight_to_fp16 = const()[name = tensor("linear15_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(394487232)))]; + tensor linear_64_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear15_2_weight_to_fp16, x = input_193_cast_fp16)[name = tensor("linear_64_cast_fp16")]; + tensor input_195_cast_fp16 = add(x = input_189_cast_fp16, y = linear_64_cast_fp16)[name = tensor("input_195_cast_fp16")]; + tensor input_197_axes_0 = const()[name = tensor("input_197_axes_0"), val = tensor([-1])]; + tensor norm16_1_weight_to_fp16 = const()[name = tensor("norm16_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(402875904)))]; + tensor norm16_1_bias_to_fp16 = const()[name = tensor("norm16_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(402878016)))]; + tensor var_4131_to_fp16 = const()[name = tensor("op_4131_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_197_cast_fp16 = layer_norm(axes = input_197_axes_0, beta = norm16_1_bias_to_fp16, epsilon = var_4131_to_fp16, gamma = norm16_1_weight_to_fp16, x = input_195_cast_fp16)[name = tensor("input_197_cast_fp16")]; + tensor attn16_in_proj_weight_to_fp16 = const()[name = tensor("attn16_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(402880128)))]; + tensor linear_65_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn16_in_proj_weight_to_fp16, x = input_197_cast_fp16)[name = tensor("linear_65_cast_fp16")]; + tensor var_4144 = const()[name = tensor("op_4144"), val = tensor([1, 3, 16, 64])]; + tensor qkv_33_cast_fp16 = reshape(shape = var_4144, x = linear_65_cast_fp16)[name = tensor("qkv_33_cast_fp16")]; + tensor x_65_begin_0 = const()[name = tensor("x_65_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_65_end_0 = const()[name = tensor("x_65_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_65_end_mask_0 = const()[name = tensor("x_65_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_65_squeeze_mask_0 = const()[name = tensor("x_65_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_65_cast_fp16 = slice_by_index(begin = x_65_begin_0, end = x_65_end_0, end_mask = x_65_end_mask_0, squeeze_mask = x_65_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("x_65_cast_fp16")]; + tensor x_67_begin_0 = const()[name = tensor("x_67_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_67_end_0 = const()[name = tensor("x_67_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_67_end_mask_0 = const()[name = tensor("x_67_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_67_squeeze_mask_0 = const()[name = tensor("x_67_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_67_cast_fp16 = slice_by_index(begin = x_67_begin_0, end = x_67_end_0, end_mask = x_67_end_mask_0, squeeze_mask = x_67_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("x_67_cast_fp16")]; + tensor v_33_begin_0 = const()[name = tensor("v_33_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_33_end_0 = const()[name = tensor("v_33_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_33_end_mask_0 = const()[name = tensor("v_33_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_33_squeeze_mask_0 = const()[name = tensor("v_33_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_33_cast_fp16 = slice_by_index(begin = v_33_begin_0, end = v_33_end_0, end_mask = v_33_end_mask_0, squeeze_mask = v_33_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("v_33_cast_fp16")]; + tensor var_4173 = const()[name = tensor("op_4173"), val = tensor([1, 1, 1])]; + tensor position16_to_fp16_dtype_0 = const()[name = tensor("position16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position16_to_fp16 = cast(dtype = position16_to_fp16_dtype_0, x = position16)[name = tensor("cast_449")]; + tensor var_4174_cast_fp16 = reshape(shape = var_4173, x = position16_to_fp16)[name = tensor("op_4174_cast_fp16")]; + tensor angles_65_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_4174_cast_fp16)[name = tensor("angles_65_cast_fp16")]; + tensor rotr_65_cast_fp16 = cos(x = angles_65_cast_fp16)[name = tensor("rotr_65_cast_fp16")]; + tensor roti_65_cast_fp16 = sin(x = angles_65_cast_fp16)[name = tensor("roti_65_cast_fp16")]; + tensor var_4182 = const()[name = tensor("op_4182"), val = tensor([1, 16, 32, 2])]; + tensor pairs_65_cast_fp16 = reshape(shape = var_4182, x = x_65_cast_fp16)[name = tensor("pairs_65_cast_fp16")]; + tensor xr_65_begin_0 = const()[name = tensor("xr_65_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_65_end_0 = const()[name = tensor("xr_65_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_65_end_mask_0 = const()[name = tensor("xr_65_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_65_squeeze_mask_0 = const()[name = tensor("xr_65_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_65_cast_fp16 = slice_by_index(begin = xr_65_begin_0, end = xr_65_end_0, end_mask = xr_65_end_mask_0, squeeze_mask = xr_65_squeeze_mask_0, x = pairs_65_cast_fp16)[name = tensor("xr_65_cast_fp16")]; + tensor xi_65_begin_0 = const()[name = tensor("xi_65_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_65_end_0 = const()[name = tensor("xi_65_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_65_end_mask_0 = const()[name = tensor("xi_65_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_65_squeeze_mask_0 = const()[name = tensor("xi_65_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_65_cast_fp16 = slice_by_index(begin = xi_65_begin_0, end = xi_65_end_0, end_mask = xi_65_end_mask_0, squeeze_mask = xi_65_squeeze_mask_0, x = pairs_65_cast_fp16)[name = tensor("xi_65_cast_fp16")]; + tensor var_4190_cast_fp16 = mul(x = xr_65_cast_fp16, y = rotr_65_cast_fp16)[name = tensor("op_4190_cast_fp16")]; + tensor var_4191_cast_fp16 = mul(x = xi_65_cast_fp16, y = roti_65_cast_fp16)[name = tensor("op_4191_cast_fp16")]; + tensor xor__65_cast_fp16 = sub(x = var_4190_cast_fp16, y = var_4191_cast_fp16)[name = tensor("xor__65_cast_fp16")]; + tensor var_4194_cast_fp16 = mul(x = xr_65_cast_fp16, y = roti_65_cast_fp16)[name = tensor("op_4194_cast_fp16")]; + tensor var_4195_cast_fp16 = mul(x = xi_65_cast_fp16, y = rotr_65_cast_fp16)[name = tensor("op_4195_cast_fp16")]; + tensor xoi_65_cast_fp16 = add(x = var_4194_cast_fp16, y = var_4195_cast_fp16)[name = tensor("xoi_65_cast_fp16")]; + tensor out_65_axis_0 = const()[name = tensor("out_65_axis_0"), val = tensor(-1)]; + tensor out_65_cast_fp16 = stack(axis = out_65_axis_0, values = (xor__65_cast_fp16, xoi_65_cast_fp16))[name = tensor("out_65_cast_fp16")]; + tensor var_4218 = const()[name = tensor("op_4218"), val = tensor([1, 16, 32, 2])]; + tensor pairs_67_cast_fp16 = reshape(shape = var_4218, x = x_67_cast_fp16)[name = tensor("pairs_67_cast_fp16")]; + tensor xr_67_begin_0 = const()[name = tensor("xr_67_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_67_end_0 = const()[name = tensor("xr_67_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_67_end_mask_0 = const()[name = tensor("xr_67_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_67_squeeze_mask_0 = const()[name = tensor("xr_67_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_67_cast_fp16 = slice_by_index(begin = xr_67_begin_0, end = xr_67_end_0, end_mask = xr_67_end_mask_0, squeeze_mask = xr_67_squeeze_mask_0, x = pairs_67_cast_fp16)[name = tensor("xr_67_cast_fp16")]; + tensor xi_67_begin_0 = const()[name = tensor("xi_67_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_67_end_0 = const()[name = tensor("xi_67_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_67_end_mask_0 = const()[name = tensor("xi_67_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_67_squeeze_mask_0 = const()[name = tensor("xi_67_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_67_cast_fp16 = slice_by_index(begin = xi_67_begin_0, end = xi_67_end_0, end_mask = xi_67_end_mask_0, squeeze_mask = xi_67_squeeze_mask_0, x = pairs_67_cast_fp16)[name = tensor("xi_67_cast_fp16")]; + tensor var_4226_cast_fp16 = mul(x = xr_67_cast_fp16, y = rotr_65_cast_fp16)[name = tensor("op_4226_cast_fp16")]; + tensor var_4227_cast_fp16 = mul(x = xi_67_cast_fp16, y = roti_65_cast_fp16)[name = tensor("op_4227_cast_fp16")]; + tensor xor__67_cast_fp16 = sub(x = var_4226_cast_fp16, y = var_4227_cast_fp16)[name = tensor("xor__67_cast_fp16")]; + tensor var_4230_cast_fp16 = mul(x = xr_67_cast_fp16, y = roti_65_cast_fp16)[name = tensor("op_4230_cast_fp16")]; + tensor var_4231_cast_fp16 = mul(x = xi_67_cast_fp16, y = rotr_65_cast_fp16)[name = tensor("op_4231_cast_fp16")]; + tensor xoi_67_cast_fp16 = add(x = var_4230_cast_fp16, y = var_4231_cast_fp16)[name = tensor("xoi_67_cast_fp16")]; + tensor out_67_axis_0 = const()[name = tensor("out_67_axis_0"), val = tensor(-1)]; + tensor out_67_cast_fp16 = stack(axis = out_67_axis_0, values = (xor__67_cast_fp16, xoi_67_cast_fp16))[name = tensor("out_67_cast_fp16")]; + tensor _inversed_4243_y_0_to_fp16 = const()[name = tensor("_inversed_4243_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_4243_cast_fp16 = mul(x = position16_to_fp16, y = _inversed_4243_y_0_to_fp16)[name = tensor("_inversed_4243_cast_fp16")]; + tensor var_4244_cast_fp16 = floor(x = _inversed_4243_cast_fp16)[name = tensor("op_4244_cast_fp16")]; + tensor var_4245_to_fp16 = const()[name = tensor("op_4245_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4246_cast_fp16 = mul(x = var_4244_cast_fp16, y = var_4245_to_fp16)[name = tensor("op_4246_cast_fp16")]; + tensor write_idx_33_cast_fp16 = sub(x = position16_to_fp16, y = var_4246_cast_fp16)[name = tensor("write_idx_33_cast_fp16")]; + tensor var_4249_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_33_cast_fp16)[name = tensor("op_4249_cast_fp16")]; + tensor var_4259 = const()[name = tensor("op_4259"), val = tensor([1, 512, 1, 1])]; + tensor onehot_65_to_fp16_dtype_0 = const()[name = tensor("onehot_65_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_4249_cast_fp16_to_fp16 = cast(dtype = onehot_65_to_fp16_dtype_0, x = var_4249_cast_fp16)[name = tensor("cast_448")]; + tensor onehot_67_cast_fp16 = reshape(shape = var_4259, x = var_4249_cast_fp16_to_fp16)[name = tensor("onehot_67_cast_fp16")]; + tensor var_4261_to_fp16 = const()[name = tensor("op_4261_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4263_cast_fp16 = sub(x = var_4261_to_fp16, y = onehot_67_cast_fp16)[name = tensor("op_4263_cast_fp16")]; + tensor k_cache16_to_fp16_dtype_0 = const()[name = tensor("k_cache16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache16_to_fp16 = cast(dtype = k_cache16_to_fp16_dtype_0, x = k_cache16)[name = tensor("cast_447")]; + tensor var_4264_cast_fp16 = mul(x = k_cache16_to_fp16, y = var_4263_cast_fp16)[name = tensor("op_4264_cast_fp16")]; + tensor var_4269 = const()[name = tensor("op_4269"), val = tensor([1, 1, 16, 64])]; + tensor var_4270_cast_fp16 = reshape(shape = var_4269, x = out_67_cast_fp16)[name = tensor("op_4270_cast_fp16")]; + tensor var_4271_cast_fp16 = mul(x = var_4270_cast_fp16, y = onehot_67_cast_fp16)[name = tensor("op_4271_cast_fp16")]; + tensor new_k_cache_33_cast_fp16 = add(x = var_4264_cast_fp16, y = var_4271_cast_fp16)[name = tensor("new_k_cache_33_cast_fp16")]; + tensor new_k_cache_33_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_33_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache16_to_fp16_dtype_0 = const()[name = tensor("v_cache16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache16_to_fp16 = cast(dtype = v_cache16_to_fp16_dtype_0, x = v_cache16)[name = tensor("cast_445")]; + tensor var_4277_cast_fp16 = mul(x = v_cache16_to_fp16, y = var_4263_cast_fp16)[name = tensor("op_4277_cast_fp16")]; + tensor var_4282 = const()[name = tensor("op_4282"), val = tensor([1, 1, 16, 64])]; + tensor var_4283_cast_fp16 = reshape(shape = var_4282, x = v_33_cast_fp16)[name = tensor("op_4283_cast_fp16")]; + tensor var_4284_cast_fp16 = mul(x = var_4283_cast_fp16, y = onehot_67_cast_fp16)[name = tensor("op_4284_cast_fp16")]; + tensor new_v_cache_33_cast_fp16 = add(x = var_4277_cast_fp16, y = var_4284_cast_fp16)[name = tensor("new_v_cache_33_cast_fp16")]; + tensor new_v_cache_33_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_33_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_16_perm_0 = const()[name = tensor("transpose_16_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_4297 = const()[name = tensor("op_4297"), val = tensor([0, 2, 1, 3])]; + tensor var_4303 = const()[name = tensor("op_4303"), val = tensor([1, 16, 1, 64])]; + tensor q4_33_cast_fp16 = reshape(shape = var_4303, x = out_65_cast_fp16)[name = tensor("q4_33_cast_fp16")]; + tensor var_4308_transpose_x_0 = const()[name = tensor("op_4308_transpose_x_0"), val = tensor(false)]; + tensor var_4308_transpose_y_0 = const()[name = tensor("op_4308_transpose_y_0"), val = tensor(false)]; + tensor transpose_16_cast_fp16 = transpose(perm = transpose_16_perm_0, x = new_k_cache_33_cast_fp16)[name = tensor("transpose_47")]; + tensor var_4308_cast_fp16 = matmul(transpose_x = var_4308_transpose_x_0, transpose_y = var_4308_transpose_y_0, x = q4_33_cast_fp16, y = transpose_16_cast_fp16)[name = tensor("op_4308_cast_fp16")]; + tensor var_4309_to_fp16 = const()[name = tensor("op_4309_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_65_cast_fp16 = mul(x = var_4308_cast_fp16, y = var_4309_to_fp16)[name = tensor("scores_65_cast_fp16")]; + tensor var_4313 = const()[name = tensor("op_4313"), val = tensor([1, 1])]; + tensor var_4314_cast_fp16 = reshape(shape = var_4313, x = position16_to_fp16)[name = tensor("op_4314_cast_fp16")]; + tensor var_4315_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_4314_cast_fp16)[name = tensor("op_4315_cast_fp16")]; + tensor mask_33_to_fp16_dtype_0 = const()[name = tensor("mask_33_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_4322_to_fp16 = const()[name = tensor("op_4322_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4315_cast_fp16_to_fp16 = cast(dtype = mask_33_to_fp16_dtype_0, x = var_4315_cast_fp16)[name = tensor("cast_443")]; + tensor var_4323_cast_fp16 = sub(x = var_4315_cast_fp16_to_fp16, y = var_4322_to_fp16)[name = tensor("op_4323_cast_fp16")]; + tensor var_4328 = const()[name = tensor("op_4328"), val = tensor([1, 1, 1, 512])]; + tensor var_4329_cast_fp16 = reshape(shape = var_4328, x = var_4323_cast_fp16)[name = tensor("op_4329_cast_fp16")]; + tensor var_4330_to_fp16 = const()[name = tensor("op_4330_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_4331_cast_fp16 = mul(x = var_4329_cast_fp16, y = var_4330_to_fp16)[name = tensor("op_4331_cast_fp16")]; + tensor scores_67_cast_fp16 = add(x = scores_65_cast_fp16, y = var_4331_cast_fp16)[name = tensor("scores_67_cast_fp16")]; + tensor var_4334 = const()[name = tensor("op_4334"), val = tensor(-1)]; + tensor weights_33_cast_fp16 = softmax(axis = var_4334, x = scores_67_cast_fp16)[name = tensor("weights_33_cast_fp16")]; + tensor attn_33_transpose_x_0 = const()[name = tensor("attn_33_transpose_x_0"), val = tensor(false)]; + tensor attn_33_transpose_y_0 = const()[name = tensor("attn_33_transpose_y_0"), val = tensor(false)]; + tensor values_33_cast_fp16 = transpose(perm = var_4297, x = new_v_cache_33_cast_fp16)[name = tensor("transpose_46")]; + tensor attn_33_cast_fp16 = matmul(transpose_x = attn_33_transpose_x_0, transpose_y = attn_33_transpose_y_0, x = weights_33_cast_fp16, y = values_33_cast_fp16)[name = tensor("attn_33_cast_fp16")]; + tensor var_4342 = const()[name = tensor("op_4342"), val = tensor([0, 2, 1, 3])]; + tensor var_4347 = const()[name = tensor("op_4347"), val = tensor([1, 1, 1024])]; + tensor var_4343_cast_fp16 = transpose(perm = var_4342, x = attn_33_cast_fp16)[name = tensor("transpose_45")]; + tensor input_199_cast_fp16 = reshape(shape = var_4347, x = var_4343_cast_fp16)[name = tensor("input_199_cast_fp16")]; + tensor attn16_out_proj_weight_to_fp16 = const()[name = tensor("attn16_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(409171648)))]; + tensor linear_66_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn16_out_proj_weight_to_fp16, x = input_199_cast_fp16)[name = tensor("linear_66_cast_fp16")]; + tensor var_4353_to_fp16 = const()[name = tensor("op_4353_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4354_cast_fp16 = add(x = position16_to_fp16, y = var_4353_to_fp16)[name = tensor("op_4354_cast_fp16")]; + tensor var_4354_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4354_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_201_cast_fp16 = add(x = input_195_cast_fp16, y = linear_66_cast_fp16)[name = tensor("input_201_cast_fp16")]; + tensor input_203_axes_0 = const()[name = tensor("input_203_axes_0"), val = tensor([-1])]; + tensor norm16_2_weight_to_fp16 = const()[name = tensor("norm16_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411268864)))]; + tensor norm16_2_bias_to_fp16 = const()[name = tensor("norm16_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411270976)))]; + tensor var_4358_to_fp16 = const()[name = tensor("op_4358_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_203_cast_fp16 = layer_norm(axes = input_203_axes_0, beta = norm16_2_bias_to_fp16, epsilon = var_4358_to_fp16, gamma = norm16_2_weight_to_fp16, x = input_201_cast_fp16)[name = tensor("input_203_cast_fp16")]; + tensor linear16_1_weight_to_fp16 = const()[name = tensor("linear16_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(411273088)))]; + tensor linear_67_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear16_1_weight_to_fp16, x = input_203_cast_fp16)[name = tensor("linear_67_cast_fp16")]; + tensor input_205_mode_0 = const()[name = tensor("input_205_mode_0"), val = tensor("EXACT")]; + tensor input_205_cast_fp16 = gelu(mode = input_205_mode_0, x = linear_67_cast_fp16)[name = tensor("input_205_cast_fp16")]; + tensor linear16_2_weight_to_fp16 = const()[name = tensor("linear16_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(419661760)))]; + tensor linear_68_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear16_2_weight_to_fp16, x = input_205_cast_fp16)[name = tensor("linear_68_cast_fp16")]; + tensor input_207_cast_fp16 = add(x = input_201_cast_fp16, y = linear_68_cast_fp16)[name = tensor("input_207_cast_fp16")]; + tensor input_209_axes_0 = const()[name = tensor("input_209_axes_0"), val = tensor([-1])]; + tensor norm17_1_weight_to_fp16 = const()[name = tensor("norm17_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(428050432)))]; + tensor norm17_1_bias_to_fp16 = const()[name = tensor("norm17_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(428052544)))]; + tensor var_4375_to_fp16 = const()[name = tensor("op_4375_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_209_cast_fp16 = layer_norm(axes = input_209_axes_0, beta = norm17_1_bias_to_fp16, epsilon = var_4375_to_fp16, gamma = norm17_1_weight_to_fp16, x = input_207_cast_fp16)[name = tensor("input_209_cast_fp16")]; + tensor attn17_in_proj_weight_to_fp16 = const()[name = tensor("attn17_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(428054656)))]; + tensor linear_69_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn17_in_proj_weight_to_fp16, x = input_209_cast_fp16)[name = tensor("linear_69_cast_fp16")]; + tensor var_4388 = const()[name = tensor("op_4388"), val = tensor([1, 3, 16, 64])]; + tensor qkv_35_cast_fp16 = reshape(shape = var_4388, x = linear_69_cast_fp16)[name = tensor("qkv_35_cast_fp16")]; + tensor x_69_begin_0 = const()[name = tensor("x_69_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_69_end_0 = const()[name = tensor("x_69_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_69_end_mask_0 = const()[name = tensor("x_69_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_69_squeeze_mask_0 = const()[name = tensor("x_69_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_69_cast_fp16 = slice_by_index(begin = x_69_begin_0, end = x_69_end_0, end_mask = x_69_end_mask_0, squeeze_mask = x_69_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("x_69_cast_fp16")]; + tensor x_71_begin_0 = const()[name = tensor("x_71_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_71_end_0 = const()[name = tensor("x_71_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_71_end_mask_0 = const()[name = tensor("x_71_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_71_squeeze_mask_0 = const()[name = tensor("x_71_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_71_cast_fp16 = slice_by_index(begin = x_71_begin_0, end = x_71_end_0, end_mask = x_71_end_mask_0, squeeze_mask = x_71_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("x_71_cast_fp16")]; + tensor v_35_begin_0 = const()[name = tensor("v_35_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_35_end_0 = const()[name = tensor("v_35_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_35_end_mask_0 = const()[name = tensor("v_35_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_35_squeeze_mask_0 = const()[name = tensor("v_35_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_35_cast_fp16 = slice_by_index(begin = v_35_begin_0, end = v_35_end_0, end_mask = v_35_end_mask_0, squeeze_mask = v_35_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("v_35_cast_fp16")]; + tensor var_4417 = const()[name = tensor("op_4417"), val = tensor([1, 1, 1])]; + tensor position17_to_fp16_dtype_0 = const()[name = tensor("position17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position17_to_fp16 = cast(dtype = position17_to_fp16_dtype_0, x = position17)[name = tensor("cast_441")]; + tensor var_4418_cast_fp16 = reshape(shape = var_4417, x = position17_to_fp16)[name = tensor("op_4418_cast_fp16")]; + tensor angles_69_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_4418_cast_fp16)[name = tensor("angles_69_cast_fp16")]; + tensor rotr_69_cast_fp16 = cos(x = angles_69_cast_fp16)[name = tensor("rotr_69_cast_fp16")]; + tensor roti_69_cast_fp16 = sin(x = angles_69_cast_fp16)[name = tensor("roti_69_cast_fp16")]; + tensor var_4426 = const()[name = tensor("op_4426"), val = tensor([1, 16, 32, 2])]; + tensor pairs_69_cast_fp16 = reshape(shape = var_4426, x = x_69_cast_fp16)[name = tensor("pairs_69_cast_fp16")]; + tensor xr_69_begin_0 = const()[name = tensor("xr_69_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_69_end_0 = const()[name = tensor("xr_69_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_69_end_mask_0 = const()[name = tensor("xr_69_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_69_squeeze_mask_0 = const()[name = tensor("xr_69_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_69_cast_fp16 = slice_by_index(begin = xr_69_begin_0, end = xr_69_end_0, end_mask = xr_69_end_mask_0, squeeze_mask = xr_69_squeeze_mask_0, x = pairs_69_cast_fp16)[name = tensor("xr_69_cast_fp16")]; + tensor xi_69_begin_0 = const()[name = tensor("xi_69_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_69_end_0 = const()[name = tensor("xi_69_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_69_end_mask_0 = const()[name = tensor("xi_69_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_69_squeeze_mask_0 = const()[name = tensor("xi_69_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_69_cast_fp16 = slice_by_index(begin = xi_69_begin_0, end = xi_69_end_0, end_mask = xi_69_end_mask_0, squeeze_mask = xi_69_squeeze_mask_0, x = pairs_69_cast_fp16)[name = tensor("xi_69_cast_fp16")]; + tensor var_4434_cast_fp16 = mul(x = xr_69_cast_fp16, y = rotr_69_cast_fp16)[name = tensor("op_4434_cast_fp16")]; + tensor var_4435_cast_fp16 = mul(x = xi_69_cast_fp16, y = roti_69_cast_fp16)[name = tensor("op_4435_cast_fp16")]; + tensor xor__69_cast_fp16 = sub(x = var_4434_cast_fp16, y = var_4435_cast_fp16)[name = tensor("xor__69_cast_fp16")]; + tensor var_4438_cast_fp16 = mul(x = xr_69_cast_fp16, y = roti_69_cast_fp16)[name = tensor("op_4438_cast_fp16")]; + tensor var_4439_cast_fp16 = mul(x = xi_69_cast_fp16, y = rotr_69_cast_fp16)[name = tensor("op_4439_cast_fp16")]; + tensor xoi_69_cast_fp16 = add(x = var_4438_cast_fp16, y = var_4439_cast_fp16)[name = tensor("xoi_69_cast_fp16")]; + tensor out_69_axis_0 = const()[name = tensor("out_69_axis_0"), val = tensor(-1)]; + tensor out_69_cast_fp16 = stack(axis = out_69_axis_0, values = (xor__69_cast_fp16, xoi_69_cast_fp16))[name = tensor("out_69_cast_fp16")]; + tensor var_4462 = const()[name = tensor("op_4462"), val = tensor([1, 16, 32, 2])]; + tensor pairs_71_cast_fp16 = reshape(shape = var_4462, x = x_71_cast_fp16)[name = tensor("pairs_71_cast_fp16")]; + tensor xr_71_begin_0 = const()[name = tensor("xr_71_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_71_end_0 = const()[name = tensor("xr_71_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_71_end_mask_0 = const()[name = tensor("xr_71_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_71_squeeze_mask_0 = const()[name = tensor("xr_71_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_71_cast_fp16 = slice_by_index(begin = xr_71_begin_0, end = xr_71_end_0, end_mask = xr_71_end_mask_0, squeeze_mask = xr_71_squeeze_mask_0, x = pairs_71_cast_fp16)[name = tensor("xr_71_cast_fp16")]; + tensor xi_71_begin_0 = const()[name = tensor("xi_71_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_71_end_0 = const()[name = tensor("xi_71_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_71_end_mask_0 = const()[name = tensor("xi_71_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_71_squeeze_mask_0 = const()[name = tensor("xi_71_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_71_cast_fp16 = slice_by_index(begin = xi_71_begin_0, end = xi_71_end_0, end_mask = xi_71_end_mask_0, squeeze_mask = xi_71_squeeze_mask_0, x = pairs_71_cast_fp16)[name = tensor("xi_71_cast_fp16")]; + tensor var_4470_cast_fp16 = mul(x = xr_71_cast_fp16, y = rotr_69_cast_fp16)[name = tensor("op_4470_cast_fp16")]; + tensor var_4471_cast_fp16 = mul(x = xi_71_cast_fp16, y = roti_69_cast_fp16)[name = tensor("op_4471_cast_fp16")]; + tensor xor__71_cast_fp16 = sub(x = var_4470_cast_fp16, y = var_4471_cast_fp16)[name = tensor("xor__71_cast_fp16")]; + tensor var_4474_cast_fp16 = mul(x = xr_71_cast_fp16, y = roti_69_cast_fp16)[name = tensor("op_4474_cast_fp16")]; + tensor var_4475_cast_fp16 = mul(x = xi_71_cast_fp16, y = rotr_69_cast_fp16)[name = tensor("op_4475_cast_fp16")]; + tensor xoi_71_cast_fp16 = add(x = var_4474_cast_fp16, y = var_4475_cast_fp16)[name = tensor("xoi_71_cast_fp16")]; + tensor out_71_axis_0 = const()[name = tensor("out_71_axis_0"), val = tensor(-1)]; + tensor out_71_cast_fp16 = stack(axis = out_71_axis_0, values = (xor__71_cast_fp16, xoi_71_cast_fp16))[name = tensor("out_71_cast_fp16")]; + tensor _inversed_4487_y_0_to_fp16 = const()[name = tensor("_inversed_4487_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_4487_cast_fp16 = mul(x = position17_to_fp16, y = _inversed_4487_y_0_to_fp16)[name = tensor("_inversed_4487_cast_fp16")]; + tensor var_4488_cast_fp16 = floor(x = _inversed_4487_cast_fp16)[name = tensor("op_4488_cast_fp16")]; + tensor var_4489_to_fp16 = const()[name = tensor("op_4489_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4490_cast_fp16 = mul(x = var_4488_cast_fp16, y = var_4489_to_fp16)[name = tensor("op_4490_cast_fp16")]; + tensor write_idx_35_cast_fp16 = sub(x = position17_to_fp16, y = var_4490_cast_fp16)[name = tensor("write_idx_35_cast_fp16")]; + tensor var_4493_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_35_cast_fp16)[name = tensor("op_4493_cast_fp16")]; + tensor var_4503 = const()[name = tensor("op_4503"), val = tensor([1, 512, 1, 1])]; + tensor onehot_69_to_fp16_dtype_0 = const()[name = tensor("onehot_69_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_4493_cast_fp16_to_fp16 = cast(dtype = onehot_69_to_fp16_dtype_0, x = var_4493_cast_fp16)[name = tensor("cast_440")]; + tensor onehot_71_cast_fp16 = reshape(shape = var_4503, x = var_4493_cast_fp16_to_fp16)[name = tensor("onehot_71_cast_fp16")]; + tensor var_4505_to_fp16 = const()[name = tensor("op_4505_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4507_cast_fp16 = sub(x = var_4505_to_fp16, y = onehot_71_cast_fp16)[name = tensor("op_4507_cast_fp16")]; + tensor k_cache17_to_fp16_dtype_0 = const()[name = tensor("k_cache17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache17_to_fp16 = cast(dtype = k_cache17_to_fp16_dtype_0, x = k_cache17)[name = tensor("cast_439")]; + tensor var_4508_cast_fp16 = mul(x = k_cache17_to_fp16, y = var_4507_cast_fp16)[name = tensor("op_4508_cast_fp16")]; + tensor var_4513 = const()[name = tensor("op_4513"), val = tensor([1, 1, 16, 64])]; + tensor var_4514_cast_fp16 = reshape(shape = var_4513, x = out_71_cast_fp16)[name = tensor("op_4514_cast_fp16")]; + tensor var_4515_cast_fp16 = mul(x = var_4514_cast_fp16, y = onehot_71_cast_fp16)[name = tensor("op_4515_cast_fp16")]; + tensor new_k_cache_35_cast_fp16 = add(x = var_4508_cast_fp16, y = var_4515_cast_fp16)[name = tensor("new_k_cache_35_cast_fp16")]; + tensor new_k_cache_35_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_35_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache17_to_fp16_dtype_0 = const()[name = tensor("v_cache17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache17_to_fp16 = cast(dtype = v_cache17_to_fp16_dtype_0, x = v_cache17)[name = tensor("cast_437")]; + tensor var_4521_cast_fp16 = mul(x = v_cache17_to_fp16, y = var_4507_cast_fp16)[name = tensor("op_4521_cast_fp16")]; + tensor var_4526 = const()[name = tensor("op_4526"), val = tensor([1, 1, 16, 64])]; + tensor var_4527_cast_fp16 = reshape(shape = var_4526, x = v_35_cast_fp16)[name = tensor("op_4527_cast_fp16")]; + tensor var_4528_cast_fp16 = mul(x = var_4527_cast_fp16, y = onehot_71_cast_fp16)[name = tensor("op_4528_cast_fp16")]; + tensor new_v_cache_35_cast_fp16 = add(x = var_4521_cast_fp16, y = var_4528_cast_fp16)[name = tensor("new_v_cache_35_cast_fp16")]; + tensor new_v_cache_35_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_35_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_17_perm_0 = const()[name = tensor("transpose_17_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_4541 = const()[name = tensor("op_4541"), val = tensor([0, 2, 1, 3])]; + tensor var_4547 = const()[name = tensor("op_4547"), val = tensor([1, 16, 1, 64])]; + tensor q4_35_cast_fp16 = reshape(shape = var_4547, x = out_69_cast_fp16)[name = tensor("q4_35_cast_fp16")]; + tensor var_4552_transpose_x_0 = const()[name = tensor("op_4552_transpose_x_0"), val = tensor(false)]; + tensor var_4552_transpose_y_0 = const()[name = tensor("op_4552_transpose_y_0"), val = tensor(false)]; + tensor transpose_17_cast_fp16 = transpose(perm = transpose_17_perm_0, x = new_k_cache_35_cast_fp16)[name = tensor("transpose_44")]; + tensor var_4552_cast_fp16 = matmul(transpose_x = var_4552_transpose_x_0, transpose_y = var_4552_transpose_y_0, x = q4_35_cast_fp16, y = transpose_17_cast_fp16)[name = tensor("op_4552_cast_fp16")]; + tensor var_4553_to_fp16 = const()[name = tensor("op_4553_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_69_cast_fp16 = mul(x = var_4552_cast_fp16, y = var_4553_to_fp16)[name = tensor("scores_69_cast_fp16")]; + tensor var_4557 = const()[name = tensor("op_4557"), val = tensor([1, 1])]; + tensor var_4558_cast_fp16 = reshape(shape = var_4557, x = position17_to_fp16)[name = tensor("op_4558_cast_fp16")]; + tensor var_4559_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_4558_cast_fp16)[name = tensor("op_4559_cast_fp16")]; + tensor mask_35_to_fp16_dtype_0 = const()[name = tensor("mask_35_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_4566_to_fp16 = const()[name = tensor("op_4566_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4559_cast_fp16_to_fp16 = cast(dtype = mask_35_to_fp16_dtype_0, x = var_4559_cast_fp16)[name = tensor("cast_435")]; + tensor var_4567_cast_fp16 = sub(x = var_4559_cast_fp16_to_fp16, y = var_4566_to_fp16)[name = tensor("op_4567_cast_fp16")]; + tensor var_4572 = const()[name = tensor("op_4572"), val = tensor([1, 1, 1, 512])]; + tensor var_4573_cast_fp16 = reshape(shape = var_4572, x = var_4567_cast_fp16)[name = tensor("op_4573_cast_fp16")]; + tensor var_4574_to_fp16 = const()[name = tensor("op_4574_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_4575_cast_fp16 = mul(x = var_4573_cast_fp16, y = var_4574_to_fp16)[name = tensor("op_4575_cast_fp16")]; + tensor scores_71_cast_fp16 = add(x = scores_69_cast_fp16, y = var_4575_cast_fp16)[name = tensor("scores_71_cast_fp16")]; + tensor var_4578 = const()[name = tensor("op_4578"), val = tensor(-1)]; + tensor weights_35_cast_fp16 = softmax(axis = var_4578, x = scores_71_cast_fp16)[name = tensor("weights_35_cast_fp16")]; + tensor attn_35_transpose_x_0 = const()[name = tensor("attn_35_transpose_x_0"), val = tensor(false)]; + tensor attn_35_transpose_y_0 = const()[name = tensor("attn_35_transpose_y_0"), val = tensor(false)]; + tensor values_35_cast_fp16 = transpose(perm = var_4541, x = new_v_cache_35_cast_fp16)[name = tensor("transpose_43")]; + tensor attn_35_cast_fp16 = matmul(transpose_x = attn_35_transpose_x_0, transpose_y = attn_35_transpose_y_0, x = weights_35_cast_fp16, y = values_35_cast_fp16)[name = tensor("attn_35_cast_fp16")]; + tensor var_4586 = const()[name = tensor("op_4586"), val = tensor([0, 2, 1, 3])]; + tensor var_4591 = const()[name = tensor("op_4591"), val = tensor([1, 1, 1024])]; + tensor var_4587_cast_fp16 = transpose(perm = var_4586, x = attn_35_cast_fp16)[name = tensor("transpose_42")]; + tensor input_211_cast_fp16 = reshape(shape = var_4591, x = var_4587_cast_fp16)[name = tensor("input_211_cast_fp16")]; + tensor attn17_out_proj_weight_to_fp16 = const()[name = tensor("attn17_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(434346176)))]; + tensor linear_70_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn17_out_proj_weight_to_fp16, x = input_211_cast_fp16)[name = tensor("linear_70_cast_fp16")]; + tensor var_4597_to_fp16 = const()[name = tensor("op_4597_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4598_cast_fp16 = add(x = position17_to_fp16, y = var_4597_to_fp16)[name = tensor("op_4598_cast_fp16")]; + tensor var_4598_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4598_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_213_cast_fp16 = add(x = input_207_cast_fp16, y = linear_70_cast_fp16)[name = tensor("input_213_cast_fp16")]; + tensor input_215_axes_0 = const()[name = tensor("input_215_axes_0"), val = tensor([-1])]; + tensor norm17_2_weight_to_fp16 = const()[name = tensor("norm17_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(436443392)))]; + tensor norm17_2_bias_to_fp16 = const()[name = tensor("norm17_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(436445504)))]; + tensor var_4602_to_fp16 = const()[name = tensor("op_4602_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_215_cast_fp16 = layer_norm(axes = input_215_axes_0, beta = norm17_2_bias_to_fp16, epsilon = var_4602_to_fp16, gamma = norm17_2_weight_to_fp16, x = input_213_cast_fp16)[name = tensor("input_215_cast_fp16")]; + tensor linear17_1_weight_to_fp16 = const()[name = tensor("linear17_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(436447616)))]; + tensor linear_71_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear17_1_weight_to_fp16, x = input_215_cast_fp16)[name = tensor("linear_71_cast_fp16")]; + tensor input_217_mode_0 = const()[name = tensor("input_217_mode_0"), val = tensor("EXACT")]; + tensor input_217_cast_fp16 = gelu(mode = input_217_mode_0, x = linear_71_cast_fp16)[name = tensor("input_217_cast_fp16")]; + tensor linear17_2_weight_to_fp16 = const()[name = tensor("linear17_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(444836288)))]; + tensor linear_72_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear17_2_weight_to_fp16, x = input_217_cast_fp16)[name = tensor("linear_72_cast_fp16")]; + tensor input_219_cast_fp16 = add(x = input_213_cast_fp16, y = linear_72_cast_fp16)[name = tensor("input_219_cast_fp16")]; + tensor input_221_axes_0 = const()[name = tensor("input_221_axes_0"), val = tensor([-1])]; + tensor norm18_1_weight_to_fp16 = const()[name = tensor("norm18_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(453224960)))]; + tensor norm18_1_bias_to_fp16 = const()[name = tensor("norm18_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(453227072)))]; + tensor var_4619_to_fp16 = const()[name = tensor("op_4619_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_221_cast_fp16 = layer_norm(axes = input_221_axes_0, beta = norm18_1_bias_to_fp16, epsilon = var_4619_to_fp16, gamma = norm18_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("input_221_cast_fp16")]; + tensor attn18_in_proj_weight_to_fp16 = const()[name = tensor("attn18_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(453229184)))]; + tensor linear_73_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn18_in_proj_weight_to_fp16, x = input_221_cast_fp16)[name = tensor("linear_73_cast_fp16")]; + tensor var_4632 = const()[name = tensor("op_4632"), val = tensor([1, 3, 16, 64])]; + tensor qkv_37_cast_fp16 = reshape(shape = var_4632, x = linear_73_cast_fp16)[name = tensor("qkv_37_cast_fp16")]; + tensor x_73_begin_0 = const()[name = tensor("x_73_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_73_end_0 = const()[name = tensor("x_73_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_73_end_mask_0 = const()[name = tensor("x_73_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_73_squeeze_mask_0 = const()[name = tensor("x_73_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_73_cast_fp16 = slice_by_index(begin = x_73_begin_0, end = x_73_end_0, end_mask = x_73_end_mask_0, squeeze_mask = x_73_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("x_73_cast_fp16")]; + tensor x_75_begin_0 = const()[name = tensor("x_75_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_75_end_0 = const()[name = tensor("x_75_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_75_end_mask_0 = const()[name = tensor("x_75_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_75_squeeze_mask_0 = const()[name = tensor("x_75_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_75_cast_fp16 = slice_by_index(begin = x_75_begin_0, end = x_75_end_0, end_mask = x_75_end_mask_0, squeeze_mask = x_75_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("x_75_cast_fp16")]; + tensor v_37_begin_0 = const()[name = tensor("v_37_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_37_end_0 = const()[name = tensor("v_37_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_37_end_mask_0 = const()[name = tensor("v_37_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_37_squeeze_mask_0 = const()[name = tensor("v_37_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_37_cast_fp16 = slice_by_index(begin = v_37_begin_0, end = v_37_end_0, end_mask = v_37_end_mask_0, squeeze_mask = v_37_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("v_37_cast_fp16")]; + tensor var_4661 = const()[name = tensor("op_4661"), val = tensor([1, 1, 1])]; + tensor position18_to_fp16_dtype_0 = const()[name = tensor("position18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position18_to_fp16 = cast(dtype = position18_to_fp16_dtype_0, x = position18)[name = tensor("cast_433")]; + tensor var_4662_cast_fp16 = reshape(shape = var_4661, x = position18_to_fp16)[name = tensor("op_4662_cast_fp16")]; + tensor angles_73_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_4662_cast_fp16)[name = tensor("angles_73_cast_fp16")]; + tensor rotr_73_cast_fp16 = cos(x = angles_73_cast_fp16)[name = tensor("rotr_73_cast_fp16")]; + tensor roti_73_cast_fp16 = sin(x = angles_73_cast_fp16)[name = tensor("roti_73_cast_fp16")]; + tensor var_4670 = const()[name = tensor("op_4670"), val = tensor([1, 16, 32, 2])]; + tensor pairs_73_cast_fp16 = reshape(shape = var_4670, x = x_73_cast_fp16)[name = tensor("pairs_73_cast_fp16")]; + tensor xr_73_begin_0 = const()[name = tensor("xr_73_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_73_end_0 = const()[name = tensor("xr_73_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_73_end_mask_0 = const()[name = tensor("xr_73_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_73_squeeze_mask_0 = const()[name = tensor("xr_73_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_73_cast_fp16 = slice_by_index(begin = xr_73_begin_0, end = xr_73_end_0, end_mask = xr_73_end_mask_0, squeeze_mask = xr_73_squeeze_mask_0, x = pairs_73_cast_fp16)[name = tensor("xr_73_cast_fp16")]; + tensor xi_73_begin_0 = const()[name = tensor("xi_73_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_73_end_0 = const()[name = tensor("xi_73_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_73_end_mask_0 = const()[name = tensor("xi_73_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_73_squeeze_mask_0 = const()[name = tensor("xi_73_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_73_cast_fp16 = slice_by_index(begin = xi_73_begin_0, end = xi_73_end_0, end_mask = xi_73_end_mask_0, squeeze_mask = xi_73_squeeze_mask_0, x = pairs_73_cast_fp16)[name = tensor("xi_73_cast_fp16")]; + tensor var_4678_cast_fp16 = mul(x = xr_73_cast_fp16, y = rotr_73_cast_fp16)[name = tensor("op_4678_cast_fp16")]; + tensor var_4679_cast_fp16 = mul(x = xi_73_cast_fp16, y = roti_73_cast_fp16)[name = tensor("op_4679_cast_fp16")]; + tensor xor__73_cast_fp16 = sub(x = var_4678_cast_fp16, y = var_4679_cast_fp16)[name = tensor("xor__73_cast_fp16")]; + tensor var_4682_cast_fp16 = mul(x = xr_73_cast_fp16, y = roti_73_cast_fp16)[name = tensor("op_4682_cast_fp16")]; + tensor var_4683_cast_fp16 = mul(x = xi_73_cast_fp16, y = rotr_73_cast_fp16)[name = tensor("op_4683_cast_fp16")]; + tensor xoi_73_cast_fp16 = add(x = var_4682_cast_fp16, y = var_4683_cast_fp16)[name = tensor("xoi_73_cast_fp16")]; + tensor out_73_axis_0 = const()[name = tensor("out_73_axis_0"), val = tensor(-1)]; + tensor out_73_cast_fp16 = stack(axis = out_73_axis_0, values = (xor__73_cast_fp16, xoi_73_cast_fp16))[name = tensor("out_73_cast_fp16")]; + tensor var_4706 = const()[name = tensor("op_4706"), val = tensor([1, 16, 32, 2])]; + tensor pairs_75_cast_fp16 = reshape(shape = var_4706, x = x_75_cast_fp16)[name = tensor("pairs_75_cast_fp16")]; + tensor xr_75_begin_0 = const()[name = tensor("xr_75_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_75_end_0 = const()[name = tensor("xr_75_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_75_end_mask_0 = const()[name = tensor("xr_75_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_75_squeeze_mask_0 = const()[name = tensor("xr_75_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_75_cast_fp16 = slice_by_index(begin = xr_75_begin_0, end = xr_75_end_0, end_mask = xr_75_end_mask_0, squeeze_mask = xr_75_squeeze_mask_0, x = pairs_75_cast_fp16)[name = tensor("xr_75_cast_fp16")]; + tensor xi_75_begin_0 = const()[name = tensor("xi_75_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_75_end_0 = const()[name = tensor("xi_75_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_75_end_mask_0 = const()[name = tensor("xi_75_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_75_squeeze_mask_0 = const()[name = tensor("xi_75_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_75_cast_fp16 = slice_by_index(begin = xi_75_begin_0, end = xi_75_end_0, end_mask = xi_75_end_mask_0, squeeze_mask = xi_75_squeeze_mask_0, x = pairs_75_cast_fp16)[name = tensor("xi_75_cast_fp16")]; + tensor var_4714_cast_fp16 = mul(x = xr_75_cast_fp16, y = rotr_73_cast_fp16)[name = tensor("op_4714_cast_fp16")]; + tensor var_4715_cast_fp16 = mul(x = xi_75_cast_fp16, y = roti_73_cast_fp16)[name = tensor("op_4715_cast_fp16")]; + tensor xor__75_cast_fp16 = sub(x = var_4714_cast_fp16, y = var_4715_cast_fp16)[name = tensor("xor__75_cast_fp16")]; + tensor var_4718_cast_fp16 = mul(x = xr_75_cast_fp16, y = roti_73_cast_fp16)[name = tensor("op_4718_cast_fp16")]; + tensor var_4719_cast_fp16 = mul(x = xi_75_cast_fp16, y = rotr_73_cast_fp16)[name = tensor("op_4719_cast_fp16")]; + tensor xoi_75_cast_fp16 = add(x = var_4718_cast_fp16, y = var_4719_cast_fp16)[name = tensor("xoi_75_cast_fp16")]; + tensor out_75_axis_0 = const()[name = tensor("out_75_axis_0"), val = tensor(-1)]; + tensor out_75_cast_fp16 = stack(axis = out_75_axis_0, values = (xor__75_cast_fp16, xoi_75_cast_fp16))[name = tensor("out_75_cast_fp16")]; + tensor _inversed_4731_y_0_to_fp16 = const()[name = tensor("_inversed_4731_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_4731_cast_fp16 = mul(x = position18_to_fp16, y = _inversed_4731_y_0_to_fp16)[name = tensor("_inversed_4731_cast_fp16")]; + tensor var_4732_cast_fp16 = floor(x = _inversed_4731_cast_fp16)[name = tensor("op_4732_cast_fp16")]; + tensor var_4733_to_fp16 = const()[name = tensor("op_4733_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4734_cast_fp16 = mul(x = var_4732_cast_fp16, y = var_4733_to_fp16)[name = tensor("op_4734_cast_fp16")]; + tensor write_idx_37_cast_fp16 = sub(x = position18_to_fp16, y = var_4734_cast_fp16)[name = tensor("write_idx_37_cast_fp16")]; + tensor var_4737_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_37_cast_fp16)[name = tensor("op_4737_cast_fp16")]; + tensor var_4747 = const()[name = tensor("op_4747"), val = tensor([1, 512, 1, 1])]; + tensor onehot_73_to_fp16_dtype_0 = const()[name = tensor("onehot_73_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_4737_cast_fp16_to_fp16 = cast(dtype = onehot_73_to_fp16_dtype_0, x = var_4737_cast_fp16)[name = tensor("cast_432")]; + tensor onehot_75_cast_fp16 = reshape(shape = var_4747, x = var_4737_cast_fp16_to_fp16)[name = tensor("onehot_75_cast_fp16")]; + tensor var_4749_to_fp16 = const()[name = tensor("op_4749_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4751_cast_fp16 = sub(x = var_4749_to_fp16, y = onehot_75_cast_fp16)[name = tensor("op_4751_cast_fp16")]; + tensor k_cache18_to_fp16_dtype_0 = const()[name = tensor("k_cache18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache18_to_fp16 = cast(dtype = k_cache18_to_fp16_dtype_0, x = k_cache18)[name = tensor("cast_431")]; + tensor var_4752_cast_fp16 = mul(x = k_cache18_to_fp16, y = var_4751_cast_fp16)[name = tensor("op_4752_cast_fp16")]; + tensor var_4757 = const()[name = tensor("op_4757"), val = tensor([1, 1, 16, 64])]; + tensor var_4758_cast_fp16 = reshape(shape = var_4757, x = out_75_cast_fp16)[name = tensor("op_4758_cast_fp16")]; + tensor var_4759_cast_fp16 = mul(x = var_4758_cast_fp16, y = onehot_75_cast_fp16)[name = tensor("op_4759_cast_fp16")]; + tensor new_k_cache_37_cast_fp16 = add(x = var_4752_cast_fp16, y = var_4759_cast_fp16)[name = tensor("new_k_cache_37_cast_fp16")]; + tensor new_k_cache_37_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_37_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache18_to_fp16_dtype_0 = const()[name = tensor("v_cache18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache18_to_fp16 = cast(dtype = v_cache18_to_fp16_dtype_0, x = v_cache18)[name = tensor("cast_429")]; + tensor var_4765_cast_fp16 = mul(x = v_cache18_to_fp16, y = var_4751_cast_fp16)[name = tensor("op_4765_cast_fp16")]; + tensor var_4770 = const()[name = tensor("op_4770"), val = tensor([1, 1, 16, 64])]; + tensor var_4771_cast_fp16 = reshape(shape = var_4770, x = v_37_cast_fp16)[name = tensor("op_4771_cast_fp16")]; + tensor var_4772_cast_fp16 = mul(x = var_4771_cast_fp16, y = onehot_75_cast_fp16)[name = tensor("op_4772_cast_fp16")]; + tensor new_v_cache_37_cast_fp16 = add(x = var_4765_cast_fp16, y = var_4772_cast_fp16)[name = tensor("new_v_cache_37_cast_fp16")]; + tensor new_v_cache_37_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_37_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_18_perm_0 = const()[name = tensor("transpose_18_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_4785 = const()[name = tensor("op_4785"), val = tensor([0, 2, 1, 3])]; + tensor var_4791 = const()[name = tensor("op_4791"), val = tensor([1, 16, 1, 64])]; + tensor q4_37_cast_fp16 = reshape(shape = var_4791, x = out_73_cast_fp16)[name = tensor("q4_37_cast_fp16")]; + tensor var_4796_transpose_x_0 = const()[name = tensor("op_4796_transpose_x_0"), val = tensor(false)]; + tensor var_4796_transpose_y_0 = const()[name = tensor("op_4796_transpose_y_0"), val = tensor(false)]; + tensor transpose_18_cast_fp16 = transpose(perm = transpose_18_perm_0, x = new_k_cache_37_cast_fp16)[name = tensor("transpose_41")]; + tensor var_4796_cast_fp16 = matmul(transpose_x = var_4796_transpose_x_0, transpose_y = var_4796_transpose_y_0, x = q4_37_cast_fp16, y = transpose_18_cast_fp16)[name = tensor("op_4796_cast_fp16")]; + tensor var_4797_to_fp16 = const()[name = tensor("op_4797_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_73_cast_fp16 = mul(x = var_4796_cast_fp16, y = var_4797_to_fp16)[name = tensor("scores_73_cast_fp16")]; + tensor var_4801 = const()[name = tensor("op_4801"), val = tensor([1, 1])]; + tensor var_4802_cast_fp16 = reshape(shape = var_4801, x = position18_to_fp16)[name = tensor("op_4802_cast_fp16")]; + tensor var_4803_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_4802_cast_fp16)[name = tensor("op_4803_cast_fp16")]; + tensor mask_37_to_fp16_dtype_0 = const()[name = tensor("mask_37_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_4810_to_fp16 = const()[name = tensor("op_4810_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4803_cast_fp16_to_fp16 = cast(dtype = mask_37_to_fp16_dtype_0, x = var_4803_cast_fp16)[name = tensor("cast_427")]; + tensor var_4811_cast_fp16 = sub(x = var_4803_cast_fp16_to_fp16, y = var_4810_to_fp16)[name = tensor("op_4811_cast_fp16")]; + tensor var_4816 = const()[name = tensor("op_4816"), val = tensor([1, 1, 1, 512])]; + tensor var_4817_cast_fp16 = reshape(shape = var_4816, x = var_4811_cast_fp16)[name = tensor("op_4817_cast_fp16")]; + tensor var_4818_to_fp16 = const()[name = tensor("op_4818_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_4819_cast_fp16 = mul(x = var_4817_cast_fp16, y = var_4818_to_fp16)[name = tensor("op_4819_cast_fp16")]; + tensor scores_75_cast_fp16 = add(x = scores_73_cast_fp16, y = var_4819_cast_fp16)[name = tensor("scores_75_cast_fp16")]; + tensor var_4822 = const()[name = tensor("op_4822"), val = tensor(-1)]; + tensor weights_37_cast_fp16 = softmax(axis = var_4822, x = scores_75_cast_fp16)[name = tensor("weights_37_cast_fp16")]; + tensor attn_37_transpose_x_0 = const()[name = tensor("attn_37_transpose_x_0"), val = tensor(false)]; + tensor attn_37_transpose_y_0 = const()[name = tensor("attn_37_transpose_y_0"), val = tensor(false)]; + tensor values_37_cast_fp16 = transpose(perm = var_4785, x = new_v_cache_37_cast_fp16)[name = tensor("transpose_40")]; + tensor attn_37_cast_fp16 = matmul(transpose_x = attn_37_transpose_x_0, transpose_y = attn_37_transpose_y_0, x = weights_37_cast_fp16, y = values_37_cast_fp16)[name = tensor("attn_37_cast_fp16")]; + tensor var_4830 = const()[name = tensor("op_4830"), val = tensor([0, 2, 1, 3])]; + tensor var_4835 = const()[name = tensor("op_4835"), val = tensor([1, 1, 1024])]; + tensor var_4831_cast_fp16 = transpose(perm = var_4830, x = attn_37_cast_fp16)[name = tensor("transpose_39")]; + tensor input_223_cast_fp16 = reshape(shape = var_4835, x = var_4831_cast_fp16)[name = tensor("input_223_cast_fp16")]; + tensor attn18_out_proj_weight_to_fp16 = const()[name = tensor("attn18_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(459520704)))]; + tensor linear_74_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn18_out_proj_weight_to_fp16, x = input_223_cast_fp16)[name = tensor("linear_74_cast_fp16")]; + tensor var_4841_to_fp16 = const()[name = tensor("op_4841_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4842_cast_fp16 = add(x = position18_to_fp16, y = var_4841_to_fp16)[name = tensor("op_4842_cast_fp16")]; + tensor var_4842_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4842_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_225_cast_fp16 = add(x = input_219_cast_fp16, y = linear_74_cast_fp16)[name = tensor("input_225_cast_fp16")]; + tensor input_227_axes_0 = const()[name = tensor("input_227_axes_0"), val = tensor([-1])]; + tensor norm18_2_weight_to_fp16 = const()[name = tensor("norm18_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(461617920)))]; + tensor norm18_2_bias_to_fp16 = const()[name = tensor("norm18_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(461620032)))]; + tensor var_4846_to_fp16 = const()[name = tensor("op_4846_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_227_cast_fp16 = layer_norm(axes = input_227_axes_0, beta = norm18_2_bias_to_fp16, epsilon = var_4846_to_fp16, gamma = norm18_2_weight_to_fp16, x = input_225_cast_fp16)[name = tensor("input_227_cast_fp16")]; + tensor linear18_1_weight_to_fp16 = const()[name = tensor("linear18_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(461622144)))]; + tensor linear_75_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear18_1_weight_to_fp16, x = input_227_cast_fp16)[name = tensor("linear_75_cast_fp16")]; + tensor input_229_mode_0 = const()[name = tensor("input_229_mode_0"), val = tensor("EXACT")]; + tensor input_229_cast_fp16 = gelu(mode = input_229_mode_0, x = linear_75_cast_fp16)[name = tensor("input_229_cast_fp16")]; + tensor linear18_2_weight_to_fp16 = const()[name = tensor("linear18_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(470010816)))]; + tensor linear_76_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear18_2_weight_to_fp16, x = input_229_cast_fp16)[name = tensor("linear_76_cast_fp16")]; + tensor input_231_cast_fp16 = add(x = input_225_cast_fp16, y = linear_76_cast_fp16)[name = tensor("input_231_cast_fp16")]; + tensor input_233_axes_0 = const()[name = tensor("input_233_axes_0"), val = tensor([-1])]; + tensor norm19_1_weight_to_fp16 = const()[name = tensor("norm19_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(478399488)))]; + tensor norm19_1_bias_to_fp16 = const()[name = tensor("norm19_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(478401600)))]; + tensor var_4863_to_fp16 = const()[name = tensor("op_4863_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_233_cast_fp16 = layer_norm(axes = input_233_axes_0, beta = norm19_1_bias_to_fp16, epsilon = var_4863_to_fp16, gamma = norm19_1_weight_to_fp16, x = input_231_cast_fp16)[name = tensor("input_233_cast_fp16")]; + tensor attn19_in_proj_weight_to_fp16 = const()[name = tensor("attn19_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(478403712)))]; + tensor linear_77_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn19_in_proj_weight_to_fp16, x = input_233_cast_fp16)[name = tensor("linear_77_cast_fp16")]; + tensor var_4876 = const()[name = tensor("op_4876"), val = tensor([1, 3, 16, 64])]; + tensor qkv_39_cast_fp16 = reshape(shape = var_4876, x = linear_77_cast_fp16)[name = tensor("qkv_39_cast_fp16")]; + tensor x_77_begin_0 = const()[name = tensor("x_77_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_77_end_0 = const()[name = tensor("x_77_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_77_end_mask_0 = const()[name = tensor("x_77_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_77_squeeze_mask_0 = const()[name = tensor("x_77_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_77_cast_fp16 = slice_by_index(begin = x_77_begin_0, end = x_77_end_0, end_mask = x_77_end_mask_0, squeeze_mask = x_77_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("x_77_cast_fp16")]; + tensor x_79_begin_0 = const()[name = tensor("x_79_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_79_end_0 = const()[name = tensor("x_79_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_79_end_mask_0 = const()[name = tensor("x_79_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_79_squeeze_mask_0 = const()[name = tensor("x_79_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_79_cast_fp16 = slice_by_index(begin = x_79_begin_0, end = x_79_end_0, end_mask = x_79_end_mask_0, squeeze_mask = x_79_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("x_79_cast_fp16")]; + tensor v_39_begin_0 = const()[name = tensor("v_39_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_39_end_0 = const()[name = tensor("v_39_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_39_end_mask_0 = const()[name = tensor("v_39_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_39_squeeze_mask_0 = const()[name = tensor("v_39_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_39_cast_fp16 = slice_by_index(begin = v_39_begin_0, end = v_39_end_0, end_mask = v_39_end_mask_0, squeeze_mask = v_39_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("v_39_cast_fp16")]; + tensor var_4905 = const()[name = tensor("op_4905"), val = tensor([1, 1, 1])]; + tensor position19_to_fp16_dtype_0 = const()[name = tensor("position19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position19_to_fp16 = cast(dtype = position19_to_fp16_dtype_0, x = position19)[name = tensor("cast_425")]; + tensor var_4906_cast_fp16 = reshape(shape = var_4905, x = position19_to_fp16)[name = tensor("op_4906_cast_fp16")]; + tensor angles_77_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_4906_cast_fp16)[name = tensor("angles_77_cast_fp16")]; + tensor rotr_77_cast_fp16 = cos(x = angles_77_cast_fp16)[name = tensor("rotr_77_cast_fp16")]; + tensor roti_77_cast_fp16 = sin(x = angles_77_cast_fp16)[name = tensor("roti_77_cast_fp16")]; + tensor var_4914 = const()[name = tensor("op_4914"), val = tensor([1, 16, 32, 2])]; + tensor pairs_77_cast_fp16 = reshape(shape = var_4914, x = x_77_cast_fp16)[name = tensor("pairs_77_cast_fp16")]; + tensor xr_77_begin_0 = const()[name = tensor("xr_77_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_77_end_0 = const()[name = tensor("xr_77_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_77_end_mask_0 = const()[name = tensor("xr_77_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_77_squeeze_mask_0 = const()[name = tensor("xr_77_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_77_cast_fp16 = slice_by_index(begin = xr_77_begin_0, end = xr_77_end_0, end_mask = xr_77_end_mask_0, squeeze_mask = xr_77_squeeze_mask_0, x = pairs_77_cast_fp16)[name = tensor("xr_77_cast_fp16")]; + tensor xi_77_begin_0 = const()[name = tensor("xi_77_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_77_end_0 = const()[name = tensor("xi_77_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_77_end_mask_0 = const()[name = tensor("xi_77_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_77_squeeze_mask_0 = const()[name = tensor("xi_77_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_77_cast_fp16 = slice_by_index(begin = xi_77_begin_0, end = xi_77_end_0, end_mask = xi_77_end_mask_0, squeeze_mask = xi_77_squeeze_mask_0, x = pairs_77_cast_fp16)[name = tensor("xi_77_cast_fp16")]; + tensor var_4922_cast_fp16 = mul(x = xr_77_cast_fp16, y = rotr_77_cast_fp16)[name = tensor("op_4922_cast_fp16")]; + tensor var_4923_cast_fp16 = mul(x = xi_77_cast_fp16, y = roti_77_cast_fp16)[name = tensor("op_4923_cast_fp16")]; + tensor xor__77_cast_fp16 = sub(x = var_4922_cast_fp16, y = var_4923_cast_fp16)[name = tensor("xor__77_cast_fp16")]; + tensor var_4926_cast_fp16 = mul(x = xr_77_cast_fp16, y = roti_77_cast_fp16)[name = tensor("op_4926_cast_fp16")]; + tensor var_4927_cast_fp16 = mul(x = xi_77_cast_fp16, y = rotr_77_cast_fp16)[name = tensor("op_4927_cast_fp16")]; + tensor xoi_77_cast_fp16 = add(x = var_4926_cast_fp16, y = var_4927_cast_fp16)[name = tensor("xoi_77_cast_fp16")]; + tensor out_77_axis_0 = const()[name = tensor("out_77_axis_0"), val = tensor(-1)]; + tensor out_77_cast_fp16 = stack(axis = out_77_axis_0, values = (xor__77_cast_fp16, xoi_77_cast_fp16))[name = tensor("out_77_cast_fp16")]; + tensor var_4950 = const()[name = tensor("op_4950"), val = tensor([1, 16, 32, 2])]; + tensor pairs_79_cast_fp16 = reshape(shape = var_4950, x = x_79_cast_fp16)[name = tensor("pairs_79_cast_fp16")]; + tensor xr_79_begin_0 = const()[name = tensor("xr_79_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_79_end_0 = const()[name = tensor("xr_79_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_79_end_mask_0 = const()[name = tensor("xr_79_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_79_squeeze_mask_0 = const()[name = tensor("xr_79_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_79_cast_fp16 = slice_by_index(begin = xr_79_begin_0, end = xr_79_end_0, end_mask = xr_79_end_mask_0, squeeze_mask = xr_79_squeeze_mask_0, x = pairs_79_cast_fp16)[name = tensor("xr_79_cast_fp16")]; + tensor xi_79_begin_0 = const()[name = tensor("xi_79_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_79_end_0 = const()[name = tensor("xi_79_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_79_end_mask_0 = const()[name = tensor("xi_79_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_79_squeeze_mask_0 = const()[name = tensor("xi_79_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_79_cast_fp16 = slice_by_index(begin = xi_79_begin_0, end = xi_79_end_0, end_mask = xi_79_end_mask_0, squeeze_mask = xi_79_squeeze_mask_0, x = pairs_79_cast_fp16)[name = tensor("xi_79_cast_fp16")]; + tensor var_4958_cast_fp16 = mul(x = xr_79_cast_fp16, y = rotr_77_cast_fp16)[name = tensor("op_4958_cast_fp16")]; + tensor var_4959_cast_fp16 = mul(x = xi_79_cast_fp16, y = roti_77_cast_fp16)[name = tensor("op_4959_cast_fp16")]; + tensor xor__79_cast_fp16 = sub(x = var_4958_cast_fp16, y = var_4959_cast_fp16)[name = tensor("xor__79_cast_fp16")]; + tensor var_4962_cast_fp16 = mul(x = xr_79_cast_fp16, y = roti_77_cast_fp16)[name = tensor("op_4962_cast_fp16")]; + tensor var_4963_cast_fp16 = mul(x = xi_79_cast_fp16, y = rotr_77_cast_fp16)[name = tensor("op_4963_cast_fp16")]; + tensor xoi_79_cast_fp16 = add(x = var_4962_cast_fp16, y = var_4963_cast_fp16)[name = tensor("xoi_79_cast_fp16")]; + tensor out_79_axis_0 = const()[name = tensor("out_79_axis_0"), val = tensor(-1)]; + tensor out_79_cast_fp16 = stack(axis = out_79_axis_0, values = (xor__79_cast_fp16, xoi_79_cast_fp16))[name = tensor("out_79_cast_fp16")]; + tensor _inversed_4975_y_0_to_fp16 = const()[name = tensor("_inversed_4975_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_4975_cast_fp16 = mul(x = position19_to_fp16, y = _inversed_4975_y_0_to_fp16)[name = tensor("_inversed_4975_cast_fp16")]; + tensor var_4976_cast_fp16 = floor(x = _inversed_4975_cast_fp16)[name = tensor("op_4976_cast_fp16")]; + tensor var_4977_to_fp16 = const()[name = tensor("op_4977_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4978_cast_fp16 = mul(x = var_4976_cast_fp16, y = var_4977_to_fp16)[name = tensor("op_4978_cast_fp16")]; + tensor write_idx_39_cast_fp16 = sub(x = position19_to_fp16, y = var_4978_cast_fp16)[name = tensor("write_idx_39_cast_fp16")]; + tensor var_4981_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_39_cast_fp16)[name = tensor("op_4981_cast_fp16")]; + tensor var_4991 = const()[name = tensor("op_4991"), val = tensor([1, 512, 1, 1])]; + tensor onehot_77_to_fp16_dtype_0 = const()[name = tensor("onehot_77_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_4981_cast_fp16_to_fp16 = cast(dtype = onehot_77_to_fp16_dtype_0, x = var_4981_cast_fp16)[name = tensor("cast_424")]; + tensor onehot_79_cast_fp16 = reshape(shape = var_4991, x = var_4981_cast_fp16_to_fp16)[name = tensor("onehot_79_cast_fp16")]; + tensor var_4993_to_fp16 = const()[name = tensor("op_4993_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4995_cast_fp16 = sub(x = var_4993_to_fp16, y = onehot_79_cast_fp16)[name = tensor("op_4995_cast_fp16")]; + tensor k_cache19_to_fp16_dtype_0 = const()[name = tensor("k_cache19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache19_to_fp16 = cast(dtype = k_cache19_to_fp16_dtype_0, x = k_cache19)[name = tensor("cast_423")]; + tensor var_4996_cast_fp16 = mul(x = k_cache19_to_fp16, y = var_4995_cast_fp16)[name = tensor("op_4996_cast_fp16")]; + tensor var_5001 = const()[name = tensor("op_5001"), val = tensor([1, 1, 16, 64])]; + tensor var_5002_cast_fp16 = reshape(shape = var_5001, x = out_79_cast_fp16)[name = tensor("op_5002_cast_fp16")]; + tensor var_5003_cast_fp16 = mul(x = var_5002_cast_fp16, y = onehot_79_cast_fp16)[name = tensor("op_5003_cast_fp16")]; + tensor new_k_cache_39_cast_fp16 = add(x = var_4996_cast_fp16, y = var_5003_cast_fp16)[name = tensor("new_k_cache_39_cast_fp16")]; + tensor new_k_cache_39_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_39_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache19_to_fp16_dtype_0 = const()[name = tensor("v_cache19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache19_to_fp16 = cast(dtype = v_cache19_to_fp16_dtype_0, x = v_cache19)[name = tensor("cast_421")]; + tensor var_5009_cast_fp16 = mul(x = v_cache19_to_fp16, y = var_4995_cast_fp16)[name = tensor("op_5009_cast_fp16")]; + tensor var_5014 = const()[name = tensor("op_5014"), val = tensor([1, 1, 16, 64])]; + tensor var_5015_cast_fp16 = reshape(shape = var_5014, x = v_39_cast_fp16)[name = tensor("op_5015_cast_fp16")]; + tensor var_5016_cast_fp16 = mul(x = var_5015_cast_fp16, y = onehot_79_cast_fp16)[name = tensor("op_5016_cast_fp16")]; + tensor new_v_cache_39_cast_fp16 = add(x = var_5009_cast_fp16, y = var_5016_cast_fp16)[name = tensor("new_v_cache_39_cast_fp16")]; + tensor new_v_cache_39_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_39_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_19_perm_0 = const()[name = tensor("transpose_19_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_5029 = const()[name = tensor("op_5029"), val = tensor([0, 2, 1, 3])]; + tensor var_5035 = const()[name = tensor("op_5035"), val = tensor([1, 16, 1, 64])]; + tensor q4_39_cast_fp16 = reshape(shape = var_5035, x = out_77_cast_fp16)[name = tensor("q4_39_cast_fp16")]; + tensor var_5040_transpose_x_0 = const()[name = tensor("op_5040_transpose_x_0"), val = tensor(false)]; + tensor var_5040_transpose_y_0 = const()[name = tensor("op_5040_transpose_y_0"), val = tensor(false)]; + tensor transpose_19_cast_fp16 = transpose(perm = transpose_19_perm_0, x = new_k_cache_39_cast_fp16)[name = tensor("transpose_38")]; + tensor var_5040_cast_fp16 = matmul(transpose_x = var_5040_transpose_x_0, transpose_y = var_5040_transpose_y_0, x = q4_39_cast_fp16, y = transpose_19_cast_fp16)[name = tensor("op_5040_cast_fp16")]; + tensor var_5041_to_fp16 = const()[name = tensor("op_5041_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_77_cast_fp16 = mul(x = var_5040_cast_fp16, y = var_5041_to_fp16)[name = tensor("scores_77_cast_fp16")]; + tensor var_5045 = const()[name = tensor("op_5045"), val = tensor([1, 1])]; + tensor var_5046_cast_fp16 = reshape(shape = var_5045, x = position19_to_fp16)[name = tensor("op_5046_cast_fp16")]; + tensor var_5047_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_5046_cast_fp16)[name = tensor("op_5047_cast_fp16")]; + tensor mask_39_to_fp16_dtype_0 = const()[name = tensor("mask_39_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_5054_to_fp16 = const()[name = tensor("op_5054_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5047_cast_fp16_to_fp16 = cast(dtype = mask_39_to_fp16_dtype_0, x = var_5047_cast_fp16)[name = tensor("cast_419")]; + tensor var_5055_cast_fp16 = sub(x = var_5047_cast_fp16_to_fp16, y = var_5054_to_fp16)[name = tensor("op_5055_cast_fp16")]; + tensor var_5060 = const()[name = tensor("op_5060"), val = tensor([1, 1, 1, 512])]; + tensor var_5061_cast_fp16 = reshape(shape = var_5060, x = var_5055_cast_fp16)[name = tensor("op_5061_cast_fp16")]; + tensor var_5062_to_fp16 = const()[name = tensor("op_5062_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_5063_cast_fp16 = mul(x = var_5061_cast_fp16, y = var_5062_to_fp16)[name = tensor("op_5063_cast_fp16")]; + tensor scores_79_cast_fp16 = add(x = scores_77_cast_fp16, y = var_5063_cast_fp16)[name = tensor("scores_79_cast_fp16")]; + tensor var_5066 = const()[name = tensor("op_5066"), val = tensor(-1)]; + tensor weights_39_cast_fp16 = softmax(axis = var_5066, x = scores_79_cast_fp16)[name = tensor("weights_39_cast_fp16")]; + tensor attn_39_transpose_x_0 = const()[name = tensor("attn_39_transpose_x_0"), val = tensor(false)]; + tensor attn_39_transpose_y_0 = const()[name = tensor("attn_39_transpose_y_0"), val = tensor(false)]; + tensor values_39_cast_fp16 = transpose(perm = var_5029, x = new_v_cache_39_cast_fp16)[name = tensor("transpose_37")]; + tensor attn_39_cast_fp16 = matmul(transpose_x = attn_39_transpose_x_0, transpose_y = attn_39_transpose_y_0, x = weights_39_cast_fp16, y = values_39_cast_fp16)[name = tensor("attn_39_cast_fp16")]; + tensor var_5074 = const()[name = tensor("op_5074"), val = tensor([0, 2, 1, 3])]; + tensor var_5079 = const()[name = tensor("op_5079"), val = tensor([1, 1, 1024])]; + tensor var_5075_cast_fp16 = transpose(perm = var_5074, x = attn_39_cast_fp16)[name = tensor("transpose_36")]; + tensor input_235_cast_fp16 = reshape(shape = var_5079, x = var_5075_cast_fp16)[name = tensor("input_235_cast_fp16")]; + tensor attn19_out_proj_weight_to_fp16 = const()[name = tensor("attn19_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(484695232)))]; + tensor linear_78_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn19_out_proj_weight_to_fp16, x = input_235_cast_fp16)[name = tensor("linear_78_cast_fp16")]; + tensor var_5085_to_fp16 = const()[name = tensor("op_5085_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5086_cast_fp16 = add(x = position19_to_fp16, y = var_5085_to_fp16)[name = tensor("op_5086_cast_fp16")]; + tensor var_5086_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5086_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_237_cast_fp16 = add(x = input_231_cast_fp16, y = linear_78_cast_fp16)[name = tensor("input_237_cast_fp16")]; + tensor input_239_axes_0 = const()[name = tensor("input_239_axes_0"), val = tensor([-1])]; + tensor norm19_2_weight_to_fp16 = const()[name = tensor("norm19_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486792448)))]; + tensor norm19_2_bias_to_fp16 = const()[name = tensor("norm19_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486794560)))]; + tensor var_5090_to_fp16 = const()[name = tensor("op_5090_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_239_cast_fp16 = layer_norm(axes = input_239_axes_0, beta = norm19_2_bias_to_fp16, epsilon = var_5090_to_fp16, gamma = norm19_2_weight_to_fp16, x = input_237_cast_fp16)[name = tensor("input_239_cast_fp16")]; + tensor linear19_1_weight_to_fp16 = const()[name = tensor("linear19_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(486796672)))]; + tensor linear_79_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear19_1_weight_to_fp16, x = input_239_cast_fp16)[name = tensor("linear_79_cast_fp16")]; + tensor input_241_mode_0 = const()[name = tensor("input_241_mode_0"), val = tensor("EXACT")]; + tensor input_241_cast_fp16 = gelu(mode = input_241_mode_0, x = linear_79_cast_fp16)[name = tensor("input_241_cast_fp16")]; + tensor linear19_2_weight_to_fp16 = const()[name = tensor("linear19_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(495185344)))]; + tensor linear_80_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear19_2_weight_to_fp16, x = input_241_cast_fp16)[name = tensor("linear_80_cast_fp16")]; + tensor input_243_cast_fp16 = add(x = input_237_cast_fp16, y = linear_80_cast_fp16)[name = tensor("input_243_cast_fp16")]; + tensor input_245_axes_0 = const()[name = tensor("input_245_axes_0"), val = tensor([-1])]; + tensor norm20_1_weight_to_fp16 = const()[name = tensor("norm20_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(503574016)))]; + tensor norm20_1_bias_to_fp16 = const()[name = tensor("norm20_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(503576128)))]; + tensor var_5107_to_fp16 = const()[name = tensor("op_5107_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_245_cast_fp16 = layer_norm(axes = input_245_axes_0, beta = norm20_1_bias_to_fp16, epsilon = var_5107_to_fp16, gamma = norm20_1_weight_to_fp16, x = input_243_cast_fp16)[name = tensor("input_245_cast_fp16")]; + tensor attn20_in_proj_weight_to_fp16 = const()[name = tensor("attn20_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(503578240)))]; + tensor linear_81_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn20_in_proj_weight_to_fp16, x = input_245_cast_fp16)[name = tensor("linear_81_cast_fp16")]; + tensor var_5120 = const()[name = tensor("op_5120"), val = tensor([1, 3, 16, 64])]; + tensor qkv_41_cast_fp16 = reshape(shape = var_5120, x = linear_81_cast_fp16)[name = tensor("qkv_41_cast_fp16")]; + tensor x_81_begin_0 = const()[name = tensor("x_81_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_81_end_0 = const()[name = tensor("x_81_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_81_end_mask_0 = const()[name = tensor("x_81_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_81_squeeze_mask_0 = const()[name = tensor("x_81_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_81_cast_fp16 = slice_by_index(begin = x_81_begin_0, end = x_81_end_0, end_mask = x_81_end_mask_0, squeeze_mask = x_81_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("x_81_cast_fp16")]; + tensor x_83_begin_0 = const()[name = tensor("x_83_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_83_end_0 = const()[name = tensor("x_83_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_83_end_mask_0 = const()[name = tensor("x_83_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_83_squeeze_mask_0 = const()[name = tensor("x_83_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_83_cast_fp16 = slice_by_index(begin = x_83_begin_0, end = x_83_end_0, end_mask = x_83_end_mask_0, squeeze_mask = x_83_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("x_83_cast_fp16")]; + tensor v_41_begin_0 = const()[name = tensor("v_41_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_41_end_0 = const()[name = tensor("v_41_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_41_end_mask_0 = const()[name = tensor("v_41_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_41_squeeze_mask_0 = const()[name = tensor("v_41_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_41_cast_fp16 = slice_by_index(begin = v_41_begin_0, end = v_41_end_0, end_mask = v_41_end_mask_0, squeeze_mask = v_41_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("v_41_cast_fp16")]; + tensor var_5149 = const()[name = tensor("op_5149"), val = tensor([1, 1, 1])]; + tensor position20_to_fp16_dtype_0 = const()[name = tensor("position20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position20_to_fp16 = cast(dtype = position20_to_fp16_dtype_0, x = position20)[name = tensor("cast_417")]; + tensor var_5150_cast_fp16 = reshape(shape = var_5149, x = position20_to_fp16)[name = tensor("op_5150_cast_fp16")]; + tensor angles_81_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_5150_cast_fp16)[name = tensor("angles_81_cast_fp16")]; + tensor rotr_81_cast_fp16 = cos(x = angles_81_cast_fp16)[name = tensor("rotr_81_cast_fp16")]; + tensor roti_81_cast_fp16 = sin(x = angles_81_cast_fp16)[name = tensor("roti_81_cast_fp16")]; + tensor var_5158 = const()[name = tensor("op_5158"), val = tensor([1, 16, 32, 2])]; + tensor pairs_81_cast_fp16 = reshape(shape = var_5158, x = x_81_cast_fp16)[name = tensor("pairs_81_cast_fp16")]; + tensor xr_81_begin_0 = const()[name = tensor("xr_81_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_81_end_0 = const()[name = tensor("xr_81_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_81_end_mask_0 = const()[name = tensor("xr_81_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_81_squeeze_mask_0 = const()[name = tensor("xr_81_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_81_cast_fp16 = slice_by_index(begin = xr_81_begin_0, end = xr_81_end_0, end_mask = xr_81_end_mask_0, squeeze_mask = xr_81_squeeze_mask_0, x = pairs_81_cast_fp16)[name = tensor("xr_81_cast_fp16")]; + tensor xi_81_begin_0 = const()[name = tensor("xi_81_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_81_end_0 = const()[name = tensor("xi_81_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_81_end_mask_0 = const()[name = tensor("xi_81_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_81_squeeze_mask_0 = const()[name = tensor("xi_81_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_81_cast_fp16 = slice_by_index(begin = xi_81_begin_0, end = xi_81_end_0, end_mask = xi_81_end_mask_0, squeeze_mask = xi_81_squeeze_mask_0, x = pairs_81_cast_fp16)[name = tensor("xi_81_cast_fp16")]; + tensor var_5166_cast_fp16 = mul(x = xr_81_cast_fp16, y = rotr_81_cast_fp16)[name = tensor("op_5166_cast_fp16")]; + tensor var_5167_cast_fp16 = mul(x = xi_81_cast_fp16, y = roti_81_cast_fp16)[name = tensor("op_5167_cast_fp16")]; + tensor xor__81_cast_fp16 = sub(x = var_5166_cast_fp16, y = var_5167_cast_fp16)[name = tensor("xor__81_cast_fp16")]; + tensor var_5170_cast_fp16 = mul(x = xr_81_cast_fp16, y = roti_81_cast_fp16)[name = tensor("op_5170_cast_fp16")]; + tensor var_5171_cast_fp16 = mul(x = xi_81_cast_fp16, y = rotr_81_cast_fp16)[name = tensor("op_5171_cast_fp16")]; + tensor xoi_81_cast_fp16 = add(x = var_5170_cast_fp16, y = var_5171_cast_fp16)[name = tensor("xoi_81_cast_fp16")]; + tensor out_81_axis_0 = const()[name = tensor("out_81_axis_0"), val = tensor(-1)]; + tensor out_81_cast_fp16 = stack(axis = out_81_axis_0, values = (xor__81_cast_fp16, xoi_81_cast_fp16))[name = tensor("out_81_cast_fp16")]; + tensor var_5194 = const()[name = tensor("op_5194"), val = tensor([1, 16, 32, 2])]; + tensor pairs_83_cast_fp16 = reshape(shape = var_5194, x = x_83_cast_fp16)[name = tensor("pairs_83_cast_fp16")]; + tensor xr_83_begin_0 = const()[name = tensor("xr_83_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_83_end_0 = const()[name = tensor("xr_83_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_83_end_mask_0 = const()[name = tensor("xr_83_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_83_squeeze_mask_0 = const()[name = tensor("xr_83_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_83_cast_fp16 = slice_by_index(begin = xr_83_begin_0, end = xr_83_end_0, end_mask = xr_83_end_mask_0, squeeze_mask = xr_83_squeeze_mask_0, x = pairs_83_cast_fp16)[name = tensor("xr_83_cast_fp16")]; + tensor xi_83_begin_0 = const()[name = tensor("xi_83_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_83_end_0 = const()[name = tensor("xi_83_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_83_end_mask_0 = const()[name = tensor("xi_83_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_83_squeeze_mask_0 = const()[name = tensor("xi_83_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_83_cast_fp16 = slice_by_index(begin = xi_83_begin_0, end = xi_83_end_0, end_mask = xi_83_end_mask_0, squeeze_mask = xi_83_squeeze_mask_0, x = pairs_83_cast_fp16)[name = tensor("xi_83_cast_fp16")]; + tensor var_5202_cast_fp16 = mul(x = xr_83_cast_fp16, y = rotr_81_cast_fp16)[name = tensor("op_5202_cast_fp16")]; + tensor var_5203_cast_fp16 = mul(x = xi_83_cast_fp16, y = roti_81_cast_fp16)[name = tensor("op_5203_cast_fp16")]; + tensor xor__83_cast_fp16 = sub(x = var_5202_cast_fp16, y = var_5203_cast_fp16)[name = tensor("xor__83_cast_fp16")]; + tensor var_5206_cast_fp16 = mul(x = xr_83_cast_fp16, y = roti_81_cast_fp16)[name = tensor("op_5206_cast_fp16")]; + tensor var_5207_cast_fp16 = mul(x = xi_83_cast_fp16, y = rotr_81_cast_fp16)[name = tensor("op_5207_cast_fp16")]; + tensor xoi_83_cast_fp16 = add(x = var_5206_cast_fp16, y = var_5207_cast_fp16)[name = tensor("xoi_83_cast_fp16")]; + tensor out_83_axis_0 = const()[name = tensor("out_83_axis_0"), val = tensor(-1)]; + tensor out_83_cast_fp16 = stack(axis = out_83_axis_0, values = (xor__83_cast_fp16, xoi_83_cast_fp16))[name = tensor("out_83_cast_fp16")]; + tensor _inversed_5219_y_0_to_fp16 = const()[name = tensor("_inversed_5219_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5219_cast_fp16 = mul(x = position20_to_fp16, y = _inversed_5219_y_0_to_fp16)[name = tensor("_inversed_5219_cast_fp16")]; + tensor var_5220_cast_fp16 = floor(x = _inversed_5219_cast_fp16)[name = tensor("op_5220_cast_fp16")]; + tensor var_5221_to_fp16 = const()[name = tensor("op_5221_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5222_cast_fp16 = mul(x = var_5220_cast_fp16, y = var_5221_to_fp16)[name = tensor("op_5222_cast_fp16")]; + tensor write_idx_41_cast_fp16 = sub(x = position20_to_fp16, y = var_5222_cast_fp16)[name = tensor("write_idx_41_cast_fp16")]; + tensor var_5225_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_41_cast_fp16)[name = tensor("op_5225_cast_fp16")]; + tensor var_5235 = const()[name = tensor("op_5235"), val = tensor([1, 512, 1, 1])]; + tensor onehot_81_to_fp16_dtype_0 = const()[name = tensor("onehot_81_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_5225_cast_fp16_to_fp16 = cast(dtype = onehot_81_to_fp16_dtype_0, x = var_5225_cast_fp16)[name = tensor("cast_416")]; + tensor onehot_83_cast_fp16 = reshape(shape = var_5235, x = var_5225_cast_fp16_to_fp16)[name = tensor("onehot_83_cast_fp16")]; + tensor var_5237_to_fp16 = const()[name = tensor("op_5237_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5239_cast_fp16 = sub(x = var_5237_to_fp16, y = onehot_83_cast_fp16)[name = tensor("op_5239_cast_fp16")]; + tensor k_cache20_to_fp16_dtype_0 = const()[name = tensor("k_cache20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache20_to_fp16 = cast(dtype = k_cache20_to_fp16_dtype_0, x = k_cache20)[name = tensor("cast_415")]; + tensor var_5240_cast_fp16 = mul(x = k_cache20_to_fp16, y = var_5239_cast_fp16)[name = tensor("op_5240_cast_fp16")]; + tensor var_5245 = const()[name = tensor("op_5245"), val = tensor([1, 1, 16, 64])]; + tensor var_5246_cast_fp16 = reshape(shape = var_5245, x = out_83_cast_fp16)[name = tensor("op_5246_cast_fp16")]; + tensor var_5247_cast_fp16 = mul(x = var_5246_cast_fp16, y = onehot_83_cast_fp16)[name = tensor("op_5247_cast_fp16")]; + tensor new_k_cache_41_cast_fp16 = add(x = var_5240_cast_fp16, y = var_5247_cast_fp16)[name = tensor("new_k_cache_41_cast_fp16")]; + tensor new_k_cache_41_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_41_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache20_to_fp16_dtype_0 = const()[name = tensor("v_cache20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache20_to_fp16 = cast(dtype = v_cache20_to_fp16_dtype_0, x = v_cache20)[name = tensor("cast_413")]; + tensor var_5253_cast_fp16 = mul(x = v_cache20_to_fp16, y = var_5239_cast_fp16)[name = tensor("op_5253_cast_fp16")]; + tensor var_5258 = const()[name = tensor("op_5258"), val = tensor([1, 1, 16, 64])]; + tensor var_5259_cast_fp16 = reshape(shape = var_5258, x = v_41_cast_fp16)[name = tensor("op_5259_cast_fp16")]; + tensor var_5260_cast_fp16 = mul(x = var_5259_cast_fp16, y = onehot_83_cast_fp16)[name = tensor("op_5260_cast_fp16")]; + tensor new_v_cache_41_cast_fp16 = add(x = var_5253_cast_fp16, y = var_5260_cast_fp16)[name = tensor("new_v_cache_41_cast_fp16")]; + tensor new_v_cache_41_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_41_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_20_perm_0 = const()[name = tensor("transpose_20_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_5273 = const()[name = tensor("op_5273"), val = tensor([0, 2, 1, 3])]; + tensor var_5279 = const()[name = tensor("op_5279"), val = tensor([1, 16, 1, 64])]; + tensor q4_41_cast_fp16 = reshape(shape = var_5279, x = out_81_cast_fp16)[name = tensor("q4_41_cast_fp16")]; + tensor var_5284_transpose_x_0 = const()[name = tensor("op_5284_transpose_x_0"), val = tensor(false)]; + tensor var_5284_transpose_y_0 = const()[name = tensor("op_5284_transpose_y_0"), val = tensor(false)]; + tensor transpose_20_cast_fp16 = transpose(perm = transpose_20_perm_0, x = new_k_cache_41_cast_fp16)[name = tensor("transpose_35")]; + tensor var_5284_cast_fp16 = matmul(transpose_x = var_5284_transpose_x_0, transpose_y = var_5284_transpose_y_0, x = q4_41_cast_fp16, y = transpose_20_cast_fp16)[name = tensor("op_5284_cast_fp16")]; + tensor var_5285_to_fp16 = const()[name = tensor("op_5285_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_81_cast_fp16 = mul(x = var_5284_cast_fp16, y = var_5285_to_fp16)[name = tensor("scores_81_cast_fp16")]; + tensor var_5289 = const()[name = tensor("op_5289"), val = tensor([1, 1])]; + tensor var_5290_cast_fp16 = reshape(shape = var_5289, x = position20_to_fp16)[name = tensor("op_5290_cast_fp16")]; + tensor var_5291_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_5290_cast_fp16)[name = tensor("op_5291_cast_fp16")]; + tensor mask_41_to_fp16_dtype_0 = const()[name = tensor("mask_41_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_5298_to_fp16 = const()[name = tensor("op_5298_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5291_cast_fp16_to_fp16 = cast(dtype = mask_41_to_fp16_dtype_0, x = var_5291_cast_fp16)[name = tensor("cast_411")]; + tensor var_5299_cast_fp16 = sub(x = var_5291_cast_fp16_to_fp16, y = var_5298_to_fp16)[name = tensor("op_5299_cast_fp16")]; + tensor var_5304 = const()[name = tensor("op_5304"), val = tensor([1, 1, 1, 512])]; + tensor var_5305_cast_fp16 = reshape(shape = var_5304, x = var_5299_cast_fp16)[name = tensor("op_5305_cast_fp16")]; + tensor var_5306_to_fp16 = const()[name = tensor("op_5306_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_5307_cast_fp16 = mul(x = var_5305_cast_fp16, y = var_5306_to_fp16)[name = tensor("op_5307_cast_fp16")]; + tensor scores_83_cast_fp16 = add(x = scores_81_cast_fp16, y = var_5307_cast_fp16)[name = tensor("scores_83_cast_fp16")]; + tensor var_5310 = const()[name = tensor("op_5310"), val = tensor(-1)]; + tensor weights_41_cast_fp16 = softmax(axis = var_5310, x = scores_83_cast_fp16)[name = tensor("weights_41_cast_fp16")]; + tensor attn_41_transpose_x_0 = const()[name = tensor("attn_41_transpose_x_0"), val = tensor(false)]; + tensor attn_41_transpose_y_0 = const()[name = tensor("attn_41_transpose_y_0"), val = tensor(false)]; + tensor values_41_cast_fp16 = transpose(perm = var_5273, x = new_v_cache_41_cast_fp16)[name = tensor("transpose_34")]; + tensor attn_41_cast_fp16 = matmul(transpose_x = attn_41_transpose_x_0, transpose_y = attn_41_transpose_y_0, x = weights_41_cast_fp16, y = values_41_cast_fp16)[name = tensor("attn_41_cast_fp16")]; + tensor var_5318 = const()[name = tensor("op_5318"), val = tensor([0, 2, 1, 3])]; + tensor var_5323 = const()[name = tensor("op_5323"), val = tensor([1, 1, 1024])]; + tensor var_5319_cast_fp16 = transpose(perm = var_5318, x = attn_41_cast_fp16)[name = tensor("transpose_33")]; + tensor input_247_cast_fp16 = reshape(shape = var_5323, x = var_5319_cast_fp16)[name = tensor("input_247_cast_fp16")]; + tensor attn20_out_proj_weight_to_fp16 = const()[name = tensor("attn20_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(509869760)))]; + tensor linear_82_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn20_out_proj_weight_to_fp16, x = input_247_cast_fp16)[name = tensor("linear_82_cast_fp16")]; + tensor var_5329_to_fp16 = const()[name = tensor("op_5329_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5330_cast_fp16 = add(x = position20_to_fp16, y = var_5329_to_fp16)[name = tensor("op_5330_cast_fp16")]; + tensor var_5330_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5330_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_249_cast_fp16 = add(x = input_243_cast_fp16, y = linear_82_cast_fp16)[name = tensor("input_249_cast_fp16")]; + tensor input_251_axes_0 = const()[name = tensor("input_251_axes_0"), val = tensor([-1])]; + tensor norm20_2_weight_to_fp16 = const()[name = tensor("norm20_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(511966976)))]; + tensor norm20_2_bias_to_fp16 = const()[name = tensor("norm20_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(511969088)))]; + tensor var_5334_to_fp16 = const()[name = tensor("op_5334_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_251_cast_fp16 = layer_norm(axes = input_251_axes_0, beta = norm20_2_bias_to_fp16, epsilon = var_5334_to_fp16, gamma = norm20_2_weight_to_fp16, x = input_249_cast_fp16)[name = tensor("input_251_cast_fp16")]; + tensor linear20_1_weight_to_fp16 = const()[name = tensor("linear20_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(511971200)))]; + tensor linear_83_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear20_1_weight_to_fp16, x = input_251_cast_fp16)[name = tensor("linear_83_cast_fp16")]; + tensor input_253_mode_0 = const()[name = tensor("input_253_mode_0"), val = tensor("EXACT")]; + tensor input_253_cast_fp16 = gelu(mode = input_253_mode_0, x = linear_83_cast_fp16)[name = tensor("input_253_cast_fp16")]; + tensor linear20_2_weight_to_fp16 = const()[name = tensor("linear20_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(520359872)))]; + tensor linear_84_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear20_2_weight_to_fp16, x = input_253_cast_fp16)[name = tensor("linear_84_cast_fp16")]; + tensor input_255_cast_fp16 = add(x = input_249_cast_fp16, y = linear_84_cast_fp16)[name = tensor("input_255_cast_fp16")]; + tensor input_257_axes_0 = const()[name = tensor("input_257_axes_0"), val = tensor([-1])]; + tensor norm21_1_weight_to_fp16 = const()[name = tensor("norm21_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(528748544)))]; + tensor norm21_1_bias_to_fp16 = const()[name = tensor("norm21_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(528750656)))]; + tensor var_5351_to_fp16 = const()[name = tensor("op_5351_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_257_cast_fp16 = layer_norm(axes = input_257_axes_0, beta = norm21_1_bias_to_fp16, epsilon = var_5351_to_fp16, gamma = norm21_1_weight_to_fp16, x = input_255_cast_fp16)[name = tensor("input_257_cast_fp16")]; + tensor attn21_in_proj_weight_to_fp16 = const()[name = tensor("attn21_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(528752768)))]; + tensor linear_85_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn21_in_proj_weight_to_fp16, x = input_257_cast_fp16)[name = tensor("linear_85_cast_fp16")]; + tensor var_5364 = const()[name = tensor("op_5364"), val = tensor([1, 3, 16, 64])]; + tensor qkv_43_cast_fp16 = reshape(shape = var_5364, x = linear_85_cast_fp16)[name = tensor("qkv_43_cast_fp16")]; + tensor x_85_begin_0 = const()[name = tensor("x_85_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_85_end_0 = const()[name = tensor("x_85_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_85_end_mask_0 = const()[name = tensor("x_85_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_85_squeeze_mask_0 = const()[name = tensor("x_85_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_85_cast_fp16 = slice_by_index(begin = x_85_begin_0, end = x_85_end_0, end_mask = x_85_end_mask_0, squeeze_mask = x_85_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("x_85_cast_fp16")]; + tensor x_87_begin_0 = const()[name = tensor("x_87_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_87_end_0 = const()[name = tensor("x_87_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_87_end_mask_0 = const()[name = tensor("x_87_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_87_squeeze_mask_0 = const()[name = tensor("x_87_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_87_cast_fp16 = slice_by_index(begin = x_87_begin_0, end = x_87_end_0, end_mask = x_87_end_mask_0, squeeze_mask = x_87_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("x_87_cast_fp16")]; + tensor v_43_begin_0 = const()[name = tensor("v_43_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_43_end_0 = const()[name = tensor("v_43_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_43_end_mask_0 = const()[name = tensor("v_43_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_43_squeeze_mask_0 = const()[name = tensor("v_43_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_43_cast_fp16 = slice_by_index(begin = v_43_begin_0, end = v_43_end_0, end_mask = v_43_end_mask_0, squeeze_mask = v_43_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("v_43_cast_fp16")]; + tensor var_5393 = const()[name = tensor("op_5393"), val = tensor([1, 1, 1])]; + tensor position21_to_fp16_dtype_0 = const()[name = tensor("position21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position21_to_fp16 = cast(dtype = position21_to_fp16_dtype_0, x = position21)[name = tensor("cast_409")]; + tensor var_5394_cast_fp16 = reshape(shape = var_5393, x = position21_to_fp16)[name = tensor("op_5394_cast_fp16")]; + tensor angles_85_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_5394_cast_fp16)[name = tensor("angles_85_cast_fp16")]; + tensor rotr_85_cast_fp16 = cos(x = angles_85_cast_fp16)[name = tensor("rotr_85_cast_fp16")]; + tensor roti_85_cast_fp16 = sin(x = angles_85_cast_fp16)[name = tensor("roti_85_cast_fp16")]; + tensor var_5402 = const()[name = tensor("op_5402"), val = tensor([1, 16, 32, 2])]; + tensor pairs_85_cast_fp16 = reshape(shape = var_5402, x = x_85_cast_fp16)[name = tensor("pairs_85_cast_fp16")]; + tensor xr_85_begin_0 = const()[name = tensor("xr_85_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_85_end_0 = const()[name = tensor("xr_85_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_85_end_mask_0 = const()[name = tensor("xr_85_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_85_squeeze_mask_0 = const()[name = tensor("xr_85_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_85_cast_fp16 = slice_by_index(begin = xr_85_begin_0, end = xr_85_end_0, end_mask = xr_85_end_mask_0, squeeze_mask = xr_85_squeeze_mask_0, x = pairs_85_cast_fp16)[name = tensor("xr_85_cast_fp16")]; + tensor xi_85_begin_0 = const()[name = tensor("xi_85_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_85_end_0 = const()[name = tensor("xi_85_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_85_end_mask_0 = const()[name = tensor("xi_85_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_85_squeeze_mask_0 = const()[name = tensor("xi_85_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_85_cast_fp16 = slice_by_index(begin = xi_85_begin_0, end = xi_85_end_0, end_mask = xi_85_end_mask_0, squeeze_mask = xi_85_squeeze_mask_0, x = pairs_85_cast_fp16)[name = tensor("xi_85_cast_fp16")]; + tensor var_5410_cast_fp16 = mul(x = xr_85_cast_fp16, y = rotr_85_cast_fp16)[name = tensor("op_5410_cast_fp16")]; + tensor var_5411_cast_fp16 = mul(x = xi_85_cast_fp16, y = roti_85_cast_fp16)[name = tensor("op_5411_cast_fp16")]; + tensor xor__85_cast_fp16 = sub(x = var_5410_cast_fp16, y = var_5411_cast_fp16)[name = tensor("xor__85_cast_fp16")]; + tensor var_5414_cast_fp16 = mul(x = xr_85_cast_fp16, y = roti_85_cast_fp16)[name = tensor("op_5414_cast_fp16")]; + tensor var_5415_cast_fp16 = mul(x = xi_85_cast_fp16, y = rotr_85_cast_fp16)[name = tensor("op_5415_cast_fp16")]; + tensor xoi_85_cast_fp16 = add(x = var_5414_cast_fp16, y = var_5415_cast_fp16)[name = tensor("xoi_85_cast_fp16")]; + tensor out_85_axis_0 = const()[name = tensor("out_85_axis_0"), val = tensor(-1)]; + tensor out_85_cast_fp16 = stack(axis = out_85_axis_0, values = (xor__85_cast_fp16, xoi_85_cast_fp16))[name = tensor("out_85_cast_fp16")]; + tensor var_5438 = const()[name = tensor("op_5438"), val = tensor([1, 16, 32, 2])]; + tensor pairs_87_cast_fp16 = reshape(shape = var_5438, x = x_87_cast_fp16)[name = tensor("pairs_87_cast_fp16")]; + tensor xr_87_begin_0 = const()[name = tensor("xr_87_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_87_end_0 = const()[name = tensor("xr_87_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_87_end_mask_0 = const()[name = tensor("xr_87_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_87_squeeze_mask_0 = const()[name = tensor("xr_87_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_87_cast_fp16 = slice_by_index(begin = xr_87_begin_0, end = xr_87_end_0, end_mask = xr_87_end_mask_0, squeeze_mask = xr_87_squeeze_mask_0, x = pairs_87_cast_fp16)[name = tensor("xr_87_cast_fp16")]; + tensor xi_87_begin_0 = const()[name = tensor("xi_87_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_87_end_0 = const()[name = tensor("xi_87_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_87_end_mask_0 = const()[name = tensor("xi_87_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_87_squeeze_mask_0 = const()[name = tensor("xi_87_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_87_cast_fp16 = slice_by_index(begin = xi_87_begin_0, end = xi_87_end_0, end_mask = xi_87_end_mask_0, squeeze_mask = xi_87_squeeze_mask_0, x = pairs_87_cast_fp16)[name = tensor("xi_87_cast_fp16")]; + tensor var_5446_cast_fp16 = mul(x = xr_87_cast_fp16, y = rotr_85_cast_fp16)[name = tensor("op_5446_cast_fp16")]; + tensor var_5447_cast_fp16 = mul(x = xi_87_cast_fp16, y = roti_85_cast_fp16)[name = tensor("op_5447_cast_fp16")]; + tensor xor__87_cast_fp16 = sub(x = var_5446_cast_fp16, y = var_5447_cast_fp16)[name = tensor("xor__87_cast_fp16")]; + tensor var_5450_cast_fp16 = mul(x = xr_87_cast_fp16, y = roti_85_cast_fp16)[name = tensor("op_5450_cast_fp16")]; + tensor var_5451_cast_fp16 = mul(x = xi_87_cast_fp16, y = rotr_85_cast_fp16)[name = tensor("op_5451_cast_fp16")]; + tensor xoi_87_cast_fp16 = add(x = var_5450_cast_fp16, y = var_5451_cast_fp16)[name = tensor("xoi_87_cast_fp16")]; + tensor out_87_axis_0 = const()[name = tensor("out_87_axis_0"), val = tensor(-1)]; + tensor out_87_cast_fp16 = stack(axis = out_87_axis_0, values = (xor__87_cast_fp16, xoi_87_cast_fp16))[name = tensor("out_87_cast_fp16")]; + tensor _inversed_5463_y_0_to_fp16 = const()[name = tensor("_inversed_5463_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5463_cast_fp16 = mul(x = position21_to_fp16, y = _inversed_5463_y_0_to_fp16)[name = tensor("_inversed_5463_cast_fp16")]; + tensor var_5464_cast_fp16 = floor(x = _inversed_5463_cast_fp16)[name = tensor("op_5464_cast_fp16")]; + tensor var_5465_to_fp16 = const()[name = tensor("op_5465_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5466_cast_fp16 = mul(x = var_5464_cast_fp16, y = var_5465_to_fp16)[name = tensor("op_5466_cast_fp16")]; + tensor write_idx_43_cast_fp16 = sub(x = position21_to_fp16, y = var_5466_cast_fp16)[name = tensor("write_idx_43_cast_fp16")]; + tensor var_5469_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_43_cast_fp16)[name = tensor("op_5469_cast_fp16")]; + tensor var_5479 = const()[name = tensor("op_5479"), val = tensor([1, 512, 1, 1])]; + tensor onehot_85_to_fp16_dtype_0 = const()[name = tensor("onehot_85_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_5469_cast_fp16_to_fp16 = cast(dtype = onehot_85_to_fp16_dtype_0, x = var_5469_cast_fp16)[name = tensor("cast_408")]; + tensor onehot_87_cast_fp16 = reshape(shape = var_5479, x = var_5469_cast_fp16_to_fp16)[name = tensor("onehot_87_cast_fp16")]; + tensor var_5481_to_fp16 = const()[name = tensor("op_5481_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5483_cast_fp16 = sub(x = var_5481_to_fp16, y = onehot_87_cast_fp16)[name = tensor("op_5483_cast_fp16")]; + tensor k_cache21_to_fp16_dtype_0 = const()[name = tensor("k_cache21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache21_to_fp16 = cast(dtype = k_cache21_to_fp16_dtype_0, x = k_cache21)[name = tensor("cast_407")]; + tensor var_5484_cast_fp16 = mul(x = k_cache21_to_fp16, y = var_5483_cast_fp16)[name = tensor("op_5484_cast_fp16")]; + tensor var_5489 = const()[name = tensor("op_5489"), val = tensor([1, 1, 16, 64])]; + tensor var_5490_cast_fp16 = reshape(shape = var_5489, x = out_87_cast_fp16)[name = tensor("op_5490_cast_fp16")]; + tensor var_5491_cast_fp16 = mul(x = var_5490_cast_fp16, y = onehot_87_cast_fp16)[name = tensor("op_5491_cast_fp16")]; + tensor new_k_cache_43_cast_fp16 = add(x = var_5484_cast_fp16, y = var_5491_cast_fp16)[name = tensor("new_k_cache_43_cast_fp16")]; + tensor new_k_cache_43_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_43_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache21_to_fp16_dtype_0 = const()[name = tensor("v_cache21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache21_to_fp16 = cast(dtype = v_cache21_to_fp16_dtype_0, x = v_cache21)[name = tensor("cast_405")]; + tensor var_5497_cast_fp16 = mul(x = v_cache21_to_fp16, y = var_5483_cast_fp16)[name = tensor("op_5497_cast_fp16")]; + tensor var_5502 = const()[name = tensor("op_5502"), val = tensor([1, 1, 16, 64])]; + tensor var_5503_cast_fp16 = reshape(shape = var_5502, x = v_43_cast_fp16)[name = tensor("op_5503_cast_fp16")]; + tensor var_5504_cast_fp16 = mul(x = var_5503_cast_fp16, y = onehot_87_cast_fp16)[name = tensor("op_5504_cast_fp16")]; + tensor new_v_cache_43_cast_fp16 = add(x = var_5497_cast_fp16, y = var_5504_cast_fp16)[name = tensor("new_v_cache_43_cast_fp16")]; + tensor new_v_cache_43_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_43_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_21_perm_0 = const()[name = tensor("transpose_21_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_5517 = const()[name = tensor("op_5517"), val = tensor([0, 2, 1, 3])]; + tensor var_5523 = const()[name = tensor("op_5523"), val = tensor([1, 16, 1, 64])]; + tensor q4_43_cast_fp16 = reshape(shape = var_5523, x = out_85_cast_fp16)[name = tensor("q4_43_cast_fp16")]; + tensor var_5528_transpose_x_0 = const()[name = tensor("op_5528_transpose_x_0"), val = tensor(false)]; + tensor var_5528_transpose_y_0 = const()[name = tensor("op_5528_transpose_y_0"), val = tensor(false)]; + tensor transpose_21_cast_fp16 = transpose(perm = transpose_21_perm_0, x = new_k_cache_43_cast_fp16)[name = tensor("transpose_32")]; + tensor var_5528_cast_fp16 = matmul(transpose_x = var_5528_transpose_x_0, transpose_y = var_5528_transpose_y_0, x = q4_43_cast_fp16, y = transpose_21_cast_fp16)[name = tensor("op_5528_cast_fp16")]; + tensor var_5529_to_fp16 = const()[name = tensor("op_5529_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_85_cast_fp16 = mul(x = var_5528_cast_fp16, y = var_5529_to_fp16)[name = tensor("scores_85_cast_fp16")]; + tensor var_5533 = const()[name = tensor("op_5533"), val = tensor([1, 1])]; + tensor var_5534_cast_fp16 = reshape(shape = var_5533, x = position21_to_fp16)[name = tensor("op_5534_cast_fp16")]; + tensor var_5535_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_5534_cast_fp16)[name = tensor("op_5535_cast_fp16")]; + tensor mask_43_to_fp16_dtype_0 = const()[name = tensor("mask_43_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_5542_to_fp16 = const()[name = tensor("op_5542_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5535_cast_fp16_to_fp16 = cast(dtype = mask_43_to_fp16_dtype_0, x = var_5535_cast_fp16)[name = tensor("cast_403")]; + tensor var_5543_cast_fp16 = sub(x = var_5535_cast_fp16_to_fp16, y = var_5542_to_fp16)[name = tensor("op_5543_cast_fp16")]; + tensor var_5548 = const()[name = tensor("op_5548"), val = tensor([1, 1, 1, 512])]; + tensor var_5549_cast_fp16 = reshape(shape = var_5548, x = var_5543_cast_fp16)[name = tensor("op_5549_cast_fp16")]; + tensor var_5550_to_fp16 = const()[name = tensor("op_5550_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_5551_cast_fp16 = mul(x = var_5549_cast_fp16, y = var_5550_to_fp16)[name = tensor("op_5551_cast_fp16")]; + tensor scores_87_cast_fp16 = add(x = scores_85_cast_fp16, y = var_5551_cast_fp16)[name = tensor("scores_87_cast_fp16")]; + tensor var_5554 = const()[name = tensor("op_5554"), val = tensor(-1)]; + tensor weights_43_cast_fp16 = softmax(axis = var_5554, x = scores_87_cast_fp16)[name = tensor("weights_43_cast_fp16")]; + tensor attn_43_transpose_x_0 = const()[name = tensor("attn_43_transpose_x_0"), val = tensor(false)]; + tensor attn_43_transpose_y_0 = const()[name = tensor("attn_43_transpose_y_0"), val = tensor(false)]; + tensor values_43_cast_fp16 = transpose(perm = var_5517, x = new_v_cache_43_cast_fp16)[name = tensor("transpose_31")]; + tensor attn_43_cast_fp16 = matmul(transpose_x = attn_43_transpose_x_0, transpose_y = attn_43_transpose_y_0, x = weights_43_cast_fp16, y = values_43_cast_fp16)[name = tensor("attn_43_cast_fp16")]; + tensor var_5562 = const()[name = tensor("op_5562"), val = tensor([0, 2, 1, 3])]; + tensor var_5567 = const()[name = tensor("op_5567"), val = tensor([1, 1, 1024])]; + tensor var_5563_cast_fp16 = transpose(perm = var_5562, x = attn_43_cast_fp16)[name = tensor("transpose_30")]; + tensor input_259_cast_fp16 = reshape(shape = var_5567, x = var_5563_cast_fp16)[name = tensor("input_259_cast_fp16")]; + tensor attn21_out_proj_weight_to_fp16 = const()[name = tensor("attn21_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(535044288)))]; + tensor linear_86_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn21_out_proj_weight_to_fp16, x = input_259_cast_fp16)[name = tensor("linear_86_cast_fp16")]; + tensor var_5573_to_fp16 = const()[name = tensor("op_5573_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5574_cast_fp16 = add(x = position21_to_fp16, y = var_5573_to_fp16)[name = tensor("op_5574_cast_fp16")]; + tensor var_5574_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5574_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_261_cast_fp16 = add(x = input_255_cast_fp16, y = linear_86_cast_fp16)[name = tensor("input_261_cast_fp16")]; + tensor input_263_axes_0 = const()[name = tensor("input_263_axes_0"), val = tensor([-1])]; + tensor norm21_2_weight_to_fp16 = const()[name = tensor("norm21_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(537141504)))]; + tensor norm21_2_bias_to_fp16 = const()[name = tensor("norm21_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(537143616)))]; + tensor var_5578_to_fp16 = const()[name = tensor("op_5578_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_263_cast_fp16 = layer_norm(axes = input_263_axes_0, beta = norm21_2_bias_to_fp16, epsilon = var_5578_to_fp16, gamma = norm21_2_weight_to_fp16, x = input_261_cast_fp16)[name = tensor("input_263_cast_fp16")]; + tensor linear21_1_weight_to_fp16 = const()[name = tensor("linear21_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(537145728)))]; + tensor linear_87_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear21_1_weight_to_fp16, x = input_263_cast_fp16)[name = tensor("linear_87_cast_fp16")]; + tensor input_265_mode_0 = const()[name = tensor("input_265_mode_0"), val = tensor("EXACT")]; + tensor input_265_cast_fp16 = gelu(mode = input_265_mode_0, x = linear_87_cast_fp16)[name = tensor("input_265_cast_fp16")]; + tensor linear21_2_weight_to_fp16 = const()[name = tensor("linear21_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(545534400)))]; + tensor linear_88_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear21_2_weight_to_fp16, x = input_265_cast_fp16)[name = tensor("linear_88_cast_fp16")]; + tensor input_267_cast_fp16 = add(x = input_261_cast_fp16, y = linear_88_cast_fp16)[name = tensor("input_267_cast_fp16")]; + tensor input_269_axes_0 = const()[name = tensor("input_269_axes_0"), val = tensor([-1])]; + tensor norm22_1_weight_to_fp16 = const()[name = tensor("norm22_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(553923072)))]; + tensor norm22_1_bias_to_fp16 = const()[name = tensor("norm22_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(553925184)))]; + tensor var_5595_to_fp16 = const()[name = tensor("op_5595_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_269_cast_fp16 = layer_norm(axes = input_269_axes_0, beta = norm22_1_bias_to_fp16, epsilon = var_5595_to_fp16, gamma = norm22_1_weight_to_fp16, x = input_267_cast_fp16)[name = tensor("input_269_cast_fp16")]; + tensor attn22_in_proj_weight_to_fp16 = const()[name = tensor("attn22_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(553927296)))]; + tensor linear_89_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn22_in_proj_weight_to_fp16, x = input_269_cast_fp16)[name = tensor("linear_89_cast_fp16")]; + tensor var_5608 = const()[name = tensor("op_5608"), val = tensor([1, 3, 16, 64])]; + tensor qkv_45_cast_fp16 = reshape(shape = var_5608, x = linear_89_cast_fp16)[name = tensor("qkv_45_cast_fp16")]; + tensor x_89_begin_0 = const()[name = tensor("x_89_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_89_end_0 = const()[name = tensor("x_89_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_89_end_mask_0 = const()[name = tensor("x_89_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_89_squeeze_mask_0 = const()[name = tensor("x_89_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_89_cast_fp16 = slice_by_index(begin = x_89_begin_0, end = x_89_end_0, end_mask = x_89_end_mask_0, squeeze_mask = x_89_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("x_89_cast_fp16")]; + tensor x_91_begin_0 = const()[name = tensor("x_91_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_91_end_0 = const()[name = tensor("x_91_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_91_end_mask_0 = const()[name = tensor("x_91_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_91_squeeze_mask_0 = const()[name = tensor("x_91_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_91_cast_fp16 = slice_by_index(begin = x_91_begin_0, end = x_91_end_0, end_mask = x_91_end_mask_0, squeeze_mask = x_91_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("x_91_cast_fp16")]; + tensor v_45_begin_0 = const()[name = tensor("v_45_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_45_end_0 = const()[name = tensor("v_45_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_45_end_mask_0 = const()[name = tensor("v_45_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_45_squeeze_mask_0 = const()[name = tensor("v_45_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_45_cast_fp16 = slice_by_index(begin = v_45_begin_0, end = v_45_end_0, end_mask = v_45_end_mask_0, squeeze_mask = v_45_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("v_45_cast_fp16")]; + tensor var_5637 = const()[name = tensor("op_5637"), val = tensor([1, 1, 1])]; + tensor position22_to_fp16_dtype_0 = const()[name = tensor("position22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position22_to_fp16 = cast(dtype = position22_to_fp16_dtype_0, x = position22)[name = tensor("cast_401")]; + tensor var_5638_cast_fp16 = reshape(shape = var_5637, x = position22_to_fp16)[name = tensor("op_5638_cast_fp16")]; + tensor angles_89_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_5638_cast_fp16)[name = tensor("angles_89_cast_fp16")]; + tensor rotr_89_cast_fp16 = cos(x = angles_89_cast_fp16)[name = tensor("rotr_89_cast_fp16")]; + tensor roti_89_cast_fp16 = sin(x = angles_89_cast_fp16)[name = tensor("roti_89_cast_fp16")]; + tensor var_5646 = const()[name = tensor("op_5646"), val = tensor([1, 16, 32, 2])]; + tensor pairs_89_cast_fp16 = reshape(shape = var_5646, x = x_89_cast_fp16)[name = tensor("pairs_89_cast_fp16")]; + tensor xr_89_begin_0 = const()[name = tensor("xr_89_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_89_end_0 = const()[name = tensor("xr_89_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_89_end_mask_0 = const()[name = tensor("xr_89_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_89_squeeze_mask_0 = const()[name = tensor("xr_89_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_89_cast_fp16 = slice_by_index(begin = xr_89_begin_0, end = xr_89_end_0, end_mask = xr_89_end_mask_0, squeeze_mask = xr_89_squeeze_mask_0, x = pairs_89_cast_fp16)[name = tensor("xr_89_cast_fp16")]; + tensor xi_89_begin_0 = const()[name = tensor("xi_89_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_89_end_0 = const()[name = tensor("xi_89_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_89_end_mask_0 = const()[name = tensor("xi_89_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_89_squeeze_mask_0 = const()[name = tensor("xi_89_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_89_cast_fp16 = slice_by_index(begin = xi_89_begin_0, end = xi_89_end_0, end_mask = xi_89_end_mask_0, squeeze_mask = xi_89_squeeze_mask_0, x = pairs_89_cast_fp16)[name = tensor("xi_89_cast_fp16")]; + tensor var_5654_cast_fp16 = mul(x = xr_89_cast_fp16, y = rotr_89_cast_fp16)[name = tensor("op_5654_cast_fp16")]; + tensor var_5655_cast_fp16 = mul(x = xi_89_cast_fp16, y = roti_89_cast_fp16)[name = tensor("op_5655_cast_fp16")]; + tensor xor__89_cast_fp16 = sub(x = var_5654_cast_fp16, y = var_5655_cast_fp16)[name = tensor("xor__89_cast_fp16")]; + tensor var_5658_cast_fp16 = mul(x = xr_89_cast_fp16, y = roti_89_cast_fp16)[name = tensor("op_5658_cast_fp16")]; + tensor var_5659_cast_fp16 = mul(x = xi_89_cast_fp16, y = rotr_89_cast_fp16)[name = tensor("op_5659_cast_fp16")]; + tensor xoi_89_cast_fp16 = add(x = var_5658_cast_fp16, y = var_5659_cast_fp16)[name = tensor("xoi_89_cast_fp16")]; + tensor out_89_axis_0 = const()[name = tensor("out_89_axis_0"), val = tensor(-1)]; + tensor out_89_cast_fp16 = stack(axis = out_89_axis_0, values = (xor__89_cast_fp16, xoi_89_cast_fp16))[name = tensor("out_89_cast_fp16")]; + tensor var_5682 = const()[name = tensor("op_5682"), val = tensor([1, 16, 32, 2])]; + tensor pairs_91_cast_fp16 = reshape(shape = var_5682, x = x_91_cast_fp16)[name = tensor("pairs_91_cast_fp16")]; + tensor xr_91_begin_0 = const()[name = tensor("xr_91_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_91_end_0 = const()[name = tensor("xr_91_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_91_end_mask_0 = const()[name = tensor("xr_91_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_91_squeeze_mask_0 = const()[name = tensor("xr_91_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_91_cast_fp16 = slice_by_index(begin = xr_91_begin_0, end = xr_91_end_0, end_mask = xr_91_end_mask_0, squeeze_mask = xr_91_squeeze_mask_0, x = pairs_91_cast_fp16)[name = tensor("xr_91_cast_fp16")]; + tensor xi_91_begin_0 = const()[name = tensor("xi_91_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_91_end_0 = const()[name = tensor("xi_91_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_91_end_mask_0 = const()[name = tensor("xi_91_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_91_squeeze_mask_0 = const()[name = tensor("xi_91_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_91_cast_fp16 = slice_by_index(begin = xi_91_begin_0, end = xi_91_end_0, end_mask = xi_91_end_mask_0, squeeze_mask = xi_91_squeeze_mask_0, x = pairs_91_cast_fp16)[name = tensor("xi_91_cast_fp16")]; + tensor var_5690_cast_fp16 = mul(x = xr_91_cast_fp16, y = rotr_89_cast_fp16)[name = tensor("op_5690_cast_fp16")]; + tensor var_5691_cast_fp16 = mul(x = xi_91_cast_fp16, y = roti_89_cast_fp16)[name = tensor("op_5691_cast_fp16")]; + tensor xor__91_cast_fp16 = sub(x = var_5690_cast_fp16, y = var_5691_cast_fp16)[name = tensor("xor__91_cast_fp16")]; + tensor var_5694_cast_fp16 = mul(x = xr_91_cast_fp16, y = roti_89_cast_fp16)[name = tensor("op_5694_cast_fp16")]; + tensor var_5695_cast_fp16 = mul(x = xi_91_cast_fp16, y = rotr_89_cast_fp16)[name = tensor("op_5695_cast_fp16")]; + tensor xoi_91_cast_fp16 = add(x = var_5694_cast_fp16, y = var_5695_cast_fp16)[name = tensor("xoi_91_cast_fp16")]; + tensor out_91_axis_0 = const()[name = tensor("out_91_axis_0"), val = tensor(-1)]; + tensor out_91_cast_fp16 = stack(axis = out_91_axis_0, values = (xor__91_cast_fp16, xoi_91_cast_fp16))[name = tensor("out_91_cast_fp16")]; + tensor _inversed_5707_y_0_to_fp16 = const()[name = tensor("_inversed_5707_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5707_cast_fp16 = mul(x = position22_to_fp16, y = _inversed_5707_y_0_to_fp16)[name = tensor("_inversed_5707_cast_fp16")]; + tensor var_5708_cast_fp16 = floor(x = _inversed_5707_cast_fp16)[name = tensor("op_5708_cast_fp16")]; + tensor var_5709_to_fp16 = const()[name = tensor("op_5709_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5710_cast_fp16 = mul(x = var_5708_cast_fp16, y = var_5709_to_fp16)[name = tensor("op_5710_cast_fp16")]; + tensor write_idx_45_cast_fp16 = sub(x = position22_to_fp16, y = var_5710_cast_fp16)[name = tensor("write_idx_45_cast_fp16")]; + tensor var_5713_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_45_cast_fp16)[name = tensor("op_5713_cast_fp16")]; + tensor var_5723 = const()[name = tensor("op_5723"), val = tensor([1, 512, 1, 1])]; + tensor onehot_89_to_fp16_dtype_0 = const()[name = tensor("onehot_89_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_5713_cast_fp16_to_fp16 = cast(dtype = onehot_89_to_fp16_dtype_0, x = var_5713_cast_fp16)[name = tensor("cast_400")]; + tensor onehot_91_cast_fp16 = reshape(shape = var_5723, x = var_5713_cast_fp16_to_fp16)[name = tensor("onehot_91_cast_fp16")]; + tensor var_5725_to_fp16 = const()[name = tensor("op_5725_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5727_cast_fp16 = sub(x = var_5725_to_fp16, y = onehot_91_cast_fp16)[name = tensor("op_5727_cast_fp16")]; + tensor k_cache22_to_fp16_dtype_0 = const()[name = tensor("k_cache22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache22_to_fp16 = cast(dtype = k_cache22_to_fp16_dtype_0, x = k_cache22)[name = tensor("cast_399")]; + tensor var_5728_cast_fp16 = mul(x = k_cache22_to_fp16, y = var_5727_cast_fp16)[name = tensor("op_5728_cast_fp16")]; + tensor var_5733 = const()[name = tensor("op_5733"), val = tensor([1, 1, 16, 64])]; + tensor var_5734_cast_fp16 = reshape(shape = var_5733, x = out_91_cast_fp16)[name = tensor("op_5734_cast_fp16")]; + tensor var_5735_cast_fp16 = mul(x = var_5734_cast_fp16, y = onehot_91_cast_fp16)[name = tensor("op_5735_cast_fp16")]; + tensor new_k_cache_45_cast_fp16 = add(x = var_5728_cast_fp16, y = var_5735_cast_fp16)[name = tensor("new_k_cache_45_cast_fp16")]; + tensor new_k_cache_45_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_45_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache22_to_fp16_dtype_0 = const()[name = tensor("v_cache22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache22_to_fp16 = cast(dtype = v_cache22_to_fp16_dtype_0, x = v_cache22)[name = tensor("cast_397")]; + tensor var_5741_cast_fp16 = mul(x = v_cache22_to_fp16, y = var_5727_cast_fp16)[name = tensor("op_5741_cast_fp16")]; + tensor var_5746 = const()[name = tensor("op_5746"), val = tensor([1, 1, 16, 64])]; + tensor var_5747_cast_fp16 = reshape(shape = var_5746, x = v_45_cast_fp16)[name = tensor("op_5747_cast_fp16")]; + tensor var_5748_cast_fp16 = mul(x = var_5747_cast_fp16, y = onehot_91_cast_fp16)[name = tensor("op_5748_cast_fp16")]; + tensor new_v_cache_45_cast_fp16 = add(x = var_5741_cast_fp16, y = var_5748_cast_fp16)[name = tensor("new_v_cache_45_cast_fp16")]; + tensor new_v_cache_45_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_45_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_22_perm_0 = const()[name = tensor("transpose_22_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_5761 = const()[name = tensor("op_5761"), val = tensor([0, 2, 1, 3])]; + tensor var_5767 = const()[name = tensor("op_5767"), val = tensor([1, 16, 1, 64])]; + tensor q4_45_cast_fp16 = reshape(shape = var_5767, x = out_89_cast_fp16)[name = tensor("q4_45_cast_fp16")]; + tensor var_5772_transpose_x_0 = const()[name = tensor("op_5772_transpose_x_0"), val = tensor(false)]; + tensor var_5772_transpose_y_0 = const()[name = tensor("op_5772_transpose_y_0"), val = tensor(false)]; + tensor transpose_22_cast_fp16 = transpose(perm = transpose_22_perm_0, x = new_k_cache_45_cast_fp16)[name = tensor("transpose_29")]; + tensor var_5772_cast_fp16 = matmul(transpose_x = var_5772_transpose_x_0, transpose_y = var_5772_transpose_y_0, x = q4_45_cast_fp16, y = transpose_22_cast_fp16)[name = tensor("op_5772_cast_fp16")]; + tensor var_5773_to_fp16 = const()[name = tensor("op_5773_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_89_cast_fp16 = mul(x = var_5772_cast_fp16, y = var_5773_to_fp16)[name = tensor("scores_89_cast_fp16")]; + tensor var_5777 = const()[name = tensor("op_5777"), val = tensor([1, 1])]; + tensor var_5778_cast_fp16 = reshape(shape = var_5777, x = position22_to_fp16)[name = tensor("op_5778_cast_fp16")]; + tensor var_5779_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_5778_cast_fp16)[name = tensor("op_5779_cast_fp16")]; + tensor mask_45_to_fp16_dtype_0 = const()[name = tensor("mask_45_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_5786_to_fp16 = const()[name = tensor("op_5786_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5779_cast_fp16_to_fp16 = cast(dtype = mask_45_to_fp16_dtype_0, x = var_5779_cast_fp16)[name = tensor("cast_395")]; + tensor var_5787_cast_fp16 = sub(x = var_5779_cast_fp16_to_fp16, y = var_5786_to_fp16)[name = tensor("op_5787_cast_fp16")]; + tensor var_5792 = const()[name = tensor("op_5792"), val = tensor([1, 1, 1, 512])]; + tensor var_5793_cast_fp16 = reshape(shape = var_5792, x = var_5787_cast_fp16)[name = tensor("op_5793_cast_fp16")]; + tensor var_5794_to_fp16 = const()[name = tensor("op_5794_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_5795_cast_fp16 = mul(x = var_5793_cast_fp16, y = var_5794_to_fp16)[name = tensor("op_5795_cast_fp16")]; + tensor scores_91_cast_fp16 = add(x = scores_89_cast_fp16, y = var_5795_cast_fp16)[name = tensor("scores_91_cast_fp16")]; + tensor var_5798 = const()[name = tensor("op_5798"), val = tensor(-1)]; + tensor weights_45_cast_fp16 = softmax(axis = var_5798, x = scores_91_cast_fp16)[name = tensor("weights_45_cast_fp16")]; + tensor attn_45_transpose_x_0 = const()[name = tensor("attn_45_transpose_x_0"), val = tensor(false)]; + tensor attn_45_transpose_y_0 = const()[name = tensor("attn_45_transpose_y_0"), val = tensor(false)]; + tensor values_45_cast_fp16 = transpose(perm = var_5761, x = new_v_cache_45_cast_fp16)[name = tensor("transpose_28")]; + tensor attn_45_cast_fp16 = matmul(transpose_x = attn_45_transpose_x_0, transpose_y = attn_45_transpose_y_0, x = weights_45_cast_fp16, y = values_45_cast_fp16)[name = tensor("attn_45_cast_fp16")]; + tensor var_5806 = const()[name = tensor("op_5806"), val = tensor([0, 2, 1, 3])]; + tensor var_5811 = const()[name = tensor("op_5811"), val = tensor([1, 1, 1024])]; + tensor var_5807_cast_fp16 = transpose(perm = var_5806, x = attn_45_cast_fp16)[name = tensor("transpose_27")]; + tensor input_271_cast_fp16 = reshape(shape = var_5811, x = var_5807_cast_fp16)[name = tensor("input_271_cast_fp16")]; + tensor attn22_out_proj_weight_to_fp16 = const()[name = tensor("attn22_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(560218816)))]; + tensor linear_90_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn22_out_proj_weight_to_fp16, x = input_271_cast_fp16)[name = tensor("linear_90_cast_fp16")]; + tensor var_5817_to_fp16 = const()[name = tensor("op_5817_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5818_cast_fp16 = add(x = position22_to_fp16, y = var_5817_to_fp16)[name = tensor("op_5818_cast_fp16")]; + tensor var_5818_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5818_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_273_cast_fp16 = add(x = input_267_cast_fp16, y = linear_90_cast_fp16)[name = tensor("input_273_cast_fp16")]; + tensor input_275_axes_0 = const()[name = tensor("input_275_axes_0"), val = tensor([-1])]; + tensor norm22_2_weight_to_fp16 = const()[name = tensor("norm22_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(562316032)))]; + tensor norm22_2_bias_to_fp16 = const()[name = tensor("norm22_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(562318144)))]; + tensor var_5822_to_fp16 = const()[name = tensor("op_5822_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_275_cast_fp16 = layer_norm(axes = input_275_axes_0, beta = norm22_2_bias_to_fp16, epsilon = var_5822_to_fp16, gamma = norm22_2_weight_to_fp16, x = input_273_cast_fp16)[name = tensor("input_275_cast_fp16")]; + tensor linear22_1_weight_to_fp16 = const()[name = tensor("linear22_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(562320256)))]; + tensor linear_91_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear22_1_weight_to_fp16, x = input_275_cast_fp16)[name = tensor("linear_91_cast_fp16")]; + tensor input_277_mode_0 = const()[name = tensor("input_277_mode_0"), val = tensor("EXACT")]; + tensor input_277_cast_fp16 = gelu(mode = input_277_mode_0, x = linear_91_cast_fp16)[name = tensor("input_277_cast_fp16")]; + tensor linear22_2_weight_to_fp16 = const()[name = tensor("linear22_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(570708928)))]; + tensor linear_92_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear22_2_weight_to_fp16, x = input_277_cast_fp16)[name = tensor("linear_92_cast_fp16")]; + tensor input_279_cast_fp16 = add(x = input_273_cast_fp16, y = linear_92_cast_fp16)[name = tensor("input_279_cast_fp16")]; + tensor input_281_axes_0 = const()[name = tensor("input_281_axes_0"), val = tensor([-1])]; + tensor norm23_1_weight_to_fp16 = const()[name = tensor("norm23_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(579097600)))]; + tensor norm23_1_bias_to_fp16 = const()[name = tensor("norm23_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(579099712)))]; + tensor var_5839_to_fp16 = const()[name = tensor("op_5839_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_281_cast_fp16 = layer_norm(axes = input_281_axes_0, beta = norm23_1_bias_to_fp16, epsilon = var_5839_to_fp16, gamma = norm23_1_weight_to_fp16, x = input_279_cast_fp16)[name = tensor("input_281_cast_fp16")]; + tensor attn23_in_proj_weight_to_fp16 = const()[name = tensor("attn23_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(579101824)))]; + tensor linear_93_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn23_in_proj_weight_to_fp16, x = input_281_cast_fp16)[name = tensor("linear_93_cast_fp16")]; + tensor var_5852 = const()[name = tensor("op_5852"), val = tensor([1, 3, 16, 64])]; + tensor qkv_cast_fp16 = reshape(shape = var_5852, x = linear_93_cast_fp16)[name = tensor("qkv_cast_fp16")]; + tensor x_93_begin_0 = const()[name = tensor("x_93_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor x_93_end_0 = const()[name = tensor("x_93_end_0"), val = tensor([1, 1, 16, 64])]; + tensor x_93_end_mask_0 = const()[name = tensor("x_93_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_93_squeeze_mask_0 = const()[name = tensor("x_93_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_93_cast_fp16 = slice_by_index(begin = x_93_begin_0, end = x_93_end_0, end_mask = x_93_end_mask_0, squeeze_mask = x_93_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("x_93_cast_fp16")]; + tensor x_begin_0 = const()[name = tensor("x_begin_0"), val = tensor([0, 1, 0, 0])]; + tensor x_end_0 = const()[name = tensor("x_end_0"), val = tensor([1, 2, 16, 64])]; + tensor x_end_mask_0 = const()[name = tensor("x_end_mask_0"), val = tensor([true, false, true, true])]; + tensor x_squeeze_mask_0 = const()[name = tensor("x_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor x_cast_fp16 = slice_by_index(begin = x_begin_0, end = x_end_0, end_mask = x_end_mask_0, squeeze_mask = x_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("x_cast_fp16")]; + tensor v_begin_0 = const()[name = tensor("v_begin_0"), val = tensor([0, 2, 0, 0])]; + tensor v_end_0 = const()[name = tensor("v_end_0"), val = tensor([1, 3, 16, 64])]; + tensor v_end_mask_0 = const()[name = tensor("v_end_mask_0"), val = tensor([true, false, true, true])]; + tensor v_squeeze_mask_0 = const()[name = tensor("v_squeeze_mask_0"), val = tensor([false, true, false, false])]; + tensor v_cast_fp16 = slice_by_index(begin = v_begin_0, end = v_end_0, end_mask = v_end_mask_0, squeeze_mask = v_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("v_cast_fp16")]; + tensor var_5881 = const()[name = tensor("op_5881"), val = tensor([1, 1, 1])]; + tensor position23_to_fp16_dtype_0 = const()[name = tensor("position23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position23_to_fp16 = cast(dtype = position23_to_fp16_dtype_0, x = position23)[name = tensor("cast_393")]; + tensor var_5882_cast_fp16 = reshape(shape = var_5881, x = position23_to_fp16)[name = tensor("op_5882_cast_fp16")]; + tensor angles_93_cast_fp16 = mul(x = rope_freqs_to_fp16, y = var_5882_cast_fp16)[name = tensor("angles_93_cast_fp16")]; + tensor rotr_93_cast_fp16 = cos(x = angles_93_cast_fp16)[name = tensor("rotr_93_cast_fp16")]; + tensor roti_93_cast_fp16 = sin(x = angles_93_cast_fp16)[name = tensor("roti_93_cast_fp16")]; + tensor var_5890 = const()[name = tensor("op_5890"), val = tensor([1, 16, 32, 2])]; + tensor pairs_93_cast_fp16 = reshape(shape = var_5890, x = x_93_cast_fp16)[name = tensor("pairs_93_cast_fp16")]; + tensor xr_93_begin_0 = const()[name = tensor("xr_93_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_93_end_0 = const()[name = tensor("xr_93_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_93_end_mask_0 = const()[name = tensor("xr_93_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_93_squeeze_mask_0 = const()[name = tensor("xr_93_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_93_cast_fp16 = slice_by_index(begin = xr_93_begin_0, end = xr_93_end_0, end_mask = xr_93_end_mask_0, squeeze_mask = xr_93_squeeze_mask_0, x = pairs_93_cast_fp16)[name = tensor("xr_93_cast_fp16")]; + tensor xi_93_begin_0 = const()[name = tensor("xi_93_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_93_end_0 = const()[name = tensor("xi_93_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_93_end_mask_0 = const()[name = tensor("xi_93_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_93_squeeze_mask_0 = const()[name = tensor("xi_93_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_93_cast_fp16 = slice_by_index(begin = xi_93_begin_0, end = xi_93_end_0, end_mask = xi_93_end_mask_0, squeeze_mask = xi_93_squeeze_mask_0, x = pairs_93_cast_fp16)[name = tensor("xi_93_cast_fp16")]; + tensor var_5898_cast_fp16 = mul(x = xr_93_cast_fp16, y = rotr_93_cast_fp16)[name = tensor("op_5898_cast_fp16")]; + tensor var_5899_cast_fp16 = mul(x = xi_93_cast_fp16, y = roti_93_cast_fp16)[name = tensor("op_5899_cast_fp16")]; + tensor xor__93_cast_fp16 = sub(x = var_5898_cast_fp16, y = var_5899_cast_fp16)[name = tensor("xor__93_cast_fp16")]; + tensor var_5902_cast_fp16 = mul(x = xr_93_cast_fp16, y = roti_93_cast_fp16)[name = tensor("op_5902_cast_fp16")]; + tensor var_5903_cast_fp16 = mul(x = xi_93_cast_fp16, y = rotr_93_cast_fp16)[name = tensor("op_5903_cast_fp16")]; + tensor xoi_93_cast_fp16 = add(x = var_5902_cast_fp16, y = var_5903_cast_fp16)[name = tensor("xoi_93_cast_fp16")]; + tensor out_93_axis_0 = const()[name = tensor("out_93_axis_0"), val = tensor(-1)]; + tensor out_93_cast_fp16 = stack(axis = out_93_axis_0, values = (xor__93_cast_fp16, xoi_93_cast_fp16))[name = tensor("out_93_cast_fp16")]; + tensor var_5926 = const()[name = tensor("op_5926"), val = tensor([1, 16, 32, 2])]; + tensor pairs_cast_fp16 = reshape(shape = var_5926, x = x_cast_fp16)[name = tensor("pairs_cast_fp16")]; + tensor xr_begin_0 = const()[name = tensor("xr_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor xr_end_0 = const()[name = tensor("xr_end_0"), val = tensor([1, 16, 32, 1])]; + tensor xr_end_mask_0 = const()[name = tensor("xr_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xr_squeeze_mask_0 = const()[name = tensor("xr_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xr_cast_fp16 = slice_by_index(begin = xr_begin_0, end = xr_end_0, end_mask = xr_end_mask_0, squeeze_mask = xr_squeeze_mask_0, x = pairs_cast_fp16)[name = tensor("xr_cast_fp16")]; + tensor xi_begin_0 = const()[name = tensor("xi_begin_0"), val = tensor([0, 0, 0, 1])]; + tensor xi_end_0 = const()[name = tensor("xi_end_0"), val = tensor([1, 16, 32, 2])]; + tensor xi_end_mask_0 = const()[name = tensor("xi_end_mask_0"), val = tensor([true, true, true, false])]; + tensor xi_squeeze_mask_0 = const()[name = tensor("xi_squeeze_mask_0"), val = tensor([false, false, false, true])]; + tensor xi_cast_fp16 = slice_by_index(begin = xi_begin_0, end = xi_end_0, end_mask = xi_end_mask_0, squeeze_mask = xi_squeeze_mask_0, x = pairs_cast_fp16)[name = tensor("xi_cast_fp16")]; + tensor var_5934_cast_fp16 = mul(x = xr_cast_fp16, y = rotr_93_cast_fp16)[name = tensor("op_5934_cast_fp16")]; + tensor var_5935_cast_fp16 = mul(x = xi_cast_fp16, y = roti_93_cast_fp16)[name = tensor("op_5935_cast_fp16")]; + tensor xor__cast_fp16 = sub(x = var_5934_cast_fp16, y = var_5935_cast_fp16)[name = tensor("xor__cast_fp16")]; + tensor var_5938_cast_fp16 = mul(x = xr_cast_fp16, y = roti_93_cast_fp16)[name = tensor("op_5938_cast_fp16")]; + tensor var_5939_cast_fp16 = mul(x = xi_cast_fp16, y = rotr_93_cast_fp16)[name = tensor("op_5939_cast_fp16")]; + tensor xoi_cast_fp16 = add(x = var_5938_cast_fp16, y = var_5939_cast_fp16)[name = tensor("xoi_cast_fp16")]; + tensor out_axis_0 = const()[name = tensor("out_axis_0"), val = tensor(-1)]; + tensor out_cast_fp16 = stack(axis = out_axis_0, values = (xor__cast_fp16, xoi_cast_fp16))[name = tensor("out_cast_fp16")]; + tensor _inversed_5951_y_0_to_fp16 = const()[name = tensor("_inversed_5951_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5951_cast_fp16 = mul(x = position23_to_fp16, y = _inversed_5951_y_0_to_fp16)[name = tensor("_inversed_5951_cast_fp16")]; + tensor var_5952_cast_fp16 = floor(x = _inversed_5951_cast_fp16)[name = tensor("op_5952_cast_fp16")]; + tensor var_5953_to_fp16 = const()[name = tensor("op_5953_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5954_cast_fp16 = mul(x = var_5952_cast_fp16, y = var_5953_to_fp16)[name = tensor("op_5954_cast_fp16")]; + tensor write_idx_cast_fp16 = sub(x = position23_to_fp16, y = var_5954_cast_fp16)[name = tensor("write_idx_cast_fp16")]; + tensor var_5957_cast_fp16 = equal(x = slot_indices_to_fp16, y = write_idx_cast_fp16)[name = tensor("op_5957_cast_fp16")]; + tensor var_5967 = const()[name = tensor("op_5967"), val = tensor([1, 512, 1, 1])]; + tensor onehot_93_to_fp16_dtype_0 = const()[name = tensor("onehot_93_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_5957_cast_fp16_to_fp16 = cast(dtype = onehot_93_to_fp16_dtype_0, x = var_5957_cast_fp16)[name = tensor("cast_392")]; + tensor onehot_cast_fp16 = reshape(shape = var_5967, x = var_5957_cast_fp16_to_fp16)[name = tensor("onehot_cast_fp16")]; + tensor var_5969_to_fp16 = const()[name = tensor("op_5969_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5971_cast_fp16 = sub(x = var_5969_to_fp16, y = onehot_cast_fp16)[name = tensor("op_5971_cast_fp16")]; + tensor k_cache23_to_fp16_dtype_0 = const()[name = tensor("k_cache23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor k_cache23_to_fp16 = cast(dtype = k_cache23_to_fp16_dtype_0, x = k_cache23)[name = tensor("cast_391")]; + tensor var_5972_cast_fp16 = mul(x = k_cache23_to_fp16, y = var_5971_cast_fp16)[name = tensor("op_5972_cast_fp16")]; + tensor var_5977 = const()[name = tensor("op_5977"), val = tensor([1, 1, 16, 64])]; + tensor var_5978_cast_fp16 = reshape(shape = var_5977, x = out_cast_fp16)[name = tensor("op_5978_cast_fp16")]; + tensor var_5979_cast_fp16 = mul(x = var_5978_cast_fp16, y = onehot_cast_fp16)[name = tensor("op_5979_cast_fp16")]; + tensor new_k_cache_cast_fp16 = add(x = var_5972_cast_fp16, y = var_5979_cast_fp16)[name = tensor("new_k_cache_cast_fp16")]; + tensor new_k_cache_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_k_cache_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor v_cache23_to_fp16_dtype_0 = const()[name = tensor("v_cache23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor v_cache23_to_fp16 = cast(dtype = v_cache23_to_fp16_dtype_0, x = v_cache23)[name = tensor("cast_389")]; + tensor var_5985_cast_fp16 = mul(x = v_cache23_to_fp16, y = var_5971_cast_fp16)[name = tensor("op_5985_cast_fp16")]; + tensor var_5990 = const()[name = tensor("op_5990"), val = tensor([1, 1, 16, 64])]; + tensor var_5991_cast_fp16 = reshape(shape = var_5990, x = v_cast_fp16)[name = tensor("op_5991_cast_fp16")]; + tensor var_5992_cast_fp16 = mul(x = var_5991_cast_fp16, y = onehot_cast_fp16)[name = tensor("op_5992_cast_fp16")]; + tensor new_v_cache_cast_fp16 = add(x = var_5985_cast_fp16, y = var_5992_cast_fp16)[name = tensor("new_v_cache_cast_fp16")]; + tensor new_v_cache_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("new_v_cache_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor transpose_23_perm_0 = const()[name = tensor("transpose_23_perm_0"), val = tensor([0, 2, 3, 1])]; + tensor var_6005 = const()[name = tensor("op_6005"), val = tensor([0, 2, 1, 3])]; + tensor var_6011 = const()[name = tensor("op_6011"), val = tensor([1, 16, 1, 64])]; + tensor q4_cast_fp16 = reshape(shape = var_6011, x = out_93_cast_fp16)[name = tensor("q4_cast_fp16")]; + tensor var_6016_transpose_x_0 = const()[name = tensor("op_6016_transpose_x_0"), val = tensor(false)]; + tensor var_6016_transpose_y_0 = const()[name = tensor("op_6016_transpose_y_0"), val = tensor(false)]; + tensor transpose_23_cast_fp16 = transpose(perm = transpose_23_perm_0, x = new_k_cache_cast_fp16)[name = tensor("transpose_26")]; + tensor var_6016_cast_fp16 = matmul(transpose_x = var_6016_transpose_x_0, transpose_y = var_6016_transpose_y_0, x = q4_cast_fp16, y = transpose_23_cast_fp16)[name = tensor("op_6016_cast_fp16")]; + tensor var_6017_to_fp16 = const()[name = tensor("op_6017_to_fp16"), val = tensor(0x1p-3)]; + tensor scores_93_cast_fp16 = mul(x = var_6016_cast_fp16, y = var_6017_to_fp16)[name = tensor("scores_93_cast_fp16")]; + tensor var_6021 = const()[name = tensor("op_6021"), val = tensor([1, 1])]; + tensor var_6022_cast_fp16 = reshape(shape = var_6021, x = position23_to_fp16)[name = tensor("op_6022_cast_fp16")]; + tensor var_6023_cast_fp16 = less_equal(x = slot_indices_to_fp16, y = var_6022_cast_fp16)[name = tensor("op_6023_cast_fp16")]; + tensor mask_to_fp16_dtype_0 = const()[name = tensor("mask_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor var_6030_to_fp16 = const()[name = tensor("op_6030_to_fp16"), val = tensor(0x1p+0)]; + tensor var_6023_cast_fp16_to_fp16 = cast(dtype = mask_to_fp16_dtype_0, x = var_6023_cast_fp16)[name = tensor("cast_387")]; + tensor var_6031_cast_fp16 = sub(x = var_6023_cast_fp16_to_fp16, y = var_6030_to_fp16)[name = tensor("op_6031_cast_fp16")]; + tensor var_6036 = const()[name = tensor("op_6036"), val = tensor([1, 1, 1, 512])]; + tensor var_6037_cast_fp16 = reshape(shape = var_6036, x = var_6031_cast_fp16)[name = tensor("op_6037_cast_fp16")]; + tensor var_6038_to_fp16 = const()[name = tensor("op_6038_to_fp16"), val = tensor(0x1.388p+13)]; + tensor var_6039_cast_fp16 = mul(x = var_6037_cast_fp16, y = var_6038_to_fp16)[name = tensor("op_6039_cast_fp16")]; + tensor scores_cast_fp16 = add(x = scores_93_cast_fp16, y = var_6039_cast_fp16)[name = tensor("scores_cast_fp16")]; + tensor var_6042 = const()[name = tensor("op_6042"), val = tensor(-1)]; + tensor weights_cast_fp16 = softmax(axis = var_6042, x = scores_cast_fp16)[name = tensor("weights_cast_fp16")]; + tensor attn_transpose_x_0 = const()[name = tensor("attn_transpose_x_0"), val = tensor(false)]; + tensor attn_transpose_y_0 = const()[name = tensor("attn_transpose_y_0"), val = tensor(false)]; + tensor values_cast_fp16 = transpose(perm = var_6005, x = new_v_cache_cast_fp16)[name = tensor("transpose_25")]; + tensor attn_cast_fp16 = matmul(transpose_x = attn_transpose_x_0, transpose_y = attn_transpose_y_0, x = weights_cast_fp16, y = values_cast_fp16)[name = tensor("attn_cast_fp16")]; + tensor var_6050 = const()[name = tensor("op_6050"), val = tensor([0, 2, 1, 3])]; + tensor var_6055 = const()[name = tensor("op_6055"), val = tensor([1, 1, 1024])]; + tensor var_6051_cast_fp16 = transpose(perm = var_6050, x = attn_cast_fp16)[name = tensor("transpose_24")]; + tensor input_283_cast_fp16 = reshape(shape = var_6055, x = var_6051_cast_fp16)[name = tensor("input_283_cast_fp16")]; + tensor attn23_out_proj_weight_to_fp16 = const()[name = tensor("attn23_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(585393344)))]; + tensor linear_94_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn23_out_proj_weight_to_fp16, x = input_283_cast_fp16)[name = tensor("linear_94_cast_fp16")]; + tensor var_6061_to_fp16 = const()[name = tensor("op_6061_to_fp16"), val = tensor(0x1p+0)]; + tensor var_6062_cast_fp16 = add(x = position23_to_fp16, y = var_6061_to_fp16)[name = tensor("op_6062_cast_fp16")]; + tensor var_6062_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_6062_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_285_cast_fp16 = add(x = input_279_cast_fp16, y = linear_94_cast_fp16)[name = tensor("input_285_cast_fp16")]; + tensor input_287_axes_0 = const()[name = tensor("input_287_axes_0"), val = tensor([-1])]; + tensor norm23_2_weight_to_fp16 = const()[name = tensor("norm23_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(587490560)))]; + tensor norm23_2_bias_to_fp16 = const()[name = tensor("norm23_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(587492672)))]; + tensor var_6066_to_fp16 = const()[name = tensor("op_6066_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_287_cast_fp16 = layer_norm(axes = input_287_axes_0, beta = norm23_2_bias_to_fp16, epsilon = var_6066_to_fp16, gamma = norm23_2_weight_to_fp16, x = input_285_cast_fp16)[name = tensor("input_287_cast_fp16")]; + tensor linear23_1_weight_to_fp16 = const()[name = tensor("linear23_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(587494784)))]; + tensor linear_95_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear23_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_95_cast_fp16")]; + tensor input_289_mode_0 = const()[name = tensor("input_289_mode_0"), val = tensor("EXACT")]; + tensor input_289_cast_fp16 = gelu(mode = input_289_mode_0, x = linear_95_cast_fp16)[name = tensor("input_289_cast_fp16")]; + tensor linear23_2_weight_to_fp16 = const()[name = tensor("linear23_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(595883456)))]; + tensor linear_96_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear23_2_weight_to_fp16, x = input_289_cast_fp16)[name = tensor("linear_96_cast_fp16")]; + tensor input_291_cast_fp16 = add(x = input_285_cast_fp16, y = linear_96_cast_fp16)[name = tensor("input_291_cast_fp16")]; + tensor input_axes_0 = const()[name = tensor("input_axes_0"), val = tensor([-1])]; + tensor out_norm_weight_to_fp16 = const()[name = tensor("out_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(604272128)))]; + tensor out_norm_bias_to_fp16 = const()[name = tensor("out_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(604274240)))]; + tensor var_6083_to_fp16 = const()[name = tensor("op_6083_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_cast_fp16 = layer_norm(axes = input_axes_0, beta = out_norm_bias_to_fp16, epsilon = var_6083_to_fp16, gamma = out_norm_weight_to_fp16, x = input_291_cast_fp16)[name = tensor("input_cast_fp16")]; + tensor input_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("input_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor out_eos_weight_to_fp16 = const()[name = tensor("out_eos_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(604276352)))]; + tensor out_eos_bias_to_fp16 = const()[name = tensor("out_eos_bias_to_fp16"), val = tensor([-0x1.5ep-3])]; + tensor linear_97_cast_fp16 = linear(bias = out_eos_bias_to_fp16, weight = out_eos_weight_to_fp16, x = input_cast_fp16)[name = tensor("linear_97_cast_fp16")]; + tensor linear_97_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("linear_97_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor is_eos = cast(dtype = linear_97_cast_fp16_to_fp32_dtype_0, x = linear_97_cast_fp16)[name = tensor("cast_384")]; + tensor transformer_out = cast(dtype = input_cast_fp16_to_fp32_dtype_0, x = input_cast_fp16)[name = tensor("cast_385")]; + tensor new_position23 = cast(dtype = var_6062_cast_fp16_to_fp32_dtype_0, x = var_6062_cast_fp16)[name = tensor("cast_386")]; + tensor new_v_cache23 = cast(dtype = new_v_cache_cast_fp16_to_fp32_dtype_0, x = new_v_cache_cast_fp16)[name = tensor("cast_388")]; + tensor new_k_cache23 = cast(dtype = new_k_cache_cast_fp16_to_fp32_dtype_0, x = new_k_cache_cast_fp16)[name = tensor("cast_390")]; + tensor new_position22 = cast(dtype = var_5818_cast_fp16_to_fp32_dtype_0, x = var_5818_cast_fp16)[name = tensor("cast_394")]; + tensor new_v_cache22 = cast(dtype = new_v_cache_45_cast_fp16_to_fp32_dtype_0, x = new_v_cache_45_cast_fp16)[name = tensor("cast_396")]; + tensor new_k_cache22 = cast(dtype = new_k_cache_45_cast_fp16_to_fp32_dtype_0, x = new_k_cache_45_cast_fp16)[name = tensor("cast_398")]; + tensor new_position21 = cast(dtype = var_5574_cast_fp16_to_fp32_dtype_0, x = var_5574_cast_fp16)[name = tensor("cast_402")]; + tensor new_v_cache21 = cast(dtype = new_v_cache_43_cast_fp16_to_fp32_dtype_0, x = new_v_cache_43_cast_fp16)[name = tensor("cast_404")]; + tensor new_k_cache21 = cast(dtype = new_k_cache_43_cast_fp16_to_fp32_dtype_0, x = new_k_cache_43_cast_fp16)[name = tensor("cast_406")]; + tensor new_position20 = cast(dtype = var_5330_cast_fp16_to_fp32_dtype_0, x = var_5330_cast_fp16)[name = tensor("cast_410")]; + tensor new_v_cache20 = cast(dtype = new_v_cache_41_cast_fp16_to_fp32_dtype_0, x = new_v_cache_41_cast_fp16)[name = tensor("cast_412")]; + tensor new_k_cache20 = cast(dtype = new_k_cache_41_cast_fp16_to_fp32_dtype_0, x = new_k_cache_41_cast_fp16)[name = tensor("cast_414")]; + tensor new_position19 = cast(dtype = var_5086_cast_fp16_to_fp32_dtype_0, x = var_5086_cast_fp16)[name = tensor("cast_418")]; + tensor new_v_cache19 = cast(dtype = new_v_cache_39_cast_fp16_to_fp32_dtype_0, x = new_v_cache_39_cast_fp16)[name = tensor("cast_420")]; + tensor new_k_cache19 = cast(dtype = new_k_cache_39_cast_fp16_to_fp32_dtype_0, x = new_k_cache_39_cast_fp16)[name = tensor("cast_422")]; + tensor new_position18 = cast(dtype = var_4842_cast_fp16_to_fp32_dtype_0, x = var_4842_cast_fp16)[name = tensor("cast_426")]; + tensor new_v_cache18 = cast(dtype = new_v_cache_37_cast_fp16_to_fp32_dtype_0, x = new_v_cache_37_cast_fp16)[name = tensor("cast_428")]; + tensor new_k_cache18 = cast(dtype = new_k_cache_37_cast_fp16_to_fp32_dtype_0, x = new_k_cache_37_cast_fp16)[name = tensor("cast_430")]; + tensor new_position17 = cast(dtype = var_4598_cast_fp16_to_fp32_dtype_0, x = var_4598_cast_fp16)[name = tensor("cast_434")]; + tensor new_v_cache17 = cast(dtype = new_v_cache_35_cast_fp16_to_fp32_dtype_0, x = new_v_cache_35_cast_fp16)[name = tensor("cast_436")]; + tensor new_k_cache17 = cast(dtype = new_k_cache_35_cast_fp16_to_fp32_dtype_0, x = new_k_cache_35_cast_fp16)[name = tensor("cast_438")]; + tensor new_position16 = cast(dtype = var_4354_cast_fp16_to_fp32_dtype_0, x = var_4354_cast_fp16)[name = tensor("cast_442")]; + tensor new_v_cache16 = cast(dtype = new_v_cache_33_cast_fp16_to_fp32_dtype_0, x = new_v_cache_33_cast_fp16)[name = tensor("cast_444")]; + tensor new_k_cache16 = cast(dtype = new_k_cache_33_cast_fp16_to_fp32_dtype_0, x = new_k_cache_33_cast_fp16)[name = tensor("cast_446")]; + tensor new_position15 = cast(dtype = var_4110_cast_fp16_to_fp32_dtype_0, x = var_4110_cast_fp16)[name = tensor("cast_450")]; + tensor new_v_cache15 = cast(dtype = new_v_cache_31_cast_fp16_to_fp32_dtype_0, x = new_v_cache_31_cast_fp16)[name = tensor("cast_452")]; + tensor new_k_cache15 = cast(dtype = new_k_cache_31_cast_fp16_to_fp32_dtype_0, x = new_k_cache_31_cast_fp16)[name = tensor("cast_454")]; + tensor new_position14 = cast(dtype = var_3866_cast_fp16_to_fp32_dtype_0, x = var_3866_cast_fp16)[name = tensor("cast_458")]; + tensor new_v_cache14 = cast(dtype = new_v_cache_29_cast_fp16_to_fp32_dtype_0, x = new_v_cache_29_cast_fp16)[name = tensor("cast_460")]; + tensor new_k_cache14 = cast(dtype = new_k_cache_29_cast_fp16_to_fp32_dtype_0, x = new_k_cache_29_cast_fp16)[name = tensor("cast_462")]; + tensor new_position13 = cast(dtype = var_3622_cast_fp16_to_fp32_dtype_0, x = var_3622_cast_fp16)[name = tensor("cast_466")]; + tensor new_v_cache13 = cast(dtype = new_v_cache_27_cast_fp16_to_fp32_dtype_0, x = new_v_cache_27_cast_fp16)[name = tensor("cast_468")]; + tensor new_k_cache13 = cast(dtype = new_k_cache_27_cast_fp16_to_fp32_dtype_0, x = new_k_cache_27_cast_fp16)[name = tensor("cast_470")]; + tensor new_position12 = cast(dtype = var_3378_cast_fp16_to_fp32_dtype_0, x = var_3378_cast_fp16)[name = tensor("cast_474")]; + tensor new_v_cache12 = cast(dtype = new_v_cache_25_cast_fp16_to_fp32_dtype_0, x = new_v_cache_25_cast_fp16)[name = tensor("cast_476")]; + tensor new_k_cache12 = cast(dtype = new_k_cache_25_cast_fp16_to_fp32_dtype_0, x = new_k_cache_25_cast_fp16)[name = tensor("cast_478")]; + tensor new_position11 = cast(dtype = var_3134_cast_fp16_to_fp32_dtype_0, x = var_3134_cast_fp16)[name = tensor("cast_482")]; + tensor new_v_cache11 = cast(dtype = new_v_cache_23_cast_fp16_to_fp32_dtype_0, x = new_v_cache_23_cast_fp16)[name = tensor("cast_484")]; + tensor new_k_cache11 = cast(dtype = new_k_cache_23_cast_fp16_to_fp32_dtype_0, x = new_k_cache_23_cast_fp16)[name = tensor("cast_486")]; + tensor new_position10 = cast(dtype = var_2890_cast_fp16_to_fp32_dtype_0, x = var_2890_cast_fp16)[name = tensor("cast_490")]; + tensor new_v_cache10 = cast(dtype = new_v_cache_21_cast_fp16_to_fp32_dtype_0, x = new_v_cache_21_cast_fp16)[name = tensor("cast_492")]; + tensor new_k_cache10 = cast(dtype = new_k_cache_21_cast_fp16_to_fp32_dtype_0, x = new_k_cache_21_cast_fp16)[name = tensor("cast_494")]; + tensor new_position9 = cast(dtype = var_2646_cast_fp16_to_fp32_dtype_0, x = var_2646_cast_fp16)[name = tensor("cast_498")]; + tensor new_v_cache9 = cast(dtype = new_v_cache_19_cast_fp16_to_fp32_dtype_0, x = new_v_cache_19_cast_fp16)[name = tensor("cast_500")]; + tensor new_k_cache9 = cast(dtype = new_k_cache_19_cast_fp16_to_fp32_dtype_0, x = new_k_cache_19_cast_fp16)[name = tensor("cast_502")]; + tensor new_position8 = cast(dtype = var_2402_cast_fp16_to_fp32_dtype_0, x = var_2402_cast_fp16)[name = tensor("cast_506")]; + tensor new_v_cache8 = cast(dtype = new_v_cache_17_cast_fp16_to_fp32_dtype_0, x = new_v_cache_17_cast_fp16)[name = tensor("cast_508")]; + tensor new_k_cache8 = cast(dtype = new_k_cache_17_cast_fp16_to_fp32_dtype_0, x = new_k_cache_17_cast_fp16)[name = tensor("cast_510")]; + tensor new_position7 = cast(dtype = var_2158_cast_fp16_to_fp32_dtype_0, x = var_2158_cast_fp16)[name = tensor("cast_514")]; + tensor new_v_cache7 = cast(dtype = new_v_cache_15_cast_fp16_to_fp32_dtype_0, x = new_v_cache_15_cast_fp16)[name = tensor("cast_516")]; + tensor new_k_cache7 = cast(dtype = new_k_cache_15_cast_fp16_to_fp32_dtype_0, x = new_k_cache_15_cast_fp16)[name = tensor("cast_518")]; + tensor new_position6 = cast(dtype = var_1914_cast_fp16_to_fp32_dtype_0, x = var_1914_cast_fp16)[name = tensor("cast_522")]; + tensor new_v_cache6 = cast(dtype = new_v_cache_13_cast_fp16_to_fp32_dtype_0, x = new_v_cache_13_cast_fp16)[name = tensor("cast_524")]; + tensor new_k_cache6 = cast(dtype = new_k_cache_13_cast_fp16_to_fp32_dtype_0, x = new_k_cache_13_cast_fp16)[name = tensor("cast_526")]; + tensor new_position5 = cast(dtype = var_1670_cast_fp16_to_fp32_dtype_0, x = var_1670_cast_fp16)[name = tensor("cast_530")]; + tensor new_v_cache5 = cast(dtype = new_v_cache_11_cast_fp16_to_fp32_dtype_0, x = new_v_cache_11_cast_fp16)[name = tensor("cast_532")]; + tensor new_k_cache5 = cast(dtype = new_k_cache_11_cast_fp16_to_fp32_dtype_0, x = new_k_cache_11_cast_fp16)[name = tensor("cast_534")]; + tensor new_position4 = cast(dtype = var_1426_cast_fp16_to_fp32_dtype_0, x = var_1426_cast_fp16)[name = tensor("cast_538")]; + tensor new_v_cache4 = cast(dtype = new_v_cache_9_cast_fp16_to_fp32_dtype_0, x = new_v_cache_9_cast_fp16)[name = tensor("cast_540")]; + tensor new_k_cache4 = cast(dtype = new_k_cache_9_cast_fp16_to_fp32_dtype_0, x = new_k_cache_9_cast_fp16)[name = tensor("cast_542")]; + tensor new_position3 = cast(dtype = var_1182_cast_fp16_to_fp32_dtype_0, x = var_1182_cast_fp16)[name = tensor("cast_546")]; + tensor new_v_cache3 = cast(dtype = new_v_cache_7_cast_fp16_to_fp32_dtype_0, x = new_v_cache_7_cast_fp16)[name = tensor("cast_548")]; + tensor new_k_cache3 = cast(dtype = new_k_cache_7_cast_fp16_to_fp32_dtype_0, x = new_k_cache_7_cast_fp16)[name = tensor("cast_550")]; + tensor new_position2 = cast(dtype = var_938_cast_fp16_to_fp32_dtype_0, x = var_938_cast_fp16)[name = tensor("cast_554")]; + tensor new_v_cache2 = cast(dtype = new_v_cache_5_cast_fp16_to_fp32_dtype_0, x = new_v_cache_5_cast_fp16)[name = tensor("cast_556")]; + tensor new_k_cache2 = cast(dtype = new_k_cache_5_cast_fp16_to_fp32_dtype_0, x = new_k_cache_5_cast_fp16)[name = tensor("cast_558")]; + tensor new_position1 = cast(dtype = var_694_cast_fp16_to_fp32_dtype_0, x = var_694_cast_fp16)[name = tensor("cast_562")]; + tensor new_v_cache1 = cast(dtype = new_v_cache_3_cast_fp16_to_fp32_dtype_0, x = new_v_cache_3_cast_fp16)[name = tensor("cast_564")]; + tensor new_k_cache1 = cast(dtype = new_k_cache_3_cast_fp16_to_fp32_dtype_0, x = new_k_cache_3_cast_fp16)[name = tensor("cast_566")]; + tensor new_position0 = cast(dtype = var_450_cast_fp16_to_fp32_dtype_0, x = var_450_cast_fp16)[name = tensor("cast_570")]; + tensor new_v_cache0 = cast(dtype = new_v_cache_1_cast_fp16_to_fp32_dtype_0, x = new_v_cache_1_cast_fp16)[name = tensor("cast_572")]; + tensor new_k_cache0 = cast(dtype = new_k_cache_1_cast_fp16_to_fp32_dtype_0, x = new_k_cache_1_cast_fp16)[name = tensor("cast_574")]; + } -> (transformer_out, is_eos, new_k_cache0, new_v_cache0, new_position0, new_k_cache1, new_v_cache1, new_position1, new_k_cache2, new_v_cache2, new_position2, new_k_cache3, new_v_cache3, new_position3, new_k_cache4, new_v_cache4, new_position4, new_k_cache5, new_v_cache5, new_position5, new_k_cache6, new_v_cache6, new_position6, new_k_cache7, new_v_cache7, new_position7, new_k_cache8, new_v_cache8, new_position8, new_k_cache9, new_v_cache9, new_position9, new_k_cache10, new_v_cache10, new_position10, new_k_cache11, new_v_cache11, new_position11, new_k_cache12, new_v_cache12, new_position12, new_k_cache13, new_v_cache13, new_position13, new_k_cache14, new_v_cache14, new_position14, new_k_cache15, new_v_cache15, new_position15, new_k_cache16, new_v_cache16, new_position16, new_k_cache17, new_v_cache17, new_position17, new_k_cache18, new_v_cache18, new_position18, new_k_cache19, new_v_cache19, new_position19, new_k_cache20, new_v_cache20, new_position20, new_k_cache21, new_v_cache21, new_position21, new_k_cache22, new_v_cache22, new_position22, new_k_cache23, new_v_cache23, new_position23); +} \ No newline at end of file diff --git a/v2.1/spanish_24l/flowlm_step_ane.mlmodelc/weights/weight.bin b/v2.1/spanish_24l/flowlm_step_ane.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..7ba8a5a22378b78d7d1c5e202ffff3e25bc3b337 --- /dev/null +++ b/v2.1/spanish_24l/flowlm_step_ane.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:53b10455916a541a403f42560a73cb3feba6b46eb984f3f27d91e510b4b47642 +size 604278464 diff --git a/v2.1/spanish_24l/flowlm_step_ane.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/spanish_24l/flowlm_step_ane.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..2cc4bcf46667502331f02e759949efe02fe74e34 --- /dev/null +++ b/v2.1/spanish_24l/flowlm_step_ane.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:14e3c4fb39b635d69ffd5b351ed2023a0b40e9c7715cac2f83ffd3ee9ac34e03 +size 545925 diff --git a/v2.1/spanish_24l/flowlm_step_ane.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/spanish_24l/flowlm_step_ane.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..7ba8a5a22378b78d7d1c5e202ffff3e25bc3b337 --- /dev/null +++ b/v2.1/spanish_24l/flowlm_step_ane.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:53b10455916a541a403f42560a73cb3feba6b46eb984f3f27d91e510b4b47642 +size 604278464 diff --git a/v2.1/spanish_24l/flowlm_step_ane.mlpackage/Manifest.json b/v2.1/spanish_24l/flowlm_step_ane.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..700872ee722da9aef340b80d1f07c282394b78fb --- /dev/null +++ b/v2.1/spanish_24l/flowlm_step_ane.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "4CEAE3E4-734E-4470-95F0-27D83096DD57": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "E2152A7D-7CE3-4CA0-9410-2BDD0A78815D": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "4CEAE3E4-734E-4470-95F0-27D83096DD57" +}